On Wed, Aug 21, 2024 at 09:12:20AM -0700, Sean Christopherson wrote: > On Tue, Aug 20, 2024, Keith Busch wrote: > > From: Keith Busch <kbusch@xxxxxxxxxx> > > > > Because people would like to use this (see "Link"), interpret the VEX > > Please explicitly call define the use case in the changelog. Yeah, I can follow > the link, but I shouldn't have to just to understand that this is the compiler > generating vmovdqu for its built-in memcpy(). Sorry about that, but yes, it's essentially a compiler using an intrinsic aware memcpy. That all works for mmio addresses from pci passthrough functions, but currently fails on emulated device addresses. > > prefix and emulate mov instrutions accordingly. The only avx > > instructions emulated here are the aligned and unaligned mov. > > Everything else will fail as before. > > > > This is new territory for me, so any feedback is appreciated. > > Heh, this is probably new territory for everyone except possibly Paolo. I don't > recall the last time KVM was effectively forced to add emulation for something > this gnarly. Thanks, I feel a little less shame admitting this patch took me longer to figure out than predicted. :)