to say educational commit. Reordering instructions doesn't improve
performance much, rather exhibits Power6 limitations.

developers responsible for -march choice.

T[ed]4 prefetch.

check for SSE bit instead of MMX, as pshufw was introduces in PIII, minor
optimization, typos...

256 byte S-box. AES_cbc_encrypt needs further work as it should also use
slow routines when processing smaller amount of data.

strong impact on decrypt performance, 20-25%. One probably should consider
switching between slower and faster routines depending on how much data
we were asked to process.

only chance for T[ed]4 to get evicted in this module is when its cache
"overlaps" with last 128 bits of key schedule.

prefetch could have been moved closer to Td4 references. Something for
later consideration...