David R. MacIver

I’ve recently been writing some collection implementations for Scala which will hopefully go in the standard library in the upcoming release. Naturally I want to make sure that I tested the hell out of them, so I’ve been writing a reasonably large number of ScalaCheck tests to test the code and make sure it does what I want it to. Earlier I realised I should probably be doing some sort of coverage checking to make sure I had tested things adequately so thought I’d see how easy it would be to plug Cobertura into a Scala project. And I have good news! The answer is: Essentially trivial.

There were a few minor issues that cropped up while doing it:

• Cobertura strips scala specific metadata out of the class files. This means that you must compile everything first, including your unit tests, before instrumenting code.
• Scala does a lot of name mangling and generates a lot of classes with weird names. Cobertura naturally doesn’t understand the mapping back to Scala code, so you’ll see the mangled names. They’re not usually that hard to figure out.
• Cobertura won’t be able to find your source files if you don’t follow Java directory conventions for packages.

Other than that, it all works fine. You get per line coverage information for your source files, decent per package and per class reports, etc. No problem.

Unfortunately my coverage wasn’t nearly as good as I hoped, so I’ll probably have to write a bunch more tests before I submit a patch.

Update: Actually, the line number reporting for Cobertura + Scala appears to be really bad, and often really hard to decipher. But it’s still a nice tool, and actually helped me catch some bugs.

Java (and consequently Scala) lack unsigned integer types. Most of the time this isn’t an issue - the majority of the operations you’re likely to care about (addition, subtraction, bitwise operations) work the same regardless of whether you treat the integer as unsigned. There are a few cases where it is a pain though - one that bit me in the context of the work I’ve been doing on and off with Java classfiles is that converting between types of different size. The other one that’s an issue is comparison - integer comparison is signed in Java and there are no operations for doing the unsigned compare.

One way to fix this is to delegate to a solution to the previous problem: Bump the integers up to a type one size up in an unsigned way so that they get converted to positive values, do a comparison on there. This works fine for the smaller values, but if you were to do this with a long you’d have to upgrade to a BigInt, which is a bit sad. I’d previously been doing it this way but earlier when I needed to do unsigned comparisons on longs I decided to look for a better route. (Full disclosure: Actually what happend is I converted some code using ints to code using long and suddenly all my unit tests broke. At that point I discovered that I’d still been trying to do an unsigned comparison using the code for doing it for integers).

Anyway, I popped into ##java to see if anyone there had any ideas for how to do it. Totally counter to form, ##java was actually incredibly useful. The best solution came from someone who went by Tamutnefret and is as follows:

  def unsignedCompare(i : Long, j : Long) = (i < j) ^ (i < 0) ^ (j < 0)

Which is really neat and avoids doing any branching.

Let’s see why this works. We’ll break it down by cases on i (there’s probably a nicer way to see that it works, but I’m blanking on what it is).

Suppose i = 0. This becomes (0 < j) ^ false ^ (j < 0), which is (0 < j) ^ (j < 0), or j != 0. Which is correct, as the only unsigned integer not > 0 is 0.

Suppose now i < 0. This becomes (i < j) ^ true ^ (j < 0).
If (j < 0) this becomes i < j, as x ^ true ^ true = x ^ false = x. If j >= 0 then we certainly have i < j, so we get true ^ true ^ false = false, which is again correct: In unsigned comparison any negative number is greater than any non-negative one.

The case i > 0 follows similarily (the standard mathematician’s weasel words).

Anyway, the explanation of why is boring. It takes less time to convince yourself of its truth than it did to read that. :-) But the point is that it’s a cute, useful formula which I’d not previously encountered so I thought I’d record it here.