[mlir] Add initial support for a binary serialization format

[rriddle]

Summary

This commit adds a new bytecode serialization format for MLIR.
The actual serialization of MLIR to binary is relatively straightforward,
given the very very general structure of MLIR. The underlying basis for
this format is a variable-length encoding for integers, which gets heavily
used for nearly all aspects of the encoding (given that most of the encoding
is just indexing into lists).

The format currently does not provide support for custom attribute/type
serialization, and thus always uses an assembly format fallback. It also
doesn't provide support for resources. These will be added in followups,
the intention for this patch is to provide something that supports the
basic cases, and can be built on top of.

https://discourse.llvm.org/t/rfc-a-binary-serialization-format-for-mlir/63518

[rriddle]

I don't have any failure tests right now because those are annoying to make/update, and I want to make sure we agree on various aspects before doing that.

[jpienaar]

Nice, just started with doc for now.

I was wondering if we should we have mlir-translate do the text to binary and vice versa, but can see the convenience when stitching together passes and it's not an external format.

mlir/docs/BytecodeFormat.md

36

OOC why not use it?

69

What is used instead?

121

Nit: [] instead of *? (I see the relational database or pointer notations of the latter but I find the former more intuitive)

128

Link to encoding section?

209

And 7 of 8 optional elements used?

214

So location is either previous or new? Unspecified doesn't mean unknown but previous. That's nice. The only other common case I could think of is successive lines in a file but that would take a bit.

247

So we could have a region with 0 blocks here that is considered not empty?

[rriddle]

mlir/docs/BytecodeFormat.md

36

The encoding described here, or more commonly referred to as PrefixVarInt, is essentially a variant of LEB. Encoding and decoding are generally faster using the prefix strategy, though. The decode for the prefix variant, for example, is effectively branchless given that you just need to count the trailing zeros, which is an intrinsic on most modern hardware. I've benchmarked a bunch of different strategies and implementations over the past week, using both a corpus of .mlir taken from various projects and just using random distributions of integers. This is what emerged from that testing.

247

Zero block regions should be marked as empty. I used codes for various "bool" like things just to make it easier to change things later on. E.g. if we wanted to change the way regions are encoded we could just have a new "kRegion2" code. Though that only matters after we start versioning things.

[mehdi_amini]

Just commenting on the doc right now.

mlir/docs/BytecodeFormat.md

26

No fixed64 ? Is the bet that we'd always use VBR for this?

36

Are you using PrefixVarInt or how does it differ? Is your variant format documented somewhere in the literature?

I'd rather have us stick to something existing, I doubt that we'll invent a revolutionary trick here somehow.

Can you position what you're proposing against varint-G8IU, PFOR, SIMD-BP12, ... ?

120

I'm not sure here why these varint are useful?

121

+1 for [] :)

142

What is kAsmForm ? Seems to refer to an enum not documented here.

148

I'm not sure yet how we dispatch to a dialect for loading a type/attribute

154

Basically this is differential encoding of the offset table, but that means you need to decompress the table here IIUC. Accessing the last attribute requires to read the entire table.
(I assume we would do it once and for-all in the "BitcodeReaderContext" or whatever you're naming it)

168

That seems overly conservative to me: what about just storing the part after the dot and add a varint for the dialect ID?

191

I assume this field is needed to allow some level of lazy loading / random access?
I'm not sure yet...

209

Actually 6? firstResultIndex and numResults seem coupled right?

248

Do we need this indirection here? You didn't provide one for the operation content.

265

block_element ?

[mehdi_amini]

mlir/docs/BytecodeFormat.md

203

Having the regions in-line makes it hard/impossible to lazily load IR I think? (at least not without decoding the entire IR section).

[rriddle]

mlir/docs/BytecodeFormat.md

26

I just didn't add it in. Technically right now we only use byte, so I dropped the rest and added a TODO to add larger widths as necessary.

36

Yeah, it's just PrefixVarInt. I completely missed explicitly saying that when rushing out the doc.

120

These are used to know how many attributes/types/operation names need to be parsed.

121

SG, I also like [].

148

We chatted a little about this offline. Attributes and Types are grouped by dialect, with each grouping emitted in the same order as the dialects in the dialect section. This allows for us to know which dialect an attribute/type belongs to based on its index (i.e. we could know attributes 0-5 are the builtin dialect, 6-9 are the func dialect and so on).

154

Yeah. We do a single pass to initialize the bytecode structure of all attributes/types, which indicates where the data is stored in the bytecode. Reading lazily after that is trivial, because we just read the previously computed data directly.

168

Agreed, I need to look into this. I went with the current thing because it was easier to bootstrap (e.g. I don't have to worry about the difference between builtin and non-builtin).

191

This field is the value index of the first result. Every value gets a number, which is what gets referred to in the operands list. For multiple results, we know the value number are consecutive, so we just need to know the first one.

203

Yeah. The idea I have right now is that the op encoding mask will indicate if the regions are inline or out-of-line. That way we just dispatch to two different code paths depending on the encoding.

209

Yeah, right now the mask uses 6 out of 8 possible bits. Whenever we do lazy loading that may bump up to 7 bits (we could get around that by encoding the "are the regions lazy" bit a different way). We could of course change from a mask to a set of values for each possible encoding type.

214

Yeah, the file locations encoding is gonna be interesting. I'm hoping that when builtin attributes have custom encodings it ends up being mostly okay (should just be a string index+two small varints for the line/col). This is something we will likely want to play around with, but thankfully it's easy to test (I have a huge IR file that inherits locations from the file).

[jpienaar]

mlir/include/mlir/Bytecode/BytecodeWriter.h

36

This feels a little bit weird ... I'd almost expect like OpPrintingFlags having the config be separate from Operation* being printed. But perhaps it makes more sense below.

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

21

mlir-bytecode-reader ?

120

Nit: I'm more used to null-terminated, with NUL being the character name.

mlir/lib/Tools/mlir-opt/MlirOptMain.cpp

156

I think yes & no. piping these will be common and other uses seem like mistake, but I don't know how foolproof this check is on all platforms and opt tool is not a user tool.

mlir/test/Bytecode/general.mlir

2

We probably end up running all non-split or -error cases through a round trip tests to check, followed by fuzzing. It would almost seem possible to enumerate all of these kind of the constructs above.

[mehdi_amini]

mlir/docs/BytecodeFormat.md

102

Producer string?

120

Actually, following our offline discussion, they are needed to be able match to type/attr back to a dialect right?

133

Worth mentioning: the first section can't be decoded without the second one as the elements in the array of attrs/types don't include a size.

159

Maybe make it explicit: , this allows to associate an attribute back to a dialect without including a dialect reference in each type/attr entry.

168

Can't you capture this as a TODO at the end of this paragraph?

191

Right, but I'm still not sure why this needs to be encoded: if you load operations in order, you could just use an ever incrementing number for each Value.

203

I think we should think about isolated region from the get go: you don't document the value numbering in the doc but I think we can use the same "local scope" as the textual parser to ensure that the value IDs stays small (and so use less varint space).

221

Is this really the common case that (other than "unknown") the locations are repeating in sequence?

I am not convinced by this choice right now because it will make future lazy loading harder (you need to stream back to find the previously defined location).

252

When is "region empty" useful? Is it legal for an operation to have an empty region?

269

I don't get this, why multiple block_arguments blocks?

What about something like:

block {
  encoding: varint, // (numOps << 1) | (hasBlockArgs)
  arguments: block_arguments?, // Optional based on encoding
  ops : op[]
}
block_arguments {
  firstArgIndex: varint,
  numArgs: varint?,
  args: block_argument[]
}
block_argument {
  typeIndexAndHasLoc: varint, // (typeIndex << 1) | (hasLoc)
  location: varint?
}

[mehdi_amini]

mlir/docs/BytecodeFormat.md

120

Also right now it isn't used at all as far as I can tell.

262

Please document numValues :)

mlir/lib/Bytecode/Encoding.h

30

I'm not clear on how you manage "codes", or what are a "builtin section codes"

58

"a" top level operation...
We parse in a block so we should have the ability to have multiple I think.

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

136

You should sanity check sectionID here I think (and make the argument type the right enum in the API)

149

Document noinline please.

157

assert(numBytes > 0 && numBytes <= 7); ?

173

Does this work on a big endian machine?

217

We should have a pointer to the BytecodeDialect here I think, should be able to set it up in initializeOffsets

249

We should be able to have an enum for code here right?

282

offsetReader? (I was confusing to me reading the code where it is used)

306

I think you should check that currentOffset does not exceed the sectionData.size(), a malformed byte code coud have offsets going beyond.

381

I remember that Attr/Type were made "mutable" to support LLVM named struct (IIRC?), but isn't this encoding and loading scheme assuming there are no cycles? How are we gonna handle this?

566

This should move to parseSection I think. (sectionID` is used in test/set already, seems unsafe)

593

Why aren't dialects lazy loaded?

616

I was thinking: could we have a stringpool top-level section and everywhere refer to strings with an id there? Mnemonic shared between op/attributes/types and across dialects would be stored once and for-all.

749

Please break the recursion :)

755

Not sure where to attached this comment, but there is something missing somewhere (unless I missed it?) to ensure that use-lists ordering is preserved.

mlir/lib/Bytecode/Writer/IRNumbering.cpp

61

We could record the number of times an attribute is used in ordre to sort them so that the most used one have a lower IDs (and have more chances to fit in one bytes) :)

mlir/lib/Tools/mlir-opt/MlirOptMain.cpp

156

I think the difference with LLVM opt is that we're not having byte code as the default, hence we may not need to warn since the user has to opt-in to get there.

mlir/test/Bytecode/general.mlir

2

Reminds me an old proposal of mine to add some flag to mlir-opt to automatically round-trip and diff, and enable this flag optionally to process the entire test-suite :)

Seems like it would be useful here as well!

[jpienaar]

mlir/test/Bytecode/general.mlir

2

Yes indeed, I was actually wondering if we had that already :)

[mehdi_amini]

mlir/test/Bytecode/general.mlir

2

https://reviews.llvm.org/D90088

[rriddle]

mlir/docs/BytecodeFormat.md

191

Originally I was thinking that you can't do that if you load lazily, but if I use your other suggestion of doing per-isolated region numbering it should be possible. This will be much nicer, thanks!

203

Great suggestion!

221

I don't think it would make lazy-loading bad, given that we could encode the last location at the start of the region. I'm going to drop this behavior though, given that I'm not sure how many cases in practice it would help. This would also free up this optional behavior for something else (if that something else was more efficient for real-world use cases).

252

Yeah? e.g. External functions have empty regions. Almost every operation that has a region that can optionally be filled use an empty region, given that regions can't be dynamically added after op construction. Cleaned up this section though, we don't need the leading code, we can just use the block count.

269

Nice! I figured you would come up with something better ;)

mlir/include/mlir/Bytecode/BytecodeWriter.h

36

There was some reason why I did this before, but I forget now. I just dropped it.

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

149

I thought I did:

This method is marked noinline to avoid pessimizing the common case of single byte encoding.

Tried to make it more obvious.

173

Right now, no. I need to setup the proper little -> native conversions. I'm deferring that to a follow up because I have to setup a virtual machine for big endian (which is annoying/time consuming). I also need to fix some things in the textual format related to big endian as well.

381

We will likely need some form of special API that can parse just the "immutable" part (i.e. the "name" in the LLVM struct case). For example, if an attribute/type *is* recursive, we could encode both its immutable and mutable encodings in one entry (with some header that has the size of the immutable part or something). Something like:

RecursiveEntry {
  immutableEncodingSize: varint,
  immutableEncoding: ...,
  mutableEncoding: ...
}

During processing we could first process the immutable entry, and then immediately process the mutable one. That way any recursive references would resolve properly, and then we'd fix the final reference afterwards. Something like:

// Parse the immutable first, so that we have something to give recursive references.
if (!(result = parseImmutable()))
  return failure();
// Parse the mutable afterwards. Pass in `result` so that it can populate the mutable bits?
if (failed(parseMutable(result))
  return failure();

Until we figure any of this out though, I'm just going to have them always use the string fallback for those attributes and types.

755

Deferring this to a follow up to help simplify this patch, added a TODO for now.

mlir/lib/Bytecode/Writer/IRNumbering.cpp

61

We would need to encode things differently in that case, i.e. if the attributes are not in order of dialect, they would each need to have an associated dialect id encoded with them. In the case of lots of attributes/types, that would be significant. Maybe we could come up a hybrid model? i.e. encode the most common 128 attributes/types, so that they fit in one byte (or two), and then encode the rest using dialect grouping.

mlir/lib/Tools/mlir-opt/MlirOptMain.cpp

156

Makes sense to me, just dropped it. We can add a warning back in if enough people trip up on this (given bytecode generation is an explicit decision).

[jpienaar]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

616

So encoding would be start and end offsets into a string table?

[mehdi_amini]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

616

String being null terminated, you don't necessarily need the end offsets. But if we have an offset section separate from the string table: we just need to point to an entry number, same mechanism as attr/type reference.

[jpienaar]

(partial scan before meeting)

mlir/docs/BytecodeFormat.md

12

I just noticed ï and not i , I mean I guess writing this by hand and the other was that we don't actually take text file as bytecode (MĻîŘ just to bikeshed :))

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

10

Nit: I'd move this lower, I expect documentation here not todos :)

227

Dialect of this Attribute or Type ? (its mostly parent that makes me think OOP more than I normally think here, up to you)

309

attribute ?

[mehdi_amini]

mlir/docs/BytecodeFormat.md

12

I don't think your bike shed is ASCII though?

[jpienaar]

Missing negative tests are a bit unfortunate, but good to hear they are coming soon and this seems like good starting point.

mlir/docs/BytecodeFormat.md

12

Yeah I marked the comment as done before sending as I wasn't serious (and I did also use an extended ASCII encoding without realizing).

But in seriousness: MLÏR wouldn't be a valid dialect name, so this looks good.

mlir/lib/Bytecode/Encoding.h

25

uint8_t here too? (I mean with 0 it doesn't matter)

79

We can now use binary literals (not sure it makes if it is more readable, but was reminded of it)

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

242

Unsigned needed?

413

Comment?

501

So this parses the string starting at front of reader? And null-terminated?

616

Indeed, null-termination means we can't have substrings referenced (not sure if that is common here, could think for error strings, but unsure about decoding cost).

mlir/lib/Bytecode/Writer/IRNumbering.cpp

61

Would sorting attributes per frequency per dialect? (Keep dialect attributes still together but just sort dialects per frequency). We could measure all three of course, doesn't require version bump ;-)

[rriddle]

mlir/lib/Bytecode/Encoding.h

25

kVersion is encoded as a varint now, so that we don't have to change it if we have some burst of changing versions (makes it easier to change version if we don't have a cap looming overhead). I suppose I could switch the general constants to use inline constexpr variables now (given we are on C++17), let me know your preference.

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

501

Yeah, the front of the reader has an index to a string defined in the string section. Updated the comment.

mlir/test/Bytecode/general.mlir

2

Do you plan on reviving that @mehdi_amini ?

