[Gomp-discuss] Re: OpenMP, HPC, and the future of GCC

[Per Bothner]

Steven Bosscher wrote:
> Nobody has said you're responsible for that.  Nobody has said you're
> required to merge it.

No, but the following (your words) comes awfully close to asking for
pre-approval:

> But if that is how the majority of the GCC developer community feels
> about OpenMP, then why would anyone spend time on developing the idea?
> I know that I'd rather spend a few hours doing some contract research
> to fund my new vendor compiler if nobody will ever accept my patches
> for GCC.

> b) That is whoever is willing and capable (by the judgement of whoever
> rejects or accepts patches) of helping with the development of GCC. 
> Clearly you don't understand this, instead you seem to be a purist with
> no sense of what is asked for in a compiler in the real world.

No, I'm saying we cannot be all things to all people, nor can we
accept all patches, even if they are useful in themselves.

> Oh is it?  I don't think so.  If you already know you would reject it
> because *you* don't like the desing, then why don't you say that now.

As to the design of the specification:  The first impression is not
positive, but perhaps with further study I could appreciate its
virtues.  In any case, it's a standard, and it seems to be popular,
so it is reasonable to implement it.

As to the design of the implementation:  That is the issue.  It may
be difficult to implement it cleanly, and the advice from Neil Booth,
David Edelsohn, Jan Hubicka, and others should be taken to heart.

> BS.
>
> I'll buy you a beer (if I ever meet you:) if you can name me just one
> such designs [for 'multi threading' or 'native parallellism'] that
> works out-of-the-box for C/C++ and Fortran and that is as populair
> as OpenMP.

Ok, I get it:  OpenMP is popular.  Of course that is no guarantee
of good design or long life.

> Also, we've said before that we don't design concurrency support around
> OpenMP, but rather we want to develop a framework to support
> concurrency.  OpenMP would be just one thing that can be built on that.

Great.

> Also, you said this morning that C++ isn't so well designed.

That is not what I said.  I said C++ is an excessingly complicated
language.  I think very few people will dispute that.  It is also
a very powerful and useful language.  I think C++ is a much better
language to write (say) compilers in than C, or Java (if you want
it to be fast and/or small).

> Why do you think more people use C++ than Java?

Are you sure?

> You probably don't understand that, ...

I have many years of C++ programming experience. I took over from
Doug Lea as the official maintainer of the GNU C++ library libg++,
and maintained it for years; if you look at /usr/include/libio.h
on a GNU/Linux system you will find my name as the author of the
original combined C stdio and C++ iostream library; and I was briefly
a member of the C++ standardization committee.  I understand very
well the virtues and usefulness of C++.

> If you want to have a compiler that only implements "quality standards"
> by your measures, then clearly GCC is just an academic project to you.

> Vectorizing is orthogonal to this discussion.  It has nothing to do with
> OpenMP.  Never has, never will.
>
> And, nobody was even talking about vectorization in this thread, let
> alone "new vectorization" whatever that may be.

I was using vectorization as an example of a HPC concern, which is
mentioned in teh subject line of this thread.

> Oh hear, is this a SC member is speaking?

Yes, but not officially.

> I'm sorry, but your whole mail just sucks, and your attitude makes me
> sick.

Oh, grow up.
--
        --Per Bothner
address@hidden   http://www.bothner.com/per/