Re: PATCH: IA64 floating-point types

[Jim Wilson <wilson at specifixinc dot com>]

Joseph S. Myers wrote:
> Where are these types specified from a language extension point of view?  

I've never had access to ia64-hpux, but I believe the only reason we 
define them is because some HPUX system header files assume they are 
available.

I don't believe they are documented anywhere from a language extension 
point of view, but I think they are really more for convenience than for FP.

The only purpose of _fpreg seems to be for generating special purpose FP 
load/store instructions (ldf.spill/stf.spill), and that support hasn't 
been implemented yet.  I can't be sure though, since I have no 
documentation for _fpreg.  It seems to be an HP specifix extension.

On Linux, where long double is __float80, I think the only possible 
operations that can work with __float128 are load/store. This is because 
linux has no support for 128 bit FP types.  However, this assumes it is 
implemented correctly, which it isn't, because we can't have both 80 and 
128 bit FP types at the same time, so the question is moot.

On HPUX, where long double is __float128, in theory one could support FP 
operations on __float80 because it is supported by the hardware, but the 
question is moot because gcc can't have both 80 and 128 bit FP types at 
the same time.  Whether conversions are available between __float80 and 
__float128 depends on whether HP defined library routines for them.  I 
don't have that info.

I see that Mark made __float128 a synonym for long double for HPUX, but 
he did not make __float80 a synonym for long double for Linux.  That 
seems wrong, but I don't care much since I wouldn't recommend use of 
these types anyways.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com