Re: [AD] div-by-zero... math3d.c

[Sven Sandberg <svens@xxxxxxxxxx>]

aj wrote:
> normalize_vector_f()
>
> in /allegro419/src/math3d.c
>
> could the vector_length_f()  return zero ?

I replied:
> Certainly, e.g., if all three arguments are zero.

Maybe I should elaborate a bit on that:  If all three arguments are 
zero, then the local variable `length' becomes infinity and 
*x=*y=*z=NaN.  However, it seems possible that, the arguments are 
initially moderately tiny but not zero, so that their squares are zero. 
 In this case `length' becomes infinity but *x=*y=*z=0.0, a little bit 
inconsistent with the previous case.

OTOH, I think there is probably a problem in the user code in such a 
case; you shouldn't try to normalize so small vectors anyway.

> Btw, I think this routine would better be implemented as
>
>    fixhypot(x, fixhypot(y,z))
>
> to avoid the current loss of precision.

Um, sorry, I now changed context, and was of course referring to the 
implementation of `vector_length()'.

Sven