Re: [PATCH] var-tracking.c speedups

[Jakub Jelinek <jakub at redhat dot com>]

On Tue, Jun 23, 2009 at 06:50:12PM +0200, Paolo Bonzini wrote:
> static inline int /* should be bool but that is a separate change */
> rtx_equal_p (const_rtx x, const_rtx y)
> {
>   if (x == y)
>     return 1;
>   if (x == NULL_RTX || y == NULL_RTX)
>     return 0;
>   if (GET_CODE (x) != GET_CODE (y))
>     return 0;
>   return rtx_equal_p_cb_1 (x, y, NULL);
> }
>
> static inline int /* should be bool but that is a separate change */
> rtx_equal_p_cb (const_rtx x, const_rtx y, yaddayadda)
> {
>   if (x == y)
>     return 1;
>   if (x == NULL_RTX || y == NULL_RTX)
>     return 0;
>   if (GET_CODE (x) != GET_CODE (y))
>     return 0;
>   return rtx_equal_p_cb_1 (x, y, yaddayadda);
> }
>
> so that you can remove the common checks from the callee, too.

While that addresses primarily the tail call and not checking/passing the
callback through recursively, I think moving some of the checks to the
callers is a good idea; I guess in many cases gcc will be able to optimize away
the NULL pointer checks in the callers at least.

BTW, var-tracking.c calls rtx_equal_p in a weird way:
                  if ((REG_P (vui[jj].lc->loc)
                       && REG_P (node->loc)
                       && REGNO (vui[jj].lc->loc) == REGNO (node->loc))
                      || rtx_equal_p (vui[jj].lc->loc, node->loc))
If that is because it doesn't care about REG mode, when it already knows
that it is a REG it shouldn't call rtx_equal_p at all.
So it could use something like:
/* Like rtx_equal_p, but assumes x and y are non-NULL and for REGs ignores
   mode.  */
static inline bool
vt_rtx_equal (const_rtx x, const_rtx y)
{
  if (x == y)
    return true;
  if (GET_CODE (x) != GET_CODE (y))
    return false;
  if (REG_P (x))
    return REGNO (x) == REGNO (y);
  return rtx_equal_p_cb_1 (x, y, NULL);
}

	Jakub