Re: Performance of records with variant parts

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

On 3/24/21 12:00 AM, John Perry wrote:
> 
> |              if denom <= 0.0 then
> |                 [constraint_error when
> |                   objects__thing_typeD3 (obj.kind)
> |                   "discriminant check failed"]
> |                 [constraint_error when
> |                   objects__thing_typeD3 (obj.kind)
> |                   "discriminant check failed"]
> |                 return (vectors__vector_dot (obj.norm, ray.start) +
> |                   obj.offset) / (-denom);
> 
> Am I right in thinking this is a bug? It shouldn't even check it, let alone check it twice.

It's not technically an error. Both Obj.Norm and Obj.Offset depend on a 
discriminant, so having two checks is technically OK. But eliminating the checks 
completely seems like a basic optimization that one would expect at -O3.

It's possible that this kind of optimization takes place later, and I guess the 
only way to check that would be to look at the assembler.

You could perhaps create a reproducer that demonstrates the timing difference 
and submit it to AdaCore.

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