Vorner's random stuff

Truly zero cost

I know it is claimed how Rust has zero cost abstractions and such and that all these levels of abstractions will just go away in a release build. But there’s a difference in hearing the theory and seeing it really happen in practice. And I don’t appreciate it because I’d consider it magic, but more because I understand how that is being done and it still looks cool.

The background

I wrote this arc-swap library that allows storing an Arc and swap it atomically while other threads keep reading it. It’s something like RwLock<Arc>, but faster and readers are never blocked. And, from time to time I try to polish it a bit more and make it a bit faster under this or that circumstance. Not that there would be some great need for that, but some people like fast cars, other people like playing with hard problems, it’s probably similar.

Anyway, there’s this thing I call a generation lock that is being used sometimes. It’s a one-sided spin-lockish thing. While readers never get blocked by anything, writers can get blocked by this. The arc-swap library offers several mechanisms how to read the data, some of it sometimes using this generation lock. And usually even when using the generation lock, it’s locked for a really short amount of time, so it’s not an issue. However, I also use it in the signal-hook library, where there’s not much choice ‒ allowed operations inside a signal handler are severely limited and the generation lock is the only option available. And the signal handler might hold it for longer.

That in itself would be mostly OK, except that performance is always about trade-offs. So, because the arc-swap tries to be fast on read operations even if many happen at the same time on the same ArcSwap, I had multiple copies (shards) of the generation lock, spreading the load and lowering the contention over the cache lines where the lock lives. But because that takes a lot of space, I had only one set of the locks shared by all the ArcSwap storages. This means that the signal handler could block other, potentially performance-sensitive parts of application from doing their thing and for a long time ‒ depending on how much the signal handler was doing. Not good.

The goal

Obviously, there’s the same tool that can be used for two very different purposes. One that needs to be fast, the other that simply needs to work in restricted environment (the signal handlers), but without much performance needs. They are slightly in conflict with each other on some little details. The signal handler thing would really like its own generation lock. It doesn’t care about the contention and it doesn’t care if that single ArcSwap will be few bytes larger. But most of the code is the same.

I could have taken the code, stripped it of all the fancy things I didn’t need for the signals and embedded this stripped down thing into signal-hook. But that didn’t feel right ‒ all that code duplication and stuff, most of the code would be identical.

How it went

Rust traits are like the hammer that makes everything look like a nail. Therefore I went and defined a trait to describe the place where the lock is stored. Then I made the ArcSwapAny (a „backend“ type with many parameters which is used by type aliases like ArcSwap) generic over the LockStore. Adding a type parameter can actually be a semver-compatible change, if a default is provided.

Currently, it can either be a zero-sized field that delegates to the global storage as before, or a field that actually keeps the generation lock inside itself.

The places in the hot-path code where I was directly accessing a global fixed-sized array got replaced by calls to methods on some generic storage, returning some associated type that could be turned into a reference to a slice.

From what the code was saying to do, it got changed from:

To:

That looks like a lot of extra work to do.

But I’m pretty sure it got just thrown away by the compiler and turned back to that one or two assembler instructions. Not that I’d be reading the assembler code that got out ‒ I’m old school, but not really that old school to read pages of machine code before bedtime. But as I said, the library cares a lot about performance, so every time I touch part of the hot code, I re-run the benchmarks. This way, I throw out about half of the optimizations again (as they make it actually slower). And the benchmarks are very sensitive to even tiny changes ‒ adding length check for the slice would probably turn into visible change in the numbers.

I’m also pretty sure that most languages would not go that far. The idea that the type plugged in has only one possible value, therefore it doesn’t need to be stored and methods on that don’t care about the self reference is pretty neat.

On the other hand, it’s kind of mind-bending, considering how much work there needs to go into throwing away most of the code I’ve written. There’s an arm race going on between the compilers, competing in how much code they are able to discard.

Hints for the compiler

Actually, I did try make it as easy for the compiler as possible. The trait implementation has #[inline] stuck on the methods, so the optimisations can happen across the function boundary. This doesn’t always help and sometimes even hurts, but this time it turned out for the better (well, at least not hurting).

The other thing is, I’m actually declaring the true type the methods are returning as an associated type, not returning a slice-reference. During the measurement, I was seeing this to have a small effect. But I can’t say if it was large enough to be significant, or if it was just a noise. Anyway, preserving critical information ‒ like a constant length of something ‒ across abstractions helps the compiler in doing its job.