Re: [Stds-754] Those horrible global flags

[Nick Maclaren <nmm1@xxxxxxxxxxxxx>]

Ivan Godard <igodard@xxxxxxxxxxx> wrote:
> Posix threads (or other user threading explicit in the program) are only 
> part of the problem. Another part is implicit, very lightweight,  
> threads that have been generated by the compiler as an optimization. 

Yes.  Those are the real problem, and I referred to them as being
associated with OpenMP and automatic parallelisation, but may not have
been entirely clear.

People who think only in terms of POSIX often think that the problem
is simpler than it is.  POSIX threads are the easy case.

> A still different problem not mentioned by Maclaren is speculation. 

Yup.  Agreed.

> This could be done easily by 
> defining that any flag-raising operation generates a NaN, and encode the 
> flags into the NaN payloads.

Well, yes, but that would make the problem of operations losing the
NaN state even more of an issue.  That is fixable, but not compatibly.

> The default 
> should be that inexact is disabled and everything else produces a NaN.

A lot of people won't like that behaviour for underflow.


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