Boost logo

Boost Users :

Subject: Re: [Boost-users] deleting operators to forbid mixed arithmetic? [multiprecision]
From: Dani Brake (danielthebrake_at_[hidden])
Date: 2017-08-13 17:54:04

> Been there, done that – fallen into the big pit awaiting the unwary mixing
> double and multiprecision – again and again… L

I am being bit by this one hard. Converted a bunch of multipliers and
tolerances from mpfr_floats to doubles in an effort to simplify some code,
and now mixed precision is introducing error.

> I think that you don’t really want to delete a family of operator*s – what
> you DO want to get a warning when there is loss of significance, similar to
> the warning when you fit a double into a float.
> It’s OK to multiply (or other ops) with a double as long as this doesn’t
> lose any bits or digits of precision.
> So multiplying by 0.5 is OK, but not by multiplying 0.1 (the
> multiprecision type needs FP_type(1)/10 to represent 0.1 as accurately as
> possible for the FP type).

Multiplying by 0.1 is a common desire in my household (adaptive precision
solving of polynomials).

If only the compiler provided a trait bool is_exactly_representable(double)
> so that a warning could be issued…

This is a bit tricky, because any double is by definition exactly
representable by a double, right?. You would want is_exactly_representable
to get the value before being transformed into a double? Another issue I
foresee having to deal with is runtime parsing. Forbid the runtime user
from feeding doubles problematic numbers.

Thanks for the help,

Boost-users list run by williamkempf at, kalb at, bjorn.karlsson at, gregod at, wekempf at