Hacker News new | comments | show | ask | jobs | submit login

Funny, I had something to do with this code back in the day! I'm guessing it was a copy+paste bug and they copied from the LambdaCompiler, which uses StrongBox<T> for its closed-over parameters[1], since StrongBox<T>.Value is a field. The idea was to have the closures be really fast.

The history of ET compiler: it started with LINQ in .NET 3.5. Originally it was pretty simple and just handled expressions. In .NET 4.0 we merged the entire codebase with the IronPython/IronRuby compiler trees, expanding the "expression trees" to handle statements. IIRC, it can generate almost any IL construct that you might need, and is usually a lot easier to work with. But we found .NET's runtime compiler (DynamicMethod) was a bit too slow for a lot of use cases. It also wasn't supported on some CLR configurations. To address this we wrote an interpreter and some heuristics to switch from interpreted to compiled. But the actual System.Linq.Expressions.Interpreter must have happened after 4.0, because I don't remember that at all. Instead we just shipped it as a shared library used by IronPython and IronRuby.

Here's the normal ExpressionQuoter: https://github.com/IronLanguages/main/blob/7be8b73e246bfb029...

And here was the interpreter. I don't see the ExpressionQuoter, so either that's a newer fork of the code that was rolled into System.Core, or maybe a completely new implementation. https://github.com/IronLanguages/main/tree/master/Runtime/Mi...

IIRC, ExpressionQuoter was mainly to support the Quote expression, and was always a bit buggy. The 3.5 version was seriously messed up, and our prerelease versions of .NET 4.0 also had various bugs, and very few tests. I tried to fix it by having it use the same reuse closure mechanism as the normal compiler. Funny that same feature caused issues later on.

[1] one might wonder: why use StrongBox<T>, essentially boxing every parameter, rather than just generating a type with only the right fields? The reason was that generating a type in .NET 4.0 timeframe was absurdly slow. Like, a few hundred per second slow. I think this has been largely fixed now, but it was a huge performance problem for Iron* language runtimes back in the day




So, do you think this is Dino's fault? :P


Hah. Or Tomas. Not sure. The S.L.E.Interpreter is after my time :)


It might be my fault... The DLR interpreter got rolled into .NET native so that ETs could be interpreted (hence the weird EETypeRva:0x01588388 error). I actually did the initial set of work on that having had experience w/ the DLR interpreter and handed that off to the .NET team. I probably did the expression quoter but I don't quite remember :(

One thing I'll point out though, it's a Field on StrongBox<T> for correctness not performance - the field needs to be capable of being passed by reference to get consistently correct semantics. That's simply not possible on .NET native using the interpreter so it will end up with copy in / copy out semantics (which could break people but it's pretty unlikely). Also StrongBox<T> pre-existed the DLR expression compiler and was originally added w/ LINQ's ETs in 3.5 so we were also just re-using what they had already done. IronPython actually had Reference<T> early on which grew into the DLR's version and then finally converged back on StrongBox<T>.


Aha! Good call on the ByRef. Totally had forgot about that. Yeah, that was very important for correctness.




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact

Search: