Re: infrastructure for suggested/required packages, and better error reporting

[Ralf Wildenhues]

Hi Bob,

* Bob Friesenhahn wrote on Sat, Jul 31, 2010 at 06:15:32PM CEST:
> On Sat, 31 Jul 2010, Ralf Wildenhues wrote:
> >
> >So, we should have something that allows an error message but also
> >continues the configure script in search for other possible errors,
> >and summarizes the suggestions and errors at the end.
> 
> Library dependencies often behave like a pyramid so the first
> failure is often likely to cause all of the other tests to fail.

Sure.  I'm not suggesting that errors suddenly cause the script to
continue, or that all library tests should use a nonfatal error.
But configure.ac authors that can handle the complication should be
able to deal with more complex situations and be able to at least
report all build requirements that can be checked independently.

And I think distro packagers that could just use some autoconf --trace
command to extract dependency sets without even running configure would
be a nice boon.

> What is sorely lacking from Autotools is a proper way to determine
> library dependencies.  Many packages uses pkg-config for this
> purpose, but it becomes quite unreliable in complex situations.  If
> libtool .la files were not deleted, then the configure script would
> have a way to query and determine the actual dependencies.

I see several non-exclusive strategies:

1) fix pkg-config,
2) fix libtool and its .la files, then get people to not remove them any
more,
3) invent yet another method.

I'd rather avoid (3), and I know (2) is going to be quite some work, but
should be done.  What things go wrong with pkg-config?  (Please adjust
the Subject: when replying to this, thanks!)

Cheers,
Ralf