14 ms·

Nick Nethercote and others have done a lot of work with traditional profiling and optimization like that. They've done a great job, but for my project the funda

by roca 6y ago

Nick Nethercote and others have done a lot of work with traditional profiling and optimization like that. They've done a great job, but for my project the fundamental problem seems to be limited parallelism, which I think requires more fundamental design changes (perhaps getting away from the crate-at-a-time approach of rustc). Though TBH I'm not an expert on the compiler and I'm not confident I know the root problems or the solutions.

kzrdude 6y ago
Performance of the generated code is always the #1 priority in both Rust and llvm, performance improvements are rarely declined, and this prioritization needs to change radically, to get a much faster compiler.
- steveklabnik 6y ago
  Prioritization is not that straightforward. Every open source project has different sub-groups with different goals. While we do care a ton about generated code, there is also movement by other folks to work on the compiler speed problem. It has been a lot of work to come this far, and it will be even more work to get to a better place, but that work is underway. That's not even to mention that since rustc is written in Rust, work to make generated code faster can also make the compiler faster. I'm referring to the rust-analyzer work, which is effectively an outside-in re-write of the compiler, if you squint at it right. At the end of the day, rustc is a very large codebase, and has already undergone major architectural changes, and will be undergoing even more. It is not fast to steer a large ship in a new direction, but that is what is happening. See https://pbs.twimg.com/media/ENC1ls6XYAEdKkv?format=jpg&name=900x900 https://pbs.twimg.com/media/ENC1ls6XYAEdKkv?format=jpg&name=... for a time I ran https://github.com/erikbern/git-of-theseus https://github.com/erikbern/git-of-theseus on the rustc codebase. Thank goodness Rust is such a good refactoring language.
  - kzrdude 6y ago
    You are talking incremental improvements. Radical is needed not just using cranelift for debug builds, but something like that for all builds. Current llvm model does not work well for delivering a fast compiler (opt and non-debug)
    - steveklabnik 6y ago
      You can deliver radical improvements in an incremental fashion. The move towards a query-based architecture, and the rust-analyzer work, are both fundamental, radical improvements. If anything, the cranelift stuff is less radical. Doesn't require any architecture changes, "only" involves implementing the backend API.
      - leeoniya 6y ago
        then there's also ongoing work in rustc_codegen_cranelift, which [as of 2 weeks ago] can now compile rustc: https://github.com/bjorn3/rustc_codegen_cranelift/issues/743#issuecomment-651108168 https://github.com/bjorn3/rustc_codegen_cranelift/issues/743...
        steveklabnik 6y ago
        Yeah to be clear, I am very excited about cranelift too!
      - kzrdude 6y ago
        Sure, that's seriously cool but I don't see how it changes the material balance so to speak... If you still want llvm to codegen the same items, we don't have any radical departure at all, just a different way to do the same thing, when we are release-compiling any particular crate, from scratch.
        steveklabnik 6y ago
        Oh sure, if you're talking purely about a final release build, then yes. But most of the time, compiles are not final release build compiles. People don't generally complain about final release build times. They complain about the build time as they're developing.
  - jbkcc 6y ago
    d
- FpUser 6y ago
  What is wrong having 2 versions: slower at runtime but super fast compile for debug and the other way around for production?
  - necubi 6y ago
    Of course, that is the status quo for rustc (as it is for all production compilers I know of). The issues are that (1) debug builds are still slow enough to be annoying, and (2) they are sometimes _so much slower_ to execute that a debug build + test run is slower than a release build + run. A lot of work is being done to address these issues. The big projects I'm aware of are a move to a query-based compiler (which should improve incremental build times) and the new cranelift backend, which should dramatically speed up debug build times.
loeg 6y ago
The slowness of building crates makes an immense impact on ordinary rust project build times, but as the Linux kernel would just be using rustc directly, and Kbuild already parallelizes CC invocations, the relevant gains here would be single-TU compilation speed.
- benesch 6y ago
  But the translation unit in Rust is the crate. Running `rustc` directly won't change that fact. Unlike in C, where you can compile TUs in any order since you've got interfaces in headers, in Rust you can't compile a crate until you've compiled all its dependencies--and if it's not Cargo that's going to sort that out for you, Kbuild will have to.
  - nitrogen 6y ago
    Given the modular nature of the kernel, how big will each "crate" be? Will they mostly end up interfacing with the rest of the kernel using the C ABI and not need to be compiled as a single unit?
  - cesarb 6y ago
    > in Rust you can't compile a crate until you've compiled all its dependencies Since Rust 1.38 (https://blog.rust-lang.org/2019/09/26/Rust-1.38.0.html#pipelined-compilation https://blog.rust-lang.org/2019/09/26/Rust-1.38.0.html#pipel...) that's no longer the case: "To compile a crate, the compiler doesn't need the dependencies to be fully built. Instead, it just needs their "metadata" (i.e. the list of types, dependencies, exports...). This metadata is produced early in the compilation process. Starting with Rust 1.38.0, Cargo will take advantage of this by automatically starting to build dependent crates as soon as metadata is ready."
    - roca 6y ago
      This is good but it doesn't work well enough in my experience. I still spend a lot of time waiting for dependencies to build before the crates that depend on them.
    - 6y ago

by roca 6y ago

kzrdude 6y ago

Performance of the generated code is always the #1 priority in both Rust and llvm, performance improvements are rarely declined, and this prioritization needs to change radically, to get a much faster compiler.

