llvm dev - May 2017 - Email list just for front end developers?

Pardon the length of this reply, but it seemed useful to be explicit
about my concerns.

On Sun, 7 May 2017 00:55:51 +0800 C Bergstrom
<cbergstrom at pathscale.com> wrote:
> You don't seem to be listening, but one last try.
I listened. I respectfully disagree.
> If you have a FE related question - Ask on cfe-dev and if you don't
> get a response I'd be surprised.
I would not be surprised.

I'll be straightforward. It is very hard to get answers to questions
on any of the lists about simple front end stuff, especially stupid
newbie questions. It is also hard to get answers to simple questions
in the IRC chat room. (I've tried, believe me.)

I can understand why. It isn't anyone's fault, and it isn't
unexpected. These questions aren't very interesting to people who are
deeply involved in the Clang infrastructure or the LLVM optimization
infrastructure, they're probably quite repetitive, and I'm sure that
sometimes they're annoying because they're so obvious and basic to
people who live and breathe LLVM. That's entirely understandable.

However, there are many people out there, maybe even a lot more than
are developing Clang in fact (though it is hard to know), who are
developers using LLVM for other languages.

Yes, I agree there are more _users_ of Clang than there are of Rust
and Julia and Rubymotion and Crystal and Clasp and the dozens of
others out there. After all, Clang is Apple's compiler, so there are
huge numbers of users of Clang. However, in terms of people
_developing_ other front ends, I get the impression there are a very
large number of us, because, frankly, right now LLVM is the most
rational choice for any compiler developer, and the world has a lot of
compilers under development. (Indeed, it has more than it used to
partially because LLVM exists.)

It makes sense to have a place for us to talk to each other where
we're the dominant traffic, where our questions for each other aren't
concealed amidst a very large amount of traffic on Clang's internals
(or the rest of LLVM's), and where and our questions unrelated to local
concerns are not viewed as an annoyance.

Again, I suspect front end developers will generally be more
sympathetic to each other's problems and questions than people who are
worried much more about Clang's internals or the backend internals.
> Additionally, if you don't get a response
It isn't a question of "a" response. I myself have a new question
almost every day. After one sees what happens when people ask
questions, one gets intimidated and does not wish to continue.

Here's an example. This post about a month ago from someone having a
great deal of trouble with very basic things got not a single answer
in response:

http://lists.llvm.org/pipermail/llvm-dev/2017-March/111526.html

I think that gentleman's situation is very typical. I've had questions
almost identical to his, too. I've been reduced to reading source code
instead of documentation, painfully, and making very slow progress
because of it.

Again, I don't blame anyone for this state of affairs. LLVM is open
source software, and no one has any duty to answer my questions, write
documentation for me, etc. I cannot reasonably ask that people who are
working hard on a new processor back end or new optimizations or what
have you take time out to help me. It isn't their job.

I would just like a place that makes self help among other people in
my position and that of the gentleman whose query I link to above more
likely to succeed.
> from there maybe then follow-up with your idea of another list.
I understand that it may be too much trouble to ask the LLVM project
to set up another mailman mailing list.

If that is the case, given that I have my own mailman setup, perhaps
I'll just create the list on my own machine and advertise it, and
perhaps create a Wiki. It would be unofficial so fewer people would
find it, which might doom it, but as I've indicated, I would like to
have such a thing, and I suspect others would, too, so perhaps that's
the route to take.

I don't want to bother people associated with the project overly much,
as frankly I'd prefer that all of you spend your time making LLVM even
better, and if the project doesn't want to create such a list I won't
push, I'll just do it on my own if it continues to seem important to me.
> What you're describing would be extremely low traffic
I very, very much doubt that will be the case.

I understand that this isn't obvious to those who are very familiar
with the way LLVM works, but again, it is challenging getting started
with LLVM once one is past the tutorial. One has endless questions
about details the tutorial doesn't cover, and the documentation sadly
doesn't help with them. Even when one is past that stage, one keeps
encountering issues one would like to ask about.

