[Bug middle-end/67438] [6 Regression] ~X op ~Y pattern relocation causes loop performance degradation on 32bit x86

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67438

--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 17 Nov 2015, ysrumyan at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67438
> 
> --- Comment #9 from Yuri Rumyantsev <ysrumyan at gmail dot com> ---
> It looks like such transformation is profitable if only def statements have a
> single use, i.e. it looks reasonable for 
>    if (255 - a) > (255 -b) /* a,b have char type.  */
> but it does not look reasonable for attached test-case since after it we missed
> min/max recognition, namely,
> 
>         c = 255 - r; /* c has mulitple uses!  */
>         m = 255 - g; /* likewise.  */
>         y = 255 - b; /* likewise.  */
>         if (c < m) 
>           k = MIN (c, y);
>         else
>           k = MIN (m, y);
>         *write++ = c - k;

Looks like we are missing the corresponding pattern for MIN/MAX
instead.

 MIN (~X, ~Y) -> ~MAX (X, Y)
 MAX (~X, ~Y) -> ~MIN (X, Y)

(for minmax (min max)
     maxmin (max min)
 (simplify
  (minmax (bit_not @0) (bit_not @1))
  (bit_not (maxmin @0 @1))))

maybe that helps.  I notice a missed optimization to combine
the test with the two MINs on the GIMPLE level anyway.