Re: Octave 3.0.1 available for ftp

[Michael Goffioul]

On Thu, Apr 24, 2008 at 3:38 PM, Tatsuro MATSUOKA <address@hidden> wrote:
> Hello David
>
>  Thank you for your reply.
>
>  --- David Bateman <address@hidden> wrote:
>  > Wouldn't it better better to have the autoconf test refuse to set
>  > HAVE_ROUND for mingw rather than hard code in the source.. Alternatively
>  > we code test the behavior of the round function in an autoconf test to
>  > see if it fails for the case near bitmax, though that would be more work
>  > to implement..
>  >
>  > D.
>
>  What you said is right.  However, I do not have enough knowledge of auto 
> tools.
>  And I cannot understand what is the problem with 'round' implemeted in the 
> MinGW g++ 3.4.5.
>  If you suggest me a autoconf modification, I will test to it and report here.

The code used internally by MinGW for "round" is the following
(http://www.koders.com/c/fidB63EAFA2C117AAFC1CF9FE9691F8DDBE4DD01A22.aspx)

double
round (double x)
{
  /* Add +/- 0.5 then then round towards zero.  */
  return trunc ( x + (x >= 0.0 ?  0.5 : -0.5));
}

The problem with this implementation is when x is the vicinity of bitmax: it
leads to floating-point overflow (in the mantissa) and produces rounding error,
For instance if x = bitmax, round(x) => bitmax+1, while it should return
bitmax.

Octave used to use the above implementation for MSVC, but because of
this problem, I switched it to use the (more robust) gnulib implementation.

Michael.