okay millert@

okay millert@.
- atoi was bogus.
- atoi was not reentrant.
- there's no reason to cast stuff to unsigned int.

to /usr/libexec/mail.local;
from millert@

ok xsa@ henning@

ok cedric@ mcbride@

o Make MD5Update() more readable
o Simplify padding in MD5Final()
o Do copy + byte flip in MD5Transform() instead of in the caller
The result is a faster md5 and the byte order foo is kept to a minimum.

functions can take a NULL buf pointer.  They will malloc memory as
needed in this case.

aligned so copy it in a way that a) is endian indepenent and b) does
not rely on alignment.  Problem found and solution tested by hshoexer@

Add constants for buffer lengths and use them.
Make bit count 64bit based on changes by niklas@ long ago.
Conver htole32n() to macros and unroll loops.

With help from avsm@

changes Niklas made to the md5 code long ago.  OK hshoexer@

re-reported by Jan Filipovsky.

previously lived in foohl.c.  The foohl.c files are now generated
via sed, though perhaps cpp could be used in the future.
Use u_int8_t instead of unsigned char for the buffers struct fooContext.
Add constants for buffer lengths and use them in function prototypes
and the man pages.
This is basically cosmetic surgery; there should be no functional changes.
OK deraadt@

mcbride and i think this diff looks good but do not feel like doing the
250 meter walk to go test it

from Dmitry Medvedev, ok deraadt@

ok deraadt@, millert@, krw@ - mdoc ok jmc@

This allows an application to only update a password entry if it is
in the state it expects.  Additionally, if the old passwd struct
is specified the new one may have a different pw_name field since
matching is done on the original.  Adapted from FreeBSD.

This allows for safe abort handling, without tripping into
false recursivity problems.
Ok tedu@, deraadt@