steveklabnik 6y ago
Prioritization is not that straightforward. Every open source project has different sub-groups with different goals. While we do care a ton about generated code, there is also movement by other folks to work on the compiler speed problem. It has been a lot of work to come this far, and it will be even more work to get to a better place, but that work is underway. That's not even to mention that since rustc is written in Rust, work to make generated code faster can also make the compiler faster. I'm referring to the rust-analyzer work, which is effectively an outside-in re-write of the compiler, if you squint at it right. At the end of the day, rustc is a very large codebase, and has already undergone major architectural changes, and will be undergoing even more. It is not fast to steer a large ship in a new direction, but that is what is happening. See https://pbs.twimg.com/media/ENC1ls6XYAEdKkv?format=jpg&name=900x900 https://pbs.twimg.com/media/ENC1ls6XYAEdKkv?format=jpg&name=... for a time I ran https://github.com/erikbern/git-of-theseus https://github.com/erikbern/git-of-theseus on the rustc codebase. Thank goodness Rust is such a good refactoring language.
- kzrdude 6y ago
  You are talking incremental improvements. Radical is needed not just using cranelift for debug builds, but something like that for all builds. Current llvm model does not work well for delivering a fast compiler (opt and non-debug)
  - steveklabnik 6y ago
    You can deliver radical improvements in an incremental fashion. The move towards a query-based architecture, and the rust-analyzer work, are both fundamental, radical improvements. If anything, the cranelift stuff is less radical. Doesn't require any architecture changes, "only" involves implementing the backend API.
    - leeoniya 6y ago
      then there's also ongoing work in rustc_codegen_cranelift, which [as of 2 weeks ago] can now compile rustc: https://github.com/bjorn3/rustc_codegen_cranelift/issues/743#issuecomment-651108168 https://github.com/bjorn3/rustc_codegen_cranelift/issues/743...
      - steveklabnik 6y ago
        Yeah to be clear, I am very excited about cranelift too!
    - kzrdude 6y ago
      Sure, that's seriously cool but I don't see how it changes the material balance so to speak... If you still want llvm to codegen the same items, we don't have any radical departure at all, just a different way to do the same thing, when we are release-compiling any particular crate, from scratch.
      - steveklabnik 6y ago
        Oh sure, if you're talking purely about a final release build, then yes. But most of the time, compiles are not final release build compiles. People don't generally complain about final release build times. They complain about the build time as they're developing.
- jbkcc 6y ago
  d
FpUser 6y ago
What is wrong having 2 versions: slower at runtime but super fast compile for debug and the other way around for production?
- necubi 6y ago
  Of course, that is the status quo for rustc (as it is for all production compilers I know of). The issues are that (1) debug builds are still slow enough to be annoying, and (2) they are sometimes _so much slower_ to execute that a debug build + test run is slower than a release build + run. A lot of work is being done to address these issues. The big projects I'm aware of are a move to a query-based compiler (which should improve incremental build times) and the new cranelift backend, which should dramatically speed up debug build times.

loeg 6y ago

The slowness of building crates makes an immense impact on ordinary rust project build times, but as the Linux kernel would just be using rustc directly, and Kbuild already parallelizes CC invocations, the relevant gains here would be single-TU compilation speed.

benesch 6y ago
But the translation unit in Rust is the crate. Running `rustc` directly won't change that fact. Unlike in C, where you can compile TUs in any order since you've got interfaces in headers, in Rust you can't compile a crate until you've compiled all its dependencies--and if it's not Cargo that's going to sort that out for you, Kbuild will have to.
- nitrogen 6y ago
  Given the modular nature of the kernel, how big will each "crate" be? Will they mostly end up interfacing with the rest of the kernel using the C ABI and not need to be compiled as a single unit?
- cesarb 6y ago
  > in Rust you can't compile a crate until you've compiled all its dependencies Since Rust 1.38 (https://blog.rust-lang.org/2019/09/26/Rust-1.38.0.html#pipelined-compilation https://blog.rust-lang.org/2019/09/26/Rust-1.38.0.html#pipel...) that's no longer the case: "To compile a crate, the compiler doesn't need the dependencies to be fully built. Instead, it just needs their "metadata" (i.e. the list of types, dependencies, exports...). This metadata is produced early in the compilation process. Starting with Rust 1.38.0, Cargo will take advantage of this by automatically starting to build dependent crates as soon as metadata is ready."
  - roca 6y ago
    This is good but it doesn't work well enough in my experience. I still spend a lot of time waiting for dependencies to build before the crates that depend on them.
  - 6y ago