Also, checking for NaNs at certain critical points (instead of everywhere) is rather cheap. Then one can throw a DomainError() or similar.
On Wed, Apr 27 2016, Stefan Karpinski wrote: > If you petition Intel to improve support for signaling NaNs, I will gladly > sign that petition. Maybe change.org <https://www.change.org/>? > > But seriously, this is not a Julia thing. It's the result of a bunch of > horse trading between hardware manufacturers and spec designers in the late > 70s. You may want to read these: > > - https://www.cs.berkeley.edu/~wkahan/ieee754status/754story.html > - http://www.cs.berkeley.edu/~wkahan/ieee754status/IEEE754.PDF (see page > 7 in particular) > > In short, "signaling NaNs" which is what you want are an optional part of > the IEEE 754 standard, not well supported in most hardware, and have > performance issues even when they are supported. > > On Tue, Apr 26, 2016 at 6:42 PM, Anonymous <[email protected]> wrote: > >> Why are NaNs allowed to be created without raising an error, and then >> allowed to propagate around your program corrupting basically all future >> computation. >>
