[bananas]gnu gcc 3.3

[Eric Greisen]

This is to warn AIPS sites about the gnu compiler version 3.3.

Previously, we announced that 3.3 appears to work, but that there were
problems with 3.2.2 included with RedHat 9.  The latter were simple
failures in optimization that made very obviously bad outputs and were
corrected by changes to OPTIMIZE.LIS.

In testing a complicated 80-facet, 74-MHz imaging problem, we have
found that versions 2.95.3 (compiled on RedHat 7.2, run on both RedHat
7.2 and 9) and 3.2.2 (compiled and run on RedHat 9) produce rather
similar answers (difference 1.3 mJy out of 12 Jy).  But, version 3.3
(compiled on 7.2 and run on 7.2 and 9 and compiled on 9 and run on 9)
produce answers that differ by more than 40 mJy.  Furthermore, when
compiled with no optimization the answer under 3.3 disagrees with that
produced with normal optimization by a similar amount.  Those two have
features that agree with each other and disagree with the older
compilers and other features which disagree with each other as well as
with the older compilers.

This is much more disturbing than the 3.2.2 problems in that even
unoptimized code has clear errors and at the same time the errors are
not gross failures to calibrate data or to place objects in roughly
the right place as with 3.2.2.  I think it is clear that the 3.3
answers are wrong (isolated positions with nearby + and - point
sources - which were probably supposed to cancel etc)

Oh dear....

Eric Greisen