What version comes after 3.9? (Was: [3.9 Release] Release plan and call for testers)

Project Infrastructure LLVM Dev List Archives
1

<snip>

Given that our releases are time-based rather than feature-based,
I
don't see a distinct major / minor version being anything other
than
arbitrary, so I'd suggest we take 4.0 as our next release, 4.1 as
the first patch release on that, 5.0 as the next release after
that,
and so on.

I completely agree with Richard here. “Breaking of IR
compatibility”
was an interesting metric for older and less mature versions of
LLVM. We can solve the same sort of challenge (the desire to eject
old autoupgrade code) by having a sliding window of versions
supported (e.g. version 4.5 supports back to version 3.6).

Let me clarify. What I’m trying to say is that:

a) LLVM has a time-based release cycle, not a schedule-based one. As
such, a simple and predictable version number makes sense.
b) The LLVM project as a whole is a lot bigger than LLVM IR, even
given its centrality to the project in some ways.
c) I think that it makes sense to keep adding 0.1 to each major
release going forward well into the future.

Now you’re making the versions sound like floating point numbers :-). Just to be clear, you are proposing we use 3.10/3.11/etc. rather than 4.0/4.1/etc.?

If so, I agree with that for a couple of reasons. First, as several people said, version numbers should not be driven primarily by IR changes: the LLVM project is a lot more than the IR, even though the IR is a fundamental component. Our releases are time-based and the predictability of that has worked pretty well.

A second reason is that major version numbers also have a useful communication value: signifying a major step forward in the system along some dimension. It just so happens that these major changes have been IR changes in the past -- and perhaps opaque pointer types will turn out to be the next major change -- but regardless of what the change is, I think there is some value in reserving major version increments (like 3.x.y to 4.0) when major changes happen.

--Vikram

On the topic of the pointer changes proposed, I really don’t think
the community is served by waiting for that. The supposition seems
to be that we’d land it *without* upgrade support, but then bump the
major version number to indicate this. If that’s the proposal, I
think that doing such a thing would be disastrous for the LLVM
community as a whole

To be clear, that's not at all what I was saying. The plan has always been to have upgrade support. My thought was that once the pointer changes land, it will mean major changes in how many frontends and IR transformations use the API. A lot of out-of-tree frontends and passes support multiple versions of LLVM, and use ifdefs to work around relevant API changes. For the most part, this works reasonably, and the changes necessary between versions are not often large (perhaps especially because we have time-based releases). I suspect, however, that the amount of code changes necessary to adapt for the pointer changes will be much larger, and so calling that a new major version seems fitting.

-Hal

-—Vikram

// Vikram S. Adve
// Professor, Department of Computer Science
// University of Illinois at Urbana-Champaign
// vadve@illinois.edu
// http://llvm.org

2

Let me clarify. What I’m trying to say is that:

a) LLVM has a time-based release cycle, not a schedule-based one. As
such, a simple and predictable version number makes sense.
b) The LLVM project as a whole is a lot bigger than LLVM IR, even
given its centrality to the project in some ways.
c) I think that it makes sense to keep adding 0.1 to each major
release going forward well into the future.

Now you’re making the versions sound like floating point numbers :-). Just to be clear, you are proposing we use 3.10/3.11/etc. rather than 4.0/4.1/etc.?

No, I’m suggesting that we continue the progression we’ve had from the start and proceed from 3.8 to 3.9 to 4.0 to 4.1.

If so, I agree with that for a couple of reasons. First, as several people said, version numbers should not be driven primarily by IR changes: the LLVM project is a lot more than the IR, even though the IR is a fundamental component. Our releases are time-based and the predictability of that has worked pretty well.

A second reason is that major version numbers also have a useful communication value: signifying a major step forward in the system along some dimension. It just so happens that these major changes have been IR changes in the past -- and perhaps opaque pointer types will turn out to be the next major change -- but regardless of what the change is, I think there is some value in reserving major version increments (like 3.x.y to 4.0) when major changes happen.

These seem like contradictory points: on the one hand you’re observing that we have an inherently schedule driven process, but are also saying that major version numbers are important to signify “major” changes. Given the scope of LLVM, I suspect we’ll never have a “major” change that lines up across all of the sub-projects, so this doesn’t seem like something we should bank on.

I don’t think that there is really a problem to solve here, but if we were sufficiently motivated to “solve” this problem, then the answer is obvious: instead of 4.0, we should just go to 40, and add one for every release after that.

