Implement C23 pown [committed]

[Joseph Myers]

On Wed, 4 Jun 2025, Vincent Lefevre wrote:

> On 2025-06-03 11:09:07 -0500, Segher Boessenkool wrote:
> > The commit message clearly says this implementation isd correct for
> > IEEE float only.  And indeed it is not for double-double, which is very
> > much not IEEE.
> > 
> > Maybe these tests should require IEEE float?  The implementation does,
> > after all!
> 
> Shouldn't this implementation rather be disabled with non-IEEE floats
> (thus not just the tests), as it is not correct?

The basic principle is that we provide the same API across architectures 
and floating-point types (modulo some obsolescent functions not being 
provided for _FloatN / _FloatNx).  So pownl is supported for all glibc 
configurations.  Having a generic implementation (that works for IBM long 
double as well as the underlying arithmetic in libgcc does, which isn't 
that well outside round-to-nearest mode) avoids needing to write five 
separate implementations for five different floating-point formats in 
order to add the feature to glibc.  (And while people might well add 
better - faster and more accurate - implementations for particular formats 
in future, for example taken from CORE-MATH if that gains pown 
implementations at some point, I very much doubt anyone is going to care 
about IBM long double enough to write a custom implementation for that 
format of any new math.h functions.)

-- 
Joseph S. Myers
josmyers@redhat.com