Re: [PATCH] Do not redirect ld stdout/stderr in collect2 with -debug

[Richard Biener]

On Sat, 21 Feb 2015, Thomas Schwinge wrote:

> Hi!
> 
> On Sat, 21 Feb 2015 17:22:40 +0100, I wrote:
> > On Mon, 10 Jun 2013 14:01:53 +0200 (CEST), Richard Biener 
> > <rguenther@suse.de> wrote:
> > > This fixes one very annoying thing collect2 does when trying to
> > > debug LTO WPA issues.  Even with -v you need to wait until all
> > > LTRANS stages completed to see the lto1 -fwpa invocation which
> > > is because collect2 buffers and replays stdout/stderr of ld
> > > (to avoid duplicating that in some cases).  But I really want
> > > to see the output immediately but there is no way to do that.
> > > The easiest is to disable the buffering with -debug (that is,
> > > -Wl,-debug to the -flto driver command line).
> > > 
> > > Tested with/without -debug.
> > > 
> > > Ok for trunk?
> > 
> > >   * collect2.c (main): Do not redirect ld stdout/stderr when
> > >   debugging.
> > 
> > > --- gcc/collect2.c        (revision 199732)
> > > +++ gcc/collect2.c        (working copy)
> > > @@ -1189,8 +1189,11 @@ main (int argc, char **argv)
> > >  #ifdef COLLECT_EXPORT_LIST
> > >    export_file = make_temp_file (".x");
> > >  #endif
> > > -  ldout = make_temp_file (".ld");
> > > -  lderrout = make_temp_file (".le");
> > > +  if (!debug)
> > > +    {
> > > +      ldout = make_temp_file (".ld");
> > > +      lderrout = make_temp_file (".le");
> > > +    }
> > >    *c_ptr++ = c_file_name;
> > >    *c_ptr++ = "-x";
> > >    *c_ptr++ = "c";
> > 
> > This change (r199936) is problematic, given the usage of ldout and
> > lderrout in gcc/tlink.c:do_tlink.  If, for -debug, they're not
> > initialized, they'll be NULL, which in do_tlink, the tlink_execute call
> > will handle fine (as I understand it), but after that, for example,
> > dump_ld_file will attempt to fopen (NULL), which will cause a SIGSEGV on
> > GNU Hurd at least.  (Correct me if I'm wrong -- I have not yet read the
> > relevant standards in detail -- but from what I remember, that's
> > appropriate: NULL is not a string naming a file.)
> > 
> > I found this when running binutils' gold testsuite:
> > 
> >     $ gcc-4.9 -W -Wall -Wstrict-prototypes -Wmissing-prototypes -Wshadow 
> > -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -fmerge-constants -g -O2 
> > -fno-use-linker-plugin -o incremental_test -Bgcctestdir/ 
> > -Wl,--incremental-full incremental_test_1.o incremental_test_2.o -v 
> > -Wl,-debug
> >     [...]
> >     gcc-4.9: internal compiler error: Segmentation fault (program collect2)
> >     [...]
> > 
> > In do_tlink, the last four uses of ldout and lderrout should be guarded
> > by NULL and empty string checks (as done in gcc/collect2.c:tool_cleanup),
> > but I'm not sure what the correct fix is in the »if (ret)« block:
> > 
> >     void
> >     do_tlink (char **ld_argv, char **object_lst ATTRIBUTE_UNUSED)
> >     {
> >       int ret = tlink_execute ("ld", ld_argv, ldout, lderrout,
> >                                HAVE_GNU_LD && at_file_supplied);
> >     
> >       tlink_init ();
> >     
> >       if (ret)
> >         {
> >           int i = 0;
> >     
> >           /* Until collect does a better job of figuring out which are 
> > object
> >              files, assume that everything on the command line could be.  */
> >           if (read_repo_files (ld_argv))
> >             while (ret && i++ < MAX_ITERATIONS)
> >               {
> >                 if (tlink_verbose >= 3)
> >                   {
> >                     dump_ld_file (ldout, stdout);
> >                     dump_ld_file (lderrout, stderr);
> >                   }
> >                 demangle_new_symbols ();
> >                 if (! scan_linker_output (ldout)
> >                     && ! scan_linker_output (lderrout))
> >                   break;
> >                 if (! recompile_files ())
> >                   break;
> >                 if (tlink_verbose)
> >                   fprintf (stderr, _("collect: relinking\n"));
> >                 ret = tlink_execute ("ld", ld_argv, ldout, lderrout,
> >                                      HAVE_GNU_LD && at_file_supplied);
> >               }
> >         }
> >     
> >       dump_ld_file (ldout, stdout);
> >       unlink (ldout);
> >       dump_ld_file (lderrout, stderr);
> >       unlink (lderrout);
> >     [...]
> 
> By the way, to make progress without having to rebuild Debian's gcc-4.9
> package, I had the idea of creating a LD_PRELOAD wrapper to wrap the
> fopen (NULL) and unlink (NULL) calls, but this turned out to be an
> interesting exercise in its own right: I relatively quickly realized that
> I actually need to wrap fopen64 instead of fopen, but it took me longer
> to realize why the unlink wrapper just didn't work.  GCC has been happily
> optimizing away my path != NULL check, because of:
> 
> /usr/include/unistd.h:
> 
>     extern int unlink (const char *__name) __THROW __nonnull ((1));
> 
> /usr/include/i386-gnu/sys/cdefs.h:
> 
>     /* The nonull function attribute allows to mark pointer parameters which
>        must not be NULL.  */
>     #if __GNUC_PREREQ (3,3)
>     # define __nonnull(params) __attribute__ ((__nonnull__ params))
>     #else
>     # define __nonnull(params)
>     #endif
> 
> Certainly this is another indication that unlink (NULL) really isn't
> meant to be done.  ;-)
> 
> Got this resolved by defusing the __nonnull macro.  (See attached.  Not
> sure if I'm using the atomic builtins correctly.)

Well, the issue should be as simple as to guard the dump_ld_file
calls properly, like collect2.c tool_cleanup does.

Richard.