[mehdi_amini]

mlir/test/Bytecode/general.mlir

2

Yeah I should. I had memory that we couldn't reach consensus on it but I may be wrong.

[mehdi_amini]

LG, great start :)

[jpienaar]

mlir/test/Bytecode/general.mlir

2

I think only question was on on by default or not (e.g., how much of testing tool mlir-opt really is, should it be used in directed testing only etc)

[jpienaar]

Nice, like the encoding change.

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

58

Why this change?

mlir/lib/Bytecode/Writer/IRNumbering.cpp

81

Could this just be a static function here?

[rriddle]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

58

So that we can load in enums.

[vitalybuka]

This fails asan https://lab.llvm.org/buildbot/#/builders/5/builds/26955

[vitalybuka]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

902

also a problem, it emplace_back may relocate container, but the for loop above uses readState which is the ref to the element of container.

926

This pop_back and then readState.isIsolatedFromAbove which from the regionStack?

[rriddle]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

902

This should be fine, given that we always return in this case (i.e. never touch to invalid reference again).

926

Thanks for catching this. I'm not sure why my local asan build didn't catch this (I'll try nuking and resetting it).

[vitalybuka]

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

902

Thanks, I see.

926

I'm not sure why my local asan build didn't catch this

Probably you don't use libc++ or instrumented libc++?

I'm not sure why my local asan build didn't catch this I'm not sure why my local asan build didn't catch this

If you can fix it quickly go for it.
If not, please let me know, I have a patch to revert it with related fixes.

[mehdi_amini]

If you need to unbreak a sanitizer bot, we can XFAIL: asan the two tests, this is pretty cheap.

[vitalybuka]

In D131747#3740323, @mehdi_amini wrote:

If you need to unbreak a sanitizer bot, we can XFAIL: asan the two tests, this is pretty cheap.

You are not scare of out of bound mem access in alive code?

[mehdi_amini]

In D131747#3740334, @vitalybuka wrote:

In D131747#3740323, @mehdi_amini wrote:

If you need to unbreak a sanitizer bot, we can XFAIL: asan the two tests, this is pretty cheap.

You are not scare of out of bound mem access in alive code?

Define "alive"? This is a new feature that has zero users and we're actively bootstrapping. So no I'm not scared by an out-of-bound here for a couple of days at most.

[vitalybuka]

Looks like it's already fixed with 96fd3f2

[rriddle]

In D131747#3740368, @vitalybuka wrote:

In D131747#3740335, @mehdi_amini wrote:

In D131747#3740334, @vitalybuka wrote:

In D131747#3740323, @mehdi_amini wrote:

If you need to unbreak a sanitizer bot, we can XFAIL: asan the two tests, this is pretty cheap.

You are not scare of out of bound mem access in alive code?

Define "alive"? This is a new feature that has zero users and we're actively bootstrapping. So no I'm not scared by an out-of-bound here for a couple of days at most.

Sure, I don't know that code. So XFAIL is OK to me if you accept implications. (having that @rriddle failed to reproduce locally, maybe UNSUPPORTED instead, in case if some asan setups will miss the issues)
But seems revert/reland safe and easy to do as well.

Also if this is the only issue

Trivial fix may work?

bool isIsolatedFromAbove = readState.isIsolatedFromAbove;
 regionStack.pop_back();
 if (isIsolatedFromAbove)
   valueScopes.pop_back();

Yeah, sorry I've been in meetings. I pushed https://github.com/llvm/llvm-project/commit/96fd3f2d5be21ded6ffed0ac75195df04ec679df an hour ago and have been watching the bot to see if that is the only issue.

[uweigand]

Hi @rriddle , as of commit https://github.com/llvm/llvm-project/commit/93cf0e8a28e8c682f65d3e5c394d1eb169ca09ce the s390x build bot is still red due to "unexpected success":

XPASS: MLIR::invalid-string_section.mlir
XPASS: MLIR::invalid_attr_type_offset_section.mlir
XPASS: MLIR::invalid_attr_type_section.mlir
XPASS: MLIR::invalid-structure.mlir
XPASS: MLIR::invalid-ir_section.mlir
XPASS: MLIR::invalid-dialect_section.mlir

(see https://lab.llvm.org/buildbot/#/builders/199/builds/8674)

Can these "invalid" tests still legitimately pass even on a big-endian platform? It seems these XFAILs should either be removed or changed into UNSUPPORTED.

[rriddle]

In D131747#3740383, @uweigand wrote:

Hi @rriddle , as of commit https://github.com/llvm/llvm-project/commit/93cf0e8a28e8c682f65d3e5c394d1eb169ca09ce the s390x build bot is still red due to "unexpected success":

XPASS: MLIR::invalid-string_section.mlir
XPASS: MLIR::invalid_attr_type_offset_section.mlir
XPASS: MLIR::invalid_attr_type_section.mlir
XPASS: MLIR::invalid-structure.mlir
XPASS: MLIR::invalid-ir_section.mlir
XPASS: MLIR::invalid-dialect_section.mlir

(see https://lab.llvm.org/buildbot/#/builders/199/builds/8674)

Can these "invalid" tests still legitimately pass even on a big-endian platform? It seems these XFAILs should either be removed or changed into UNSUPPORTED.

@uweigand Thanks for the ping, it's possible big-endian is fine up to the point at which some of the tests fail (I haven't had time to setup a venv to test everything out). UNSUPPORTED is likely a better check than XFAIL there (I just copied from our other s390x broken test)

[uweigand]

In D131747#3740387, @rriddle wrote:

@uweigand Thanks for the ping, it's possible big-endian is fine up to the point at which some of the tests fail (I haven't had time to setup a venv to test everything out). UNSUPPORTED is likely a better check than XFAIL there (I just copied from our other s390x broken test)

As of commit df4e637ca7ef4ef17b662845120864921e65bb67 the build bot is green again on s390x. Thanks!

[RVP]

mlir/lib/Bytecode/Writer/BytecodeWriter.cpp

186

Is this parenthesized correctly?

[jpienaar]

mlir/lib/Bytecode/Writer/BytecodeWriter.cpp

186

This is checking if the value post shift is 0 (and relies on this function being called only when multi byte), what issue did you run into with this?

[RVP]

mlir/lib/Bytecode/Writer/BytecodeWriter.cpp

186

Shouldn't LLVM_LIKELY be around the whole condition instead of the shift expression? Isn't == 0 the likely case and not the shift result being non-zero?

[RVP]

mlir/lib/Bytecode/Writer/BytecodeWriter.cpp

186

I didn't see any issues. Was looking at the code and this question popped. I now saw that emitVarInt specially handles the common case (... >> 7) == 0. Maybe a comment here as well would have avoided the question. Thanks.

mlir/docs/BytecodeFormat.md

+# MLIR Bytecode Format
+
+This documents describes the MLIR bytecode format and its encoding.
+
+[TOC]
+
+## Magic Number
+
+MLIR uses the following four-byte magic number to indicate bytecode files:
+
+'\[‘M’<sub>8</sub>, ‘L’<sub>8</sub>, ‘ï’<sub>8</sub>, ‘R’<sub>8</sub>\]'
+
+## Format Overview
+
+An MLIR Bytecode file is comprised of a byte stream, with a few simple
+structural concepts layered on top.
+
+### Primitives
+
+#### Fixed-Width Integers
+
+```
+  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
+
+Variable width integers, or `VarInt`s, 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")](https://en.wikipedia.org/wiki/LEB128)
+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
[phone]: 56 value bits, the encoding uses 8 bytes
+00000000: 64 value bits, the encoding uses 9 bytes
+```
+
+#### Strings
+
+Strings are blobs of characters with an associated length.
+
+### Sections
+
+```
+section {
+  id: byte
+  length: varint
+}
+```
+
+Sections are a mechanism for grouping data within the bytecode. The enable
+delayed processing, which is useful for out-of-order processing of data,
+lazy-loading, and more. Each section contains a Section ID and a length (which
+allowing for skipping over the section).
+
+TODO: Sections should also carry an optional alignment. Add this when necessary.
+
+## MLIR Encoding
+
+Given the generic structure of MLIR, the bytecode encoding is actually fairly
+simplistic. It effectively maps to the core components of MLIR.
+
+### Top Level Structure
+
+The top-level structure of the bytecode contains the 4-byte "magic number", a
+version number, 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[]
+}
+```
+
+### String 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.
+
+### Dialect Section
+
+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 {
+  dialects: dialect[]  
+}
+
+dialect {
+  name: varint,
+  numAttrs: varint,
+  numTypes: varint,
+  numOpNames: varint,
+  opNames: varint[]
+}
+```
+
+The name and operation names of the dialect are indexes to strings within the
+string section. `numAttrs`, `numTypes`, and `numOpNames` indicate how many of
+each component belonging to the dialect are encoded within the bytecode.
+
+### Attribute/Type Sections
+
+Attributes and types are encoded using two [sections](#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 {
+  offset: varint[]
+}
+
+attribute {
+  code: byte, // kAsmForm
+  encoding: ...
+}
+type {
+  code: byte, // kAsmForm
+  encoding: ...
+}
+```
+
+Each `offset` in the `attr_type_offset_section` above is the size of the
+encoding for the attribute or type. 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 dialect
+grouping in the same order of the dialects within the
+[dialect section](#dialect-section), this allows to associate an attribute back
+to a dialect without including a dialect reference in each type/attr entry.
+
+#### Attribute/Type Encodings
+
+In the previous section, the forms of `attribute` and `type` both start with a
+`code` field. This field indicates how the attribute or type was encoded. 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.
+
+##### Assembly Format Fallback
+
+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.
+
+##### Dialect Defined Encoding
+
+TODO: This is not yet supported.
+
+### IR Section
+
+The IR section contains the encoded form of operations within the bytecode.
+
+#### Operation Encoding
+
+```
+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.
+
+##### Location
+
+The location is encoded as the index of the location within the attribute table.
+
+##### Attributes
+
+If the operation has attribues, the index of the operation attribute dictionary
+within the attribute table is encoded.
+
+##### Results
+
+If the operation has results, the number of results and the indexes of the
+result types within the type table are encoded.
+
+##### Operands
+
+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.
+
+##### Successors
+
+If the operation has successors, the number of successors and the indexes of the
+successor blocks within the parent region are encoded.
+
+##### Regions
+
+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 Encoding
+
+```
+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
+
+```
+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.

mlir/include/mlir/Bytecode/BytecodeReader.h

+//===- BytecodeReader.h - MLIR Bytecode Reader ------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This header defines interfaces to read MLIR bytecode files/streams.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_BYTECODE_BYTECODEREADER_H
+#define MLIR_BYTECODE_BYTECODEREADER_H
+
+#include "mlir/IR/AsmState.h"
+#include "mlir/Support/LLVM.h"
+
+namespace llvm {
+class MemoryBufferRef;
+} // namespace llvm
+
+namespace mlir {
+/// Returns true if the given buffer starts with the magic bytes that signal
+/// MLIR bytecode.
+bool isBytecode(llvm::MemoryBufferRef buffer);
+
+/// Read the operations defined within the given memory buffer, containing MLIR
+/// bytecode, into the provided block.
+LogicalResult readBytecodeFile(llvm::MemoryBufferRef buffer, Block *block,
+                               const ParserConfig &config);
+} // namespace mlir
+
+#endif // MLIR_BYTECODE_BYTECODEREADER_H

mlir/include/mlir/Bytecode/BytecodeWriter.h

+//===- BytecodeWriter.h - MLIR Bytecode Writer ------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This header defines interfaces to write MLIR bytecode files/streams.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_BYTECODE_BYTECODEWRITER_H
+#define MLIR_BYTECODE_BYTECODEWRITER_H
+
+#include "mlir/Support/LLVM.h"
+#include "llvm/ADT/StringRef.h"
+
+namespace mlir {
+class Operation;
+
+//===----------------------------------------------------------------------===//
+// Entry Points
+//===----------------------------------------------------------------------===//
+
+/// Write the bytecode for the given operation to the provided output stream.
+/// For streams where it matters, the given stream should be in "binary" mode.
+/// `producer` is an optional string that can be used to identify the producer
+/// of the bytecode when reading. It has no functional effect on the bytecode
+/// serialization.
+void writeBytecodeToFile(Operation *op, raw_ostream &os,
+                         StringRef producer = "MLIR" LLVM_VERSION_STRING);
+
+} // namespace mlir
+
+#endif // MLIR_BYTECODE_BYTECODEWRITER_H

mlir/include/mlir/IR/OperationSupport.h

struct OperationState {
   /// Regions that the op will hold.
   SmallVector<std::unique_ptr<Region>, 1> regions;
 
 public:
   OperationState(Location location, StringRef name);
   OperationState(Location location, OperationName name);
 
   OperationState(Location location, OperationName name, ValueRange operands,
-                 TypeRange types, ArrayRef<NamedAttribute> attributes,
+                 TypeRange types, ArrayRef<NamedAttribute> attributes = {},
                  BlockRange successors = {},
                  MutableArrayRef<std::unique_ptr<Region>> regions = {});
   OperationState(Location location, StringRef name, ValueRange operands,
-                 TypeRange types, ArrayRef<NamedAttribute> attributes,
+                 TypeRange types, ArrayRef<NamedAttribute> attributes = {},
                  BlockRange successors = {},
                  MutableArrayRef<std::unique_ptr<Region>> regions = {});
 
   void addOperands(ValueRange newOperands);
 
   void addTypes(ArrayRef<Type> newTypes) {
     types.append(newTypes.begin(), newTypes.end());
   }

mlir/include/mlir/Tools/mlir-opt/MlirOptMain.h

 /// "expected-(error|note|remark|warning)" are parsed in the input and matched
 /// against emitted diagnostics.
 /// - verifyPasses enables the IR verifier in-between each pass in the pipeline.
 /// - allowUnregisteredDialects allows to parse and create operation without
 /// registering the Dialect in the MLIRContext.
 /// - preloadDialectsInContext will trigger the upfront loading of all
 ///   dialects from the global registry in the MLIRContext. This option is
 ///   deprecated and will be removed soon.
+/// - emitBytecode will generate bytecode output instead of text.
 LogicalResult MlirOptMain(llvm::raw_ostream &outputStream,
                           std::unique_ptr<llvm::MemoryBuffer> buffer,
                           const PassPipelineCLParser &passPipeline,
                           DialectRegistry &registry, bool splitInputFile,
                           bool verifyDiagnostics, bool verifyPasses,
                           bool allowUnregisteredDialects,
-                          bool preloadDialectsInContext = false);
+                          bool preloadDialectsInContext = false,
+                          bool emitBytecode = false);
 
 /// Support a callback to setup the pass manager.
 /// - passManagerSetupFn is the callback invoked to setup the pass manager to
 ///   apply on the loaded IR.
 LogicalResult MlirOptMain(llvm::raw_ostream &outputStream,
                           std::unique_ptr<llvm::MemoryBuffer> buffer,
                           PassPipelineFn passManagerSetupFn,
                           DialectRegistry &registry, bool splitInputFile,
                           bool verifyDiagnostics, bool verifyPasses,
                           bool allowUnregisteredDialects,
-                          bool preloadDialectsInContext = false);
+                          bool preloadDialectsInContext = false,
+                          bool emitBytecode = false);
 
