Jeff and Katja:

Our code runs on the Origin ute with the following
fpe environment:

  setenv TRAP_FPE
  setenv TRAP_FPE "${TRAP_FPE}UNDERFL=FLUSH_ZERO"
  setenv TRAP_FPE "${TRAP_FPE};OVERFL=MAX"
  setenv TRAP_FPE "${TRAP_FPE};DIVZERO=ABORT,TRACE"
  setenv TRAP_FPE "${TRAP_FPE};INVALID=ABORT,TRACE"

What is the best way to simulate this situation on
the IBM blackforest? I have set the following for
serial debug runs of the code on the IBM:

FC = xlf90
FFLAGS= -qfixed -qarch=pwr3 -qtune=pwr3 -qrealsize=8 \
       -qalias=noaryovrlp -O -WF,-DAIX \
       -qflttrap=zero:invalid:enable -qsigtrap=xl__trcedump \
       -I/usr/local/include
LIBS = -L /usr/local/lib32/r4i4 -lmass -lnetcdf -lmss

The only env vars I set on IBM are related to mpi, etc,
and these are serial runs.  With this compiler setup, 
it looks like underflows flush to zero by default, and 
overflows go to INF or -INF. Is there a way to tell 
blackforest to flush overflows to some MAX as in the Origin?

When *not* trapping invalids, the code runs 3-6 hours model
time (3-minute time step) on the IBM, but then starts producing 
NaNs (NaNQ's according to netcdf) in certain areas of
the geographic grid. I can save histories such that I
can get it to produce NaNs in a single time step, so
repeated short debug runs are possible.

When I do enable invalid trapping, as above, it stops in an 
expression using the intrinsic exp(). Many of the arguments 
to exp are large and negative -- by the time it signals
an fpe (presumably an invalid), these were successfully 
being flushed to zero (e.g., exp(-1.e64)=0.).

A small test code shows that blackforest will flush
exp(-big) to zero, and exp(big) to INF, and that it will
not trap in exp() itself because its a lib routine. I 
could see the problem if it were getting INF from exp(big),
and that referencing the INF's later would produce the
NaN's. But the exp's I've looked at so far (3 routines in 
our code, flagged by trapping invalids) use exp(-big), rather 
than exp(big). I think the problem arises with the argument
is "close" to the abs(maximum) argument before overflow or
underflow, i.e., about 709.78 or -745.13.