SSE min/max take2 - documentation

[Geert Bosch]

On 25-Feb-2001, Jan Hubicka <jh@suse.cz> wrote:
|   @itemx (smax:@var{m} @var{x} @var{y})
|   Represents the smaller (for @code{smin}) or larger (for @code{smax}) of
|   @var{x} and @var{y}, interpreted as signed integers in mode @var{m}.
| + The code can be also used to for floating point values.  In that case
| + result on two zeros or operands where one is NaN is undefined.

On Sun, 25 Feb 2001, Fergus Henderson replied:
|The wording there is somewhat misleading;
|saying that the result for two zeros is undefined
|implies that returning any random number, e.g. 1.23, is OK.
|But I'm pretty sure that you only want it to be unspecified which of
|the two zeros is returned (in case their sign differs).

I'm still not clear on the intended use of the min/max codes
for floating point values.

It would seem that we'd like these to be usable to translate the
following to min/max if the target has such an instruction:

  float x = -0.0;
  float y = +0.0;
  float p, q, r, s;
  p =  (x >  y) ? x : y;  /* +0.0 */
  q =  (x >= y) ? x : y;  /* -0.0 */
  r =  (x <  y) ? x : y;  /* +0.0 */
  s =  (x <= y) ? x : y;  /* -0.0 */

In case either x or y is a NaN, the value of y should be returned.

Since the SSE instructions have been defined in a way that allow
them to be used for the above expressions, would think GCC should
use this definition. For targets that have min/max instructions that
do not preserve these properties, these sequence should only be
enabled if strict IEEE compliance is not required.

  -Geert