hardware float stack corruption

Hi All

I think I have seen inexplicable stack errors on both the floating point stack and the data stack. Two years ago I saw a problem that was intermittent and bizarre. The code I was running used the AMOEBA.F nonlinear minimizer from the unsupported FORTH scientific library. Periodically, when writing intermediate values to the screen I would sometimes get a F.P. stack underflow error. If I pushed 3 or 4 values on the stack, before starting the calculation, to see how deep the problem might be, the error would go away and all my extra values remained untouched. Two years later in an unrelated problem I have converted AMOEBA to work in fixed point using only the data stack. The same type of behavior has surfaced, but this time on the data stack. Most of the time the code works just fine with no errors, even when displaying intermediate results. Again it was only when writing to the screen and it would disappear if the stack had extra values. The extra values would remain unsullied after the calculation. The stack
 errors would also disappear if I changed the displayed values. The 'fix' for me was just to put values on the stack when this occurs. I do suspect an interesting interaction in TTY.DLL but my code is not production code, it is experiment analysis, so I have no need to chase it any further.

AMOEBA runs can really tie up the CPU while seeking a solution and can run for seconds even on modern multi-gHz processors. This means that many unrelated IO interrupts are occurring during the processing. I am pretty sure it is not in AMOEBA because things look right when the extra values are on the stack.

Ken B

Dr. Kenneth B Butterfield

phone: 505 667 8944

PO Box 1663
ms J562
Los Alamos National Laboratory
Los Alamos New Mexico 87545

"I do not trust a computer that I am not carrying!"

----------------------------------------------------------------------
sftalk_at_forth.com The SwiftForth programming discussion email list
To unsubscribe, send subject "unsubscribe" to sftalk-request_at_forth.com
For list command help, send subject "help" to sftalk-request_at_forth.com
Message archives are located at http://www.forth.com/archive/sftalk
----------------------------------------------------------------------
This list is a forum for SwiftForth users. For product support and bug
reports, please send email to support_at_forth.com
----------------------------------------------------------------------
Received on Wed Oct 26 2005 - 08:44:37 PDT