Re: Inline round for IA64

[Richard Henderson <rth at redhat dot com>]

On Sun, Jan 18, 2004 at 12:44:51PM -0800, Geoff Keating wrote:
> Suppose the number involved is something like 2^-10, represented as
> (2^-10, 0).  You'll add 2^106 to it, which will be represented as
> (2^106, 2^-10).  Then subtract 2^106, and you get (2^-10, 0) again,
> but the correct answer is (0, 0).

Ah.  I would have expected the library to enforce a constant number
of bits in the fraction, making that second number ill-formed.

I can't even imagine what the lack of such constancy does to
analysis of floating point algorithms...


r~