Re: Performance of records with variant parts

["Jeffrey R. Carter" <spam.jrcarter.not@spam.not.acm.org>]

On 3/22/21 11:11 PM, John Perry wrote:
> 
> Current optimization is -O3 -funroll-loops -gnatn, but the behavior remains consistent.
> 
> |  function Object_Intersect( obj: in Thing_Type; ray: in Ray_Type )
> |                            return Long_Float
> |  is
> |  begin
> |     case obj.kind is
> |     when Sphere =>
> |        declare
> |           eo: Vector := obj.center - ray.start;
> |           v: Long_Float := eo * ray.dir;
> |        begin
> |           if v > 0.0 then
> |              declare
> |                 disc: Long_Float := obj.radius2 - ( eo * eo - v * v );
> |              begin
> |                 if disc > 0.0 then
> |                    return v - Sqrt(disc);
> |                 end if;
> |              end;
> |           end if;
> |        end;
> |     when Plane =>
> |        declare
> |           denom: Long_Float := obj.norm * ray.dir;
> |        begin
> |           if denom <= 0.0 then
> |              return ( obj.norm * ray.start + obj.offset ) / ( -denom );
> |           end if;
> |        end;
> |        when None =>
> |           return far_away;
> |     end case;
> |     return Far_Away;
> |  end Object_Intersect;

Interesting. With that optimization I would not expect this to differ noticeably 
from a record without discriminants.

I note that all your nested declarations could be declared constant.

Use of Long_Float is completely non-portable.

If you want to look into this further, IIRC, there is an option that will output 
a low-level, Ada-like description of the IR that the front end produces. You'll 
have to look in the top-secret GNAT documentation to find it.

-- 
Jeff Carter
"[I]f you ask, 'Why does Ada do/have this?',
the answer often makes you a better programmer."
Chad R. Meiners
177