Re: pitfalls of verifying floating-point computations -- register spills

[Nick Maclaren <nmm1@xxxxxxxxxxxxx>]

A damn good question.  Let's ignore my usual legalism, and simply
stick with what most (but not all) of SC22WG14 seem to intend.

Assuming that __STDC_IEC_559__ is not set to 1, the compiler is
C90 or C++, etc.

>       e = (a+b)*c;

That may either reduce the bits after the addition or not, at whim.
In extreme cases, not every execution of the same statement need
do the same thing.

>       d = a+b;
>       e = d*c;

That must reduce it, always.

>       register double t;
>       t = a+b;
>       e = t*c;

There are more dissenters here, but the consensus is that register
is merely a form of decoration.

Assuming that __STDC_IEC_559__ is set to 1 and the compiler is C99.

All of the examples must reduce it.


Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QH, England.
Email:  nmm1@xxxxxxxxx
Tel.:  +44 1223 334761    Fax:  +44 1223 334679