Re: MPS: Forwording symbols

[Gerd Möllmann]

Gerd Möllmann <gerd.moellmann@gmail.com> writes:

> Gerd Möllmann <gerd.moellmann@gmail.com> writes:
>
>> Helmut Eller <eller.helmut@gmail.com> writes:
>>
>>> On Mon, Jun 17 2024, Gerd Möllmann wrote:
>>>
>>>> The whole dumping of these structs looks highly dubious to me,
>>>> independent of MPS. They are constants, always have been, AFAICT, even
>>>> from what I remember from the 90s.
>>>
>>> The patch below creates the necessary relocs so that no forwarding
>>> structs end up in the dump.
>>
>> Thanks. pushed.
>>
>>>
>>>> Looking forward to the time when the mirror code is gone :-).
>>>
>>> I'm not so optimistic.  The first collection is very slow:
>>>
>>>    time ./emacs -batch -f igc--collect
>>>    
>>>    real    0m12.555s
>>>    user    0m11.708s
>>>    sys     0m0.097s
>>
>> That's pretty slow, indeed. The version of Emacs I'm currently running,
>> which is my fork which does not contain your recent changes, and is an
>> optimized build with -lmps says
>>
>> .../emacs/github/igc % /usr/bin/time src/emacs -batch -f igc--collect
>>         0.25 real         0.14 user         0.11 sys
>>
>>> That's not good.  Maybe there is some unfixed problem that is causing
>>> this slowness, but it could also be that MPS simply is so slow for this
>>> allocation pattern.
>>
>> I guess it is what you suspect because I run into an assertion when dumping:
>>
>>   Dumping fingerprint: 
>> 79b503a407de48b18f4c304971c234d89bf236c259a1e7cd857559f919943b91
>>
>>   igc.c:345: Emacs fatal error: assertion failed: h->obj_type == IGC_OBJ_PAD 
>> || nbytes >= sizeof (struct igc_fwd)
>>   Fatal error 6: Aborted
>>
>> That's a build starting from git clean -xdf, optimized, -lmps, and no
>> native compilation here of course.
>>
>> I'll try to debug this, but I think I have to read the new code first a
>> bit to understand where to start.
>
> Anyway, maybe the following gives you a hint already. The abort happens
> with this backtrace while writing the dump
>
> * thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 5.2
>     frame #0: 0x00000001001e3cf8 bootstrap-emacs`emacs_abort at 
> sysdep.c:2391:3
>     frame #1: 0x00000001003ec57c bootstrap-emacs`ns_term_shutdown(sig=6) at 
> nsterm.m:5889:5
>     frame #2: 0x00000001001a2dbc bootstrap-emacs`shut_down_emacs(sig=6, 
> stuff=(struct Lisp_Symbol *) $2 = 0x0000000100ca5010) at emacs.c:3162:3
>     frame #3: 0x00000001001a27a4 
> bootstrap-emacs`terminate_due_to_signal(sig=6, backtrace_limit=2147483647) at 
> emacs.c:464:11
>     frame #4: 0x0000000100399fa4 
> bootstrap-emacs`igc_assert_fail(file="igc.c", line=345, msg="h->obj_type == 
> IGC_OBJ_PAD || nbytes >= sizeof (struct igc_fwd)") at igc.c:82:3
>     frame #5: 0x000000010039cde8 
> bootstrap-emacs`obj_size(h=0x00000001298085f8) at igc.c:345:3
>     frame #6: 0x000000010039cc6c 
> bootstrap-emacs`igc_dump_finish_obj(client=0x0000000129808600, 
> type=IGC_OBJ_DUMPED_BUFFER_TEXT, base="", end="") at igc.c:3721:7
>     frame #7: 0x0000000100277fc0 
> bootstrap-emacs`dump_igc_finish_obj(ctx=0x000000016fdfc3e8) at 
> pdumper.c:910:26
>   * frame #8: 0x00000001002840d0 
> bootstrap-emacs`dump_cold_buffer(ctx=0x000000016fdfc3e8, data=(struct buffer 
> *) $3 = 0x000000010e4d3308) at pdumper.c:3694:3
>
> here:
>   char *
>   igc_dump_finish_obj (void *client, enum igc_obj_type type,
>                        char *base, char *end)
>   {
>     if (client == NULL)
>       return end;
>
>     struct igc_header *out = (struct igc_header *) base;
>     if (is_mps (client))
>       {
>         struct igc_header *h = client_to_base (client);
>         if (h->obj_type == IGC_OBJ_VECTOR_WEAK)
>           igc_assert (
>             (type == IGC_OBJ_VECTOR && h->obj_type == IGC_OBJ_VECTOR_WEAK)
>             || h->obj_type == type);
>         igc_assert (base + obj_size (h) >= end);
>
> Looks like the client object passed in is bogus
>
>   (lldb) p *h
>   (igc_header)  (obj_type = IGC_OBJ_INVALID, hash = 128931976, nwords = 1)
>   (lldb)

And that is because for the 2nd buffer that is dumped in
dump_cold_buffer

  (lldb) p is_mps (b->own_text.beg)
  (bool) true

which makes igc_dump_finish_obj assume it is an MPS object which I don't
think it is, or is it?