differences between implementations of IEEE-754 basic operators

[Michel Hack <hack@xxxxxxxxxxxxxx>]

If we're talking 754-1985, that's it for basic operations (taking
Vincent's addition of "don't look at the bits" into account.  The
latter also covers issues of NaN propagation and NaN signs.

For non-basic operations, watch out for 754-1985 string conversions,
which need be correctly-rounding only for a restricted exponent and
precision range.  754-1985 also makes SNaN exceptions optional for
copy, copysign, and absolute_value.

The just-approved 754-2008 adds fma (fused multiply-add) to the basic
operations, and allows one subtle variation, namely whether a QNaN
addend suppresses an Invalid Operation due to the multiplication.
The new standard tightens requirements for conversions and copy-type
operations, and explicitly defines rules for reproducible results.

Another subtle issue:  min(-0,+0) could be either -0 or +0 (ditto
for max().  This becomes noticeable in 1/min(-0,+0) of course.

Michel.
Sent: 2008-07-04 10:35:26 UTC