Re: coreutils-5.0.90 problems on Crays

[Paul Eggert]

Wendy Palm <address@hidden> writes:

> with the change i proposed, the only error i have with the tests is
> date-tests and i haven't had time to check what that problem is yet.

I doubt whether the tests exercise the md5 and sha code a lot, and I
suspect that there will be latent bugs, unless someone carefully inspects
the code and makes sure that there are no problems.

> i'd be interested to know what programs are potentially problems with >32bits.
> possibly some shifting assumptions may be messed up.

For example, here's the first use of md5_uint32 in md5.c:

  ((md5_uint32 *) resbuf)[0] = SWAP (ctx->A);

SWAP is defined as follows:

#ifdef WORDS_BIGENDIAN
# define SWAP(n)                                                        \
    (((n) << 24) | (((n) & 0xff00) << 8) | (((n) >> 8) & 0xff00) | ((n) >> 24))
#else
# define SWAP(n) (n)
#endif

If Cray is a WORDS_BIGENDIAN host, this will put garbage bits into the
high-order part of ((md5_uint32 *) resbuf)[0].

This is just the first occurrence that I found.  Basically, someone
needs to check every instance of md5_uint32 and make sure that it
works OK.

>          cc  -hnointrinsics  -Wl,-Dmsglevel=334:fatal -o [  lbracket.o 
> ../lib/libfetish.a  ../lib/libfetish.a -lgen
> GETTMC: invalid primary machine name on TARGET statement

Does it work to change the "-o [" to "-o lbracket", and then to
issue "mv [ lbracket" afterwards?