[PATCH] D20260: IR: Introduce local_unnamed_addr attribute.

Wed May 18 09:53:37 PDT 2016

On Wed, May 18, 2016 at 8:22 AM, Mehdi Amini <mehdi.amini at apple.com> wrote:

> Hi David,
>
>
> On May 17, 2016, at 11:47 PM, David Blaikie <dblaikie at gmail.com> wrote:
>
>
>
> On Tue, May 17, 2016 at 11:40 PM, Mehdi Amini <mehdi.amini at apple.com>
> wrote:
>
>>
>> On May 17, 2016, at 11:33 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>
>>
>>
>> On Tue, May 17, 2016 at 11:24 PM, Mehdi Amini <mehdi.amini at apple.com>
>> wrote:
>>
>>>
>>> On May 17, 2016, at 11:11 PM, David Blaikie <dblaikie at gmail.com> wrote:
>>>
>>>
>>>
>>> On Tue, May 17, 2016 at 9:58 PM, Mehdi Amini via llvm-commits <
>>> llvm-commits at lists.llvm.org> wrote:
>>>
>>>>
>>>> > On May 17, 2016, at 9:48 PM, Justin Bogner <mail at justinbogner.com>
>>>> wrote:
>>>> >
>>>> > Mehdi Amini via llvm-commits <llvm-commits at lists.llvm.org> writes:
>>>> >>    On May 17, 2016, at 5:21 PM, Chandler Carruth <
>>>> chandlerc at gmail.com> wrote:
>>>> >>
>>>> >>    On Tue, May 17, 2016 at 5:08 PM Mehdi Amini via llvm-commits <
>>>> >>    llvm-commits at lists.llvm.org> wrote:
>>>> >>
>>>> >>        Seriously I absolutely don't get the "start a new revision"
>>>> thing
>>>> >>        because the initial email was not on llvm-commit. Really what
>>>> is the
>>>> >>        rational?
>>>> >>
>>>> >>        We are losing the history and the threading for the review,
>>>> this is
>>>> >>        really annoying.
>>>> >>
>>>> >>        So big -1
>>>> >>
>>>> >>    I don't care very much about starting a new revision.
>>>> >>
>>>> >>    Justin and others who don't use phabricator asked for this
>>>> because that is
>>>> >>    what causes an email with a patch file to be sent to
>>>> llvm-commits, and so
>>>> >>    I've been trying to encourage it based on their request.
>>>> >>
>>>> >> There *has been* an email with a patch sent to llvm-commit after the
>>>> first
>>>> >> update to the diff:
>>>> http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160516/357163.html
>>>> >> Just "having a patch sent to llvm-commit" can be accomplished with
>>>> an manual
>>>> >> email answering to the revision instead of create a new one from
>>>> scratch.
>>>> >
>>>> > Yes, there's a patch, and it explains what's happening as "Minimise
>>>> the
>>>> > diff a little" and "Add assembler/bitcode test". The description of
>>>> > what's happening is completely missing. The comments so far aren't on
>>>> > the mailing list at all
>>>>
>>>> Re-creating a revision is not gonna help to put this history back on
>>>> the mailing list.
>>>>
>>>>
>>>> > , and somebody who reads this mail first has no
>>>> > idea what's going on. Is the patch something interesting? Should I
>>>> spend
>>>> > my time following the link and catching up? It also has a subject that
>>>> > starts with "Re:", which makes me and my email filter think I've
>>>> already
>>>> > read the first patch and decided whether or not to pay attention.
>>>> >
>>>> > Essentially, patches that start without llvm commits don't look like
>>>> > mails that need to be read.
>>>>
>>>> I judge based on the title, and then click on the link.
>>>>
>>>> >
>>>> > The first email in a patch review thread is incredibly important. It
>>>> > summarizes the point of the patch, provides context, and has the
>>>> patch.
>>>> > Coming into the conversation half way through is very difficult.
>>>>
>>>> We all agree that it is *better* to have llvm-commit as an initial
>>>> subscriber, but unless someone put the manpower in to get it enforced by
>>>> phabricator, we'll have to live with the fact that occasionally it will be
>>>> missing.
>>>>
>>>
>>> I'm pretty sure the warning that turns up in Phab suggests creating a
>>> new code review with *-commits on from the start. I'd encourage this to be
>>> the path we encourage/request when we see this (for myself, it's really
>>> obvious when someone creates a review with me as a reviewer but without the
>>> mailing list - such a mail ends up in my inbox, whereas if it had the
>>> mailing list on it it'd be in my mailing list bucket). I think we should do
>>> it early and often. Yes, fixing Phab would be better. But I think
>>> restarting the review is better than continuing it without the
>>> original/initial context - earlier the better, before lots of conversation
>>> has already happened off list.
>>>
>>>
>>> That's the point: *after* some round of reviews were already conducted
>>> on Phabricator, the most important context for the review *is* now this
>>> review history and this is what is important to preserve, and this is why
>>> restarting a new revision in this condition is a non-starter to me.
>>>
>>
>> Except none of that is on the mailing list archive or otherwise preserved
>> in LLVM's system of record. It was pretty well discussed when Phab was
>> introduced that the mailing list would be the system of record and Phab
>> just a tool. If it's not on the list, it's not here.
>>
>>
>> Except it exists, and in the present time it is useful
>> context/information, so "it's not here" is definitely not an accurate
>> representation of the situation.
>>
>>
>> In the case where a bunch of review without the list - we should
>> reintroduce any of that context/discussion.
>>
>>
>> As said before, opening a new review is not gonna make this appear
>> magically.
>>
>
> I'm not suggesting it appear by magic. I'm suggesting that the person
> (Peter) who creates the new review, summarize the existing off-list (&
> partly on list as well) discussion in that new review, so everyone's got
> the context to engage in this review. (& so it's recorded on the mailing
> list for archival purposes, etc)
>
>
> But why does it have to be in a *new* review? How forwarding the original
> email that started the thread not enough to get the initial context?
>
>
> I don't know whether a new thread's especially necessary, but at least an
>> authoritative email with all that context being summarized into this thread
>> is called for. It'll be harder to follow the thread (people expect the
>> first email to have the introductory context, etc) but not impossible.
>>
>>
>> I'd probably favor a new thread, with the old one referenced for
>> archaeology - the current state can be summarized, any outstanding comments
>> can be replied to in the new thread with any context quoted as needed.
>>
>>
>> If I spent time reviewing a revision, using inline comment on
>> Phabricator, and a further round of review is done with a different
>> revision, I don't see my inline comments and I can't figure if they were
>> addressed, and I can't really see what changed between the two revisions.
>>
>
> That just brings us back to where we would be if we didn't have Phab -
> that's what we all had to do anyway, read the patch, read someone's
> unstructured comments, etc. And other people in the community are seeing a
> worse situation - because they're not using/looking at Phab they're missing
> all that context .
>
>
>
>> This, as a reviewer, is most important than anything else for me, and I
>> consider that opening a new revision in this condition is a waste of *my*
>> time as a reviewer, and this is for benefit that sounds hand-wavy to me. So
>> in the future I'll probably resign as a reviewer on the any new revision
>> created in this condition.
>>
>
> This last bit seems like it's entirely the point - if you notice a review
> created in this condition if the first thing you do is say "this isn't in a
> good condition, please create one with the mailing list on it, I'll review
> that one" - then we get exactly what we all seem to want, right?
>
>
> You didn't get my last bit the way I meant it apparently, let me try
> again:
>
> 1) A revision is created without llvm-commits as a subscriber
> 2) I put inline comments in the review and we go through one or multiple
> round of reviews.
> 3) At some point during the process (before 2 or between two rounds of
> review, llvm-commits gets added)
> 4) People complains because there didn't get the previous emails and ask
> to start a new revision
>
> Of course if we catch it during 1, we can still act on it (forward the
> email to llvm-commits, or nuke the revision and start a new one, same
> effect to me), and it seems to me that it is what you're saying, right?
> However, the case I am complaining about here, is when we don't catch it
> initially and we end up in 4. What do we do now? Restarting a new revision
> nukes the review history and all my inline comments are lost, i.e. I'm not
> OK to restart on a new revision.
>

