Re: [PATCH][4.3] Expand lround inline for x86_64/i?86 SSE math

[Brooks Moses <bmoses at stanford dot edu>]

Steve Kargl wrote:
> On Thu, Oct 19, 2006 at 03:34:23PM -0700, Steve Kargl wrote:
> > It states that NINT(X) = INT(X+0.5) for X > 0.  You then
> > find INT(A) = 0 for |A| < 1.  A quick scan over the section
> > on numeric binary operators did not reveal the normal 
> > weasel words, but I suspect that the result of X+0.5 is a
> > processor-dependent approximation and round-to-nearest
> > meets that criterium.
> >
> > OTOH, I think it is a bug because we can choose a better
> > algorithm to determine if x+0.5 is truly less than 1.
> > See for example round[f] in intrinsics/c99_functions.c
>
> Upon further review, the standard says NINT(X) = INT(X+0.5)
> for X > 0.  So, if and when someone makes a change to the
> inlining of NINT(X), its result needs to agree with
> INT(X+0.5).  

Note, though, the complete change of language in the definition of NINT 
in the Fortran 2003 standard.  In the pathological cases where they 
might arguably be held to require different things, I think we should be 
following the later language.

Beyond that, I am also somewhat dubious about the idea that the language 
you cite was really meant to mean that NINT(X) must be subject to 
exactly the same rounding errors as INT(X+0.5).  I think that's the same 
sort of splitting too fine a hair that Richard Maine told me I was doing 
when I claimed a few months ago that DATA I/.../ and I=... were required 
to treat the largest negative integer differently.  :)

- Brooks