Re: Simple optimization for read_avail_input()

[Kim F. Storm]

"Dmitry Antipov" <address@hidden> writes:

> Kim F. Storm wrote:
> 
> > Nice idea, but in its current form it fails because the buf array is 
> > allocated on
> > the stack. It may work to just declare it static though.
> 
> Bups. I agree. But, in any case, we should avoid to zero 40k each time when 
> read_avail_input() is called.

I agree, this is wasteful.

> It's not quite clear for me why this array should be so large (4k), btw.
> 
> > I think it's ok, if we can guarantee that read_avail_input is never called
> > recursively (I haven't checked).
> 
> It happens at least 1 time sometimes after creating X frame. Is it the same 
> as expected ?

I don't know.

> 
> > But I really wonder why it is necessary to initialize the array at all.
> 
>   IMHO it's necessary. For example, if read_socket_hook is NULL, buf[i].x and 
> buf[i].y
> are untouched. But they are Lisp_Objects, and if Fgarbage_collect() happens 
> immediately
> after kbd_buffer_store_event() (from read_avail_input()), we will got invalid 
> Lisp_Objects
> for mark_object() from mark_kboards (). I can reproduce this situation 
> sometimes  (when
> my fingers are blazingly fast :-)).
>   The same for X events - not all of them touches x and y.

I wasn't advocating for never initializing input_events -- I would
just postpone the initialization until we actually use one of those
input_events, i.e. also before we do kbd_buffer_store_event...

> 
> > BTW, it definitely is unnecessary to init the whole array if
> > read_socket_hook is NULL; we could just init the first nread
> > elements in that case.
> 
> Agree. For XTread_socket and others, we probably should do EVENT_INIT when 
> really needed.

It's the same principle...

-- 
Kim F. Storm <address@hidden> http://www.cua.dk