[Bug lto/54962] New: Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[Bug lto/54962] New: Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[dmalcolm at redhat dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

             Bug #: 54962
           Summary: Strange-looking diagnostics from
                    diagnostic_report_current_module() from warnings
                    emitted during LTO
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: lto
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: dmalcolm@redhat.com


I'm working on a whole-program static analyzer that runs from a gcc plugin
during lto1 when invoked via
  -flto -flto-partition=none
as an IPA_PASS registered before "whole-program".

I'm currently emitting errors using error_at() with inform() to give extra
information on the (possibly interprocedural) execution path taken to reach the
error.  This works, but leads to unusual-looking output.

For example, with:
  $ gcc -flto -flto-partition=none [...various plugin args...] input-f.c
input-g.c input-h.c

my code detects a double-free involving an execution path between input-h.c and
input-g.c and calls
  error_at("double-free of r")
on a location_t from input-h.c and gets this strange output on stderr:

  In file included from test.h:1:0,
                   from /usr/include/stdlib.h:489,
                   from input-f.c:14,
                   from :3:
  input-h.c:13:9: error: double-free of r

it then calls
    inform("q passed to free()")
on a location_t from input-g.c (to show the location of the 1st free call) and
gets this on stderr:

  In file included from test.h:1:0,
                   from /usr/include/stdlib.h:489,
                   from input-f.c:14,
                   from :3:
  input-g.c:11:7: note: q passed to free()

In both cases, the "In file included from " hierarchies look bogus: the files
are being directly compiled, not included.

Debugging, it comes from gcc/diagnostic.c:diagnostic_report_current_module()
where I see this code:

  275      if (map && diagnostic_last_module_changed (context, map))
  276        {
  277          diagnostic_set_last_module (context, map);
  278          if (! MAIN_FILE_P (map))
  279        {
  280          map = INCLUDED_FROM (line_table, map);
  281          if (context->show_column)
  282            pp_verbatim (context->printer,
  283                 "In file included from %s:%d:%d",
  284                 LINEMAP_FILE (map),
  285                 LAST_SOURCE_LINE (map), LAST_SOURCE_COLUMN (map));
  286          else
  287            pp_verbatim (context->printer,
  288                 "In file included from %s:%d",
  289                 LINEMAP_FILE (map), LAST_SOURCE_LINE (map));
  290          while (! MAIN_FILE_P (map))
  291            {
  292              map = INCLUDED_FROM (line_table, map);

This code appears to assume (via MAIN_FILE_P) that there is a concept of a main
file, and that everything else is included from it.  However in the context of
diagnostics emitted during LTO it's not clear that this concept makes sense (or
maybe I'm missing something?  I also wonder if by attempting to issue
diagnostics from within lto1 if I'm in unexplored territory here?)

This is called by default_tree_diagnostic_starter FWIW; perhaps lto1 needs its
own implementation of this?

[Bug lto/54962] Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[manu at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

Manuel López-Ibáñez <manu at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |manu at gcc dot gnu.org

--- Comment #1 from Manuel López-Ibáñez <manu at gcc dot gnu.org> 2012-10-17 20:43:51 UTC ---
Where doe the line_maps that LTO use come from? Perhaps they are not
saved/restored correctly?

[Bug lto/54962] Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-10-18
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> 2012-10-18 09:55:45 UTC ---
(In reply to comment #0)
>   290          while (! MAIN_FILE_P (map))
>   291            {
>   292              map = INCLUDED_FROM (line_table, map);
> 
> This code appears to assume (via MAIN_FILE_P) that there is a concept of a main
> file, and that everything else is included from it.  However in the context of
> diagnostics emitted during LTO it's not clear that this concept makes sense (or
> maybe I'm missing something?  I also wonder if by attempting to issue
> diagnostics from within lto1 if I'm in unexplored territory here?)

You definitely are, if you are expecting everything to work just fine ;)

> This is called by default_tree_diagnostic_starter FWIW; perhaps lto1 needs its
> own implementation of this?

Maybe yes.  LTO should print the "main" file as well (so it doesn't have
a MAIN_FILE_P), but not as included-from but in some other way.

Patches welcome ;)

[Bug lto/54962] Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[manu at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

--- Comment #3 from Manuel López-Ibáñez <manu at gcc dot gnu.org> 2012-10-18 10:39:08 UTC ---
(In reply to comment #2)
> > This is called by default_tree_diagnostic_starter FWIW; perhaps lto1 needs its
> > own implementation of this?
> 
> Maybe yes.  LTO should print the "main" file as well (so it doesn't have
> a MAIN_FILE_P), but not as included-from but in some other way.
> 

How does the line_map works in LTO? Is it saved and restored or re-build from
scratch?

[Bug lto/54962] Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> 2012-10-18 10:40:34 UTC ---
(In reply to comment #3)
> (In reply to comment #2)
> > > This is called by default_tree_diagnostic_starter FWIW; perhaps lto1 needs its
> > > own implementation of this?
> > 
> > Maybe yes.  LTO should print the "main" file as well (so it doesn't have
> > a MAIN_FILE_P), but not as included-from but in some other way.
> > 
> 
> How does the line_map works in LTO? Is it saved and restored or re-build from
> scratch?

We stream the expanded location and allocate new line-map entries at LTO
read time.

[Bug lto/54962] Strange-looking diagnostics from diagnostic_report_current_module() from warnings emitted during LTO

[manu at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54962

--- Comment #5 from Manuel López-Ibáñez <manu at gcc dot gnu.org> 2012-10-18 11:38:33 UTC ---
(In reply to comment #4)
> 
> We stream the expanded location and allocate new line-map entries at LTO
> read time.

Where? 

I guess this precludes any knowledge of what is included from where and for
sure no tracking of macro expansion, but it still doesn't explain why
MAIN_FILE_P and INCLUDED_FROM do not work reliably. They should be true for
every map and return NULL, respectively. That seems like a bug.