numtestf is broken
Tue Feb 4 18:23:00 GMT 2014
Ah good to know. I'm working on a project with a newlib based runtime,
apparently the maintainers there didn't see the README and copied mathfp
into their project...
Applying the patch shouldn't hurt, but I'll definitely try to get our
project switched over to the stuff in the "math" subdir.
On Tue, 2014-02-04 at 12:26 -0500, Craig Howland wrote:
> Just in case you did not know, mathfp is not supposed to be used any more (being
> pretty well messed up), and supposedly you can't even choose it from configure.
> So unless you were planning to try and fix the whole thing, you probably should
> save your time.
> This does not necessarily mean that the patch cannot be checked in. It looks
> good to me. (Although, while glancing at it, the routine could be considered to
> be flawed in execution order in that it does a floating point comparison for
> zero before the NAN and INF checks, which can produce a floating point exception
> (invalid, if the argument is NAN). Check routines are supposed to avoid those.
> But this only matters if the user cares about FP exceptions.)
> This brings up something that apparently has been true for at least a couple of
> years (c.f. https://sourceware.org/ml/newlib/2011/msg00381.html): the newlib
> email archive search engine fails to find "mathfp" at all. (I just tried,
> altering plenty of the options. I ended up just using Google with
> "site:https://sourceware.org/ml/newlib mathfp" to find stuff.) How can this be
> On 02/04/2014 07:13 AM, Matthias Braun wrote:
> > Attached is a patch for numtestf which did not detect NaN and INF
> > correctly.
> > Greetings
> > Matthias Braun
More information about the Newlib