Re: [PATCH] tui: replace deprecated_register_changed_hook with observer

[Pedro Alves <palves@redhat.com>]

On 07/08/2015 01:30 PM, Patrick Palka wrote:
> On Wed, Jul 8, 2015 at 7:41 AM, Pedro Alves <palves@redhat.com> wrote:
>> On 07/06/2015 02:17 AM, Patrick Palka wrote:
>>> This is a straightforward replacement of the TUI's use of the
>>> aforementioned hook with the register_changed observer.  Since this was
>>> the only user of the hook, this patch also removes the hook.
>>>
>>> [ I am not sure if the changes to the function tui_register_changed are
>>>   correct.  In particular, the inputted frame argument is now passed down
>>>   to tui_check_data_values instead of the frame returned by
>>>   get_selected_frame.  The frame argument passed to each register_changed
>>>   observer corresponds to the VALUE_FRAME_ID of the register being
>>>   modified within a register assignment, e.g. the $rax in "print $rax =
>>>   FOO".  When would the frame corresponding to the VALUE_FRAME_ID of a
>>>   register not be the currently selected frame?  ]
>>>
>>
>> Grepping for value_assign callers finds e.g., varobjs:
>>
>>   varobj.c:      val = value_assign (var->value, value);
>>
>> Adding an assertion like this:
>>
>> @@ -1169,6 +1169,7 @@ value_assign (struct value *toval, struct value *fromval)
>>               }
>>           }
>>
>> +       gdb_assert (frame == get_selected_frame (NULL));
>>         observer_notify_register_changed (frame, value_reg);
>>         if (deprecated_register_changed_hook)
>>           deprecated_register_changed_hook (-1);
>>
>> and playing with varobjs shows the assertion failing:
>>
>>  (gdb) interpreter-exec mi "-var-create - * $rax"
>>  ^done,name="var1",numchild="0",value="6295640",type="int64_t",has_more="0"
>>  (gdb) up
>>  #1  0x000000000040082a in thread_function0 (arg=0x0) at threads.c:69
>>  69              usleep (1);  /* Loop increment.  */
>>  (gdb) up
>>  #2  0x0000003616a07ee5 in start_thread (arg=0x7ffff7fc1700) at pthread_create.c:309
>>  309           THREAD_SETMEM (pd, result, CALL_THREAD_FCT (pd));
>>  (gdb) interpreter-exec mi "-var-assign var1 1"
>>  ~"/home/pedro/gdb/mygit/build/../src/gdb/valops.c:1172: internal-error: value_assign: Assertion `frame == get_selected_frame (NULL)' failed.\nA problem internal to GDB has been detected,\nfurther debugging may prove unreliable.\nQuit this debugging session? (y or n) "
>>
>> The TUI doesn't use MI, but there are probably other similar cases
>> in the tree.  E.g., I'd assume you can create a register Value with Python,
>> and then assign to it when the selected frame is not
>> the register's frame.
> 
> Ah okay.. So it seems to me that if the frame argument !=
> get_selected_frame, then we should not update the register window at
> all since the register window is supposed to show the register values
> of the currently selected frame.

Yes, I think so.

> Or instead, just ignore the frame argument and always pass
> get_selected_frame to tui_check_data_values, even if frame !=
> get_selected_frame.  Seems to me that this is the safest option.

That'd be a 1-1 with the current code.  Though, I believe
that results in spuriously clearing the highlight of
previously changed registers (of the selected frame), because
nothing will have changed.  So seems like the other option
actually fixes a bug.

Thanks,
Pedro Alves