Re: Low-order uncertainty in 87 math processing

Stephen Schumacher wrote:


> x:=y;
> [code not changing x or y]
> q:=x-y;
> [ why is q not zero]


John's suggestions (esp about NAN) are good. As you say it only happens
occasionally, without seeing all the code and knowing the input values,
I would have to rate this the most likely explanation.


Here is another. Unfortunately, you did not give the disassembly, so
I cannot check if this is happening.


If x is getting stored and y is being retained in a floating register,
you lose some precision of x but not y, and the difference in q is the
round-off between the values. (Of course, it might be y that is stored
instead of x, too)


MNF (Minnesota University Fortran) used to print out a warning message
on all attempts to compare floating point numbers for equality.
It went along the lines of "Can this really be expected to compare
equal?"
and was delivered with a stern wagging of fingers.


David Moore.
[Given that the wierd behavior is repeatable but goes away sometimes until
a reboot, I'd still be looking at those persistent x87 control flags. -John]
--

Return to the comp.compilers page.
Search the comp.compilers archives again.