Why is std.math slower than the C baseline?

[jmh530]

On Friday, 5 June 2020 at 19:39:26 UTC, Andrei Alexandrescu wrote:
> [snip]
>
> This needs to change. It's one thing to offer more precision to 
> the user who consciously uses 80-bit reals, and it's an 
> entirely different thing to force that bias on the user who's 
> okay with double precision.

I agree with you that more precision should be opt-in.

However, I have always been sympathetic to Walter's argument in 
favor doing intermediates at the highest precision. There are 
many good reasons why critical calculations need to be done at 
the highest precision possible.

In addition, the current PR to remove this behavior will make 
opt-ing in more difficult. Users will need to either put their 
own casts everwhere, or basically re-implement what is currently 
in std.math. It may be a good idea to release std.math in its 
current state as a separate module, similar to stdx.allocator, 
(the real versions can just call the real versions in phobos to 
reduce the maintenance cost). Shame to just get rid of it if 
there are some people who might use it as is.