Problem 1
=========
Multipath doesn't work for optimized BLACS on sun PVM and CM5.

Solution
========
Compiler error for high levels of optimization.  Rewrote mpath_bs for
CM-5 and PVM to avoid this problem.


Problem 2
=========
CM-5 fails some tests, usually multipath.

Solution
========
Was running out of asynchronous messages.  I rewrote Asend routines to detect
this, and loop until the system catches up.

Problem 3
=========
Unaligned data access on alpha.

Solution
========
I fixed the tester routine SAFEINDEX, which I had screwed up.

Problem 4
=========
Failing some SUM tests on HP, CM-5 and INTEL.

Solution
========
Error bounds in tester too tight.  Changed from 2*nnodes*eps to 
2 * (nnodes - 1) * eps * max(0.5, abs(ans)).


Problem 5a
==========
Sun4 prints out :
[t40003] BLACS WARNING 'Trying to change ID range after context definition'
[t40003] from {-1,-1}, pnum=1, Contxt=-1, on line 27 of file 'blacs_set_.c'.

Problem 5b
==========
Mixed sun and other architectures don't work.

Solution
========
I use a save and data statement (in tester routine ibtmsgid) to avoid 
multiple calls to blacs_set.  It appears that for -O4 optimization, on the
sun the save and data no longer work quite right, at least when everything
is in one big file.  If I compile all files seperately, or with no 
optimization, everything works OK.  For now, I just force no optimization
on tester compilation.  This results in SUN getting good IDs, so mixed 
arch runs work as well (solving prob 5b).

Problem 6
=========
Mixed rs6k and anything else generates errors in SINGLE and SINGLE COMPLEX
SDRV and BSBR.

Solution
========
Problem seems to be that slarnd gives eps-different answers.  Fixed here
by using sbtran = real( dlarnd(2, iseed) ).  Note this is going to kill
us if we go on cray . . .

Problem 7
=========
RS6K code hangs (place it hangs varies) for verb=0.  If verb=2, it finishes.
When it hangs, pvm also hangs.  Other platforms finish OK.

Solution
========
None.  It is an RS6K/pvm problem.  Bob has looked at it and is aware of it.

Problem 8
=========
16 node SP1 failed for some sum checks.

Solution
========
Expand error bound, of course.  Happened only for sums with small results.
error bound now:  2 * eps * nnodes * (biggest [partial] result sum could produce).

Problem 9
=========
IBM reported 3 errors: allocating 2*length in emergency buff, nonsensical call
to mpc_task_set, and blacs_exit not resetting all values.

Solution
========
All problems fixed, but note that calling BLACS after blacs_exit is illegal
anyway; won't work on PVM at all . . .