Fix valgrind: Operate only in little endian

[rsaxvc]

Previously, memcmp() switched the processor into
big-endian mode so that vectorized compares could
be done more efficiently. However, this makes
emulation quite difficult, and breaks valgrind.

Usually I would fix valgrind, but this was the
only place I could find setend being used.

This patch removes 2 SETEND instructions,
and adds 8 REV instructions if there is a
difference detected. Both SETEND and REV
should take 1 cycle on ARM11.

[bavison]

Thanks for this patch, though I had to think a bit to convince myself that it was valid to re-do all four CMPs again at the end in all cases (and I've updated the test harness to be doubly sure).

As you say, it should save 2 cycles on ARM11 in the no-difference-found case, and gain 10 when there is a difference, but that should be neither here nor there in the grand scheme of things. I've benchmarked it on ARM11, Cortex-A7 and Cortex-A53 and there's no statistically significant timing difference that I can detect (so no unexpected interactions with caches etc).

I'm aware that the SETEND trick I used also caused grief for people wanting to run the binaries under qemu, so they should appreciate this patch too.