Deps: compiler_builtins #15
Comments
|
I believe that almost all intrinsics rustc uses have implementations in rust now at this point. It's largely a matter of time until they're as optimal as the C versions or more are translated. I think it's probably safe to conclude here that defaulting to rust implementations is the way to go and eventually supporting and opt-in C implementation is possible |
|
@alexcrichton Do you mean defaulting to rust implementations for all builds, including binaries shipped through rustup? If not, having a different default between those binaries and std built locally on demand by Cargo contradicts the goal that they should be as close to each other as possible, except for spending bandwidth v.s. CPU time. I think this should be a goal, see the last part of my comment #5 (comment) |
|
Oh nah official rustup builds I think would still use C/assembly since the assembly versions are assumed to be faster and we haven't optimized the Rust ones. Additionally the not all C intrinsics have Rust versions yet, so there may be some esoteric things which haven't been ported which we need to fill out 100% of everything. This may be a topic for a separate issue, but I would personally take it as a given that the official build of libstd and the cargo-built libstd would be different from the get-go. I think it'd be good to unify them in the long run but allowing for differences over time I think is what we'll need to tweak things in libstd as issues come up. |
I'm sympathetic, but at the same time we dont want to expose "use the default binaries" in Is it the case right now that |
Basic standard library support. This is not intended to be useful to anyone. If people want to try it, that's great, but do not rely on this. This is only for experimenting and setting up for future work. This adds a flag `-Zbuild-std` to build the standard library with a project. The flag can also take optional comma-separated crate names, like `-Zbuild-std=core`. Default is `std,core,panic_unwind,compiler_builtins`. Closes rust-lang/wg-cargo-std-aware#10. Note: I can probably break some of the refactoring into smaller PRs if necessary. ## Overview The general concept here is to use two resolvers, and to combine everything in the Unit graph. There are a number of changes to support this: - A synthetic workspace for the standard library is created to set up the patches and members correctly. - Decouple `unit_dependencies` from `Context` to make it easier to manage. - Add `features` to `Unit` to keep it unique and to remove the need to query a resolver. - Add a `UnitDep` struct which encodes the edges between `Unit`s. This removes the need to query a resolver for `extern_crate_name` and `public`. - Remove `Resolver` from `BuildContext` to avoid any confusion and to keep the complexity focused in `unit_dependencies`. - Remove `Links` from `Context` since it used the resolver. Adjusted so that instead of checking links at runtime, they are all checked at once in the beginning. Note that it does not check links for the standard lib, but it should be safe? I think `compiler-rt` is the only `links`? I currently went with a strategy of linking the standard library dependencies using `--extern` (instead of `--sysroot` or `-L`). This has some benefits but some significant drawbacks. See below for some questions. ## For future PRs - Add Cargo.toml support. See rust-lang/wg-cargo-std-aware#5 - Source is not downloaded. It assumes you have run `rustup component add rust-src`. See rust-lang/wg-cargo-std-aware#11 - `cargo metadata` does not include any information about std. I don't know how this should work. - `cargo clean` is not std-aware. - `cargo fetch` does not fetch std dependencies. - `cargo vendor` does not vendor std dependencies. - `cargo pkgid` is not std-aware. - `--target` is required on the command-line. This should default to host-as-target. - `-p` is not std aware. - A synthetic `Cargo.toml` workspace is created which has to know about things like `rustc-std-workspace-core`. Perhaps rust-lang/rust should publish the source with this `Cargo.toml` already created? - `compiler_builtins` uses default features (pure Rust implementation, etc.). See rust-lang/wg-cargo-std-aware#15 - `compiler_builtins` may need to be built without debug assertions, see [this](https://github.com/rust-lang/rust/blob/8e917f48382c6afaf50568263b89d35fba5d98e4/src/bootstrap/bin/rustc.rs#L210-L214). Could maybe use profile overrides. - Panic issues: - `panic_abort` is not yet supported, though it should probably be easy. It could maybe look at the profile to determine which panic implementation to use? This requires more hard-coding in Cargo to know about rustc implementation