Hey Martin,

Thanks for running these and pointing this out. I've replicated the
results with tcrypt and fixed some issues, and the next patch series
should be a lot closer to what you'd expect, instead of the regression
you noticed. Most of the slowdown happened as a result of over-eager
XSAVEs, which I've now rectified. I'm still working on a few other
facets of it, but I believe v5 will be more satisfactory when posted.

Regards,
Jason