Compiler design sprint summary

12 February 2017

This last week we had the rustc compiler team design sprint. This was our second rustc compiler team sprint; the first one (last year) we simply worked on pushing various projects over the finish line (for example, in an epic effort, arielb1 completed dynamic drop during that sprint).

This sprint was different: we had the goal of talking over many of the big design challenges that we’d like to tackle in the upcoming year and making sure that the compiler team was roughly on board with the best way to implement them.

I or others will be trying to write up many of the details in various forums, either on this blog or perhaps on internals etc, but I thought it’d be fun to start with a quick post that describes the overall topics of discussion. For each one, I’ll give a quick summary and, where possible, point you at the minutes and notes that we took.

On-demand processing and incremental compilation

The first topic of discussion was perhaps the most massive, in terms of its impact on the codebase. The goal is to reorient how rustc works internally completely. Right now, like many compilers, rustc works by running a series of passes, one after the other. So for example we first parse, then do macro expansion and name resolution (these used to be distinct, but have now become interwoven as part of the work on macros 2.0), then type-checking, and so forth. This is a time-honored approach, but it’s beginning to show its age:

  • Some parts of the compiler front-end cannot be so neatly separated. I already mentioned how macro expansion and name resolution are now interdependent (you have to resolve the path that leads to a macro to know which macro to expand). Similar things arise in type-checking, particularly as we aim to support constant expressions in types. In that case, we have to type-check the constant expression, but it must also be part of a type, and so forth.
  • For better IDE support, it is desirable to be able to compile just what is needed to type-check a particular function (we can come back and cleanup the rest later).
  • Things like impl Trait make the type-checking of some functions partially dependent on the results of others, so the old approach of type-checking all function bodies in an arbitrary order doesn’t work.

The idea is to replace it with on-demand compilation, which basically means that we will have a graph of “things we might want to compute” (for example, “does the function foo type-check”). We can “demand” any one of these “queries”, and the compiler will go and do what it has to do to figure out the answer. That may involve satisfying other queries internally (hopefully without cycles). In the end, your entire type-check will complete, but the order in which we do the compiler will be far less specified.

This idea for on-demand compilation naturally dovetails with the plans for the next generation of incremental compilation. The current design is similar to make: when a change is made, we eagerly propagate the effect of that change, throwing away any old results that might have been affected. Often, though, we don’t know that the old results would have been affected. It frequently happens that one makes changes which only affect some parts of a result: e.g., a change to a fn body that just renames some variables might still wind up generating precisely the same MIR in the end.

Under the newer scheme, the idea is to limit the spread of changes. If the inputs to a particular computation change, we do indeed have to re-run the computation, but we can check if its output is different from the output we have saved. If not, we don’t have to dirty things that were dependent on the computation. (The scheme we wound up with can be considered a specialized variant of Adapton, which is a very cool Rust and Ocaml library for doing generic incrementalized computation.)


Supporting alternate backends

We spent some time discussing how to integrate alternate backends (e.g., Cretonne, WASM, and – in its own way – miri.). Now that we have MIR, a lot of the hard work is done: the translation from MIR to LLVM is fairly straightforward, and the translation from MIR to Cretonne or WASM might be even more simple (particularly since eddyb already made the code that computes field and struct layouts be independent from LLVM).

There are still some parts of the system that we will need to factor out from librustc_trans. For example, the “collector”, which is the bit of code that determines what monomorphizations we need to generate of each function, is independent from LLVM.

The goal with Cretonne, as discussed on internals, is ultimately to use it as the debug-mode backend. It promises to offer a very fast, “decent quality” compilation experience, with LLVM sticking around as the heavyweight compiler (and to support more architectures). The plan for Cretonne integration is (most likely) to begin with a stateless REPL, similar to or the playbot on IRC. The idea would be to take a complete Rust program (i.e., with a main() function), compile it to a buffer, and execute that. This avoids the need to generate .o files from Cretonne, since that code does not exist (Cretonne’s first consumer is going to be a JIT, after all).

After we had finished admiring stoklund’s admirable job of writing clean, documented code in Cretonne, we also dug into some of the details of how it works. There are still a number of things that are needed before we can really get this project off the ground (notably: a register allocator), but in general it is a very nice match with MIR and also our plans around constant evaluation via miri (discussed in an upcoming part of this blog post). We discussed how best to maintain debuginfo, and in particular some of stoklund’s very cool ideas to use the same feature that JITs use to perform de-optimization to track debuginfo values (which would then guarantee perfect fidelity).

We had the idea that we might enable different backends per codegen-unit (i.e., per module, in incremental compilation), so that we can use LLVM to accommodate some of the more annoying features (e.g., inline assembly) that may not appear in Cretonne any time soon.


MIR Optimization

We spent some time – not as much as I might have liked – digging into the idea of optimizing MIR and trying to form an overall strategy. Almost any optimization we might do requires some notion of unsafe code guidelines to justify, so one of the things we talked about was how to “separate out” that part of the system so that it can be evolved and tightened as we get a more firm idea of what unsafe code can and cannot do. The general conclusion was that this could be done primarily by having some standard dataflow analyses that try to detect when values “escape” and so forth – we would probably start with a VERY conservative notion that any local which has ever been borrowed may be mutated by any pointer write or function call, for example, and then gradually tighten up.

