[Bug libfortran/93871] COTAN is slow for complex types

[thenlich at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93871

--- Comment #38 from Thomas Henlich <thenlich at gcc dot gnu.org> ---
(In reply to Thomas Henlich from comment #37)
> It would make sense to keep optimization in mind:
> 
> Several calls to conversions of the same value should be performed only once.
> 
> As a special case: Calls to compute sind(x) and cosd(x) should be optimized
> to a conversion, followed by a call to sincos.
> 
> The conversion could also be provided as a user function, for example in
> code like y=tand(x)-deg2rad(x) the conversion would need to be performed
> only once.

I just realized this is probably a lot harder than it sounds, since it is
incompatible with some of the range reduction steps, or at the least requires
some further folding into x <= 45 for sincos to get sufficient accuracy.

On the other hand side, always folding sind(45...90) to cosd(45...0) and
cosd(45...90) to sind(45...0) probably wouldn't be such a bad thing.