https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103623
--- Comment #5 from Bill Schmidt <wschmidt at gcc dot gnu.org> --- Agreed that this needs a new attribute, and digging through the macros used to guard the associated patterns, this really does need to be restricted to the case when long double is implemented by IBM-128. Safest is to use FLOAT128_2REG_P (TFmode), i.e., applied to the mode used for long double, which is what enables the corresponding pack<mode> and unpack<mode> patterns.