Re: printf("%.0f", 0.5) and round-to-even

[Bruno Haible]

Hi Eric,

> An interesting question popped up on the bug-gnu-utils list:
> http://lists.gnu.org/archive/html/bug-gnu-utils/2008-09/msg00055.html
> 
> POSIX states that with %f, the low-order digit is rounded in an 
> implementation-
> defined manner.  However, since 0.5 is exactly half-way between 0 and 1, and 
> most machines default to IEEE round-to-even, should we make the gnulib printf-
> posix check enforce that printf("%.0f", 0.5) outputs "0" per round-to-even 
> rules?  If so, then at least cygwin 1.5.x fails this check, by outputting "1".

On one hand, having deterministic round-to-even output is certainly a feature.

On the other hand:
  - If POSIX says it is implementation-defined, then programmers should
    know that they are relying on behaviour beyond POSIX.
  - It requires extra code that most likely not everyone who just wants a
    POSIX printf is willing to bear.

I would therefore propose a module fprintf-posix-ext (or fprintf-gnu) that
provides this POSIX behaviour with extensions. And so on for the other
*printf variants.

Does glibc do round-to-even here generally (not just in the one special case
you mentioned?) If so, the module name should be called fprintf-gnu, otherwise
fprintf-posix-ext.

Bruno