 /// Implementation for tools like `mlir-opt`.
 /// - toolName is used for the header displayed by `--help`.
 /// - registry should contain all the dialects that can be parsed in the source.
 /// - preloadDialectsInContext will trigger the upfront loading of all
 ///   dialects from the global registry in the MLIRContext. This option is
 ///   deprecated and will be removed soon.
 LogicalResult MlirOptMain(int argc, char **argv, llvm::StringRef toolName,

mlir/lib/Bytecode/CMakeLists.txt

+add_subdirectory(Reader)
+add_subdirectory(Writer)

mlir/lib/Bytecode/Encoding.h

+//===- Encoding.h - MLIR binary format encoding information -----*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This header defines enum values describing the structure of MLIR bytecode
+// files.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LIB_MLIR_BYTECODE_ENCODING_H
+#define LIB_MLIR_BYTECODE_ENCODING_H
+
+#include <cstdint>
+
+namespace mlir {
+namespace bytecode {
+//===----------------------------------------------------------------------===//
+// General constants
+//===----------------------------------------------------------------------===//
+
+enum {
+  /// The current bytecode version.
+  kVersion = 0,
+};
+
+//===----------------------------------------------------------------------===//
+// Sections
+//===----------------------------------------------------------------------===//
+
+namespace Section {
+enum ID : uint8_t {
+  /// This section contains strings referenced within the bytecode.
+  kString = 0,
+
+  /// This section contains the dialects referenced within an IR module.
+  kDialect = 1,
+
+  /// This section contains the attributes and types referenced within an IR
+  /// module.
+  kAttrType = 2,
+
+  /// This section contains the offsets for the attribute and types within the
+  /// AttrType section.
+  kAttrTypeOffset = 3,
+
+  /// This section contains the list of operations serialized into the bytecode,
+  /// and their nested regions/operations.
+  kIR = 4,
+
+  /// The total number of section types.
+  kNumSections = 5,
+};
+} // namespace Section
+
+//===----------------------------------------------------------------------===//
+// AttrType Section
+//===----------------------------------------------------------------------===//
+
+enum class AttrTypeCode : uint8_t {
+  /// This code represents an attribute or type represented in the textual
+  /// assembly format.
+  kAsmForm,
+};
+
+//===----------------------------------------------------------------------===//
+// IR Section
+//===----------------------------------------------------------------------===//
+
+/// This enum represents a mask of all of the potential components of an
+/// operation. This mask is used when encoding an operation to indicate which
+/// components are present in the bytecode.
+namespace OpEncodingMask {
+enum : uint8_t {
+  kHasAttrs = 1 << 0,
+  kHasResults = 1 << 1,
+  kHasOperands = 1 << 2,
+  kHasSuccessors = 1 << 3,
+  kHasInlineRegions = 1 << 4,
+};
+} // namespace OpEncodingMask
+
+} // namespace bytecode
+} // namespace mlir
+
+#endif