details. - [This](https://github.com/rust-lang/rust/blob/8e917f48382c6afaf50568263b89d35fba5d98e4/src/bootstrap/bin/rustc.rs#L186-L201) should probably be handled where `panic` is set for `panic_abort` and `compiler_builtins`. I would like to get a test case for it. This can maybe be done with profile overrides? - Using two resolvers is quite messy and causes a lot of complications. It would be ideal if it could only use one, though that may not be possible for the foreseeable future. See rust-lang/wg-cargo-std-aware#12 - Features are hard-coded. See rust-lang/wg-cargo-std-aware#13 - Lots of various platform-specific support is not included (musl, wasi, windows-gnu, etc.). - Default `backtrace` is used with C compiler. See rust-lang/wg-cargo-std-aware#16 - Sanitizers are not built. See rust-lang/wg-cargo-std-aware#17 - proc_macro has some hacky code to synthesize its dependencies. See rust-lang/wg-cargo-std-aware#18. This may not be necessary if this uses `--sysroot` instead. - Profile overrides cause weird linker errors. That is: ```toml [profile.dev.overrides.std] opt-level = 2 ``` Using `[profile.dev.overrides."*"]` works. I tried fiddling with it, but couldn't figure it out. We may also want to consider altering the syntax for profile overrides. Having to repeat the same profile for `std` and `core` and `alloc` and everything else would not be ideal. - ~~`Context::unit_deps` does not handle build overrides, see #7215.~~ FIXED ## Questions for this PR - I went with the strategy of using `--extern` to link the standard lib. This seems to work, and I haven't found any problems, but it seems risky. It also forces Cargo to know about certain implicit dependencies like `compiler_builtins` and `panic_*`. The alternative is to create a sysroot and copy all the crates to that directory and pass `--sysroot`. However, this is complicated by pipelining, which would require special support to copy `.rmeta` files when they are generated. Let me know if you think I should use a different strategy. I'm on the fence here, and I think using `--sysroot` may be safer, but adds more complexity. - As an aside, if rustc ever tries to grab a crate from sysroot that was not passed in via `--extern`, then it results in duplicate lang items. For example, saying `extern crate proc_macro;` without specifying `proc_macro` as a dependency. We could prevent rustc from ever trying by passing `--sysroot=/nonexistent` to prevent it from trying. Or add an equivalent flag to rustc. - How should this be tested? I added a janky integration test, but it has some drawbacks. It requires internet access. It is slow. Since it is slow, it reuses the same target directory for multiple tests which makes it awkward to work with. - What interesting things are there to test? - We may want to disable the test before merging if it seems too annoying to make it the default. It requires rust-src to be downloaded, and takes several minutes to run, and are somewhat platform-dependent. - How to test that it is actually linking the correct standard library? I did tests locally with a modified libcore, but I can't think of a good way to do that in the test suite. - I did not add `__CARGO_DEFAULT_LIB_METADATA` to the hash. I had a hard time coming up with a test case where it would matter. - My only thought is that it is a problem because libstd includes a dylib, which prevents the hash from being added to the filename. It does cause recompiles when switching between compilers, for example, when it normally wouldn't. - Very dumb question: Why exactly does libstd include a dylib? This can cause issues (see rust-lang/rust#56443). - This should probably change, but I want to better understand it first. - The `bin_nostd` test needs to link libc on linux, and I'm not sure I understand why. I'm concerned there is something wrong there. libstd does not do that AFAIK.
ehuss
added
plan before stabilization
This issue is for working through the implementation issues for compiling
compiler_builtins.The
compiler_builtinscrate provides various intrinsics needed by the compiler. This presents some portability and compile-time issues when building your own standard library.There is the
cfeature of the crate that enables using a C compiler to build optimized versions of some intrinsics. This is the default for rustc.There is the
memfeature which demangles some memory functions, used in no-std builds of alloc. (rust-lang/rust#56825)It wants a different codegen-units: https://github.com/rust-lang/rust/blob/1a1cc050d8efc906ede39f444936ade1fdc9c6cb/library/Cargo.toml#L14-L25
I'm not sure exactly how this will fit in the std-aware Cargo story.
The text was updated successfully, but these errors were encountered: