patch: alpha haifa abort

[Paul Koning]

>>>>> "John" == John Carr <jfc@tiac.net> writes:

 >> or do you want both __load_big_endian() and __load_little_endian()
 >> with GCC creating a no-op load for the case matching the target's
 >> byte-endianness?

 John> I want explicitly big- and little-endian loads and stores.  I
 John> think it is rare to want to swap bytes regardless of the
 John> current byte order.  Usually one is doing networking
 John> (big-endian) or PCI (little-endian) and wants to write in a
 John> specific byte order.

A different approach that might be nice to have: rather than have big
and little endian load/store intrinsics, how about explicitly little
and big endian integer data types?  Or, I suppose, it's best to have
both, since the intrinsic is a way to make "ntoh" faster, and the
datatype is a way to avoid that syntactic clutter in the first place.

I did this sort of thing recently for a device driver, by defining
"little endian int" as a class with appropriate conversion operators.
Works almost as nicely as having it builtin, the major drawback is
that the debug display is ugly.

	paul