What to do about divMod'.
dons at galois.com
Fri Nov 14 23:00:42 EST 2008
> Data.Fixed contains divMod' :: (Real a, Integral b) => a -> a -> (b, a)
> as well as div' and mod'.
> There are two problems:
> 1) Implementation of this type signature requires conversion to Rational,
> which is extremely slow. It really isn't usable in any kind of
> significant loop at all.
> 2) This has nothing specific to do with fixed precision arithmetic
> despite being in the Data.Fixed module.
> I played with writing a patch but I think this requires discussion. I
> believe the original author was aware of the problems and didn't consider
> them urgent.
> Should there be a RULES to address the performance issues? Should this be
> based on RealFrac or specific to Float, Double, etc? Alternately OR
> additionally should there be a separate set of functions created based on
Almost certainly. Things like fromIntegral and truncate already have
extensive type-based rules.
> If we use RULES, this may make programs that depend on it effectively
> unusable without the optimizer. Whatever we do will probably change the
> semantics a bit in terms of infinite/NaN, but I don't think that is likely
> to matter.
Well, we make them the same as they currently are.
> I personally prefer changing the signature of the divMod' family, but I
> anticipate objections.
> I also think that we should make sure that the divMod' family gets inlined
> as my understanding is this will avoid dictionary lookups and aid
> the strictness analyzer, but I'm not 100% certain on this point.
Check the core!
More information about the Libraries