How (not) to convert CDouble to Double

Published on

What’s wrong with the following code?

If you use QuickCheck to test the equivalence of Acos.acos and Prelude.acos, you’ll quickly find a counterexample:

You might think this is a difference in the semantics of Haskell acos vs C acos, but the acos(3) manpage disproves that:

If x is outside the range [-1, 1], a domain error occurs, and a NaN is returned.

Moreover, you’ll notice the discrepancy only when compiling the Haskell program with -O0. If you compile with -O1 or higher, both versions will result in NaN. So what’s going on here?

What turns the NaN turned into the Infinity is realToFrac. It is defined as follows:

Unlike Double, Rational, which is defined as a ratio of two Integers, has no way to represent special values such as NaN. Instead, toRational (acos 1.1) results in a fraction with some ridiculously large numerator, which turns into Infinity when converted back to Double.

When you compile with -O1 or higher, the following rewrite rules fire and avoid the round trip through Rational:

Unfortunately, the Haskell 2010 Report doesn’t give you any reliable way to convert between Double and CDouble. According to the Report, CDouble is an abstract newtype, about which all you know is the list of instances, including Real and Fractional. So if you want to stay portable, realToFrac seems to be the only solution available.

However, if you only care about GHC and its base library (which pretty much everyone is using nowadays), then you can take advantage of the fact that the constructor of the CDouble newtype is exported. You can use coerce from Data.Coerce or apply the data constructor CDouble directly.

So here’s a reliable, but not portable, version of the Acos module above: