[Cfp-interest] problem with fmaxmag code

[Jim Thomas]

The sample implementations of fmax and round have similar problems.

-Jim

On Oct 15, 2012, at 2:38 PM, Jim Thomas wrote:

> Our Part 1 draft has
> 
> ------------------------------------------------
> [3] The body of the fmaxmag function might be[1] 
>        
> 
> { return (isgreaterequal(fabs(x), fabs(y)) || isnan(y)) ? x : y; }
> 
> 
> [1] This implementation does not handle signaling NaNs as required of implementations that define FP_SNANS_ALWAYS_SIGNAL.
> 
> ------------------------------------------------
> 
> isgreaterequal might or might not raise the "invalid" floating-point exception if the implementation does not define FP_SNANS_ALWAYS_SIGNAL. Suppose it does. Suppose also that same-format conversions are copy operations. Then if x is a signaling NaN, the code would raise "invalid" but return a signaling NaN, violating F.2.1[4]. A fix would be to change the code to
> 
> { return canonicalize((isgreaterequal(fabs(x), fabs(y)) || isnan(y)) ? x : y); }
> 
> Then we could remove the footnote. The change would also serve to produce a canonical result.
> 
> Any better ideas?
> 
> -Jim

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20121015/859665d5/attachment-0001.html