Dear all,
this is not really a question but a sort of "summary before the
question".
Some of you might have been using -fast quite happily since it does
generate some pretty fast FP code on Alphas, in particularly we've
noticed major performance improvements on EV56 based machines.
Well, in the man pages for cc there is also a warning about the
"assumptions" the compiler makes. Having just tacked down a pretty
amazing runtime bug to -fast I think I would like to recommend
reading the paragraph carefully.
Our EV5 machine (8200, on sale if anyone is interested... we're
getting DS20's ;-)) was running -fast code quite happily for an year
or so, we just compiled -fast on EV56 PWSs and we are now littered
with FPE problems all over the place. Fundamentally the code now
produces probabilities > 1 and other assorted mathematical rubbish. DU
version 4.0D on both boxes, DXML same release level, etc. So clearly
the EV56 optimiser does something more daring which breaks our
code.
Now, before you all say "well, you shouldn't use -fast", we have taken
-fast out of the EV56 Makefiles, in fact we've taken it out of the EV5
too for good measure, and we are buying EV6 to get the speed back...
I just felt I should warn people - this problem only appeared at
runtime (clearly) and after quite a few hours of runtime (= wasted
machine time).
Hope this helps someone else,
Arrigo
--
Arrigo Triulzi <arrigo_at_albourne.com> - Peripatetic Wizard
Albourne Partners Ltd. - London, UK
APL Financial Services (Overseas) Ltd. - Nicosia, Cyprus
Received on Thu May 13 1999 - 19:14:11 NZST