[daip] Re: IBM compiler
Eric Greisen
egreisen at nrao.edu
Tue May 24 19:38:21 EDT 2005
Lincoln writes:
> Shoshana told me (I just remembered) that
> the Absoft compiler supports 64 bit and IBM
> only 32 bit. That sounded great - and scary.
> If we buy absoft, we would want to make comparisons
> of Absoft and gnu computing results. Would you
> recommend a particular set of exercises?
>
AIPS has no use for 64-bit so some care needs to be taken. I suspect
that one gets the 32-bit data okay, but 64-bit addresses.
I then worry that the 64-bit might do what one crazy system in Italy
did - give dynamic memory at one end of the address range and regular
program memory at the other. In that case, the way we currently do
dynamic memory breaks - we assume that a 32-bit integer subscript may
be used to represent the offset in words. I do have code I was
looking at that would fix this except that I took advantage of the way
the pseudo-AP is constructed (one array with many offsets in it) to
allow for a dynamic memory pseudo-AP in FRING. To make all pseudo-AP
subscripts be LONGINT (and all subscripts in the Q routines) would be
a very big job. I would have to drop the dynamic memory in FRING at
least for 64-bit address space machines, even those that do not do the
crazy malloc's.
Thus it is a bit of a gamble. Use Y2K large or DDT large for the
initial tests.
Eric Greisen
More information about the Daip
mailing list