bug#70914: 29.3; Crashes often on Windows

[Simen Endsjø]

> Please try this with a healthy Emacs process before you do it after
> the crash, to make sure this procedure works.  Here's my attempt to
> validate this technique:

Looks like I needed a binary with debugging symbols, so I used my previous
build. It's as good as any other. But looks like there are two stack frames with
zero pointers, so the technique doesn't work?

    Thread 1 received signal SIGSEGV, Segmentation fault.
    0x0000000000000000 in ?? ()
    (gdb) thread 1
    [Switching to thread 1 (Thread 19884.0x7bf8)]
    #0  0x0000000000000000 in ?? ()
    (gdb) p/x *(uintptr_t *)$sp
    $1 = 0x0
    (gdb) list *$
    (gdb) bt
    #0  0x0000000000000000 in ?? ()
    #1  0x0000000000000000 in ?? ()
    (gdb)

On Mon, May 20, 2024 at 3:47 PM Eli Zaretskii <eliz@gnu.org> wrote:
>
> > From: Simen Endsjø <simendsjo@gmail.com>
> > Date: Sun, 19 May 2024 20:38:24 +0200
> > Cc: 70914@debbugs.gnu.org, corwin@bru.st
> >
> > *****************************| System Information 
> > |*****************************
>
> Thanks.  I think, given that every other avenue of approach has
> failed, we should try the direct one: try to determine which code
> called the zero PC address.  I think the following should work for
> you, after Emacs crashes due to zero address:
>
>   (gdb) thread 1
>   (gdb) p/x *(uintptr_t *)$sp
>   $1 = 0x1234567887654321
>   (gdb) list *$
>
> The "0x1234567887654321" stands for some 64-bit address that GDB will
> show in your case, which is the address pointed by the stack pointer
> register.  AFAIU, that address should hold the return address of the
> function which called the "zero address", and the "list" command
> should show its source code (assuming it's some Emacs code).
>
> Please try this with a healthy Emacs process before you do it after
> the crash, to make sure this procedure works.  Here's my attempt to
> validate this technique:
>
>   gdb ./emacs.exe
>   ...
>   (gdb) break Frecursive_edit
>   Breakpoint 2 at 0x115dc2f: file emacs.c, line 2621.
>   (gdb) run -Q
>   Thread 1 hit Breakpoint 2, main (argc=2, argv=0x7ab2570) at emacs.c:2621
>   2621      Frecursive_edit ();
>   (gdb) si
>   Frecursive_edit () at keyboard.c:808
>   808     {
>   (gdb) p/x *(uintptr_t *)$sp
>   $4 = 0x76dc34
>   (gdb) list *$
>   0x76dc34 is in main (emacs.c:2622).
>   2617    #endif
>   2618
>   2619      /* Enter editor command loop.  This never returns.  */
>   2620      set_initial_minibuffer_mode ();
>   2621      Frecursive_edit ();
>   2622      eassume (false);
>   2623    }
>   2624    ^L
>   2625    /* Sort the args so we can find the most important ones
>   2626       at the beginning of argv.  */
>   (gdb)
>
> The above is in a 32-bit build of Emacs, not 64-bit build as in your
> case.  The above tells us that Frecursive_edit was called from a line
> before 2622 (since the return address on the stack is the address of
> the first function _after_ Frecursive_edit).
>
> Note that I used the "si" command (stepi) to step 1 machine
> instruction inside Frecursive_edit and stop immediately after the
> call, before the function's preamble pushes local variables onto the
> stack, so as to ensure that the stack pointer points to the return
> address.
>
> I hope using this technique we will be able to find the immediate
> caller of the "zero address".  Fingers crossed.