mlir/lib/Bytecode/Reader/BytecodeReader.cpp

+//===- BytecodeReader.cpp - MLIR Bytecode Reader --------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// TODO: Support for big-endian architectures.
+// TODO: Properly preserve use lists of values.
+
+#include "mlir/Bytecode/BytecodeReader.h"
+#include "../Encoding.h"
+#include "mlir/AsmParser/AsmParser.h"
+#include "mlir/IR/BuiltinDialect.h"
+#include "mlir/IR/BuiltinOps.h"
+#include "mlir/IR/OpImplementation.h"
+#include "mlir/IR/Verifier.h"
+#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/ScopeExit.h"
+#include "llvm/ADT/SmallString.h"
+#include "llvm/Support/MemoryBufferRef.h"
+#include "llvm/Support/SaveAndRestore.h"
+
+#define DEBUG_TYPE "mlir-bytecode-reader"
+
+using namespace mlir;
+
+//===----------------------------------------------------------------------===//
+// EncodingReader
+//===----------------------------------------------------------------------===//
+
+namespace {
+class EncodingReader {
+public:
+  explicit EncodingReader(ArrayRef<uint8_t> contents, Location fileLoc)
+      : dataIt(contents.data()), dataEnd(contents.end()), fileLoc(fileLoc) {}
+  explicit EncodingReader(StringRef contents, Location fileLoc)
+      : EncodingReader({reinterpret_cast<const uint8_t *>(contents.data()),
+                        contents.size()},
+                       fileLoc) {}
+
+  /// Returns true if the entire section has been read.
+  bool empty() const { return dataIt == dataEnd; }
+
+  /// Returns the remaining size of the bytecode.
+  size_t size() const { return dataEnd - dataIt; }
+
+  /// Emit an error using the given arguments.
+  template <typename... Args>
+  LogicalResult emitError(Args &&...args) const {
+    return ::emitError(fileLoc).append(std::forward<Args>(args)...);
+  }
+
+  /// Parse a single byte from the stream.
+  template <typename T>
+  ParseResult parseByte(T &value) {
+    if (empty())
+      return emitError("attempting to parse a byte at the end of the bytecode");
+    value = static_cast<T>(*dataIt++);
+    return success();
+  }
+  /// Parse a range of bytes of 'length' into the given result.
+  ParseResult parseBytes(size_t length, ArrayRef<uint8_t> &result) {
+    if (length > size()) {
+      return emitError("attempting to parse ", length, " bytes when only ",
+                       size(), " remain");
+    }
+    result = {dataIt, length};
+    dataIt += length;
+    return success();
+  }
+  /// Parse a range of bytes of 'length' into the given result, which can be
+  /// assumed to be large enough to hold `length`.
+  ParseResult parseBytes(size_t length, uint8_t *result) {
+    if (length > size()) {
+      return emitError("attempting to parse ", length, " bytes when only ",
+                       size(), " remain");
+    }
+    memcpy(result, dataIt, length);
+    dataIt += length;
+    return success();
+  }
+
+  /// Parse a variable length encoded integer from the byte stream. The first
+  /// encoded byte contains a prefix in the low bits indicating the encoded
+  /// length of the value. This length prefix is a bit sequence of '0's followed
+  /// by a '1'. The number of '0' bits indicate the number of _additional_ bytes
+  /// (not including the prefix byte). All remaining bits in the first byte,
+  /// along with all of the bits in additional bytes, provide the value of the
+  /// integer encoded in little-endian order.
+  ParseResult parseVarInt(uint64_t &result) {
+    // Parse the first byte of the encoding, which contains the length prefix.
+    if (parseByte(result))
+      return failure();
+
+    // Handle the overwhelmingly common case where the value is stored in a
+    // single byte. In this case, the first bit is the `1` marker bit.
+    if (LLVM_LIKELY(result & 1)) {
+      result >>= 1;
+      return success();
+    }
+
+    // Handle the overwhelming uncommon case where the value required all 8
+    // bytes (i.e. a really really big number). In this case, the marker byte is
+    // all zeros: `00000000`.
+    if (LLVM_UNLIKELY(result == 0))
+      return parseBytes(sizeof(result), reinterpret_cast<uint8_t *>(&result));
+    return parseMultiByteVarInt(result);
+  }
+
+  /// Skip the first `length` bytes within the reader.
+  ParseResult skipBytes(size_t length) {
+    if (length > size()) {
+      return emitError("attempting to skip ", length, " bytes when only ",
+                       size(), " remain");
+    }
+    dataIt += length;
+    return success();
+  }
+
+  /// Parse a null-terminated string into `result` (without including the NUL
+  /// terminator).
+  ParseResult parseNullTerminatedString(StringRef &result) {
+    const char *startIt = (const char *)dataIt;
+    const char *nulIt = (const char *)memchr(startIt, 0, size());
+    if (!nulIt)
+      return emitError("malformed null-terminated string, no NUL found");
+
+    result = StringRef(startIt, nulIt - startIt);
+    dataIt = (const uint8_t *)nulIt + 1;
+    return success();
+  }
+
+  /// Parse a section header, placing the kind of section in `sectionID` and the
+  /// contents of the section in `sectionData`.
+  ParseResult parseSection(bytecode::Section::ID &sectionID,
+                           ArrayRef<uint8_t> &sectionData) {
+    size_t length;
+    if (parseByte(sectionID) || parseVarInt(length))
+      return failure();
+    if (sectionID >= bytecode::Section::kNumSections)
+      return emitError("invalid section ID: ", sectionID);
+
+    // Parse the actua section data now that we have its length.
+    return parseBytes(length, sectionData);
+  }
+
+private:
+  /// Parse a variable length encoded integer from the byte stream. This method
+  /// is a fallback when the number of bytes used to encode the value is greater
+  /// than 1, but less than the max (9). The provided `result` value can be
+  /// assumed to already contain the first byte of the value.
+  /// NOTE: This method is marked noinline to avoid pessimizing the common case
+  /// of single byte encoding.
+  LLVM_ATTRIBUTE_NOINLINE ParseResult parseMultiByteVarInt(uint64_t &result) {
+    // Count the number of trailing zeros in the marker byte, this indicates the
+    // number of trailing bytes that are part of the value. We use `uint32_t`
+    // here because we only care about the first byte, and so that be actually
+    // get ctz intrinsic calls when possible (the `uint8_t` overload uses a loop
+    // implementation).
+    uint32_t numBytes =
+        llvm::countTrailingZeros<uint32_t>(result, llvm::ZB_Undefined);
+    assert(numBytes > 0 && numBytes <= 7 &&
+           "unexpected number of trailing zeros in varint encoding");
+
+    // Parse in the remaining bytes of the value.
+    if (parseBytes(numBytes, reinterpret_cast<uint8_t *>(&result) + 1))
+      return failure();
+
+    // Shift out the low-order bits that were used to mark how the value was
+    // encoded.
+    result >>= (numBytes + 1);
+    return success();
+  }
+
+  /// The current data iterator, and an iterator to the end of the buffer.
+  const uint8_t *dataIt, *dataEnd;
+
+  /// A location for the bytecode used to report errors.
+  Location fileLoc;
+};
+} // namespace
+
+//===----------------------------------------------------------------------===//
+// BytecodeDialect
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// This struct represents a dialect entry within the bytecode.
+struct BytecodeDialect {
+  BytecodeDialect(Dialect *dialect, StringRef name, unsigned numAttrs,
+                  unsigned numTypes)
+      : dialect(dialect), name(name), numAttrs(numAttrs), numTypes(numTypes) {}
+
+  /// The loaded dialect entry, if available, otherwise nullptr.
+  Dialect *dialect;
+
+  /// The name of the dialect.
+  StringRef name;
+
+  /// The number of attributes owned by this dialect in the bytecode.
+  unsigned numAttrs;
+
+  /// The number of types owned by this dialect in the bytecode.
+  unsigned numTypes;
+};
+} // namespace
+
+//===----------------------------------------------------------------------===//
+// Attribute/Type Reader
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// This class provides support for reading attribute and type entries from the
+/// bytecode. Attribute and Type entries are read lazily on demand, so we use
+/// this reader to manage when to actually parse them from the bytecode.
+class AttrTypeReader {
+  /// This class represents a single attribute or type entry.
+  template <typename T>
+  struct Entry {
+    /// The entry, or null if it hasn't been resolved yet.
+    T entry = {};
+    /// The parent dialect of this entry.
+    BytecodeDialect *dialect = nullptr;
+    /// The raw data of this entry in the bytecode.
+    ArrayRef<uint8_t> data;
+  };
+  using AttrEntry = Entry<Attribute>;
+  using TypeEntry = Entry<Type>;
+
+public:
+  AttrTypeReader(Location fileLoc) : fileLoc(fileLoc) {}
+
+  /// Initialize the attribute and type information within the reader.
+  LogicalResult initialize(ArrayRef<BytecodeDialect> dialects,
+                           ArrayRef<uint8_t> sectionData,
+                           ArrayRef<uint8_t> offsetSectionData);
+
+  /// Resolve the attribute or type at the given index. Returns nullptr on
+  /// failure.
+  Attribute resolveAttribute(unsigned index) {
+    return resolveEntry(attributes, index, "Attribute");
+  }
+  Type resolveType(unsigned index) {
+    return resolveEntry(types, index, "Type");
+  }
+
+private:
+  /// Initialize the offsets for the attribute and type entries.
+  LogicalResult initializeOffsets(ArrayRef<uint8_t> sectionData,
+                                  ArrayRef<uint8_t> offsetSectionData);
+
+  /// Resolve the given entry at `index`.
+  template <typename T>
+  T resolveEntry(SmallVectorImpl<Entry<T>> &entries, unsigned index,
+                 StringRef entryType);
+
+  /// Parse the value defined within the given reader. `code` indicates how the
+  /// entry was encoded.
+  LogicalResult parseEntry(EncodingReader &reader, bytecode::AttrTypeCode code,
+                           Attribute &result);
+  LogicalResult parseEntry(EncodingReader &reader, bytecode::AttrTypeCode code,
+                           Type &result);
+
+  /// The set of attribute and type entries.
+  SmallVector<AttrEntry> attributes;
+  SmallVector<TypeEntry> types;
+
+  /// A location used for error emission.
+  Location fileLoc;
+};
+} // namespace
+
+LogicalResult AttrTypeReader::initialize(ArrayRef<BytecodeDialect> dialects,
+                                         ArrayRef<uint8_t> sectionData,
+                                         ArrayRef<uint8_t> offsetSectionData) {
+  // Initialize the entries using the dialect information.
+  unsigned numAttrs = 0, numTypes = 0;
+  for (const BytecodeDialect &dialect : dialects) {
+    numAttrs += dialect.numAttrs;
+    numTypes += dialect.numTypes;
+  }
+  attributes.resize(numAttrs);
+  types.resize(numTypes);
+
+  // With the entries initialized, we can process the offsets.
+  return initializeOffsets(sectionData, offsetSectionData);
+}
+
+LogicalResult
+AttrTypeReader::initializeOffsets(ArrayRef<uint8_t> sectionData,
+                                  ArrayRef<uint8_t> offsetSectionData) {
+  EncodingReader offsetReader(offsetSectionData, fileLoc);
+
+  // A functor used to accumulate the offsets for the entries in the given
+  // range.
+  uint64_t currentOffset = 0;
+  auto accumulateOffsets = [&](auto &&range) {
+    for (auto &entry : range) {
+      uint64_t entrySize;
+      if (offsetReader.parseVarInt(entrySize))
+        return failure();
+      if (currentOffset + entrySize > sectionData.size())
+        return offsetReader.emitError(
+            "Attribute or Type entry offset points past the end of section");
+      entry.data = sectionData.slice(currentOffset, entrySize);
+      currentOffset += entrySize;
+    }
+    return success();
+  };
+
+  // Process each of the attributes, and then the types.
+  if (failed(accumulateOffsets(attributes)) || failed(accumulateOffsets(types)))
+    return failure();
+
+  // Ensure that we read everything from the section.
+  if (!offsetReader.empty()) {
+    return offsetReader.emitError(
+        "unexpected trailing data in the Attribute/Type offset section");
+  }
+  return success();
+}
+
+template <typename T>
+T AttrTypeReader::resolveEntry(SmallVectorImpl<Entry<T>> &entries,
+                               unsigned index, StringRef entryType) {
+  if (index >= entries.size()) {
+    emitError(fileLoc) << "invalid " << entryType << "index:" << index;
+    return {};
+  }
+
+  // If the entry has already been resolved, there is nothing left to do.
+  Entry<T> &entry = entries[index];
+  if (entry.entry)
+    return entry.entry;
+
+  // Parse the entry. Each entry starts with a specific code that indicates how
+  // it is represented.
+  EncodingReader reader(entry.data, fileLoc);
+  bytecode::AttrTypeCode code;
+  if (reader.parseByte(code) || failed(parseEntry(reader, code, entry.entry)))
+    return T();
+  if (!reader.empty()) {
+    (void)reader.emitError("unexpected trailing bytes after " + entryType +
+                           " entry");
+    return T();
+  }
+  return entry.entry;
+}
+
+LogicalResult AttrTypeReader::parseEntry(EncodingReader &reader,
+                                         bytecode::AttrTypeCode code,
+                                         Attribute &result) {
+  // Handle the fallback case, where the attribute was encoded using its
+  // assembly format.
+  if (code == bytecode::AttrTypeCode::kAsmForm) {
+    StringRef attrStr;
+    if (failed(reader.parseNullTerminatedString(attrStr)))
+      return failure();
+
+    size_t numRead = 0;
+    if (!(result = parseAttribute(attrStr, fileLoc->getContext(), numRead)))
+      return failure();
+    if (numRead != attrStr.size()) {
+      return reader.emitError(
+          "trailing characters found after Attribute assembly format: ",
+          attrStr.drop_front(numRead));
+    }
+    return success();
+  }
+
+  return reader.emitError("unexpected Attribute encoding: ",
+                          static_cast<unsigned>(code));
+}
+
+LogicalResult AttrTypeReader::parseEntry(EncodingReader &reader,
+                                         bytecode::AttrTypeCode code,
+                                         Type &result) {
+  // Handle the fallback case, where the type was encoded using its
+  // assembly format.
+  if (code == bytecode::AttrTypeCode::kAsmForm) {
+    StringRef typeStr;
+    if (failed(reader.parseNullTerminatedString(typeStr)))
+      return failure();
+
+    size_t numRead = 0;
+    if (!(result = parseType(typeStr, fileLoc->getContext(), numRead)))
+      return failure();
+    if (numRead != typeStr.size()) {
+      return reader.emitError(
+          "trailing characters found after Type assembly format: " +
+          typeStr.drop_front(numRead));
+    }
+    return success();
+  }
+
+  return reader.emitError("unexpected Type encoding: ",
+                          static_cast<unsigned>(code));
+}
+
+//===----------------------------------------------------------------------===//
+// Bytecode Reader
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// This class is used to read a bytecode buffer and translate it into MLIR.
+class BytecodeReader {
+public:
+  BytecodeReader(Location fileLoc, const ParserConfig &config)
+      : config(config), fileLoc(fileLoc), attrTypeReader(fileLoc),
+        forwardRefOpState(UnknownLoc::get(config.getContext()),
+                          "builtin.unrealized_conversion_cast", ValueRange(),
+                          NoneType::get(config.getContext())) {}
+
+  /// Read the bytecode defined within `buffer` into the given block.
+  LogicalResult read(llvm::MemoryBufferRef buffer, Block *block);
+
+private:
+  /// Return the context for this config.
+  MLIRContext *getContext() const { return config.getContext(); }
+
+  /// Parse the bytecode version.
+  LogicalResult parseVersion(EncodingReader &reader);
+
+  //===--------------------------------------------------------------------===//
+  // Dialect Section
+
+  LogicalResult parseDialectSection(ArrayRef<uint8_t> sectionData);
+
+  /// Parse an operation name reference using the given reader.
+  FailureOr<OperationName> parseOpName(EncodingReader &reader);
+
+  //===--------------------------------------------------------------------===//
+  // Attribute/Type Section
+
+  /// Parse an attribute or type using the given reader. Returns nullptr in the
+  /// case of failure.
+  Attribute parseAttribute(EncodingReader &reader);
+  Type parseType(EncodingReader &reader);
+
+  template <typename T>
+  T parseAttribute(EncodingReader &reader) {
+    if (Attribute attr = parseAttribute(reader)) {
+      if (auto derivedAttr = attr.dyn_cast<T>())
+        return derivedAttr;
+      (void)reader.emitError("expected attribute of type: ",
+                             llvm::getTypeName<T>(), ", but got: ", attr);
+    }
+    return T();
+  }
+
+  //===--------------------------------------------------------------------===//
+  // IR Section
+
+  /// This struct represents the current read state of a range of regions. This
+  /// struct is used to enable iterative parsing of regions.
+  struct RegionReadState {
+    RegionReadState(Operation *op, bool isIsolatedFromAbove)
+        : RegionReadState(op->getRegions(), isIsolatedFromAbove) {}
+    RegionReadState(MutableArrayRef<Region> regions, bool isIsolatedFromAbove)
+        : curRegion(regions.begin()), endRegion(regions.end()),
+          isIsolatedFromAbove(isIsolatedFromAbove) {}
+
+    /// The current regions being read.
+    MutableArrayRef<Region>::iterator curRegion, endRegion;
+
+    /// The number of values defined immediately within this region.
+    unsigned numValues = 0;
+
+    /// The current blocks of the region being read.
+    SmallVector<Block *> curBlocks;
+    Region::iterator curBlock = {};
+
+    /// The number of operations remaining to be read from the current block
+    /// being read.
+    unsigned numOpsRemaining = 0;
+
+    /// A flag indicating if the regions being read are isolated from above.
+    bool isIsolatedFromAbove = false;
+  };
+
+  LogicalResult parseIRSection(ArrayRef<uint8_t> sectionData, Block *block);
+  LogicalResult parseRegions(EncodingReader &reader,
+                             std::vector<RegionReadState> &regionStack,
+                             RegionReadState &readState);
+  FailureOr<Operation *> parseOpWithoutRegions(EncodingReader &reader,
+                                               RegionReadState &readState,
+                                               bool &isIsolatedFromAbove);
+
+  LogicalResult parseRegion(EncodingReader &reader, RegionReadState &readState);
+  LogicalResult parseBlock(EncodingReader &reader, RegionReadState &readState);
+  LogicalResult parseBlockArguments(EncodingReader &reader, Block *block);
+
+  //===--------------------------------------------------------------------===//
+  // String Section
+
+  LogicalResult parseStringSection(ArrayRef<uint8_t> sectionData);
+
+  /// Parse a shared string from the string section.
+  LogicalResult parseSharedString(EncodingReader &reader, StringRef &result);
+
+  //===--------------------------------------------------------------------===//
+  // Value Processing
+
+  /// Parse an operand reference using the given reader. Returns nullptr in the
+  /// case of failure.
+  Value parseOperand(EncodingReader &reader);
+
+  /// Sequentially define the given value range.
+  LogicalResult defineValues(EncodingReader &reader, ValueRange values);
+
+  /// Create a value to use for a forward reference.
+  Value createForwardRef();
+
+  //===--------------------------------------------------------------------===//
+  // Fields
+
+  /// This class represents a single value scope, in which a value scope is
+  /// delimited by isolated from above regions.
+  struct ValueScope {
+    /// Push a new region state onto this scope, reserving enough values for
+    /// those defined within the current region of the provided state.
+    void push(RegionReadState &readState) {
+      nextValueIDs.push_back(values.size());
+      values.resize(values.size() + readState.numValues);
+    }
+
+    /// Pop the values defined for the current region within the provided region
+    /// state.
+    void pop(RegionReadState &readState) {
+      values.resize(values.size() - readState.numValues);
+      nextValueIDs.pop_back();
+    }
+
+    /// The set of values defined in this scope.
+    std::vector<Value> values;
+
+    /// The ID for the next defined value for each region current being
+    /// processed in this scope.
+    SmallVector<unsigned, 4> nextValueIDs;
+  };
+
+  /// The configuration of the parser.
+  const ParserConfig &config;
+
+  /// A location to use when emitting errors.
+  Location fileLoc;
+
+  /// The reader used to process attribute and types within the bytecode.
+  AttrTypeReader attrTypeReader;
+
+  /// The version of the bytecode being read.
+  uint64_t version = 0;
+
+  /// The producer of the bytecode being read.
+  StringRef producer;
+
+  /// The table of IR units referenced within the bytecode file.
