[cfp-interest 3466] Re: [[reproducible]] EXAMPLE 3 in 6.7.13.8

[Jim Thomas]

JeanHeyd,

CFP discussed this at its May meeting. Two problems with the sequence I suggested:

1. It’s not practical. With three accesses to the floating-point environment, the performance overhead would be worth it only if the result value were reused exceedingly many times. The user could write the code to save the result and then access the saved value, which would avoid performance dependency on support for the attribute and on the optimizer.

2. The functions accessing the dynamic floating-point environment are allowed to fail (they return zero or nonzero according as they succeed or fail), which undermines the reproducibility of the sequence.

Changing the computation in current EXAMPLE 3 to sqrt(fabs(x)) would avoid “overflow" and “underflow" floating-point exceptions, but not “inexact”. And, it wouldn’t compute a distance.

Another problem with the current example is that a non-Annex F implementation might not define FE_TONEAREST and there is no rounding mode guaranteed to be supported by all implementations. 

- Jim Thomas

> On Apr 17, 2025, at 7:01 PM, JeanHeyd Meneide <phdofthehouse at gmail.com> wrote:
> 
> On Thu, Apr 17, 2025  Jim Thomas <jaswthomas at sbcglobal.net <mailto:jaswthomas at sbcglobal.net>> wrote:
>> This matter was on the updated agenda for the CFP meeting this week but we didn’t get to it. We will carry it over to the 12 May CFP meeting and come up with a recommendation. JeanHeyd, is that suitable?
> 
>      Sounds good to me! Just let me know how it goes.
> 
> Sincerely,
> JeanHeyd

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20250521/586655aa/attachment.htm>