From: James Amundson <amundson at fnal>
To: Richard Fateman <fateman@cs.berkeley.edu>
On Thu, 2002-09-19 at 16:06, Richard Fateman wrote:
> It is entirely possible that the 68020 etc uses "extended"
> precision (using 64 bit fractions instead of 53 bits) for
> floating point registers.
Yes. In days long gone I used to program 68K Macintoshes. I remember
there was always something funny about the way floating-point precision
worked there, but I don't remember any of the details.
IIRC the 68881 chip internally has 80-bit float registers and does
calculations with all the bits. There are about 8 registers, and if
the if the code generator is smart enough to keep intermediate results
in the registers, the difference if low-order-bit roundoff in chained
calculations can be quite visible.