+  SmallVector<BytecodeDialect> dialects;
+  SmallVector<OperationName> opNames;
+
+  /// The table of strings referenced within the bytecode file.
+  SmallVector<StringRef> strings;
+
+  /// The current set of available IR value scopes.
+  std::vector<ValueScope> valueScopes;
+  /// A block containing the set of operations defined to create forward
+  /// references.
+  Block forwardRefOps;
+  /// A block containing previously created, and no longer used, forward
+  /// reference operations.
+  Block openForwardRefOps;
+  /// An operation state used when instantiating forward references.
+  OperationState forwardRefOpState;
+};
+} // namespace
+
+LogicalResult BytecodeReader::read(llvm::MemoryBufferRef buffer, Block *block) {
+  EncodingReader reader(buffer.getBuffer(), fileLoc);
+
+  // Skip over the bytecode header, this should have already been checked.
+  if (reader.skipBytes(StringRef("ML\xefR").size()))
+    return failure();
+  // Parse the bytecode version and producer.
+  if (failed(parseVersion(reader)) ||
+      failed(reader.parseNullTerminatedString(producer)))
+    return failure();
+
+  // Add a diagnostic handler that attaches a note that includes the original
+  // producer of the bytecode.
+  ScopedDiagnosticHandler diagHandler(getContext(), [&](Diagnostic &diag) {
+    diag.attachNote() << "in bytecode file produced by: " << producer;
+    return failure();
+  });
+
+  // Parse the raw data for each of the top-level sections of the bytecode.
+  Optional<ArrayRef<uint8_t>> sectionDatas[bytecode::Section::kNumSections];
+  while (!reader.empty()) {
+    // Read the next section from the bytecode.
+    bytecode::Section::ID sectionID;
+    ArrayRef<uint8_t> sectionData;
+    if (reader.parseSection(sectionID, sectionData))
+      return failure();
+
+    // Check for duplicate sections, we only expect one instance of each.
+    if (sectionDatas[sectionID])
+      return reader.emitError("duplicate top-level section ID: ", sectionID);
+    sectionDatas[sectionID] = sectionData;
+  }
+  // Check that all of the sections were found.
+  for (int i = 0; i < bytecode::Section::kNumSections; ++i)
+    if (!sectionDatas[i])
+      return reader.emitError("missing data for top-level section: ", i);
+
+  // Process the string section first.
+  if (failed(parseStringSection(*sectionDatas[bytecode::Section::kString])))
+    return failure();
+
+  // Process the dialect section.
+  if (failed(parseDialectSection(*sectionDatas[bytecode::Section::kDialect])))
+    return failure();
+
+  // Process the attribute and type section.
+  if (failed(attrTypeReader.initialize(
+          dialects, *sectionDatas[bytecode::Section::kAttrType],
+          *sectionDatas[bytecode::Section::kAttrTypeOffset])))
+    return failure();
+
+  // Finally, process the IR section.
+  return parseIRSection(*sectionDatas[bytecode::Section::kIR], block);
+}
+
+LogicalResult BytecodeReader::parseVersion(EncodingReader &reader) {
+  if (reader.parseVarInt(version))
+    return failure();
+
+  // Validate the bytecode version.
+  if (version < bytecode::kVersion) {
+    return reader.emitError(
+        "bytecode version ", version, " is older than the current version of ",
+        bytecode::kVersion, ", and upgrade is not supported.");
+  }
+  if (version > bytecode::kVersion) {
+    return reader.emitError("bytecode version ", version,
+                            " is newer than the current version ",
+                            bytecode::kVersion, ".");
+  }
+  return success();
+}
+
+//===----------------------------------------------------------------------===//
+// Dialect Section
+
+LogicalResult
+BytecodeReader::parseDialectSection(ArrayRef<uint8_t> sectionData) {
+  MLIRContext *ctx = getContext();
+
+  EncodingReader sectionReader(sectionData, fileLoc);
+  while (!sectionReader.empty()) {
+    // Read the name of the next dialect.
+    StringRef dialectName;
+    if (failed(parseSharedString(sectionReader, dialectName)))
+      return failure();
+
+    // Parse the attribute and type counts.
+    uint64_t attrCount, typeCount;
+    if (sectionReader.parseVarInt(attrCount) ||
+        sectionReader.parseVarInt(typeCount))
+      return failure();
+
+    // Try to load the dialect.
+    Dialect *dialect = ctx->getOrLoadDialect(dialectName);
+    if (!dialect && !ctx->allowsUnregisteredDialects()) {
+      return sectionReader.emitError(
+          "dialect '", dialectName,
+          "' is unknown. If this is intended, please call "
+          "allowUnregisteredDialects() on the MLIRContext, or use "
+          "-allow-unregistered-dialect with the MLIR tool used.");
+    }
+    dialects.emplace_back(dialect, dialectName, attrCount, typeCount);
+
+    // Parse the operation names of the dialect.
+    uint64_t numOpNames;
+    if (sectionReader.parseVarInt(numOpNames))
+      return failure();
+    SmallString<32> opNameStorage({dialectName, "."});
+    while (numOpNames--) {
+      StringRef opName;
+      if (failed(parseSharedString(sectionReader, opName)))
+        return failure();
+
+      opNameStorage.resize(dialectName.size() + 1);
+      opNameStorage.append(opName);
+      opNames.push_back(OperationName(opNameStorage, ctx));
+    }
+  }
+  return success();
+}
+
+FailureOr<OperationName> BytecodeReader::parseOpName(EncodingReader &reader) {
+  uint64_t opNameIdx;
+  if (reader.parseVarInt(opNameIdx))
+    return failure();
+
+  if (opNameIdx >= opNames.size())
+    return reader.emitError("invalid operation name index: ", opNameIdx);
+  return opNames[opNameIdx];
+}
+
+//===----------------------------------------------------------------------===//
+// Attribute/Type Section
+
+Attribute BytecodeReader::parseAttribute(EncodingReader &reader) {
+  uint64_t attrIdx;
+  if (reader.parseVarInt(attrIdx))
+    return Attribute();
+  return attrTypeReader.resolveAttribute(attrIdx);
+}
+
+Type BytecodeReader::parseType(EncodingReader &reader) {
+  uint64_t typeIdx;
+  if (reader.parseVarInt(typeIdx))
+    return Type();
+  return attrTypeReader.resolveType(typeIdx);
+}
+
+//===----------------------------------------------------------------------===//
+// IR Section
+
+LogicalResult BytecodeReader::parseIRSection(ArrayRef<uint8_t> sectionData,
+                                             Block *block) {
+  EncodingReader reader(sectionData, fileLoc);
+
+  // A stack of operation regions currently being read from the bytecode.
+  std::vector<RegionReadState> regionStack;
+
+  // Parse the top-level block using a temporary module operation.
+  OwningOpRef<ModuleOp> moduleOp = ModuleOp::create(fileLoc);
+  regionStack.emplace_back(*moduleOp, /*isIsolatedFromAbove=*/true);
+  regionStack.back().curBlocks.push_back(moduleOp->getBody());
+  regionStack.back().curBlock = regionStack.back().curRegion->begin();
+  if (failed(parseBlock(reader, regionStack.back())))
+    return failure();
+  valueScopes.emplace_back(ValueScope());
+  valueScopes.back().push(regionStack.back());
+
+  // Iteratively parse regions until everything has been resolved.
+  while (!regionStack.empty())
+    if (failed(parseRegions(reader, regionStack, regionStack.back())))
+      return failure();
+  if (!forwardRefOps.empty()) {
+    return reader.emitError(
+        "not all forward unresolved forward operand references");
+  }
+
+  // Verify that the parsed operations are valid.
+  if (failed(verify(*moduleOp)))
+    return failure();
+
+  // Splice the parsed operations over to the provided top-level block.
+  auto &parsedOps = moduleOp->getBody()->getOperations();
+  auto &destOps = block->getOperations();
+  destOps.splice(destOps.empty() ? destOps.end() : std::prev(destOps.end()),
+                 parsedOps, parsedOps.begin(), parsedOps.end());
+  return success();
+}
+
+LogicalResult
+BytecodeReader::parseRegions(EncodingReader &reader,
+                             std::vector<RegionReadState> &regionStack,
+                             RegionReadState &readState) {
+  // Read the regions of this operation.
+  for (; readState.curRegion != readState.endRegion; ++readState.curRegion) {
+    // If the current block hasn't been setup yet, parse the header for this
+    // region.
+    if (readState.curBlock == Region::iterator()) {
+      if (failed(parseRegion(reader, readState)))
+        return failure();
+
+      // If the region is empty, there is nothing to more to do.
+      if (readState.curRegion->empty())
+        continue;
+    }
+
+    // Parse the blocks within the region.
+    do {
+      while (readState.numOpsRemaining--) {
+        // Read in the next operation. We don't read its regions directly, we
+        // handle those afterwards as necessary.
+        bool isIsolatedFromAbove = false;
+        FailureOr<Operation *> op =
+            parseOpWithoutRegions(reader, readState, isIsolatedFromAbove);
+        if (failed(op))
+          return failure();
+
+        // If the op has regions, add it to the stack for processing.
+        if ((*op)->getNumRegions()) {
+          regionStack.emplace_back(*op, isIsolatedFromAbove);
+
+          // If the op is isolated from above, push a new value scope.
+          if (isIsolatedFromAbove)
+            valueScopes.emplace_back(ValueScope());
+          return success();
+        }
+      }
+
+      // Move to the next block of the region.
+      if (++readState.curBlock == readState.curRegion->end())
+        break;
+      if (failed(parseBlock(reader, readState)))
+        return failure();
+    } while (true);
+
+    // Reset the current block and any values reserved for this region.
+    readState.curBlock = {};
+    valueScopes.back().pop(readState);
+  }
+
+  // When the regions have been fully parsed, pop them off of the read stack. If
+  // the regions were isolated from above, we also pop the last value scope.
+  regionStack.pop_back();
+  if (readState.isIsolatedFromAbove)
+    valueScopes.pop_back();
+  return success();
+}
+
+FailureOr<Operation *>
+BytecodeReader::parseOpWithoutRegions(EncodingReader &reader,
+                                      RegionReadState &readState,
+                                      bool &isIsolatedFromAbove) {
+  // Parse the name of the operation.
+  FailureOr<OperationName> opName = parseOpName(reader);
+  if (failed(opName))
+    return failure();
+
+  // Parse the operation mask, which indicates which components of the operation
+  // are present.
+  uint8_t opMask;
+  if (reader.parseByte(opMask))
+    return failure();
+
+  /// Parse the location.
+  LocationAttr opLoc = parseAttribute<LocationAttr>(reader);
+  if (!opLoc)
+    return failure();
+
+  // With the location and name resolved, we can start building the operation
+  // state.
+  OperationState opState(opLoc, *opName);
+
+  // Parse the attributes of the operation.
+  if (opMask & bytecode::OpEncodingMask::kHasAttrs) {
+    DictionaryAttr dictAttr = parseAttribute<DictionaryAttr>(reader);
+    if (!dictAttr)
+      return failure();
+    opState.attributes = dictAttr;
+  }
+
+  /// Parse the results of the operation.
+  if (opMask & bytecode::OpEncodingMask::kHasResults) {
+    uint64_t numResults;
+    if (reader.parseVarInt(numResults))
+      return failure();
+    opState.types.resize(numResults);
+    for (int i = 0, e = numResults; i < e; ++i)
+      if (!(opState.types[i] = parseType(reader)))
+        return failure();
+  }
+
+  /// Parse the operands of the operation.
+  if (opMask & bytecode::OpEncodingMask::kHasOperands) {
+    uint64_t numOperands;
+    if (reader.parseVarInt(numOperands))
+      return failure();
+    opState.operands.resize(numOperands);
+    for (int i = 0, e = numOperands; i < e; ++i)
+      if (!(opState.operands[i] = parseOperand(reader)))
+        return failure();
+  }
+
+  /// Parse the successors of the operation.
+  if (opMask & bytecode::OpEncodingMask::kHasSuccessors) {
+    uint64_t numSuccs;
+    if (reader.parseVarInt(numSuccs))
+      return failure();
+    opState.successors.reserve(numSuccs);
+    for (int i = 0, e = numSuccs; i < e; ++i) {
+      uint64_t succID;
+      if (reader.parseVarInt(succID))
+        return failure();
+      if (succID >= readState.curBlocks.size())
+        return reader.emitError("invalid successor index: ", succID);
+      opState.successors.push_back(readState.curBlocks[succID]);
+    }
+  }
+
+  /// Parse the regions of the operation.
+  if (opMask & bytecode::OpEncodingMask::kHasInlineRegions) {
+    uint64_t numRegions;
+    if (reader.parseVarInt(numRegions))
+      return failure();
+
+    // Extract if the regions are isolated from the numRegions varint, which is
+    // encoded in the low bit.
+    isIsolatedFromAbove = numRegions & 1;
+    numRegions >>= 1;
+
+    opState.regions.reserve(numRegions);
+    for (int i = 0, e = numRegions; i < e; ++i)
+      opState.regions.push_back(std::make_unique<Region>());
+  }
+
+  // Create the operation at the back of the current block.
+  Operation *op = Operation::create(opState);
+  readState.curBlock->push_back(op);
+
+  // If the operation had results, update the value references.
+  if (op->getNumResults() && failed(defineValues(reader, op->getResults())))
+    return failure();
+
+  return op;
+}
+
+LogicalResult BytecodeReader::parseRegion(EncodingReader &reader,
+                                          RegionReadState &readState) {
+  // Parse the number of blocks in the region.
+  uint64_t numBlocks;
+  if (reader.parseVarInt(numBlocks))
+    return failure();
+
+  // If the region is empty, there is nothing else to do.
+  if (numBlocks == 0)
+    return success();
+
+  // Parse the number of values defined in this region.
+  uint64_t numValues;
+  if (reader.parseVarInt(numValues))
+    return failure();
+  readState.numValues = numValues;
+
+  // Create the blocks within this region. We do this before processing so that
+  // we can rely on the blocks existing when creating operations.
+  readState.curBlocks.clear();
+  readState.curBlocks.reserve(numBlocks);
+  for (uint64_t i = 0; i < numBlocks; ++i) {
+    readState.curBlocks.push_back(new Block());
+    readState.curRegion->push_back(readState.curBlocks.back());
+  }
+
+  // Prepare the current value scope for this region.
+  valueScopes.back().push(readState);
+
+  // Parse the entry block of the region.
+  readState.curBlock = readState.curRegion->begin();
+  return parseBlock(reader, readState);
+}
+
+LogicalResult BytecodeReader::parseBlock(EncodingReader &reader,
+                                         RegionReadState &readState) {
+  // Parse the number of operations in the block.
+  uint64_t numOps;
+  if (reader.parseVarInt(numOps))
+    return failure();
+  // Extract out the low-bit of the operation count, which indicates if the
+  // block has arguments.
+  bool hasArgs = numOps & 1;
+  readState.numOpsRemaining = numOps >> 1;
+
+  // Parse the arguments of the block.
+  if (hasArgs && failed(parseBlockArguments(reader, &*readState.curBlock)))
+    return failure();
+
+  // We don't parse the operations of the block here, that's done elsewhere.
+  return success();
+}
+
+LogicalResult BytecodeReader::parseBlockArguments(EncodingReader &reader,
+                                                  Block *block) {
+  // Parse the value ID for the first argument, and the number of arguments.
+  uint64_t numArgs;
+  if (reader.parseVarInt(numArgs))
+    return failure();
+
+  SmallVector<Type> argTypes;
+  SmallVector<Location> argLocs;
+  argTypes.reserve(numArgs);
+  argLocs.reserve(numArgs);
+
+  while (numArgs--) {
+    Type argType = parseType(reader);
+    if (!argType)
+      return failure();
+    LocationAttr argLoc = parseAttribute<LocationAttr>(reader);
+    if (!argLoc)
+      return failure();
+
+    argTypes.push_back(argType);
+    argLocs.push_back(argLoc);
+  }
+  block->addArguments(argTypes, argLocs);
+  return defineValues(reader, block->getArguments());
+}
+
+//===----------------------------------------------------------------------===//
+// String Section
+
+LogicalResult
+BytecodeReader::parseStringSection(ArrayRef<uint8_t> sectionData) {
+  EncodingReader stringReader(sectionData, fileLoc);
+
+  // Parse the number of strings in the section.
+  uint64_t numStrings;
+  if (failed(stringReader.parseVarInt(numStrings)))
+    return failure();
+  strings.resize(numStrings);
+
+  // Parse each of the strings. The sizes of the strings are encoded in reverse
+  // order, so that's the order we populate the table.
+  size_t stringDataEndOffset = sectionData.size();
+  size_t totalStringDataSize = 0;
+  for (StringRef &string : llvm::reverse(strings)) {
+    uint64_t stringSize;
+    if (failed(stringReader.parseVarInt(stringSize)))
+      return failure();
+    if (stringDataEndOffset < stringSize) {
+      return stringReader.emitError(
+          "string size exceeds the available data size");
+    }
+
+    // Extract the string from the data, dropping the null character.
+    size_t stringOffset = stringDataEndOffset - stringSize;
+    string = StringRef(
+        reinterpret_cast<const char *>(sectionData.data() + stringOffset),
+        stringSize - 1);
+    stringDataEndOffset = stringOffset;
+
+    // Update the total string data size.
+    totalStringDataSize += stringSize;
+  }
+
+  // Check that the only remaining data was for the strings
+  if (stringReader.size() != totalStringDataSize) {
+    return stringReader.emitError("unexpected trailing data between the "
+                                  "offsets for strings and their data");
+  }
+  return success();
+}
+
+LogicalResult BytecodeReader::parseSharedString(EncodingReader &reader,
+                                                StringRef &result) {
+  uint64_t stringIdx;
+  if (failed(reader.parseVarInt(stringIdx)))
+    return failure();
+  if (stringIdx >= strings.size())
+    return reader.emitError("invalid string index: ", stringIdx);
+  result = strings[stringIdx];
+  return success();
+}
+
+//===----------------------------------------------------------------------===//
+// Value Processing
+
+Value BytecodeReader::parseOperand(EncodingReader &reader) {
+  uint64_t valueIdx;
+  if (failed(reader.parseVarInt(valueIdx)))
+    return nullptr;
+  std::vector<Value> &values = valueScopes.back().values;
+  if (valueIdx >= values.size())
+    return (void)reader.emitError("invalid value index: ", valueIdx), Value();
+
+  // Resolve it, or create a new forward reference if necessary.
+  Value &value = values[valueIdx];
+  if (!value)
+    value = createForwardRef();
+  return value;
+}
+
+LogicalResult BytecodeReader::defineValues(EncodingReader &reader,
+                                           ValueRange newValues) {
+  ValueScope &valueScope = valueScopes.back();
+  std::vector<Value> &values = valueScope.values;
+
+  unsigned &valueID = valueScope.nextValueIDs.back();
+  unsigned valueIDEnd = valueID + newValues.size();
+  if (valueIDEnd > values.size()) {
+    return reader.emitError(
+        "value index range was outside of the expected range for "
+        "the parent region, got [",
+        valueID, ", ", valueIDEnd, "), but the maximum index was ",
+        values.size() - 1);
+  }
+
+  // Assign the values and update any forward references.
+  for (unsigned i = 0, e = newValues.size(); i != e; ++i, ++valueID) {
+    Value newValue = newValues[i];
+
+    // Check to see if a definition for this value already exists.
+    if (Value oldValue = std::exchange(values[valueID], newValue)) {
+      Operation *forwardRefOp = oldValue.getDefiningOp();
+      if (!forwardRefOp || forwardRefOp->getBlock() != &forwardRefOps) {
+        return reader.emitError("value index ", valueID,
+                                " was already defined");
+      }
+
+      oldValue.replaceAllUsesWith(newValue);
+      forwardRefOp->moveBefore(&openForwardRefOps, openForwardRefOps.end());
+    }
+  }
+  return success();
+}
+
+Value BytecodeReader::createForwardRef() {
+  // Check for an avaliable existing operation to use. Otherwise, create a new
+  // fake operation to use for the reference.
+  if (!openForwardRefOps.empty()) {
+    Operation *op = &openForwardRefOps.back();
+    op->moveBefore(&forwardRefOps, forwardRefOps.end());
+  } else {
+    forwardRefOps.push_back(Operation::create(forwardRefOpState));
+  }
+  return forwardRefOps.back().getResult(0);
+}
+
+//===----------------------------------------------------------------------===//
+// Entry Points
+//===----------------------------------------------------------------------===//
+
+bool mlir::isBytecode(llvm::MemoryBufferRef buffer) {
+  return buffer.getBuffer().startswith("ML\xefR");
+}
+
+LogicalResult mlir::readBytecodeFile(llvm::MemoryBufferRef buffer, Block *block,
+                                     const ParserConfig &config) {
+  Location sourceFileLoc =
+      FileLineColLoc::get(config.getContext(), buffer.getBufferIdentifier(),
+                          /*line=*/0, /*column=*/0);
+  if (!isBytecode(buffer)) {
+    return emitError(sourceFileLoc,
+                     "input buffer is not an MLIR bytecode file");
+  }
+
+  BytecodeReader reader(sourceFileLoc, config);
+  return reader.read(buffer, block);
+}

