> On May 15, 2016, at 4:20 PM, David Holland <dholland-t...@netbsd.org> wrote: > > Uh what? On essentially all architectures the kernel can tell which > programs use the FPU, and presumably if they use it the values in it > can't be allowed to randomly disappear. > > (In fact, the mechanism for this is why the FPU can't readily be used > in the kernel.)
Reliably? I'd rather not guess. A deterministic API would remove the guess-work. Especially considering how hardware-specific this is. Backward compatibility is a bit of a pain, but I still think a simple, useable, API and implementation could be built.