Fix the destination register actions for VNNI and VBMI2 concatenate and
variable shift instructions `VPSH{L,R}DV{W,D,Q}`. In the case of mask merging,
the action should be `RW` not `W`.
Prior to this change, the bug manifests as incorrect vector register
scheduling when `avo` doesn't recognize that these instructions have a data
dependency on the destination register.
See:
https://www.felixcloutier.com/x86/vpdpbusd
https://www.felixcloutier.com/x86/vpdpbusds
https://www.felixcloutier.com/x86/vpshldv
https://www.felixcloutier.com/x86/vpshrdv
2.6 MiB
2.6 MiB