In general, we don’t expect rustc to be doing a lot of aggressive optimization, as we prefer to leave that to the backends like LLVM. However, we would like to generate better code primarily for the purposes of improving compilation time. This works because optimizing MIR is just plain simpler and faster than other IRs, since it is higher-level, and because it is pre-monomorphization. If we do a good enough job, it can also help to close the gap between the performance of debug mode and release mode builds, thus also helping with compilation time by allowing people to use debug more builds more often.

Finally, we discussed aatch’s inlining PR, and iterated around different designs. In particular, we considered an “on the fly” inlining design where we did inlining more like a JIT does it, during the lowering to LLVM (or Cretonne, etc) IR. Ultimately we deciding that the current plan (inlining in MIR) seemed best, even though it involves potentially allocating more data-structures, because it enables us to optimize (A) before monomorphization, multiplying the benefit and (B) we can remove a lot of temporaries and so forth, in particular around small functions like Deref::deref, whereas if we do the inlining as we lower, we are ultimately leaving that to LLVM to do.

Unsafe code guidelines

We spent quite a while discussing various aspects of the intersection of (theoretical) unsafe code guidelines and the compiler. I’ll be writing up some detailed posts on this topic, so I won’t go into much detail, but I’ll leave some high-level notes:

  • We discussed exhaustiveness and made up plans for how to incorporate the ! type there.
  • We discussed how to ensure that we can still optimize safe code even in the presence of unsafe code, and what kinds of guarantees we need to require.
    • Likely the kinds of assertions I was describing in my most recent post on the topic aren’t quite right, and we want the “locking” approach I began with, but modified to account for privacy.
  • We looked some at how LLVM handles dependence analysis and so forth, and what kinds of rules we would need to ensure that LLVM is not doing more aggressive optimization than our rules would permit.
    • The LLVM rules we looked at all seem to fall under the rubrik of “LLVM will consider a local variable to have escaped unless it can prove that it hasn’t”. What I wonder about is the extent to which other optimizations might take advantage of the ways that the C standard technically forbid you to transmute a pointer to a usize and then back again (or at least forbid you from using the resulting pointer). Apparently gcc will do some amount of optimization on this basis, but perhaps not LLVM, though more investigation is warranted.


Macros 2.0, hygiene, spans

jseyfried called in and filled us in on some of the latest progress around Macros 2.0. We discussed the best way to track hygiene information – in particular, whether we could do it using the same spans that we use to track line number and column information. In general I think there was consensus that this could work. =) We also discussed some of the interactions with privacy and hygiene that arise when you try to be smarter than our current macro system.


Diagnostic improvements

While talking about spans, we discussed some of the ways we could address some shortcomings in our current diagnostic output. For example, we’d like to avoid highlighting multiple lines when citing a method, and instead just underlyine the method name, and that sort of thing. We’d also like to print out types using identifiers local to the site of the error (i.e., Option<T> and not ::std::option::Option<T>). Hopefully we’ll be converting those rough plans into mentoring instructions, as these seem like good starter projects for someone wanting to learn more about how rustc works.


miri integration

We discussed integrating the miri interpreter. The initial plan is to have it play a very limited role: simply replacing the current constant evaluator that lowers to LLVM constants. Since miri produces basically a big binary blob (possibly with embedded pointers called “redirections”), but LLVM wants a higher-level thing, we have to use some bitcasts and so forth to encode it. This is actually an area where Cretonne’s level of abstraction, which is lower than LLVM, is probably a better fit. But it should all work out fine in any case.

This initial step of using miri as constant evaluator would not change in any way the set of programs that are accepted, except in so far as it makes them work better and more reliably. But it does give us the tools to start handling constants in the front-end as well as a much wider range of const fn bodies and so forth (possibly even including limited amounts of unsafe code).


Variable length arrays and allocas

We discussed the desire to support allocas (RFC 1808) coupled with the desire to support unsized types in more locations (in particular as the types of parameters). We worked through how we would implement this and what some of the complications might be, and drew up a rough plan for an extension to the language that would be expressive, efficiently implementable, and avoid unpredictable rampant stack growth. This will hopefully makes its way into an RFC soon.


Non-lexical lifetimes

We spent quite a while iterating on the design for non-lexical lifetimes. I plan to write this up shortly in another blog post, but the summary is that we think we have a design that we are quite happy with. It addresses (I believe) all the known examples and even extends to support nested method calls where the outer call has an &mut self argument (e.g., vec.push(vec.len()), which today do not compile.



Those were the main topics of discussion – pretty exciting stuff! I can’t wait to see these changes play out over the next year. Thanks to all the attendees, and particularly those who dialed in remotely at indecent hours of the day and night (notably jseyfried and nrc) to accommodate the Parisian time zone.

Comments? Check out the internals thread.