On Jan 8, 2009, at 9:30 AM, Hans-Martin Mosner wrote:

> AFAIK, the Croquet project needs to have exactly reproducible float
> arithmetics on all platforms. Would these rounding effects on the x86
> affect that goal (perhaps with denormal intermediate results)?

Yes.

> Maybe it would be nice to write a unit test which would uncover
> different rounding behaviors between purely-double-precision FP 
> hardware
> and extended-precision hardware.

I'd have to play around a bit to get this into a Squeak unit test, but 
here's a test vector which shows the effect:

	3ff3208a25e04e87 * 000316dd1d02d1ae

	x86 result:		0003b16ef930a76e
	powerPC result:	0003b16ef930a76f

These are IEEE double-precision floating-point numbers, specified in 
hex (big-endian) so that they are bit exact (no conversion error from a 
decimal representation).  The first number is between 1.0 and 2.0, 
while the second number is a denormal value.  The product shows an lsb 
difference between an x86 platform (extended-precision FPU) and a PPC 
platform (double-precision FPU), even when the x86 floating-point 
control word is set to use double-precision rounding for all results.

	-- Tim Olson