On 21 April 2014 18:43, Paolo Bonzini <pbonz...@redhat.com> wrote: > This series adds setend support by using the MO_BE/MO_LE that Richard > implemented. This is used by Raspbian's faster memcmp implementation.
There are also "by hand" loads done in linux-user/main.c's implementation of ldrex/strex which need to honour CPSR.E now. We seem to be OK for the ldrd/strd since we implement these as a pair of 32 bit loads (unlike the pseudocode which does a 64 bit load and then parcels the result out to Rt/Rt2 so it gets atomic 64 bit access semantics). I think we get the various neon load/store insns right, but did you check the pseudocode against our implementation or are you just hoping that we never try to combine or split a load/store compared to the pseudocode spec? thanks -- PMM
