For both the [4, 7] and [8,15] length cases, we can perform the same
technique: perform two overlapped loads, zero-extend, then perform two
overlapped stores. The 8-character case could be done in a single
load/store pair, but is not worth the extra conditionals. And it should
have the exact same performance numbers whether we use non-overlapping
4-character operations or completely-overlapping 8-character ones (I
*think* the full overlap is actually better).
The 4-character operation is new in this commit. That reduces the
non-vectorized, unrolled to at most 3 characters.
Change-Id: Ib42b3adc93bf4d43bd55fffd16c257ada774236a
Reviewed-by: Lars Knoll <lars@knoll.priv.no>
3ef43ca837