This prevent callers from having to save SSE registers to the stack.

It is for instance the case in MSVCRT__towlower_l, which is called on
every character by MSVCRT__wcsicmp_l.
Signed-off-by: Rémi Bernon <rbernon@codeweavers.com>
Signed-off-by: Piotr Caban <piotr@codeweavers.com>
Signed-off-by: Alexandre Julliard <julliard@winehq.org>