Slow memcmp for aligned strings on Pentium 3

[Bonzini]

> I was doing some tests with with memcmp and it seams that gcc 3.2.2
> always uses cmps, even for aligned strings.

I also had a very fast memcmp that also worked word-by-word in the unaligned case
(using some bit-twiddling tricks with SHRD), but I never had the time to polish it for
inclusion in glibc.  It turned out however that CMPS is as fast as mine (or even
faster) on AMD chips.

I do think anyway that memcmp expansion in gcc should be disabled unless optimizing for
size or unless the size is known to be small, say 16 bytes.

See these posts:

http://gcc.gnu.org/ml/gcc/2002-10/msg01616.html (me)
http://gcc.gnu.org/ml/gcc/2002-10/msg01637.html (Roger)
http://sources.redhat.com/ml/libc-alpha/2002-10/msg00511.html (source)

Paolo