On Thursday, 17 September 2015 at 23:53:30 UTC, Anthony Di Franco
wrote:
Whether this library should be part of the standard library, I
don't know. It would seem to depend on how much people want the
standard library to support verified numerical computing. If it
is clear that verified numerical computing needs good support
in the standard library, something like unums should be there,
maybe even with some other techniques built on top of them
(Taylor model or Levi-Civita for example).
I don't think you should expect D to support verifiable
programming. The only person here that has pushed for it
consistently is Bearophile, but he is not a dev (and where is
he?).
Andrei has previously voiced the opinion that interval
arithmetics as defined is ad-hoc and that D should do it
differently:
http://forum.dlang.org/post/[email protected]
Walter, Andrei and many others have previously argued that you
can turn asserts into assumes (basically assuming that they hold)
without wrecking total havoc to the correctness of the program
after optimization.
It has also been argued that signalling NaNs are useless and that
reproducible floating point math (IEEE754-2008) is not going in
based on some pragmatic assumptions that I never quite
understood. The current definition of D floats is fundamentally
incompatible with IEEE 754-2008. So I am not even sure if you can
implement IEEE 1788 (interval arithmetics) as a plain D library.
D also only have modular integer math so you cannot detect
overflow by adding a compiler switch since libraries may depend
on modular arithmetic behaviour.
D is created by hackers who enjoy hacking. They don't have the
focus on correctness that verifiable-anything requires. So if you
enjoy hacking you'll have fun. If are into reliability, stability
and correctness you'll get frustrated. I'm not even sure you can
have it both ways (both have a hacker mindset and a correctness
mindset in the same design process).
