Re: Autoconf and Solaris Lint

[Ralf Wildenhues]

Hello Jason,

* Jason Curl wrote on Wed, Jun 20, 2007 at 11:55:54PM CEST:
>
> I've searched in vain on the Web how I might run configure to use Solaris' 
> 'lint' program. It appears that their 'lint' is very much like a compiler 
> where it produces objects that can be linked together to form a final 
> result.
>
> The command below doesn't work.
> CC=lint ./configure --host=sparc-sun-solaris2.10

I'd be interested to see which tests come out wrong with the above.
Could you take a look at config.log and search for obviously wrong
results?

Next, it would help to see how this fares:
  ./configure
  env CC=lint make -e

but of course this is a bit more cumbersome to use.

FWIW, I have access to a Solaris 2.8 host that has a lint installed,
and, just as an example package, I got CVS HEAD Libtool configured with
something like
  .../configure CC=lint LD=/usr/ccs/bin/ld cross_compile=yes

The build then fails due to the ugliness that is ltdl.c  ;-)
and issues listed below, but otherwise looks like it should more or
less work for other packages.

False test result are experienced with AC_CHECK_LIB, AC_SEARCH_LIBS,
even if you have previously installed (or compiled) lint libraries
somewhere and pointed to them with LDFLAGS=-L/path/to/lint/libs.
(This lint has means to create files that describe APIs for object files
and libraries, which it calls 'lint libraries'; such pre-generated
libraries are available for libc and many others.)  However, the syntax
this lint uses (`-o' for naming the lint library to be created, and also
turning off undefined functions warnings) is incompatible with Autoconf
tests that use `-o conftest' and later check for this program.
Similar issues exist of course for AC_CHECK_FUNC*.

The default compiler output name is detected wrongly as b.out
(but that is mostly inconsequential as we're talking configure
into a cross-compile situation anyway).

I did not see other obvious false test results.  It would help to get
exposure by trying more packages (and yes, I have not yet tried to
compare Autoconf's test suite's output with this).

All of these failures may be worked around by pre-seeding configure
in a config.site file with the correct values for the system in
question (say, taken from a config.cache file generated by a configure
run that uses CC=cc).

(Unrelated side note FYI: if you pass CC as argument rather than in the
environment, then config.status will remember it when it triggers a
recheck.)

> Anybody know of a project that has Solaris' lint incorporated that I can 
> use as an example? I'm guessing it needs changes to the Makefile.am files I 
> have to add new targets.

Actually, I would hope that we could patch Autoconf (and maybe some
third-party macros) a little so that it would not need any changes to
your projects.

> Right now to capture any mistakes I make I run GCC 
> and Solaris' CC. GCC doesn't capture some of the errors, especially type 
> related ones.

FWIW, other lint-style checkers that I've found helpful before:
<http://www.splint.org/> (development has stalled a bit here),
<http://www.kernel.org/pub/software/devel/sparse/>.

> As I'm pretty new to extending Automake/Autoconf, if somebody could hint me 
> to a framework that would be most helpful.

I hope this helps a bit.

Cheers,
Ralf