-Chris

3

Let me clarify. What I’m trying to say is that:

a) LLVM has a time-based release cycle, not a schedule-based one. As
such, a simple and predictable version number makes sense.
b) The LLVM project as a whole is a lot bigger than LLVM IR, even
given its centrality to the project in some ways.
c) I think that it makes sense to keep adding 0.1 to each major
release going forward well into the future.

Now you’re making the versions sound like floating point numbers :-). Just to be clear, you are proposing we use 3.10/3.11/etc. rather than 4.0/4.1/etc.?

No, I’m suggesting that we continue the progression we’ve had from the start and proceed from 3.8 to 3.9 to 4.0 to 4.1.

If so, I agree with that for a couple of reasons. First, as several people said, version numbers should not be driven primarily by IR changes: the LLVM project is a lot more than the IR, even though the IR is a fundamental component. Our releases are time-based and the predictability of that has worked pretty well.

A second reason is that major version numbers also have a useful communication value: signifying a major step forward in the system along some dimension. It just so happens that these major changes have been IR changes in the past -- and perhaps opaque pointer types will turn out to be the next major change -- but regardless of what the change is, I think there is some value in reserving major version increments (like 3.x.y to 4.0) when major changes happen.

These seem like contradictory points: on the one hand you’re observing that we have an inherently schedule driven process, but are also saying that major version numbers are important to signify “major” changes.

Not contradictory at all. The schedule is for *when* we release new versions; it is unrelated to any numbering scheme. The major vs. minor version numbers capture something about how important a release is. I don’t really have a preference one way or another, but I do think people are missing a basic point about the numbering.

Given the scope of LLVM, I suspect we’ll never have a “major” change that lines up across all of the sub-projects, so this doesn’t seem like something we should bank on.

Agreed, but you could introduce a new major version when there is something really major in any major part of the ecosystem. E.g., making Clang the primary C/C++ compiler instead of llvm-gcc could have qualified.

I don’t think that there is really a problem to solve here, but if we were sufficiently motivated to “solve” this problem, then the answer is obvious: instead of 4.0, we should just go to 40, and add one for every release after that.

Sounds like a stretch to me, but whatever!

-Chris

—Vikram

4

Let me clarify. What I’m trying to say is that:

a) LLVM has a time-based release cycle, not a schedule-based one. As
such, a simple and predictable version number makes sense.
b) The LLVM project as a whole is a lot bigger than LLVM IR, even
given its centrality to the project in some ways.
c) I think that it makes sense to keep adding 0.1 to each major
release going forward well into the future.

Now you’re making the versions sound like floating point numbers :-). Just to be clear, you are proposing we use 3.10/3.11/etc. rather than 4.0/4.1/etc.?

No, I’m suggesting that we continue the progression we’ve had from the start and proceed from 3.8 to 3.9 to 4.0 to 4.1.

I’m sympathetic to the idea that we should move to a sliding window for IR compatibility if that’s useful.

However, I think it is a really big mistake to have a major version and just increment it arbitrarily without any associated “major” change. Lacking us explaining what we consider to be a major release, people will make assumptions that will inevitably lead to missed expectations.

If we want to have a major/minor split, I think we need to have some guidance for what will constitute a major jump. I don’t think it needs to necessarily be formulaic (like bitcode compatibility), it could be when a feature happens to be ready and happens to be decided by the community to be “big enough” we flip the major version and announce that feature happened to be ready. But we shouldn’t just go from 3.9 to 4.0 because of some decimal correspondence. Too many people will assume it had significance and become disappointed.

If we don’t want a major/minor split, then I believe Richard’s suggestion is the correct one: just have integers to mark the temporal releases, and dot releases for updates to that release branch.

I personally don’t have strong feelings about whether we should or shouldn’t have a major/minor split. I see arguments on both sides. But I really do not want us to have a meaningless split.

-Chandler

5

However, I think it is a really big mistake to *have* a major version and
just increment it arbitrarily without any associated "major" change. Lacking
us explaining what we consider to be a major release, people will make
assumptions that will inevitably lead to missed expectations.

Precisely.

I personally don't have strong feelings about whether we should or shouldn't
have a major/minor split. I see arguments on both sides. But I really do not
want us to have a *meaningless* split.

That is how I feel too. But I think we need to take am informed
decision before July (ie. not continue the status quo).

Continuing (3.9 -> 4.0) is ok for some of us because we've seen 1.9 ->
2.0 and 2.9 -> 3.0, but it's not logical to most people out there,
including many in our community today.

Moving to an OSS-friendly can be very helpful, but not without a good
rule on what's "major". Not having a rule is also *not* a reason to
continue the status quo.

I agree we don't need hard rules, and we can always argue for months
if the next one will be major or not, but I'd rather spend my time
elsewhere.

