llvm dev - May 2021 - LLVM COBOL front end question [Re: A friendly question]

Hi,

I've renamed the thread so that interested folks might see it.

I think there are a bunch of somewhat conflated here and it's probably 
worth unpicking them a bit.  They seem to be:

  - Is anyone interested in writing a COBOL front end for LLVM?
  - Would a COBOL front end be considered for integration with the LLVM 
repository?
  - Would projects to work on an LLVM COBOL front end be considered for 
the GSoC or similar?

I can take a stab at answering these:

  - There are almost certainly some people interested in COBOL.  Most 
folks on this list probably don't care about Fortran, but that hasn't 
prevented flang from being written and integrated into the repository. 
Some of the IBM people might be interested.  I think the main question 
is what the differentiation is from gnucobol.  From a quick skim, 
gnucobol translates COBOL to C, so you can presumably already use it 
with clang and get an LLVM back end.  Writing a new front-end is a 
nontrivial task and so you'd have to explain what the benefit is of a 
new one.  Most of the codegen-related benefits don't apply (presumably 
you can already use clang + gnucobol to compile COBOL to IR and do LTO 
with it and C/C++/Fortran/Whatever code).  Would you want to reuse 
Clang's diagnostic infrastructure and provide better error messages?  Is 
it just a question of the license?

  - There's a difference of opinion in the community about the 
importance of being in-tree.  I'd personally prefer that clang and flang 
were both moved out of the monorepo to force us to think more about 
library interfaces.  The more LLVM developers are also working on 
out-of-tree projects, the better our libraries become.  That said, if 
you want a project to be considered for eventual inclusion in the tree, 
then making sure that the license is the same as the rest of the project 
and that the coding style is the same is a good first step.  You can 
then defer this decision until the front end is more mature and you see 
some clear benefit in being upstream.

  - For a project to be considered for the GSoC, the only real 
requirement is finding someone willing to mentor it.  That means that 
you need to find someone who is an existing LLVM contributor who is 
interested in COBOL and you need to provide a good answer to the 'why 
isn't gnucobol good enough?' question.

David


On 05/05/2021 22:44, pawel k. via llvm-dev wrote:
> Hello,
> Not sure there will be enough use for it but I was wondering whether 
> there would be enough interest in making llvm-based gnucobol project 
> clone (possibly launching summer of code project for it).
> 
> For now it works in such a way that it processes cobol sourcecode and 
> produces source in C.
> 
> In such way we could try to beat both gnu cobol and ibm releasing lately 
> cobol compiler on x86 linux (if im not mistaken wrt to OS).
> 
> Best regards,
> Pawel Kunio
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>

> I think there are a bunch of somewhat conflated here and it's probably
> worth unpicking them a bit.  They seem to be:
> 
>   - Is anyone interested in writing a COBOL front end for LLVM?
>   - Would a COBOL front end be considered for integration with the LLVM
> repository?
>   - Would projects to work on an LLVM COBOL front end be considered for
> the GSoC or similar?
Actually my first reaction was, retarget gnucobol to emit LLVM IR instead
of C, and run the LLVM backend on it.  (I'd speculated to myself about
doing this as a fun post-retirement project.)  There's probably licensing
questions to resolve there first.  I could see compilation speed getting
better as a result, although runtime performance is unlikely to be better
than routing through C.  Not clear what the effect would be on debugging;
it looks like there are a couple of packages out there that let you debug
COBOL using the existing gnucobol setup, and gdb doesn't understand COBOL
so if anything the debugging situation is likely to get worse.

I've retargeted a COBOL front-end in the past (not to LLVM), and I'd
estimate that it is significantly more work than one GSoC project.
(Not that I've actually looked inside gnucobol to see how it works.)

As for who's interested in a COBOL front-end:
The OpenVMS people (vmssoftware.com) are definitely headed in the direction 
of a native LLVM-speaking COBOL frontend, although I'm sure it would remain
proprietary.  That direction would presumably include adding DIBuilder 
features for the COBOL data types, and getting LLVM to emit the proper DWARF 
descriptions.  Haven't seen any signs of that happening upstream, though.

--paulr

On Thu, 6 May 2021 at 10:20, David Chisnall via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
>   - There's a difference of opinion in the community about the
> importance of being in-tree.

This "conflict" should now be somewhat reduced with the introduction
of
incubation projects, which is more official than a "user" project but
less
than in the monorepo.

As a first step, a GSoC project could create a side-project seed (on a
private repo) with some support for the language, using close to trunk LLVM.

Later, if the community picks up, there could be enough momentum for an
incubation project in the LLVM Github repo, where more people collaborate
and is under our management.

Only if the project really takes off and has a large subset of our
community caring about it, and frequently testing large amounts of code
against trunk, that it should be considered in the monorepo.

Not small steps, but smaller than in the SVN times... :D

--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210510/d5477e48/attachment.html>