On Wed, Dec 17, 2025 at 11:57:30AM +0100, Tomasz Kaminski wrote:
> On Wed, Dec 17, 2025 at 11:50 AM Jakub Jelinek <[email protected]> wrote:
>
> > On Wed, Dec 17, 2025 at 11:44:31AM +0100, Tomasz Kaminski wrote:
> > > Still we will need to check if we are causing any regressions here,
> > > but the old implementation static_assert is_floating_point_v so we should
> > > not cause any, unless that allows __float128.
> >
> Oh, I wasn't clear enough. This patch caused regressions for use
> of generate cannonical long double, double (mantissa bigger than 64) for
> architectures that do not have __int128,
> which we need to address. No discussion about that.
>
> I am looking into it, but I think I will do a two phase approach, of using
> 32bit, 64bit integers when the power of the generate range greater than
> mantissa
> (that does not have 64bits) will fit in them. And separate one, one we need
> to do
> 128 bit operations.
>
> I was committing on regression on __float128, on system where it is not used
> as double, long double.
s390-linux is a 32-bit platform with IEEE 754 quad long double (at least in
the usual configuration).
powerpc-linux is a 32-bit platform with IBM double double as long double
(i.e. 53 to 2000-ish precision, often assumed to be 106 bit).
Jakub
