https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106763
--- Comment #13 from Richard Earnshaw <rearnsha at gcc dot gnu.org> --- I don't think it would hurt. With this change, a float-16 instruction that was encountered on older cores would enable the VFP unit if it wasn't previously enabled and then fault again when the retried instruction failed. That's what likely already happens today if you have an instruction that isn't legal but still falls in the top-level decoding space for 32-bit and 64-bit FP instructions. Perviously we would unconditionally raise a SIGILL, which is what you're seeing right now. However, I'm not a kernel expert and I haven't tested any of the above, so all the caveats that that implies should be assumed.