There are also a lot of people out there working on LLVM based front
end projects, far more even than LLVM web site indicates. There are
also whole classes of undergrads using LLVM in compiler courses and I
suspect they have questions too.

And, once again, most of us learn very quickly that simple newbie
questions generally don't get answered by busy people who have other
interests. (And again, this is understandable.)

That's fine, I understand why it is the case. I don't want to bother
others who have other concerns. I just want a place to ask
questions. If such a list existed, I would be able to contribute quite
constructively -- I've learned the answers to many of my own questions
"the hard way" by having to read the sources, and I could help others
to avoid having to do the same thing with things I understand. I would
also probably feel less put off asking simple questions on such a
list. I bet one of the developers of Rust or the JavaScript JIT or the
Common Lisp front end or Crystal or Julia or any of the literally
dozens of others would probably be happy to answer my questions,
having had the same steep learning curve themselves.
> and I personally doubt you'd get as much expertise interested in
> order to keep it viable.
Then it will fail. That's a risk with any new mailing list. Luckily,
the cost of mailing lists is very low, so the only risk is that the
list gets no traffic. If that happens, well, so be it.

Perry
-- 
Perry E. Metzger		perry at piermont.com

Hi,

On Sat, May 6, 2017 at 11:31 AM, Perry E. Metzger via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> It makes sense to have a place for us to talk to each other where
> we're the dominant traffic, where our questions for each other
aren't
> concealed amidst a very large amount of traffic on Clang's internals
> (or the rest of LLVM's), and where and our questions unrelated to local
> concerns are not viewed as an annoyance.
Why not just filter traffic on llvm-dev with "emails that haven't been
replied to within a week"?  Right now llvm-dev is, say, ~220 emails a
week, so after that filter you should a small enough set of threads
left which you can manually inspect.

I don't think people will be annoyed by frontend specific questions on
llvm-dev unless there is a very high number of them (unlikely), at
which point we can //maybe// consider splitting out a new mailing list
(but see below).
> I would just like a place that makes self help among other people in
> my position and that of the gentleman whose query I link to above more
> likely to succeed.
I agree that we do not do the best possible job in answering beginner
questions, but is there something preventing you from changing that
trend on llvm-dev, and making llvm-dev be that place where beginner
questions are reliably answered?
>> and I personally doubt you'd get as much expertise interested in
>> order to keep it viable.
>
> Then it will fail. That's a risk with any new mailing list. Luckily,
> the cost of mailing lists is very low, so the only risk is that the
> list gets no traffic. If that happens, well, so be it.
There are other risks too - fragmentation (I have to sign up for yet
another mailing list), discoverability (at this point many people
already know about llvm-dev, how do they find out about the
llvm-frontend-dev?).

-- Sanjoy

On Sat, 6 May 2017 12:06:03 -0700 Sanjoy Das
<sanjoy at playingwithpointers.com> wrote:
> Hi,
> 
> On Sat, May 6, 2017 at 11:31 AM, Perry E. Metzger via llvm-dev
> <llvm-dev at lists.llvm.org> wrote:
> > It makes sense to have a place for us to talk to each other where
> > we're the dominant traffic, where our questions for each other
> > aren't concealed amidst a very large amount of traffic on
Clang's
> > internals (or the rest of LLVM's), and where and our questions
> > unrelated to local concerns are not viewed as an annoyance.  
> 
> Why not just filter traffic on llvm-dev with "emails that haven't
> been replied to within a week"?  Right now llvm-dev is, say, ~220
> emails a week, so after that filter you should a small enough set
> of threads left which you can manually inspect.
I'm not sure that will work well, because people won't usually wait a
week to get their questions before moving on. If you're actively
working full time on a project, you probably want answers in much less
time.