mlir/lib/Bytecode/Reader/CMakeLists.txt

+add_mlir_library(MLIRBytecodeReader
+  BytecodeReader.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Bytecode
+
+  LINK_LIBS PUBLIC
+  MLIRAsmParser
+  MLIRIR
+  MLIRSupport
+  )

mlir/lib/Bytecode/Writer/BytecodeWriter.cpp

+//===- BytecodeWriter.cpp - MLIR Bytecode Writer --------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Bytecode/BytecodeWriter.h"
+#include "../Encoding.h"
+#include "IRNumbering.h"
+#include "mlir/IR/BuiltinDialect.h"
+#include "mlir/IR/OpImplementation.h"
+#include "llvm/ADT/CachedHashString.h"
+#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/SmallString.h"
+#include "llvm/Support/Debug.h"
+#include <random>
+
+#define DEBUG_TYPE "mlir-bytecode-writer"
+
+using namespace mlir;
+using namespace mlir::bytecode::detail;
+
+//===----------------------------------------------------------------------===//
+// EncodingEmitter
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// This class functions as the underlying encoding emitter for the bytecode
+/// writer. This class is a bit different compared to other types of encoders;
+/// it does not use a single buffer, but instead may contain several buffers
+/// (some owned by the writer, and some not) that get concatted during the final
+/// emission.
+class EncodingEmitter {
+public:
+  EncodingEmitter() = default;
+  EncodingEmitter(const EncodingEmitter &) = delete;
+  EncodingEmitter &operator=(const EncodingEmitter &) = delete;
+
+  /// Write the current contents to the provided stream.
+  void writeTo(raw_ostream &os) const;
+
+  /// Return the current size of the encoded buffer.
+  size_t size() const { return prevResultSize + currentResult.size(); }
+
+  //===--------------------------------------------------------------------===//
+  // Emission
+  //===--------------------------------------------------------------------===//
+
+  /// Return a raw pointer into the result buffer at the specified offset.
+  uint8_t *getRawPointer(uint64_t offset) {
+    assert(offset < size() && offset >= prevResultSize &&
+           "cannot get pointer to previously emitted data");
+    return currentResult.data() + (offset - prevResultSize);
+  }
+
+  //===--------------------------------------------------------------------===//
+  // Integer Emission
+
+  /// Emit a single byte.
+  template <typename T>
+  void emitByte(T byte) {
+    currentResult.push_back(static_cast<uint8_t>(byte));
+  }
+
+  /// Emit a range of bytes.
+  void emitBytes(ArrayRef<uint8_t> bytes) {
+    llvm::append_range(currentResult, bytes);
+  }
+
+  /// Emit a variable length integer. The first encoded byte contains a prefix
+  /// in the low bits indicating the encoded length of the value. This length
+  /// prefix is a bit sequence of '0's followed by a '1'. The number of '0' bits
+  /// indicate the number of _additional_ bytes (not including the prefix byte).
+  /// All remaining bits in the first byte, along with all of the bits in
+  /// additional bytes, provide the value of the integer encoded in
+  /// little-endian order.
+  void emitVarInt(uint64_t value) {
+    // In the most common case, the value can be represented in a single byte.
+    // Given how hot this case is, explicitly handle that here.
+    if ((value >> 7) == 0)
+      return emitByte((value << 1) | 0x1);
+    emitMultiByteVarInt(value);
+  }
+
+  //===--------------------------------------------------------------------===//
+  // String Emission
+
+  /// Emit the given string as a nul terminated string.
+  void emitNulTerminatedString(StringRef str) {
+    emitString(str);
+    emitByte(0);
+  }
+
+  /// Emit the given string without a nul terminator.
+  void emitString(StringRef str) {
+    emitBytes({reinterpret_cast<const uint8_t *>(str.data()), str.size()});
+  }
+
+  //===--------------------------------------------------------------------===//
+  // Section Emission
+
+  /// Emit a nested section of the given code, whose contents are encoded in the
+  /// provided emitter.
+  void emitSection(bytecode::Section::ID code, EncodingEmitter &&emitter) {
+    // Emit the section code and length.
+    emitByte(code);
+    emitVarInt(emitter.size());
+
+    // Push our current buffer and then merge the provided section body into
+    // ours.
+    appendResult(std::move(currentResult));
+    for (std::vector<uint8_t> &result : emitter.prevResultStorage)
+      appendResult(std::move(result));
+    appendResult(std::move(emitter.currentResult));
+  }
+
+private:
+  /// Emit the given value using a variable width encoding. This method is a
+  /// fallback when the number of bytes needed to encode the value is greater
+  /// than 1. We mark it noinline here so that the single byte hot path isn't
+  /// pessimized.
+  LLVM_ATTRIBUTE_NOINLINE void emitMultiByteVarInt(uint64_t value);
+
+  /// Append a new result buffer to the current contents.
+  void appendResult(std::vector<uint8_t> &&result) {
+    prevResultSize += result.size();
+    prevResultStorage.emplace_back(std::move(result));
+    prevResultList.emplace_back(prevResultStorage.back());
+  }
+
+  /// The result of the emitter currently being built. We refrain from building
+  /// a single buffer to simplify emitting sections, large data, and more. The
+  /// result is thus represented using multiple distinct buffers, some of which
+  /// we own (via prevResultStorage), and some of which are just pointers into
+  /// externally owned buffers.
+  std::vector<uint8_t> currentResult;
+  std::vector<ArrayRef<uint8_t>> prevResultList;
+  std::vector<std::vector<uint8_t>> prevResultStorage;
+
+  /// An up-to-date total size of all of the buffers within `prevResultList`.
+  /// This enables O(1) size checks of the current encoding.
+  size_t prevResultSize = 0;
+};
+
+/// A simple raw_ostream wrapper around a EncodingEmitter. This removes the need
+/// to go through an intermediate buffer when interacting with code that wants a
+/// raw_ostream.
+class raw_emitter_ostream : public raw_ostream {
+public:
+  explicit raw_emitter_ostream(EncodingEmitter &emitter) : emitter(emitter) {
+    SetUnbuffered();
+  }
+
+private:
+  void write_impl(const char *ptr, size_t size) override {
+    emitter.emitBytes({reinterpret_cast<const uint8_t *>(ptr), size});
+  }
+  uint64_t current_pos() const override { return emitter.size(); }
+
+  /// The section being emitted to.
+  EncodingEmitter &emitter;
+};
+} // namespace
+
+void EncodingEmitter::writeTo(raw_ostream &os) const {
+  for (auto &prevResult : prevResultList)
+    os.write((const char *)prevResult.data(), prevResult.size());
+  os.write((const char *)currentResult.data(), currentResult.size());
+}
+
+void EncodingEmitter::emitMultiByteVarInt(uint64_t value) {
+  // Compute the number of bytes needed to encode the value. Each byte can hold
+  // up to 7-bits of data. We only check up to the number of bits we can encode
+  // in the first byte (8).
+  uint64_t it = value >> 7;
+  for (size_t numBytes = 2; numBytes < 9; ++numBytes) {
+    if (LLVM_LIKELY(it >>= 7) == 0) {
+      uint64_t encodedValue = (value << 1) | 0x1;
+      encodedValue <<= (numBytes - 1);
+      emitBytes({reinterpret_cast<uint8_t *>(&encodedValue), numBytes});
+      return;
+    }
+  }
+
+  // If the value is too large to encode in a single byte, emit a special all
+  // zero marker byte and splat the value directly.
+  emitByte(0);
+  emitBytes({reinterpret_cast<uint8_t *>(&value), sizeof(value)});
+}
+
+//===----------------------------------------------------------------------===//
+// Bytecode Writer
+//===----------------------------------------------------------------------===//
+
+namespace {
+class BytecodeWriter {
+public:
+  BytecodeWriter(Operation *op) : numberingState(op) {}
+
+  /// Write the bytecode for the given root operation.
+  void write(Operation *rootOp, raw_ostream &os, StringRef producer);
+
+private:
+  //===--------------------------------------------------------------------===//
+  // Dialects
+
+  void writeDialectSection(EncodingEmitter &emitter);
+
+  //===--------------------------------------------------------------------===//
+  // Attributes and Types
+
+  void writeAttrTypeSection(EncodingEmitter &emitter);
+
+  //===--------------------------------------------------------------------===//
+  // Operations
+
+  void writeBlock(EncodingEmitter &emitter, Block *block);
+  void writeOp(EncodingEmitter &emitter, Operation *op);
+  void writeRegion(EncodingEmitter &emitter, Region *region);
+  void writeIRSection(EncodingEmitter &emitter, Operation *op);
+
+  //===--------------------------------------------------------------------===//
+  // Strings
+
+  void writeStringSection(EncodingEmitter &emitter);
+
+  /// Emit the given string as an index to a shared string, contained within
+  /// the string section. If an inline string is desired, emit it using the
+  /// emitter directly instead of calling this method.
+  void emitSharedString(EncodingEmitter &emitter, StringRef str);
+
+  //===--------------------------------------------------------------------===//
+  // Fields
+
+  /// The IR numbering state generated for the root operation.
+  IRNumberingState numberingState;
+
+  /// A set of strings referenced within the bytecode. The value of the map is
+  /// unused.
+  llvm::MapVector<llvm::CachedHashStringRef, size_t> strings;
+};
+} // namespace
+
+void BytecodeWriter::write(Operation *rootOp, raw_ostream &os,
+                           StringRef producer) {
+  EncodingEmitter emitter;
+
+  // Emit the bytecode file header. This is how we identify the output as a
+  // bytecode file.
+  emitter.emitString("ML\xefR");
+
+  // Emit the bytecode version.
+  emitter.emitVarInt(bytecode::kVersion);
+
+  // Emit the producer.
+  emitter.emitNulTerminatedString(producer);
+
+  // Emit the dialect section.
+  writeDialectSection(emitter);
+
+  // Emit the attributes and types section.
+  writeAttrTypeSection(emitter);
+
+  // Emit the IR section.
+  writeIRSection(emitter, rootOp);
+
+  // Emit the string section.
+  writeStringSection(emitter);
+
+  // Write the generated bytecode to the provided output stream.
+  emitter.writeTo(os);
+}
+
+//===----------------------------------------------------------------------===//
+// Dialects
+
+void BytecodeWriter::writeDialectSection(EncodingEmitter &emitter) {
+  EncodingEmitter dialectEmitter;
+
+  // Emit the referenced dialects.
+  for (DialectNumbering &dialect : numberingState.getDialects()) {
+    // Emit the dialect name.
+    emitSharedString(dialectEmitter, dialect.name);
+
+    // Emit the number of attributes and types emitted for this dialect.
+    dialectEmitter.emitVarInt(dialect.attributes.size());
+    dialectEmitter.emitVarInt(dialect.types.size());
+
+    // Emit the referenced operation names of this dialect.
+    dialectEmitter.emitVarInt(dialect.opNames.size());
+    for (OpNameNumbering *opName : dialect.opNames)
+      emitSharedString(dialectEmitter, opName->name.stripDialect());
+  }
+
+  emitter.emitSection(bytecode::Section::kDialect, std::move(dialectEmitter));
+}
+
+//===----------------------------------------------------------------------===//
+// Attributes and Types
+
+void BytecodeWriter::writeAttrTypeSection(EncodingEmitter &emitter) {
+  EncodingEmitter attrTypeEmitter;
+  EncodingEmitter offsetEmitter;
+
+  // A functor used to emit an attribute or type entry.
+  uint64_t prevOffset = 0;
+  auto emitAttrOrType = [&](auto value) {
+    // Emit the entry using the textual format.
+    // TODO: Allow dialects to provide more optimal implementations of attribute
+    // and type encodings.
+    attrTypeEmitter.emitByte(bytecode::AttrTypeCode::kAsmForm);
+    raw_emitter_ostream(attrTypeEmitter) << value;
+    attrTypeEmitter.emitByte(0);
+
+    // Record the offset of this entry.
+    uint64_t curOffset = attrTypeEmitter.size();
+    offsetEmitter.emitVarInt(curOffset - prevOffset);
+    prevOffset = curOffset;
+  };
+
+  // Emit the attribute and type entries for each dialect.
+  for (DialectNumbering &dialect : numberingState.getDialects())
+    for (AttributeNumbering *attr : dialect.attributes)
+      emitAttrOrType(attr->getValue());
+  for (DialectNumbering &dialect : numberingState.getDialects())
+    for (TypeNumbering *type : dialect.types)
+      emitAttrOrType(type->getValue());
+
+  // Emit the sections to the stream.
+  emitter.emitSection(bytecode::Section::kAttrTypeOffset,
+                      std::move(offsetEmitter));
+  emitter.emitSection(bytecode::Section::kAttrType, std::move(attrTypeEmitter));
+}
+
+//===----------------------------------------------------------------------===//
+// Operations
+
+void BytecodeWriter::writeBlock(EncodingEmitter &emitter, Block *block) {
+  ArrayRef<BlockArgument> args = block->getArguments();
+  bool hasArgs = !args.empty();
+
+  // Emit the number of operations in this block, and if it has arguments. We
+  // use the low bit of the operation count to indicate if the block has
+  // arguments.
+  unsigned numOps = numberingState.getOperationCount(block);
+  emitter.emitVarInt((numOps << 1) | (hasArgs ? 1 : 0));
+
+  // Emit the arguments of the block.
+  if (hasArgs) {
+    emitter.emitVarInt(args.size());
+    for (BlockArgument arg : args) {
+      emitter.emitVarInt(numberingState.getNumber(arg.getType()));
+      emitter.emitVarInt(numberingState.getNumber(arg.getLoc()));
+    }
+  }
+
+  // Emit the operations within the block.
+  for (Operation &op : *block)
+    writeOp(emitter, &op);
+}
+
+void BytecodeWriter::writeOp(EncodingEmitter &emitter, Operation *op) {
+  emitter.emitVarInt(numberingState.getNumber(op->getName()));
+
+  // Emit a mask for the operation components. We need to fill this in later
+  // (when we actually know what needs to be emitted), so emit a placeholder for
+  // now.
+  uint64_t maskOffset = emitter.size();
+  uint8_t opEncodingMask = 0;
+  emitter.emitByte(0);
+
+  // Emit the location for this operation.
+  emitter.emitVarInt(numberingState.getNumber(op->getLoc()));
+
+  // Emit the attributes of this operation.
+  DictionaryAttr attrs = op->getAttrDictionary();
+  if (!attrs.empty()) {
+    opEncodingMask |= bytecode::OpEncodingMask::kHasAttrs;
+    emitter.emitVarInt(numberingState.getNumber(op->getAttrDictionary()));
+  }
+
+  // Emit the result types of the operation.
+  if (unsigned numResults = op->getNumResults()) {
+    opEncodingMask |= bytecode::OpEncodingMask::kHasResults;
+    emitter.emitVarInt(numResults);
+    for (Type type : op->getResultTypes())
+      emitter.emitVarInt(numberingState.getNumber(type));
+  }
+
+  // Emit the operands of the operation.
+  if (unsigned numOperands = op->getNumOperands()) {
+    opEncodingMask |= bytecode::OpEncodingMask::kHasOperands;
+    emitter.emitVarInt(numOperands);
+    for (Value operand : op->getOperands())
+      emitter.emitVarInt(numberingState.getNumber(operand));
+  }
+
+  // Emit the successors of the operation.
+  if (unsigned numSuccessors = op->getNumSuccessors()) {
+    opEncodingMask |= bytecode::OpEncodingMask::kHasSuccessors;
+    emitter.emitVarInt(numSuccessors);
+    for (Block *successor : op->getSuccessors())
+      emitter.emitVarInt(numberingState.getNumber(successor));
+  }
+
+  // Check for regions.
+  unsigned numRegions = op->getNumRegions();
+  if (numRegions)
+    opEncodingMask |= bytecode::OpEncodingMask::kHasInlineRegions;
+
+  // Update the mask for the operation.
+  *emitter.getRawPointer(maskOffset) = opEncodingMask;
+
+  // With the mask emitted, we can now emit the regions of the operation. We do
+  // this after mask emission to avoid offset complications that may arise by
+  // emitting the regions first (e.g. if the regions are huge, backpatching the
+  // op encoding mask is more annoying).
+  if (numRegions) {
+    bool isIsolatedFromAbove = op->hasTrait<OpTrait::IsIsolatedFromAbove>();
+    emitter.emitVarInt((numRegions << 1) | (isIsolatedFromAbove ? 1 : 0));
+
+    for (Region &region : op->getRegions())
+      writeRegion(emitter, &region);
+  }
+}
+
+void BytecodeWriter::writeRegion(EncodingEmitter &emitter, Region *region) {
+  // If the region is empty, we only need to emit the number of blocks (which is
+  // zero).
+  if (region->empty())
+    return emitter.emitVarInt(/*numBlocks*/ 0);
+
+  // Emit the number of blocks and values within the region.
+  unsigned numBlocks, numValues;
+  std::tie(numBlocks, numValues) = numberingState.getBlockValueCount(region);
+  emitter.emitVarInt(numBlocks);
+  emitter.emitVarInt(numValues);
+
+  // Emit the blocks within the region.
+  for (Block &block : *region)
+    writeBlock(emitter, &block);
+}
+
+void BytecodeWriter::writeIRSection(EncodingEmitter &emitter, Operation *op) {
+  EncodingEmitter irEmitter;
+
+  // Write the IR section the same way as a block with no arguments. Note that
+  // the low-bit of the operation count for a block is used to indicate if the
+  // block has arguments, which in this case is always false.
+  uint64_t numOps = 1;
+  irEmitter.emitVarInt(numOps << 1);
+
+  // Emit the operations.
+  writeOp(irEmitter, op);
+
+  emitter.emitSection(bytecode::Section::kIR, std::move(irEmitter));
+}
+
+//===----------------------------------------------------------------------===//
+// Strings
+
+void BytecodeWriter::writeStringSection(EncodingEmitter &emitter) {
+  EncodingEmitter stringEmitter;
+  stringEmitter.emitVarInt(strings.size());
+
+  // Emit the sizes in reverse order, so that we don't need to backpatch an
+  // offset to the string data or have a separate section.
+  for (const auto &it : llvm::reverse(strings))
+    stringEmitter.emitVarInt(it.first.size() + 1);
+  // Emit the string data itself.
+  for (const auto &it : strings)
+    stringEmitter.emitNulTerminatedString(it.first.val());
+
+  emitter.emitSection(bytecode::Section::kString, std::move(stringEmitter));
+}
+
+void BytecodeWriter::emitSharedString(EncodingEmitter &emitter, StringRef str) {
+  auto it = strings.insert({llvm::CachedHashStringRef(str), strings.size()});
+  return emitter.emitVarInt(it.first->second);
+}
+
+//===----------------------------------------------------------------------===//
+// Entry Points
+//===----------------------------------------------------------------------===//
+
+void mlir::writeBytecodeToFile(Operation *op, raw_ostream &os,
+                               StringRef producer) {
+  BytecodeWriter writer(op);
+  writer.write(op, os, producer);
+}

