The vext.8 instructions only need to produce a single d register each,
making more registers available as scratch space, allowing to hide
latencies more, and group the vmul/vmla in the form that is beneficial
for in-order cores (with a special forwarding path for such patterns).