Usually a question is inspired because you're trying to do something
immediately (say, in my own case, when I was working on real constants
and realized I couldn't find much documentation for what formats
APFloat understands when you pass the constructor a string), and
laying tools down for that long isn't an option. If no one answers in
less than a day, you probably have to go and slog through the
implementation on your own, which is of course sub-optimal.
> I don't think people will be annoyed by frontend specific questions
> on llvm-dev unless there is a very high number of them (unlikely),
> at which point we can //maybe// consider splitting out a new
> mailing list (but see below).
I doubt that there ever would be a lot of them as it stands, because I
think people find that they don't get answered and that they feel
weird about asking more questions of the same form. After a while a
culture gets entrenched.

Note that it isn't even something weird or inexplicable, I understand
why this sort of thing happens. llvm-dev serves a vital need already,
which is letting the llvm developers communicate with each other on
technical matters. Unusual traffic that isn't really any one person's
area gets less mental space.

So you get a dual problem either here or on the cfe lists, which is
that the normal traffic isn't traffic you understand (or care about)
and that the traffic that you are interested in isn't of much interest
to the overwhelming bulk of the list.
> > I would just like a place that makes self help among other people
> > in my position and that of the gentleman whose query I link to
> > above more likely to succeed.
>
> I agree that we do not do the best possible job in answering
> beginner questions, but is there something preventing you from
> changing that trend on llvm-dev, and making llvm-dev be that place
> where beginner questions are reliably answered?
Well, partially, I think, that it is hard to change a culture,
partially that llvm-dev isn't geared towards it (I myself don't
understand much of the traffic since I don't know anything past
IRBuilder well and am unlikely to dive into that soon, partially
because LLVM does such a good job at separating concerns that I don't
_have_ to understand the middle and back end work.)

However, if people think that just making a more determined effort to
answer newbie questions here is the best way to go, we can try, and
I'll certainly help, since I've been learning what many of them are
the hard way.

Again, though, I didn't even notice that message I posted asking for
help until long after that guy probably didn't care any more. It
probably remains hard answer such questions in an environment where
there's an impedance mismatch between the topics. As I noted, most of
the traffic here isn't even about stuff I understand well, I'm here
(and on the Clang lists) only because they're the closest thing to
what I actually _do_ need.

And again, the impedance mismatch is unlikely to be malice against
newcomers or any such thing. I think it's because, de facto, this is
where people discuss middle and back end development, and it is the
only good channel for that, and there's apparently a lot of that to
discuss.
> >> and I personally doubt you'd get as much expertise interested
in
> >> order to keep it viable.
> >
> > Then it will fail. That's a risk with any new mailing list.
> > Luckily, the cost of mailing lists is very low, so the only risk
> > is that the list gets no traffic. If that happens, well, so be
> > it.
>
> There are other risks too - fragmentation (I have to sign up for yet
> another mailing list), discoverability (at this point many people
> already know about llvm-dev, how do they find out about the
> llvm-frontend-dev?).
Fragmentation is a real problem, especially since the people who are
often best equipped to answer the questions are people working on the
middle and back end. That could absolutely make this fail
hard. However, as it stands, such people often don't have much time to
answer anyway, or so it seems. To be sure, though, such an experiment
is most likely to fail for that reason.

I'm less worried about discoverability. One good post to the
announcements list and bit on LLVM weekly, and having it documented in
the normal places (lists.llvm.org, the various resources web pages)
should be enough.

And, if it fails, one can always turn it off, merge it back in, and
try something else. The cost is low.

But again, that said, if people really think that this is the right
place for such things and not a new list, that can probably be given a
good faith effort. My perception was that it probably wouldn't work
well because this list is de facto for something else, but I'm willing
to try if that's a strong consensus.

*That* said, I still think a Wiki would also be nice. When one finds a
hard-fought answer to a question, putting that answer somewhere beats
having the information only in people's heads. Yes, for sure, the
"right thing" is to submit a documentation patch, but there's a
high
barrier involved for an outsider to do that, and putting the
information somewhere is better than having it nowhere. The only real
problem with a Wiki is it requires policing, they acquire spam at a
prodigious rate.

Perry
-- 
Perry E. Metzger		perry at piermont.com