mlir/lib/Bytecode/Writer/CMakeLists.txt

+add_mlir_library(MLIRBytecodeWriter
+  BytecodeWriter.cpp
+  IRNumbering.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Bytecode
+
+  LINK_LIBS PUBLIC
+  MLIRIR
+  MLIRSupport
+  )

mlir/lib/Bytecode/Writer/IRNumbering.h

+//===- IRNumbering.h - MLIR bytecode IR numbering ---------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains various utilities that number IR structures in preparation
+// for bytecode emission.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LIB_MLIR_BYTECODE_WRITER_IRNUMBERING_H
+#define LIB_MLIR_BYTECODE_WRITER_IRNUMBERING_H
+
+#include "mlir/IR/OperationSupport.h"
+#include "llvm/ADT/MapVector.h"
+
+namespace mlir {
+class BytecodeWriterConfig;
+
+namespace bytecode {
+namespace detail {
+struct DialectNumbering;
+
+//===----------------------------------------------------------------------===//
+// Attribute and Type Numbering
+//===----------------------------------------------------------------------===//
+
+/// This class represents a numbering entry for an Attribute or Type.
+struct AttrTypeNumbering {
+  AttrTypeNumbering(PointerUnion<Attribute, Type> value) : value(value) {}
+
+  /// The concrete value.
+  PointerUnion<Attribute, Type> value;
+
+  /// The number assigned to this value.
+  unsigned number = 0;
+
+  /// The dialect of this value.
+  DialectNumbering *dialect = nullptr;
+};
+struct AttributeNumbering : public AttrTypeNumbering {
+  AttributeNumbering(Attribute value) : AttrTypeNumbering(value) {}
+  Attribute getValue() const { return value.get<Attribute>(); }
+};
+struct TypeNumbering : public AttrTypeNumbering {
+  TypeNumbering(Type value) : AttrTypeNumbering(value) {}
+  Type getValue() const { return value.get<Type>(); }
+};
+
+//===----------------------------------------------------------------------===//
+// OpName Numbering
+//===----------------------------------------------------------------------===//
+
+/// This class represents the numbering entry of an operation name.
+struct OpNameNumbering {
+  OpNameNumbering(OperationName name) : name(name) {}
+
+  /// The concrete name.
+  OperationName name;
+
+  /// The number assigned to this name.
+  unsigned number = 0;
+};
+
+//===----------------------------------------------------------------------===//
+// Dialect Numbering
+//===----------------------------------------------------------------------===//
+
+/// This class represents a numbering entry for an Dialect.
+struct DialectNumbering {
+  DialectNumbering(StringRef name, unsigned number)
+      : name(name), number(number) {}
+
+  /// The namespace of the dialect.
+  StringRef name;
+
+  /// The number assigned to the dialect.
+  unsigned number;
+
+  /// The loaded dialect, or nullptr if the dialect isn't loaded.
+  Dialect *dialect = nullptr;
+
+  /// Numbered sub-components of the dialect to be emitted.
+  std::vector<OpNameNumbering *> opNames;
+  std::vector<AttributeNumbering *> attributes;
+  std::vector<TypeNumbering *> types;
+};
+
+//===----------------------------------------------------------------------===//
+// IRNumberingState
+//===----------------------------------------------------------------------===//
+
+/// This class manages numbering IR entities in preparation of bytecode
+/// emission.
+class IRNumberingState {
+public:
+  IRNumberingState(Operation *op);
+
+  /// Return the numbered dialects.
+  auto getDialects() {
+    return llvm::make_pointee_range(llvm::make_second_range(dialects));
+  }
+
+  /// Return the number for the given IR unit.
+  unsigned getNumber(Attribute attr) {
+    assert(attrs.count(attr) && "attribute not numbered");
+    return attrs[attr]->number;
+  }
+  unsigned getNumber(Block *block) {
+    assert(blockIDs.count(block) && "block not numbered");
+    return blockIDs[block];
+  }
+  unsigned getNumber(OperationName opName) {
+    assert(opNames.count(opName) && "opName not numbered");
+    return opNames[opName]->number;
+  }
+  unsigned getNumber(Type type) {
+    assert(types.count(type) && "type not numbered");
+    return types[type]->number;
+  }
+  unsigned getNumber(Value value) {
+    assert(valueIDs.count(value) && "value not numbered");
+    return valueIDs[value];
+  }
+
+  /// Return the block and value counts of the given region.
+  std::pair<unsigned, unsigned> getBlockValueCount(Region *region) {
+    assert(regionBlockValueCounts.count(region) && "value not numbered");
+    return regionBlockValueCounts[region];
+  }
+
+  /// Return the number of operations in the given block.
+  unsigned getOperationCount(Block *block) {
+    assert(blockOperationCounts.count(block) && "block not numbered");
+    return blockOperationCounts[block];
+  }
+
+private:
+  /// Number the given IR unit for bytecode emission.
+  void number(Attribute attr);
+  void number(Block &block);
+  DialectNumbering &numberDialect(Dialect *dialect);
+  DialectNumbering &numberDialect(StringRef dialect);
+  void number(Operation &op);
+  void number(OperationName opName);
+  void number(Region &region);
+  void number(Type type);
+
+  /// Mapping from IR to the respective numbering entries.
+  llvm::MapVector<Attribute, AttributeNumbering *> attrs;
+  llvm::MapVector<OperationName, OpNameNumbering *> opNames;
+  llvm::MapVector<Type, TypeNumbering *> types;
+  llvm::MapVector<Dialect *, DialectNumbering *> registeredDialects;
+  llvm::MapVector<StringRef, DialectNumbering *> dialects;
+
+  /// Allocators used for the various numbering entries.
+  llvm::SpecificBumpPtrAllocator<AttributeNumbering> attrAllocator;
+  llvm::SpecificBumpPtrAllocator<DialectNumbering> dialectAllocator;
+  llvm::SpecificBumpPtrAllocator<OpNameNumbering> opNameAllocator;
+  llvm::SpecificBumpPtrAllocator<TypeNumbering> typeAllocator;
+
+  /// The value ID for each Block and Value.
+  DenseMap<Block *, unsigned> blockIDs;
+  DenseMap<Value, unsigned> valueIDs;
+
+  /// The number of operations in each block.
+  DenseMap<Block *, unsigned> blockOperationCounts;
+
+  /// A map from region to the number of blocks and values within that region.
+  DenseMap<Region *, std::pair<unsigned, unsigned>> regionBlockValueCounts;
+
+  /// The next value ID to assign when numbering.
+  unsigned nextValueID = 0;
+};
+} // namespace detail
+} // namespace bytecode
+} // namespace mlir
+
+#endif

mlir/lib/Bytecode/Writer/IRNumbering.cpp

+//===- IRNumbering.cpp - MLIR Bytecode IR numbering -----------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "IRNumbering.h"
+#include "mlir/Bytecode/BytecodeWriter.h"
+#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/IR/OpDefinition.h"
+
+using namespace mlir;
+using namespace mlir::bytecode::detail;
+
+//===----------------------------------------------------------------------===//
+// IR Numbering
+//===----------------------------------------------------------------------===//
+
+IRNumberingState::IRNumberingState(Operation *op) {
+  // Number the root operation.
+  number(*op);
+
+  // Push all of the regions of the root operation onto the worklist.
+  SmallVector<std::pair<Region *, unsigned>, 8> numberContext;
+  for (Region &region : op->getRegions())
+    numberContext.emplace_back(&region, nextValueID);
+
+  // Iteratively process each of the nested regions.
+  while (!numberContext.empty()) {
+    Region *region;
+    std::tie(region, nextValueID) = numberContext.pop_back_val();
+    number(*region);
+
+    // Traverse into nested regions.
+    for (Operation &op : region->getOps()) {
+      // Isolated regions don't share value numbers with their parent, so we can
+      // start numbering these regions at zero.
+      unsigned opFirstValueID =
+          op.hasTrait<OpTrait::IsIsolatedFromAbove>() ? 0 : nextValueID;
+      for (Region &region : op.getRegions())
+        numberContext.emplace_back(&region, opFirstValueID);
+    }
+  }
+
+  // Walk and number the recorded components within each dialect.
+  unsigned attrID = 0, opNameID = 0, typeID = 0;
+  for (DialectNumbering *dialect : llvm::make_second_range(dialects)) {
+    for (AttributeNumbering *attr : dialect->attributes)
+      attr->number = attrID++;
+    for (OpNameNumbering *opName : dialect->opNames)
+      opName->number = opNameID++;
+    for (TypeNumbering *type : dialect->types)
+      type->number = typeID++;
+  }
+}
+
+void IRNumberingState::number(Attribute attr) {
+  auto it = attrs.insert({attr, nullptr});
+  if (!it.second)
+    return;
+  auto *numbering = new (attrAllocator.Allocate()) AttributeNumbering(attr);
+  it.first->second = numbering;
+
+  // Check for OpaqueAttr, which is a dialect-specific attribute that didn't
+  // have a registered dialect when it got created. We don't want to encode this
+  // as the builtin OpaqueAttr, we want to encode it as if the dialect was
+  // actually loaded.
+  if (OpaqueAttr opaqueAttr = attr.dyn_cast<OpaqueAttr>())
+    numbering->dialect = &numberDialect(opaqueAttr.getDialectNamespace());
+  else
+    numbering->dialect = &numberDialect(&attr.getDialect());
+
+  numbering->dialect->attributes.push_back(numbering);
+}
+
+void IRNumberingState::number(Block &block) {
+  // Number the arguments of the block.
+  for (BlockArgument arg : block.getArguments()) {
+    valueIDs.try_emplace(arg, nextValueID++);
+    number(arg.getLoc());
+    number(arg.getType());
+  }
+
+  // Number the operations in this block.
+  unsigned &numOps = blockOperationCounts[&block];
+  for (Operation &op : block) {
+    number(op);
+    ++numOps;
+  }
+}
+
+auto IRNumberingState::numberDialect(Dialect *dialect) -> DialectNumbering & {
+  DialectNumbering *&numbering = registeredDialects[dialect];
+  if (!numbering) {
+    numbering = &numberDialect(dialect->getNamespace());
+    numbering->dialect = dialect;
+  }
+  return *numbering;
+}
+
+auto IRNumberingState::numberDialect(StringRef dialect) -> DialectNumbering & {
+  DialectNumbering *&numbering = dialects[dialect];
+  if (!numbering) {
+    numbering = new (dialectAllocator.Allocate())
+        DialectNumbering(dialect, dialects.size() - 1);
+  }
+  return *numbering;
+}
+
+void IRNumberingState::number(Region &region) {
+  if (region.empty())
+    return;
+  size_t firstValueID = nextValueID;
+
+  // Number the blocks within this region.
+  size_t blockCount = 0;
+  for (auto &it : llvm::enumerate(region)) {
+    blockIDs.try_emplace(&it.value(), it.index());
+    number(it.value());
+    ++blockCount;
+  }
+
+  // Remember the number of blocks and values in this region.
+  regionBlockValueCounts.try_emplace(&region, blockCount,
+                                     nextValueID - firstValueID);
+}
+
+void IRNumberingState::number(Operation &op) {
+  // Number the components of an operation that won't be numbered elsewhere
+  // (e.g. we don't number operands, regions, or successors here).
+  number(op.getName());
+  for (OpResult result : op.getResults()) {
+    valueIDs.try_emplace(result, nextValueID++);
+    number(result.getType());
+  }
+  number(op.getAttrDictionary());
+  number(op.getLoc());
+}
+
+void IRNumberingState::number(OperationName opName) {
+  OpNameNumbering *&numbering = opNames[opName];
+  if (numbering)
+    return;
+  DialectNumbering *dialectNumber = nullptr;
+  if (Dialect *dialect = opName.getDialect())
+    dialectNumber = &numberDialect(dialect);
+  else
+    dialectNumber = &numberDialect(opName.getDialectNamespace());
+  numbering = new (opNameAllocator.Allocate()) OpNameNumbering(opName);
+  dialectNumber->opNames.emplace_back(numbering);
+}
+
+void IRNumberingState::number(Type type) {
+  auto it = types.insert({type, nullptr});
+  if (!it.second)
+    return;
+  auto *numbering = new (typeAllocator.Allocate()) TypeNumbering(type);
+  it.first->second = numbering;
+
+  // Check for OpaqueType, which is a dialect-specific type that didn't have a
+  // registered dialect when it got created. We don't want to encode this as the
+  // builtin OpaqueType, we want to encode it as if the dialect was actually
+  // loaded.
+  if (OpaqueType opaqueType = type.dyn_cast<OpaqueType>())
+    numbering->dialect = &numberDialect(opaqueType.getDialectNamespace());
+  else
+    numbering->dialect = &numberDialect(&type.getDialect());
+
+  numbering->dialect->types.push_back(numbering);
+}

mlir/lib/CMakeLists.txt

 # Enable errors for any global constructors.
 add_flag_if_supported("-Werror=global-constructors" WERROR_GLOBAL_CONSTRUCTOR)
 
 add_subdirectory(Analysis)
 add_subdirectory(AsmParser)
+add_subdirectory(Bytecode)
 add_subdirectory(Conversion)
 add_subdirectory(Dialect)
 add_subdirectory(IR)
 add_subdirectory(Interfaces)
 add_subdirectory(Parser)
 add_subdirectory(Pass)
 add_subdirectory(Reducer)
 add_subdirectory(Rewrite)
 add_subdirectory(Support)
 add_subdirectory(TableGen)
 add_subdirectory(Target)
 add_subdirectory(Tools)
 add_subdirectory(Transforms)
 add_subdirectory(ExecutionEngine)

mlir/lib/Parser/CMakeLists.txt

 add_mlir_library(MLIRParser
   Parser.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Parser
 
