The first bug I decided to tackle involved incorrect type annotations when passing sum-typed variables to a pattern matched function (say that ten times fast). The fix was to look more carefully at the original compiler's logic and try to replicate it more exactly in the Epoch implementation.
That turned out to be trickier than I first thought; as has often happened with this phase of work, I fixed some cases but made others worse. Thankfully, I did manage to get a compact repro case for this particular bug, so it shouldn't be as agonizing to fix as some of the others.
One frustrating discovery is that some of my earliest "fixes" were actually masking other problems, and complicating still more bugs. So a lot of the tail-chasing I've done in the past several hours of work has been wasted time. Of course, after finding that out, I'm back to trying to fix the original bugs correctly... which is proving challenging.
I decided to punt on this one by just rewriting the code to avoid tripping the miscompile. It's kind of disappointing to do that, but at least this way I can make progress on other bugs.
And then, it happened.
At 1:04PM PST, the Epoch compiler emitted a binary of itself that successfully executed, and promptly complained that I hadn't passed it any command line arguments (which is true). The next step is to run the compiler test suite on it...
Sadly, something is still terribly wrong. The self-hosted binary is not parsing any of the test suite programs correctly.
My first hunch is that it's related to a difference in the way the two compilers handle structure value semantics. The C++ compiler emits code that never copies objects by value, which is technically broken. The Epoch compiler, on the other hand, emits code that copies objects when passed by value, which is how it should be - but I suspect that one of my optimization hacks relies on the broken behavior!
That indeed appears to be the case, as screwing up the Epoch compiler's behavior causes the self-hosted compiler to start parsing correctly. However, now there seems to be an issue with pattern matching, because a few panic-asserts are going off when trying to parse even trivial programs. Sad.
At this point, it's time for a break.
After coming back, I smacked my head against the wall for several hours, chasing a couple of the same bugs over and over again with minimal luck. Eventually I realized that one of the test suite programs I was using to verify the compiler had a bug of its own! It wasn't failing due to miscompiles, but rather because the test itself was faulty. This had cost me a lot of running around in circles over the course of the afternoon.
My head is starting to hurt and I need food. I've already burned almost an entire day on this, so I'm not sure if I'll be back later or not.
... but of course I'm an obsessive bastard, and couldn't leave it alone, so I'm back again hacking away on the problems. It looks like one of the last remaining bugs involves pattern matching; it seems that the compiler is only emitting some branches of pattern matchers and not all of them, resulting in failed pattern match assertions when the compiler runs on any nontrivial code.
That turned out to be fairly straight-forward to solve; the next problem to crop up is a bogus pointer dereference somewhere deep in one of the automatically generated constructor functions. Joy.
After poking at it for a little while, I decided that my brain is too mushy and I'm liable to make the problem worse instead of better.
So it's time for some rest.
