fma without exception or rounding mode support

[Joseph S. Myers]

On Tue, 15 May 2012, Bruno Haible wrote:

>     If it was an action item for me, then do you want me to provide
>     this replacement for math/s_fma{,f,l}.c now, for glibc 2.16 [1]? Or
>     is it less urgent?

It's not urgent, it's just always nice to get bugs fixed; if it's ready on 
time I see no reason it shouldn't go in 2.16.

>   - The proposed portable fma implementation indeed does not require
>     exception nor rounding mode support. But we did not talk about this
>     in the scope of bug 13304. In fact, it's the first time I hear that
>     glibc would want to support such systems without providing a complete
>     <fenv.h>.

It's such systems, plus providing something better than the present 
fallback for ldbl-128ibm, that it's useful for, given Jakub's 
implementations following Boldo and Melquiond that provide the functions 
when FE_TOWARDZERO and FE_INEXACT are supported.  While exceptions and 
rounding modes can be supported with software floating point (only done in 
glibc for powerpc-nofpu in ports at present) it's not clear this would be 
particularly useful to most users of systems with software floating point, 
and an integer implementation of fma would still likely be faster on them 
than one designed in terms of intermediate floating-point operations being 
fast.

-- 
Joseph S. Myers
joseph@codesourcery.com