![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
TMI: When it rains...
Mar. 15th, 2013 05:52 pmI spent too much of today dealing with an apparent hardware failure on -- irritatingly enough -- the laptop I'd borrowed from IT to replace the laptop I was using for the past almost-a-year, which just before I left for Vancouver decided it didn't have a wifi card anymore. Most of yesterday I was trying to debug an issue where I was getting non-deterministic segfaulting and also occasional not-seeing-the-bug-I-was-looking for. Graydon suggested that I run memtest86, and so I did... and it reported 193 million errors before I decided that was enough. So then I sshed into a buildbot that Graydon decommissioned temporarily, and moved my working directories onto it.
After which, I found the bug! I'm not happy with my fix, though. So you may recall that a while ago, I introduced a ty_err type to represent a value that can't be typed and allow the typechecker to typecheck other parts of the program before finally reporting all the errors. The fix I was working on was to never represent a type like () -> ty_err -- a function type where the result type is erroneous -- because this should just normalize to ty_err; a type with a component that was erroneous is erroneous itself. The problem is that a type might look like () -> α, where α is a type variable, and if α gets unified with ty_err at some later point in time, the way the Rust typechecker works makes it hard to see how that information gets propagated back to anyone who might have a copy of () -> α sitting around.
My solution was just to call a function that resolves all type variables in the type at the point where we've just checked the body of a closure, and write the resolved type back in for the closure. But this is unsatisfactory to me because I haven't really thought about whether this is the only point in the typechecker where that needs to happen.
The whole approach is probably wrong and misguided in some way I don't quite understand, and then I decided it was a good idea to rebase my branch for some reason, which resulted in conflicts and quite possibly a bad merge by me. I think that means it's time to leave for the weekend.
After which, I found the bug! I'm not happy with my fix, though. So you may recall that a while ago, I introduced a ty_err type to represent a value that can't be typed and allow the typechecker to typecheck other parts of the program before finally reporting all the errors. The fix I was working on was to never represent a type like () -> ty_err -- a function type where the result type is erroneous -- because this should just normalize to ty_err; a type with a component that was erroneous is erroneous itself. The problem is that a type might look like () -> α, where α is a type variable, and if α gets unified with ty_err at some later point in time, the way the Rust typechecker works makes it hard to see how that information gets propagated back to anyone who might have a copy of () -> α sitting around.
My solution was just to call a function that resolves all type variables in the type at the point where we've just checked the body of a closure, and write the resolved type back in for the closure. But this is unsatisfactory to me because I haven't really thought about whether this is the only point in the typechecker where that needs to happen.
The whole approach is probably wrong and misguided in some way I don't quite understand, and then I decided it was a good idea to rebase my branch for some reason, which resulted in conflicts and quite possibly a bad merge by me. I think that means it's time to leave for the weekend.