So, things that *could* influence a major change:

* IR backwards compatibility: this one is obvious
* Major internal refactoring: new pass manager, new selection dag,
(ex. when the MC layer was introduced)
* Major new language support: c++17, openMP complete, etc.
* DCE: When we delete whole back-ends (which obviously will be
backwards incompatible)

But in the end, it boils down to someone asking to up the number, and
a month long thread to agree or disagree, which is ok.

All in all, we need reasons to either side we pick.

cheers,
--renato

6

Pretty much exactly my thinking.

–Vikram
http://llvm.org

7

I'm sympathetic to the idea that we should move to a sliding window for IR
compatibility if that's useful.

However, I think it is a really big mistake to *have* a major version and
just increment it arbitrarily without any associated "major" change. Lacking
us explaining what we consider to be a major release, people will make
assumptions that will inevitably lead to missed expectations.

If we want to have a major/minor split, I think we need to have *some*
guidance for what will constitute a major jump. I don't think it needs to
necessarily be formulaic (like bitcode compatibility), it could be when a
feature happens to be ready and happens to be decided by the community to be
"big enough" we flip the major version and announce that feature happened to
be ready. But we shouldn't just go from 3.9 to 4.0 because of some decimal
correspondence. Too many people will assume it had significance and become
disappointed.

If we *don't* want a major/minor split, then I believe Richard's suggestion
is the correct one: just have integers to mark the temporal releases, and
dot releases for updates to that release branch.

I like just having just integer. IMHO it is simply not worth the time
to try to figure out what is "major" in a project with so many
different uses.

So my vote goes to
* The bitcode compatibility promise is changed to use real time (at
least 5 years?) instead of revision numbers.
* The next release is 4.0.
* The next bugfix release is 4.1.
* After 6 months or so we release 5.0.

Cheers,
Rafael

8

Regardless of the numbering scheme debate, this part sounds somewhat problematic to me.

The way I understand the current IR backward compatibility guarantee is that the major version change is a notification strategy, not a commitment to hold the IR stable for any given amount of time. That is, it’s “If we need to change the IR, we’ll bump the major version - but this may happen on the next release”. Not “The release schedule for major versions is 5 years, so if the current minor is 3.4, you have 3 more years of stable IR to go”.

This, I think, is a fairly good model. Promising IR stability N years into the future is something we may easily come to regret.

9

I can see your point about our current naming scheme being confusable with semantic versioning, but I don’t think the risk is high and I don’t think the problem is solving. That is just MHO of course, and I totally respect and understand that reasonable people would disagree with that.

What I care most about is aligning our versioning scheme with the practical realities of how we manage the software, and giving our users some idea of what we ship. I do not think it is reasonable at all to go to “3.10” after “3.9”, because we will never get to “4.0”. The GCC community has been through this (as one example) where they could never decide on when to bump to 5.0. They resolved the issue by going to the completely insane (again, IMVHO :slight_smile: version scheme they have now.

The desire to align the version with how we release the product is what leads me to think that the only rationale solution is to increment the version number predictably with every release: we have a time based release. I don’t care if that is by adding 0.1 or by adding 1.

That is why I suggested that - if you believe that consistently adding 0.1 is confusing - then the logical answer is to jump to "version 40” with the post-3.9 major release and then add 1 on every other release after that. Here are the advantages I see:

  • This communicates clearly and obviously to users that we changed policy with the version number scheme
  • Offers continuity with the prior approach (3.9 → 40 is something that people will be able to remember)
  • Aligns the new version number scheme with how the project is managed
  • Aligns the version numbers with the expectations of semantic versioning (since every release of LLVM is API breaking). Compatible bugfix/patch releases can then be “dot” releases of the major number, again aligning with semantic versioning.

The only downside I see to this is that people will think that it is crazy that LLVM has had so many incompatible releases, but hey, why hide the truth? :slight_smile:

-Chris

10

Hm. Intriguing. Then each release would have to specify (somehow) what its specific IR compatibility range is.

Release 47 supports bitcode back to release 3.9 (say), release 54 compatible back to release 44, etc where the “back to” number is not clearly predictable.

(This would be easier to implement if bitcode reliably had a produced-by-version number in it.)

–paulr

11

This sounds like over-engineering a problem that has been solved
everywhere else with predictable version numbers... :slight_smile:

--renato

12

It has since 3.8.
I added a record that contains it as a string (we use internally to reject bitcode that we know are incompatible during LTO for example).
Right now, there Is not code upstream that using this field for anything else than a diagnostic when the bitcode can’t be read, but it could be extended in the future.