Hi Kewen & Segher,
Thanks so much for your review comments.
On 22/9/2022 上午 10:28, Kewen.Lin wrote:
> on 2022/9/22 05:56, Segher Boessenkool wrote:
>> Hi!
>>
>> On Fri, Jun 24, 2022 at 10:02:19AM +0800, HAO CHEN GUI wrote:
>>> This patch also binds __builtin_vsx_xs[min/max]dp to fmin/max instead
>>> of smin/max. So the builtins always generate xs[min/max]dp on all
>>> platforms.
>>
>> But how does this not blow up with -ffast-math?
>
> Indeed. Since it guards with "TARGET_VSX && !flag_finite_math_only",
> the bifs seem to cause ICE at -ffast-math.
>
> Haochen, could you double check it?
I tested it with "-ffast-math". fmin/max functions are converted to
MIN/MAX_EXPR in gimple lower pass. But the built-ins are not and hit the
ICE. I thought the built-ins are folded to MIN/MAX_EXPR like vec_ versions'
when fast-math is set. In fact they're not. Sorry for that.
I made a patch to fold these two built-ins to MIN/MAX_EXPR when fast-math
is set. Then the built-ins are converted to MIN/MAX_EXPR and expanded to
smin/max.
Thanks for pointing out the problem!
>
>>
>> In the other direction I am worried that the unspecs will degrade
>> performance (relative to smin/smax) when -ffast-math *is* active (and
>> this new builtin code and pattern doesn't blow up).
>
> For fmin/fmax it would be fine, since they are transformed to {MAX,MIN}
> EXPR in middle end, and yes, it can degrade for the bifs, although IMHO
> the previous expansion to smin/smax contradicts with the bif names (users
> expect to map them to xs{min,max}dp than others).
>
>>
>> I still think we should get RTL codes for this, to have access to proper
>> floating point min/max semantics always and everywhere. "fmin" and
>> "fmax" seem to be good names :-)
>
> It would be good, especially if we have observed some uses of these bifs
> and further opportunities around them. :)
>
Shall we submit a PR to add fmin/fmax to RTL codes?
> BR,
> Kewen
