[llvm-dev] LLVM Releases: Upstream vs. Downstream / Distros
Stephen Hines via llvm-dev
llvm-dev at lists.llvm.org
Thu May 12 08:22:57 PDT 2016
Just some relatively quick thoughts, since I am still on vacation. +Luis
from ChromeOS toolchain.
On Thu, May 12, 2016 at 7:56 AM, Kristof Beyls <Kristof.Beyls at arm.com>
> FWIW, for our ARM Compiler product, we follow top-of-trunk, not the
> Next to picking up new functionality quicker, it also allows us to detect
> in LLVM against our in-house testing quickly, not 6 months later. We find
> that when
> we find a regression within 24 to 48 hours of the commit introducing it,
> it's much
> cheaper to get it fixed.
> In my opinion, it would be better overall for the LLVM project if
> top-of-trunk is
> tested as much as possible, if testing resources are so scarce that a
> has to be made between testing top-of-trunk or testing a release branch.
I am 100% in agreement here. TOT regressions cost more than people think. I
have seen firsthand how this hurts the non-LLVM parts of Android. In
Android, more testing resources are devoted to release branches, and thus
the development branch picks up bizarre regressions that are harder to
track down and fix than they should be. Since I don't update the compiler
in the release branch (at least not without good reason), this means that
compiler regressions are harder for me to track down, because nobody pays
attention to dev results until it is hurting them in the next release. :( I
have some ideas to make this better in the future, but nothing to share
> On 12 May 2016, at 02:10, Robinson, Paul <paul.robinson at sony.com> wrote:
> -----Original Message-----
> From: Renato Golin [mailto:renato.golin at linaro.org
> <renato.golin at linaro.org>]
> Sent: Wednesday, May 11, 2016 11:11 AM
> To: Hans Wennborg
> Cc: LLVM Dev; Clang Dev; Quentin Colombet; Tom Stellard; Robinson, Paul;
> Jim Grosbach; Kristof Beyls; Frédéric Richez; Reid Kleckner; Philip
> Reames; Matthias Braun; Bernhard Rosenkränzer; Sylvestre Ledru; Matthias
> Klose; Stephen Hines; Jeff Law; Ed Maste; Behan Webster
> Subject: Re: LLVM Releases: Upstream vs. Downstream / Distros
> On 11 May 2016 at 17:16, Hans Wennborg <hans at chromium.org> wrote:
> This is a long email :-) I've made some comments inline, but I'll
> summarize my thoughts here:
> Thanks Hans!
> I'll respond them inline, below.
> - I think we should use the bug tracker to capture issues that affect
> releases. It would be cool if a commit hook could update bugzilla
> entries that refer to it.
> That seems like a simple hook.
> At least for the major releases, I think we're doing pretty well on
> timing in terms of predictability: since 3.6, we have release every
> six months: first week of March and first week of September (+- a few
> days). Branching has been similarly predictive: mid-January and
> Indeed, we got a lot better more recently (last 2y), and mostly thanks
> to you. :)
> Absolutely. It's enough of a track record to allow reasonable planning.
> We used to vary 3 months +-, and now we're down to a few days.
> Whatever we decide, I think we should make it official but putting it
> out somewhere, so people can rely on that.
> A public commitment to the future release schedule would be that much
> more justification for planning to participate.
> Right now, even if you're extra awesome, there's nothing telling the
> distros and LLVM-based products that it will be so if someone else
> takes over the responsibility, so they can't adapt.
> That's what I meant by "quasi-chaotic".
> If there are many downstream releases for which shifting this schedule
> would be useful, I suppose we could do that, but it seems unlikely
> that there would be agreement on this, and changing the schedule is
> disruptive for those who depend on it.
> That's the catch. If we want them to participate, the process has to
> have some meaning to them. The fact that not many people do, is clear
> to me what it means.
> We also need to know better how many other releases already depend on
> the upstream process (not just Chromium, for obvious reasons), to be
> able to do an informed choice of dates and frequency.
> The more well positioned and frequent we are, the more people will
> help, but there's a point where the curve bends down, and the cost is
> just too great. We need to find the inflection point, and that will
> require some initial investigations and guesses, and a lot of fine
> tuning later. But if we're all on the same page, I think we can do
> that, even if it takes time.
> I'm particularly concerned with Android, because they not only have
> their own tree with heavily modified LLVM components (ex.