   LINK_LIBS PUBLIC
   MLIRAsmParser
+  MLIRBytecodeReader
   MLIRIR
   )

mlir/lib/Parser/Parser.cpp

 //===- Parser.cpp - MLIR Unified Parser Interface -------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the parser for the MLIR textual form.
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Parser/Parser.h"
 #include "mlir/AsmParser/AsmParser.h"
+#include "mlir/Bytecode/BytecodeReader.h"
 #include "llvm/Support/SourceMgr.h"
 
 using namespace mlir;
 
 LogicalResult mlir::parseSourceFile(const llvm::SourceMgr &sourceMgr,
                                     Block *block, const ParserConfig &config,
                                     LocationAttr *sourceFileLoc) {
   const auto *sourceBuf = sourceMgr.getMemoryBuffer(sourceMgr.getMainFileID());
   if (sourceFileLoc) {
     *sourceFileLoc = FileLineColLoc::get(config.getContext(),
                                          sourceBuf->getBufferIdentifier(),
                                          /*line=*/0, /*column=*/0);
   }
+  if (isBytecode(*sourceBuf))
+    return readBytecodeFile(*sourceBuf, block, config);
   return parseAsmSourceFile(sourceMgr, block, config);
 }
 
 LogicalResult mlir::parseSourceFile(llvm::StringRef filename, Block *block,
                                     const ParserConfig &config,
                                     LocationAttr *sourceFileLoc) {
   llvm::SourceMgr sourceMgr;
   return parseSourceFile(filename, sourceMgr, block, config, sourceFileLoc);

mlir/lib/Tools/mlir-opt/CMakeLists.txt

 add_mlir_library(MLIROptLib
   MlirOptMain.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Tools/mlir-opt
 
   LINK_LIBS PUBLIC
+  MLIRBytecodeWriter
   MLIRPass
   MLIRParser
   MLIRSupport
   )

mlir/lib/Tools/mlir-opt/MlirOptMain.cpp

 //===- MlirOptMain.cpp - MLIR Optimizer Driver ----------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This is a utility that runs an optimization pass and prints the result back
 // out. It is designed to support unit testing.
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Tools/mlir-opt/MlirOptMain.h"
+#include "mlir/Bytecode/BytecodeWriter.h"
 #include "mlir/IR/AsmState.h"
 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/Diagnostics.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/Location.h"
 #include "mlir/IR/MLIRContext.h"
 #include "mlir/Parser/Parser.h"

 /// within the specified context.
 ///
 /// This typically parses the main source file, runs zero or more optimization
 /// passes, then prints the output.
 ///
 static LogicalResult performActions(raw_ostream &os, bool verifyDiagnostics,
                                     bool verifyPasses, SourceMgr &sourceMgr,
                                     MLIRContext *context,
-                                    PassPipelineFn passManagerSetupFn) {
+                                    PassPipelineFn passManagerSetupFn,
+                                    bool emitBytecode) {
   DefaultTimingManager tm;
   applyDefaultTimingManagerCLOptions(tm);
   TimingScope timing = tm.getRootScope();
 
   // Disable multi-threading when parsing the input file. This removes the
   // unnecessary/costly context synchronization when parsing.
   bool wasThreadingEnabled = context->isMultithreadingEnabled();
   context->disableMultithreading();
if (failed(passManagerSetupFn(pm)))
     return failure();
 
   // Run the pipeline.
   if (failed(pm.run(*module)))
     return failure();
 
   // Print the output.
   TimingScope outputTiming = timing.nest("Output");
+  if (emitBytecode) {
+    writeBytecodeToFile(module->getOperation(), os);
+  } else {
     module->print(os);
     os << '\n';
+  }
   return success();
 }
 
 /// Parses the memory buffer.  If successfully, run a series of passes against
 /// it and print the result.
 static LogicalResult
 processBuffer(raw_ostream &os, std::unique_ptr<MemoryBuffer> ownedBuffer,
               bool verifyDiagnostics, bool verifyPasses,
               bool allowUnregisteredDialects, bool preloadDialectsInContext,
-              PassPipelineFn passManagerSetupFn, DialectRegistry &registry,
-              llvm::ThreadPool *threadPool) {
+              bool emitBytecode, PassPipelineFn passManagerSetupFn,
+              DialectRegistry &registry, llvm::ThreadPool *threadPool) {
   // Tell sourceMgr about this buffer, which is what the parser will pick up.
   SourceMgr sourceMgr;
   sourceMgr.AddNewSourceBuffer(std::move(ownedBuffer), SMLoc());
 
   // Create a context just for the current buffer. Disable threading on creation
   // since we'll inject the thread-pool separately.
   MLIRContext context(registry, MLIRContext::Threading::DISABLED);
   if (threadPool)
     context.setThreadPool(*threadPool);
 
   // Parse the input file.
   if (preloadDialectsInContext)
     context.loadAllAvailableDialects();
   context.allowUnregisteredDialects(allowUnregisteredDialects);
   if (verifyDiagnostics)
     context.printOpOnDiagnostic(false);
   context.getDebugActionManager().registerActionHandler<DebugCounter>();
 
   // If we are in verify diagnostics mode then we have a lot of work to do,
   // otherwise just perform the actions without worrying about it.
   if (!verifyDiagnostics) {
     SourceMgrDiagnosticHandler sourceMgrHandler(sourceMgr, &context);
     return performActions(os, verifyDiagnostics, verifyPasses, sourceMgr,
-                          &context, passManagerSetupFn);
+                          &context, passManagerSetupFn, emitBytecode);
   }
 
   SourceMgrDiagnosticVerifierHandler sourceMgrHandler(sourceMgr, &context);
 
   // Do any processing requested by command line flags.  We don't care whether
   // these actions succeed or fail, we only care what diagnostics they produce
   // and whether they match our expectations.
   (void)performActions(os, verifyDiagnostics, verifyPasses, sourceMgr, &context,
-                       passManagerSetupFn);
+                       passManagerSetupFn, emitBytecode);
 
   // Verify the diagnostic handler to make sure that each of the diagnostics
   // matched.
   return sourceMgrHandler.verify();
 }
 
 LogicalResult mlir::MlirOptMain(raw_ostream &outputStream,
                                 std::unique_ptr<MemoryBuffer> buffer,
                                 PassPipelineFn passManagerSetupFn,
                                 DialectRegistry &registry, bool splitInputFile,
                                 bool verifyDiagnostics, bool verifyPasses,
                                 bool allowUnregisteredDialects,
-                                bool preloadDialectsInContext) {
+                                bool preloadDialectsInContext,
+                                bool emitBytecode) {
   // The split-input-file mode is a very specific mode that slices the file
   // up into small pieces and checks each independently.
   // We use an explicit threadpool to avoid creating and joining/destroying
   // threads for each of the split.
   ThreadPool *threadPool = nullptr;
 
   // Create a temporary context for the sake of checking if
   // --mlir-disable-threading was passed on the command line.
   // We use the thread-pool this context is creating, and avoid
   // creating any thread when disabled.
   MLIRContext threadPoolCtx;
   if (threadPoolCtx.isMultithreadingEnabled())
     threadPool = &threadPoolCtx.getThreadPool();
 
   auto chunkFn = [&](std::unique_ptr<MemoryBuffer> chunkBuffer,
                      raw_ostream &os) {
     return processBuffer(os, std::move(chunkBuffer), verifyDiagnostics,
                          verifyPasses, allowUnregisteredDialects,
-                         preloadDialectsInContext, passManagerSetupFn, registry,
-                         threadPool);
+                         preloadDialectsInContext, emitBytecode,
+                         passManagerSetupFn, registry, threadPool);
   };
   return splitAndProcessBuffer(std::move(buffer), chunkFn, outputStream,
                                splitInputFile, /*insertMarkerInOutput=*/true);
 }
 
 LogicalResult mlir::MlirOptMain(raw_ostream &outputStream,
                                 std::unique_ptr<MemoryBuffer> buffer,
                                 const PassPipelineCLParser &passPipeline,
                                 DialectRegistry &registry, bool splitInputFile,
                                 bool verifyDiagnostics, bool verifyPasses,
                                 bool allowUnregisteredDialects,
-                                bool preloadDialectsInContext) {
+                                bool preloadDialectsInContext,
+                                bool emitBytecode) {
   auto passManagerSetupFn = [&](PassManager &pm) {
     auto errorHandler = [&](const Twine &msg) {
       emitError(UnknownLoc::get(pm.getContext())) << msg;
       return failure();
     };
     return passPipeline.addToPipeline(pm, errorHandler);
   };
   return MlirOptMain(outputStream, std::move(buffer), passManagerSetupFn,
                      registry, splitInputFile, verifyDiagnostics, verifyPasses,
-                     allowUnregisteredDialects, preloadDialectsInContext);
+                     allowUnregisteredDialects, preloadDialectsInContext,
+                     emitBytecode);
 }
 
 LogicalResult mlir::MlirOptMain(int argc, char **argv, llvm::StringRef toolName,
                                 DialectRegistry &registry,
                                 bool preloadDialectsInContext) {
   static cl::opt<std::string> inputFilename(
       cl::Positional, cl::desc("<input file>"), cl::init("-"));
 
LogicalResult mlir::MlirOptMain(int argc, char **argv, llvm::StringRef toolName,
   static cl::opt<bool> allowUnregisteredDialects(
       "allow-unregistered-dialect",
       cl::desc("Allow operation with no registered dialects"), cl::init(false));
 
   static cl::opt<bool> showDialects(
       "show-dialects", cl::desc("Print the list of registered dialects"),
       cl::init(false));
 
+  static cl::opt<bool> emitBytecode(
+      "emit-bytecode", cl::desc("Emit bytecode when generating output"),
+      cl::init(false));
+
   InitLLVM y(argc, argv);
 
   // Register any command line options.
   registerAsmPrinterCLOptions();
   registerMLIRContextCLOptions();
   registerPassManagerCLOptions();
   registerDefaultTimingManagerCLOptions();
   DebugCounter::registerCLOptions();
LogicalResult mlir::MlirOptMain(int argc, char **argv, llvm::StringRef toolName,
   auto output = openOutputFile(outputFilename, &errorMessage);
   if (!output) {
     llvm::errs() << errorMessage << "\n";
     return failure();
   }
 
   if (failed(MlirOptMain(output->os(), std::move(file), passPipeline, registry,
                          splitInputFile, verifyDiagnostics, verifyPasses,
-                         allowUnregisteredDialects, preloadDialectsInContext)))
+                         allowUnregisteredDialects, preloadDialectsInContext,
+                         emitBytecode)))
     return failure();
 
   // Keep the output file if the invocation of MlirOptMain was successful.
   output->keep();
   return success();
 }

mlir/test/Bytecode/general.mlir

+// RUN: mlir-opt -allow-unregistered-dialect -emit-bytecode %s | mlir-opt -allow-unregistered-dialect | FileCheck %s
+
+// CHECK-LABEL: "bytecode.test1"
+// CHECK-NEXT:    "bytecode.empty"() : () -> ()
+// CHECK-NEXT:    "bytecode.attributes"() {attra = 10 : i64, attrb = #bytecode.attr} : () -> ()
+// CHECK-NEXT:    %[[RESULTS:.*]]:3 = "bytecode.results"() : () -> (i32, i64, i32)
+// CHECK-NEXT:    "bytecode.operands"(%[[RESULTS]]#0, %[[RESULTS]]#1, %[[RESULTS]]#2) : (i32, i64, i32) -> ()
+// CHECK-NEXT:    "bytecode.branch"()[^[[BLOCK:.*]]] : () -> ()
+// CHECK-NEXT:  ^[[BLOCK]](%[[ARG0:.*]]: i32, %[[ARG1:.*]]: !bytecode.int, %[[ARG2:.*]]: !pdl.operation):
+// CHECK-NEXT:    "bytecode.regions"() ({
+// CHECK-NEXT:      "bytecode.operands"(%[[ARG0]], %[[ARG1]], %[[ARG2]]) : (i32, !bytecode.int, !pdl.operation) -> ()
+// CHECK-NEXT:      "bytecode.return"() : () -> ()
+// CHECK-NEXT:    }) : () -> ()
+// CHECK-NEXT:    "bytecode.return"() : () -> ()
+// CHECK-NEXT:  }) : () -> ()
+
+"bytecode.test1"() ({
+  "bytecode.empty"() : () -> ()
+  "bytecode.attributes"() {attra = 10, attrb = #bytecode.attr} : () -> ()
+  %results:3 = "bytecode.results"() : () -> (i32, i64, i32)
+  "bytecode.operands"(%results#0, %results#1, %results#2) : (i32, i64, i32) -> ()
+  "bytecode.branch"()[^secondBlock] : () -> ()
+
+^secondBlock(%arg1: i32, %arg2: !bytecode.int, %arg3: !pdl.operation):
+  "bytecode.regions"() ({
+    "bytecode.operands"(%arg1, %arg2, %arg3) : (i32, !bytecode.int, !pdl.operation) -> ()
+    "bytecode.return"() : () -> ()
+  }) : () -> ()
+  "bytecode.return"() : () -> ()
+}) : () -> ()