On Mon, Mar 23, 2015 at 8:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Michael Paquier <michael.paquier@gmail.com> writes:
>> On Tue, Mar 24, 2015 at 11:27 AM, Michael Paquier wrote:
>>> On a Windows 7 box with code compiled with MSVC 2010 I am seeing the
>>> same behavior as Rich. This looks like a bug in ~9.3 that meritates
>>> some attention assuming that the latter behavior is legit.
>
>> And MinGW outputs the latter, similarly to other platforms... I'll investigate.
>
> Look for something about setting the IEEE float rounding mode. "Round to
> nearest even" is standard in most places, but it would not astonish me
> to hear that Microsoft got that wrong.
>
> If it is wrong I don't know that we'd want to back-patch a behavioral
> change, but I'd definitely vote for conforming to the norm in 9.5
> and later.
I have done more testing (done only 9.3 with MinGW and MSVC before)
and the failure can be reproduced on master and REL9_4_STABLE as well
because visibly with MSVC 2010 the version of rint used is the one of
src/port, which is defined like that:
return (x >= 0.0) ? floor(x + 0.5) : ceil(x - 0.5);
For example by passing 2.5, we would get 3.0, and not 2.0 (nearest
even number). So isn't the problem here?
--
Michael