> Compiler-RT), but they also build differently and so their process are
> completely alien to ours. One of the key reasons why these things
> happened is because:
I am not sure why you think Android's compiler-rt is an example of a
"heavily modified" component. As I see it, our compiler-rt matches upstream
almost exactly (with one minor mistake from a duplicate merge that results
in extra copies of some static functions that we don't even build). We do
have 3 cherry-picks for some MIPS ASan patches, but all of those come
directly from TOT master.
There are some Android-only patches in LLVM and Clang (thanks to
RenderScript, and not being able to easily upstream these - i.e. the RS
frontend doesn't live as an LLVM upstream project, so it is hard to write
tests for these modifications). None of those patches should impact things
greatly. Assuming you sync to the same CL (plus <10 cherry-picks to all
projects), any C/C++ regression should be reproducible with a normal TOT
build. This has been true since we did upstream our most divergent patch
(re: calling conventions for ARM vectors) a few months ago.
> * They couldn't rely on our releases, as fixing bugs and back-porting
> wasn't a thing back then
> The real problem is retaining history. Release branches don't make this
very nice, and necessitate that we swap out an entire chunk of history for
a different chunk of history every time we change releases. That is pretty
obnoxious for keeping a good idea of what has happened, and thus following
master makes this easier for us (and I assume others too). We did try using
a release branch back for LLVM 3.5, but the maintenance cost for then
updating things when we wanted to move past 3.5 showed me why I don't want
to really look
> * They already had their own release schedule, so aligning with ours
> brought no extra benefit
> It is not 100% clear that Android will want to be dependent on LLVM's
release schedule. I think that there are definitely benefits to having
everyone do extra validation, but I am unconvinced that it is the "same"
validation for everyone that is valuable, hence this might not make things
go much smoother/faster for those groups.
> * We always expected people to work off trunk, and everyone had to
> create their own process
> I don't want to change how people work, just to add one more valid way
> of working, which is most stable for upstream releases. :)
> The only reasonable way I see of aligning upstream releases with
> downstream schedules would be to release much more often. This works
> well in Chromium where there's a 6-week staged release schedule. This
> would mean there's always a branch going for the next release, and
> important bug fixes would get merged to that.
> Full validation every 6 weeks is just not possible. But a multiple of
> that, say every 3~4 months, could be much easier to work around.
> In Chromium we drive
> this from the bug tracker -- it would be very hard to scan each commit
> for things to cherry-pick. This kind of process has a high cost
> though, there has to be good infrastructure for it (buildbots on the
> branch for all targets, for example), developers have to be aware, and
> even then it's a lot of work for those doing the releases. I'm not
> sure we'd want to take this on. I'm also not sure it would be suitable
> for a compiler, where we want the releases to have long life-time.
> This works because you have a closed system. As you say, Chromium is
> mostly final product, not a tool to develop other products, and the
> validation is a lot simpler.
> With Clang, we'd want to involve external releases into it, and it
> simply wouldn't scale.
> For the major releases, I've tried to do this. We could certainly
> formalize it by posting it on the web page though.
> I think that'd be the first step, yes. But I wanted to start with a
> good number. 2 times a year? Would 3 times improve things that much
> for the outsiders? Or just moving the dates would be enough for most
> That's why I copied so many outsiders, so they can chime in and let us
> know what would be good for *them*.
> Data point: At Sony we ship our stuff every 6 months and that isn't
> going to change. Our policy has been to base our releases on upstream
> releases, and given our current lead times, the current upstream release
> schedule is actually not bad, especially if it has stopped drifting.
> Having four releases pretty consistently follow the current schedule is
> extremely positive, thanks!
> It would help our internal planning to publish the schedule for future
> releases. You never know, we might even be able to fork an internal
> branch in time to help with the release testing, although that is not
> in any way a lightweight process (so no promises).
> There has been some talk about moving toward releasing from upstream
> trunk, or more precisely to streamlining our internal testing process
> in order to allow us to seriously contemplate releasing from upstream
> trunk. I can't argue with the streamlining part, for sure. Whether
> the rest of it works out must remain to be seen.
> Most importantly, those folks should get involved :-)
> In practice, we kind of have this for at least some of the targets.
> Maybe we should write this down somewhere instead of me asking for
> (the same) volunteers each time the release process starts?
> I give consent to mark me as the ARM/AArch64 release tester for the
> foreseeable future. :)
> I can also help Sylvestre, Doko, Ed, Jeff, Bero etc. to test on their
> system running on ARM/AArch64 hardware.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev