Re: value-preserving transformations
P1788 members and Baker
As a member, I see an interesting discussion developing on what
manipulations of expressions (or "value-preserving optimizations" are valid
in interval computation.
With my Tech Ed hat on, I find it valuable as well as interesting. I suggest that
those participants who wish, form themselves into an informal Working
Subgroup with the aim of writing a position paper intended to become a
section of the standard, similar in purpose to 10.4 of P754.
I suggest this because it seems a topic whose issues can be clarified at this
early stage - though detail may need to be revisited later.
Baker, as acting Chair I think it would fall to you to give such a group a formal
status in due course, as specified in the P&P document. Does this seem a
good idea to you?
John Pryce
-----
Scott Ferson wrote on 2008/11/22 Sat AM 12:30:20 GMT
> So, if I understand you correctly,
>
> [-2,3] * [-2,3] should *not *be optimized to [-2,3]**2
>
> but
>
> A * A *should *be optimized to A**2.
>
> I think this makes good sense. In the former case, the compiler has no
> way to know that the first and second instance of [-2,3] refer to the
> same quantity, whereas, in the latter case, the programmer has asserted
> that they do by choosing the one variable name.
> Christian Keil wrote:
> > I think we should be very cautious with such transformations.
> > Currently writing an interval library or some rigorous algorithm is
> > quite difficult because of compilers being too smart.
Michel Hack wrote:
> >> Interval arithmetic presents some unique challenges compared to
> >> plain floating-point arithmetic in this regard.
> >>
> >> On (Thu, 20 Nov 2008 11:43:47 +0100) Arnold Neumaier wrote,
...
> >>> In addition, it will specify the following about value-changing
> >>> optimizations:
> >>> These should be handled similar to Section 10.4 of IEEE-754-2008.
...
-----------------------------------------
Email sent from www.virginmedia.com/email
Virus-checked using McAfee(R) Software and scanned for spam