Re: MPS: unable to build due to assertion violation in igc_dump_check_object_starts

[Gerd Möllmann]

Eli Zaretskii <eliz@gnu.org> writes:

>> From: Gerd Möllmann <gerd.moellmann@gmail.com>
>> Cc: pipcet@protonmail.com,  eller.helmut@gmail.com,  emacs-devel@gnu.org
>> Date: Tue, 23 Jul 2024 16:52:42 +0200
>> 
>> Could you do
>> 
>>   p header_type(start)
>> 
>> please, that should show the IGC_OBJ_xy type of the offending header
>> 
>>   p header_nwords(start)
>> 
>> is the, presumably wrong, size in words, but I think we have already
>> seen it's too large.
>
> Compliance!
>
>   (gdb) p header_type(start)
>   $7 = 58
>   (gdb) p (enum igc_obj_type)header_type(start)
>   $8 = 58
>   (gdb) p IGC_OBJ_NUM_TYPES+0
>   $9 = 31
>   (gdb) p header_nwords(start)
>   $10 = 200437507

So the igc_header is completely kaputt.

But we can try to use the value of start_off at the time the assertion
is triggered. The list containing the start/end offsets should be
created here:

  static void
  dump_igc_finish_obj (struct dump_context *ctx)
  {
    if (ctx->flags.dump_object_contents)
      {
        char *base = (char *) ctx->buf + ctx->igc_base_offset;
        char *end = (char *) ctx->buf + ctx->offset;
        eassert (end > base);
        char *should_end = igc_dump_finish_obj (ctx->igc_obj_dumped, 
ctx->igc_type, base, end);
        eassert (should_end >= end);
        dump_write_zero (ctx, should_end - end);
        if (ctx->flags.record_object_starts)
          dump_push (&ctx->igc_object_starts,
                     list2 (dump_off_to_lisp (ctx->igc_base_offset),
                            dump_off_to_lisp (ctx->offset)));
Here                 ^^^^^^

      }

    ctx->igc_obj_dumped = NULL;
    ctx->igc_type = IGC_OBJ_INVALID;
    ctx->igc_base_offset = -1;
  }
  # endif // HAVE_MPS

If I'm not completely mistaken, start_off when asserting should ==
ctx->igc_base_offset at some point whle dumping. Could you perhaps put a
conditional breakpoint there? We could then go up the stack and see
which dump_xy function that is.

Maybe a short overview what's happening here, as an orientation. In
pdumper object dumping happens by first calling dump_start_object, then
dumping the fields or whatever that object contains, and finally calling
dump_finish_object. For MPS, dump_start_object writes an empty
igc_header, and dump_finish_object fills out the igc_header, since we
then know the size of the object and so on. It also records the object
start/end in ctx->igc_object_starts, which is sort of a remnant from the
itmes that the dump was a root, and not in MPS. But the object starts
are apparently still helpful.