OK, thanks for clarifying. Why would this be unacceptable to you? This is
how a non-Phab review would progress, right? (unstructured replies -
usually in infix email style, responding to your comments - the same thing
could be fairly easily replicated in an email response (by the original
author - quoting your off list review emails and interspersing their reply,
etc) to the new thread)

> So while it is unfortunate that it was not caught earlier, I argue that
> the inconvenience for the reviewer should prevail on the inconvenience for
> the hypothetical people that may engage in the review: if sending them the
> original email is not enough, either they'll have to use a browser, or
> they'll still be able to post-commit review.
>

It's not just a matter of hypothetical reviewers, but also of documenting
the review in the public system of record for participants, lurkers, and
archaeologists who may need to discover why certain choices were made after
the fact. This was always the plan for Phab - that email was still the
system of record.

I suppose it comes down to this: I think the best thing would be for a new
thread to be created, and the feedback/context ported over from here. So
it's got a fresh start and is easy to access for current and future readers
(rather than having to come in half way, trawl through the thread, find the
original context provided part-way through, etc). There would be some
inconvenience for existing reviewers, but much of that could be mitigated
by the original patch author porting over the context into that new thread.

Not as good (for the review, the community, etc), but probably sufficient,
would be to add the original context to this thread to get everyone up to
speed on the mailing list.

- Dave

>
> --
> Mehdi
>
>
>
>
>>
>>>
>>> A revision on Phabricator that has no history can be nuked without
>>> consequence of course...
>>>
>>> --
>>> Mehdi
>>>
>>>
>>>
>>>
>>>>
>>>> >> That said the patch is always a mere two-clicks away from the email:
>>>> http://
>>>> >> reviews.llvm.org/D20260?download=true
>>>> >> One could even write a plugin for his email client to act on link in
>>>> the email
>>>> >> "http://reviews.llvm.org/D20260" and automatically curl the latest
>>>> patch.
>>>> >
>>>> > How do I know if I need or want to look at it? Is "Minimise the diff a
>>>> > little" an interesting patch?
>>>>
>>>> Seriously?? The email title says "IR: Introduce local_unnamed_addr
>>>> attribute", that pretty clear IMO.
>>>>
>>>> >
>>>> >>    There is still the fact that even the initial phab revision has
>>>> >>    essentially no context. I've skimmed thin three or four times and
>>>> I'm not
>>>> >>    sure yet what the motivation is... I'm sure it has one, I just
>>>> can't find
>>>> >>    it.
>>>> >>
>>>> >> s/Initial phab revision/current phab revision/
>>>> >>
>>>> >> Yes, I'd expect a phabricator revision to have a correct description.
>>>>
>>>> _______________________________________________
>>>> llvm-commits mailing list
>>>> llvm-commits at lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160518/43c92248/attachment.html>