Re: incomplete warning message

[Eric Blake]

On 08/30/2010 01:50 PM, Eric Blake wrote:
> > $ autoconf
> > configure.ac:63: warning: gl_LIBUNISTRING_LIBSOURCE was called before
> > gl_LIBUNISTRING
> > m4/libunistring-base.m4:21: gl_LIBUNISTRING_LIBSOURCE is expanded from...
> > m4/gnulib-comp.m4:148: gl_INIT is expanded from...
> > configure.ac:63: the top level
> > configure.ac:63: warning: gl_LIBUNISTRING_LIBHEADER was called before
> > gl_LIBUNISTRING
> > m4/libunistring-base.m4:53: gl_LIBUNISTRING_LIBHEADER is expanded from...
> >
> > The last line looks like the stack trace would include at least one
> > more line.
> > But it ends abruptly.
>
> I'm wondering if somehow the mere act of printing the backtrace is
> deleting the stacktrace. I think this one should be easy enough to track
> down before 2.68, so I'm reviving it on my list of hot TODO items this
> week.

The backtrace produced by m4 is just fine.  For proof, look at 
autom4te.cache/traces.n (for the appropriate n).  I think what is 
happening is that autom4te is then passing these messages on through 
perl, and perl is doing some sort of internal caching about whether a 
particular warning line has already been printed, in which case it omits 
the second printout.  Since autom4te is handling each location line as a 
separate warning, rather than passing the entire warning, location stack 
and all, as a single entity, this means that all duplicated location 
lines are getting squelched at the perl layer.

> > How to reproduce:
> > $ wget http://www.haible.de/bruno/gnu/autoconf-bug-20100531.tar.gz
> > $ tar xvfz autoconf-bug-20100531.tar.gz
> > $ cd autoconf-bug-20100531
> > $ autoconf
>
> I saw it as well last week when cleaning up some gnulib
> AC_COMPILE_IFELSE usage patterns.

This is the area of autom4te that is introducing the problem:

handle_traces ($req, "$tmp/warnings",
               ('_m4_warn' => "\$1::\$f:\$l::\$2::\$3$separator"));
# Swallow excessive newlines.
for (split (/\n*$separator\n*/o, contents ("$tmp/warnings")))
{
  # The message looks like:
  # | syntax::input.as:5::ouch
  # | ::input.as:4: baz is expanded from...
  # | input.as:2: bar is expanded from...
  # | input.as:3: foo is expanded from...
  # | input.as:5: the top level
  my ($cat, $loc, $msg, $stacktrace) = split ('::', $_, 4);
  msg $cat, $loc, "warning: $msg";
  for (split /\n/, $stacktrace)
    {
      my ($loc, $trace) = split (': ', $_, 2);
      msg $cat, $loc, $trace;
    }
}

However, my perl is rather weak, so I'm not sure if I can quickly 
correct it.  Help would be appreciated.

Hmm, reading lib/Autom4te/Channels.pm is proving to be enlightening:

> =item C<partial =E<gt> 0>
>
> When set, indicates a partial message that should
> be output along with the next message with C<partial> unset.
> Several partial messages can be stacked this way.
>
> Duplicate filtering will apply to the I<global> message resulting from
> all I<partial> messages, using the options from the last (non-partial)
> message.  Linking associated messages is the main reason to use this
> option.
>
> For instance the following messages
>
>   msg 'channel', 'foo:2', 'redefinition of A ...';
>   msg 'channel', 'foo:1', '... A previously defined here';
>   msg 'channel', 'foo:3', 'redefinition of A ...';
>   msg 'channel', 'foo:1', '... A previously defined here';
>
> will result in
>
>  foo:2: redefinition of A ...
>  foo:1: ... A previously defined here
>  foo:3: redefinition of A ...
>
> where the duplicate "I<... A previously defined here>" has been
> filtered out.
>
> Linking these messages using C<partial> as follows will prevent the
> fourth message to disappear.
>
>   msg 'channel', 'foo:2', 'redefinition of A ...', partial => 1;
>   msg 'channel', 'foo:1', '... A previously defined here';
>   msg 'channel', 'foo:3', 'redefinition of A ...', partial => 1;
>   msg 'channel', 'foo:1', '... A previously defined here';
>
> Note that because the stack of C<partial> messages is printed with the
> first non-C<partial> message, most options of C<partial> messages will
> be ignored.

Maybe the solution is just to teach autom4te that warning outputs 
consist of partial messages.

--
Eric Blake   address@hidden    +1-801-349-2682
Libvirt virtualization library http://libvirt.org