For 32-bit PPC targets, we translated:

evmergelo rX, rX, rY

as:

rX-lo = rY-lo
rX-hi = rX-lo

which is wrong, because we should be transferring rX-lo first.  This
problem is fixed by swapping the order in which we write the parts of
rX.

Similarly, we translated:

evmergelohi rX, rX, rY

as:

rX-lo = rY-hi
rX-hi = rX-lo

In this case, we can't swap the assignment statements, because that
would just cause problems for:

evmergelohi rX, rY, rX

Instead, we detect the first case and save rX-lo in a temporary
variable:

tmp = rX-lo
rX-lo = rY-hi
rX-hi = tmp

These problems don't occur on PPC64 targets because we don't split the
SPE registers into hi/lo parts for such targets.

Signed-off-by: Nathan Froyd <froydnj@codesourcery.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>