Re: V3 PATCH: IRIX configuration

[Jakub Jelinek <jakub at redhat dot com>]

On Mon, Oct 30, 2000 at 11:52:01AM +0100, Gabriel Dos Reis wrote:
> Mark Mitchell <mark@codesourcery.com> writes:
> 
> | >>>>> "Gabriel" == Gabriel Dos Reis <gdr@codesourcery.com> writes:
> | 
> |     Gabriel> Using a higher precision to compute a result, followed by
> |     Gabriel> a truncation to the input precision is something
> |     Gabriel> generally harmeless (it is just waste of ressource in
> |     Gabriel> case computing with higher precision is not necessary nor
> |     Gabriel> useful).  That technique is usually used in routines
> |     Gabriel> which are sentitive to noise.
> | 
> | But what about `__complex__ long double'?  That one gets truncated
> | first, right?
> 
> Hmm, I'm unsure about what you mean here.
> 
> For __complex__ float and __complex__ double, the promotion to
> __complex__ long double followed by the truncation to float or double
> after calculations is safe.

But from reading complex.cc it looks like __complex__ float is promoted to
__complex__ double and __complex__ long double is truncated to __complex__
double, then the complex function is called and then it is truncated resp.
promoted back to the original file.
E.g.

  template<>
  FLT
  abs(const complex<FLT>& __x)
  { return cabs(__x._M_value); }

is using cabs (ie. double precision), not cabsl (long double precision).
But using cabsl and the like in complex.cc is not a good idea IMHO either,
considering architectures with software emulated IEEE quad long double
(sparc64 and soon alpha, ppc and others).

	Jakub