double rounding in x86

[Tom Lynch]

here is an augument to the contrary:

 the answer placed in memory *is not* the infinitely precise
 value rounded into the destination format.

 The standard is pretty clear about this.  Even the foot note
 about expanded exponent range disallows double roundings.

 Also, the spirit of the standard, same bit strings for the
 same operations on conforming platforms, is violated.

-tom