Re: status of quadruple precision arithmetic in g95 and gfortran?
James Giles <email@example.com> wrote:
> If the compiler make *inconsistent* use of extended precision,
> Kahan's method can actually be less precise than the simple
> sum. This is because Kahan's algorithm is keeping track of
> the discrepancy between the simple sum and what should be
> calculated and adding it back in at each step.
x87 implementations are well known for inconsistent use.
It is expensive to save/reload data frequently, and there
aren't enough x87 registers to keep complicated expressions
completely in registers.
The original 8087 was supposed to be designed to have an
infinite stack, with registers being spilled to memory when
needed. It turned out not to be possible to actually write
the needed software in the end. I don't know if the problems
were fixed in later 80x87 implementations, but it doesn't seem
that it has been done yet, as far as I know.