Some special notes about the 64-bit implementation:

general:

*) The TYPEPUN mode is not implemented because if words are 64 bits, then
   then the addresses of two successive words differ by 8, or 3 bits. But
   using a path size of 6 bits, to index into the appropriate array would
   reguire 9 bits of information (because there are 2^6 words, so 6 bits of
   address, plus the 3 bits of offset separating each word address).  So
   you would need a 9-bit byte.  As the other arrays are all larger (the
   pc1 array uses 7 bits (2^7 entries), pc2 uses 8 (2^8 entries), ipe and
   eipinv use 8 bits (2^8 entries), and spe12 uses 12 bits (2^12 entries),
   it didn't seem worthwhile.  The hooks are there id you want to add it;
   you'll need to add something to undo the final salting (see the des32
   routine if you want to do this).

machine specific:

Cray YMP:
   This does not have a TWOBITS type (unsigned char is 8 bits, unsigned
   short is 32) so PACKBITS == BYTES will not work.

Convex C-1:
   This is not "true" 64 bit mode, but is really a pipelining of two 32
   bit words.  For that reason, bitfields cannot cross the 32-bit boundary;
   so you cannot divide a 64 bit "word" into fields of 4 bits, 20 bits,
   20 bits, and 20 bits, as the second 20-bit field covers bits 32 and 33.
   So PACKBYTES == BITFIELDS will not work.