Note that a CodeSmell is a hint that something might be wrong, not a certainty. A perfectly good idiom may be considered a CodeSmell because it's often misused, or because there's a simpler alternative that works in most cases. Calling something a CodeSmell is not an attack; it's simply a sign that a closer look is warranted.

That may sound like semantics, but I think it's an important distinction.

Geez, you've sure drunk the kool-aide, haven't you?

You say that, Reality has corner cases, code doesn't. But in the real world there are cases where people need optimizations. (Less than most people think, but it does happen.) Optimizations naturally lead to having corner cases. Those corner cases should generally not be visible in the interface. Therefore code sometimes does need corner cases that aren't part of reality.

Now you may think that we should never optimize. But just today I wrote about an example where not having the optimization leads to bugs. Underscoring my point even more, that optimization should be invisible in the interface, and the initial implementation of this optimization had bugs in its corner cases. (Interestingly, the corner cases where it had bugs would not have been visible to a code coverage test like Devel::Cover, underscoring even more the need for white-box tests on it!) Therefore if your ideal is fewer bugs, then we have a concrete case where you really want to have corner cases that are not visible in the interface.

In short, you're being blinded by your fanaticism. Real code sometimes does have corner cases that aren't visible in the interface. Those corner cases should either be removed or tested with white-box tests. I agree that often it should be removed, but you go too far when you say that they should always be removed.

I'll admit that I've been enjoying some TDD Kool-Aid of late. In part, this is a reaction to my attitude towards testing several years ago and how it's come back to haunt me now. Specifically, with PDF::Template.

In my view of the world (however limited in scope that might be), optimizations are an implementation detail, save where they are required by spec. For example, "This process must happen in 0.1 seconds or less per record." Then, you write tests to verify that not only does this occur for the 1-record situation, but that the process scales linearly as you run 100, 1000, and 1_000_000 records. (Obviously, the latter is an weekend-only test.)

When they are not required by spec, I begin to wonder why optimizations are needed. Assuming that they are, then it's just a different implementation that meets the same specification (as enforced by the unit-tests).

In your case, the original implementation had bugs in it. When you find the bug, it becomes part of the spec (through the testsuite) and you fix the implementation to pass the tests. If that results in a quicker runtime or a reduced memory usage, that's bonus. It's not an optimization - it's correctly functioning code.

Think of this another way - if your code has corner-cases, then your interface also has corner-cases. If the interface doesn't specify to the user that these corner-cases exist, then how will s/he know to take them into account? Or, let's say that there's what you would call a corner-case because of VMS. File::Spec is a good example of this. Well, the spec now includes "This will run on VMS." Whatever is needed to meet the spec is what is needed to be done. And in the File::Spec testsuite, there's a section that deals with VMS, as there should be.

---

My criteria for good software:

1. Does it work?
2. Can someone else come in, make a change, and be reasonably certain no bugs were introduced?

I think Ovid gave perfect examples to disprove his point. Those are all corner cases that need to be in the spec and so could be subject to black-box testing.

However, I don't agree with the extreme point of view that every corner case needs to be in the spec. Because... You Are Not Going To Need It.

If you spend all of your time gold-plating and turd-polishing your spec, then you run into the same problems with doing that to your code.

We save a lot of time by leaving a lot of corner cases unspec'ed because they are implementation details and whether X or Y happens in that corner case is not going to matter for 99.9% of them. The 0.1% of cases where a corner case that was found during implementation mattered in the finished product are caught in other phases of testing or based on judgement. Most code corner cases don't need the attention of revisiting and rehashing the spec.

But we test those corner cases in our unit tests. Usually we either want the corner case to blow up (so we know when it becomes a case that matters) or we want to make sure it doesn't blow up (because the outcome doesn't matter but you want to know that the code doesn't contain a stupid and fatal mistake just in case the "impossible" isn't as impossible as first thought).

Corner cases do not always mean extra code. If the only corner cases in code were places where you could just remove the code for that corner case because You Are Not Going To Need It, then you could almost make your high-ground claim of "no corner cases unless in the spec" (other than optimizations, as tilly pointed out).

That is why languages have "undefined" behaviors. Because it is a huge waste of time to over-specify every tiny detail of behavior such that you've completely tied up the hands of the implementors for the sake of guarantees that don't matter (except they matter now because your implementation sucks because it is forced to meet the micro-managing spec).

And yet, part of the spec is that the language never dumps core. So you need to test every line of code even though some of those lines were a means (implementation detail) to an end (spec'd requirement).