Skip to content

Add chapter on libs and metadata. #1044

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Feb 4, 2021
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions src/SUMMARY.md
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,7 @@
- [Debugging LLVM](./backend/debugging.md)
- [Backend Agnostic Codegen](./backend/backend-agnostic.md)
- [Implicit Caller Location](./backend/implicit-caller-location.md)
- [Libraries and Metadata](./backend/libs-and-metadata.md)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps instead put this under "High-level compiler architecture"? That's where queries and serialization are discussed, which are somewhat related.

- [Profile-guided Optimization](./profile-guided-optimization.md)
- [LLVM Source-Based Code Coverage](./llvm-coverage-instrumentation.md)
- [Sanitizers Support](./sanitizers.md)
Expand Down
192 changes: 192 additions & 0 deletions src/backend/libs-and-metadata.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,192 @@
# Libraries and Metadata

When the compiler sees a reference to an external crate, it needs to load some
information about that crate. This chapter gives an overview of that process,
and the supported file formats for crate libraries.

## Libraries

A crate dependency can be loaded from an `rlib`, `dylib`, or `rmeta` file. A
key point of these file formats is that they contain `rustc`-specific
[*metadata*](#metadata). This metadata allows the compiler to discover enough
information about the external crate to understand the items it contains,
which macros it exports, and *much* more.

### rlib

An `rlib` is an [archive file], which is similar to a tar file. This file
format is specific to `rustc`, and may change over time. This file contains:

* Object code, which is the result of code generation. This is used during
regular linking. There is a separate `.o` file for each [codegen unit]. The
codegen step can be skipped with the [`-C
linker-plugin-lto`][linker-plugin-lto] CLI option, which means each `.o`
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does rustc reuse .o as the extension for bitcode files? I would expect .ll or something.

Copy link
Member

@camelid camelid Feb 4, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What do you mean? My understanding is this is talking about object code (not LLVM bitcode), so .o is correct.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, it confused me that your comment only extended until the line before :)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In the past, the bitcode was stored as a separate, compressed file in the archive (with the .bc.z extension). This was changed to simplify the implementation and to match the behavior of clang, which always places the bitcode in the .o file.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Kind of like .o for "output" :)

file will only contain LLVM bitcode.
* [LLVM bitcode], which is a binary representation of LLVM's intermediate
representation, which is embedded as a section in the `.o` files. This can
be used for [Link Time Optimization] (LTO). This can be removed with the
[`-C embed-bitcode=no`][embed-bitcode] CLI option to improve compile times
and reduce disk space if LTO is not needed.
* `rustc` [metadata], in a file named `lib.rmeta`.
* A symbol table, which is generally a list of symbols with offsets to the
object file that contain that symbol. This is pretty standard for archive
files.

[archive file]: https://en.wikipedia.org/wiki/Ar_(Unix)
[LLVM bitcode]: https://llvm.org/docs/BitCodeFormat.html
[Link Time Optimization]: https://llvm.org/docs/LinkTimeOptimization.html
[codegen unit]: ../backend/codegen.md
[embed-bitcode]: https://doc.rust-lang.org/rustc/codegen-options/index.html#embed-bitcode
[linker-plugin-lto]: https://doc.rust-lang.org/rustc/codegen-options/index.html#linker-plugin-lto

### dylib

A `dylib` is a platform-specific shared library. It includes the `rustc`
[metadata] in a special link section called `.rustc` in a compressed format.

### rmeta

An `rmeta` file is custom binary format that contains the [metadata] for the
crate. This file can be used for fast "checks" of a project by skipping all
code generation (as is done with `cargo check`), collecting enough information
for documentation (as is done with `cargo doc`), or for
[pipelining](#pipelining). This file is created if the
[`--emit=metadata`][emit] CLI option is used.

`rmeta` files do not support linking, since they do not contain compiled
object files.

[emit]: https://doc.rust-lang.org/rustc/command-line-arguments.html#option-emit

## Metadata

The metadata contains a wide swath of different elements. This guide will not
go into detail of every field it contains. You are encouraged to browse the
[`CrateRoot`] definition to get a sense of the different elements it contains.
Everything about metadata encoding and decoding is in the [`rustc_metadata`]
package.

Here are a few highlights of things it contains:

* The version of the `rustc` compiler. The compiler will refuse to load files
from any other version.
* The [Strict Version Hash](#strict-version-hash) (SVH). This helps ensure the
correct dependency is loaded.
* The [Crate Disambiguator](#crate-disambiguator). This is a hash used
to disambiguate between different crates of the same name.
* Information about all the source files in the library. This can be used for
a variety of things, such as diagnostics pointing to sources in a
dependency.
* Information about exported macros, traits, types, and items. Generally,
anything that's needed to be known when a path references something inside a
crate dependency.
* Encoded [MIR]. This is optional, and only encoded if needed for code
generation. `cargo check` skips this for performance reasons.

[`CrateRoot`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/rmeta/struct.CrateRoot.html
[`rustc_metadata`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/index.html
[MIR]: ../mir/index.md

### Strict Version Hash

The Strict Version Hash ([SVH], also known as the "crate hash") is a 64-bit
hash that is used to ensure that the correct crate dependencies are loaded. It
is possible for a directory to contain multiple copies of the same dependency
built with different settings, or built from different sources. The crate
loader will skip any crates that have the wrong SVH.

The SVH is also used for the [incremental compilation] session filename,
though that usage is mostly historic.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What's used instead? I'm not sure what you mean by historic here.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In the past, rustc would use the incremental directories of dependencies for resolving queries, and would use the SVH in the directory name to ensure it was querying the correct dependency. That was removed (it no longer looks for dependencies), and AFAIK it no longer looks at the SVH in the filename. I think it could be replaced with a random number and have the same effect (it uses a random number before it knows the svh, it gets renamed at the end).

The incremental naming convention is described at https://doc.rust-lang.org/nightly/nightly-rustc/rustc_incremental/persist/fs/index.html.


The hash includes a variety of elements:

* Hashes of the HIR nodes.
* All of the upstream crate hashes.
* All of the source filenames.
* Hashes of certain command-line flags (like `-C metadata` via the [Crate
Disambiguator](#crate-disambiguator), and all CLI options marked with
`[TRACKED]`).

See [`finalize_and_compute_crate_hash`] for where the hash is actually
computed.

[SVH]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_data_structures/svh/struct.Svh.html
[incremental compilation]: ../queries/incremental-compilation.md
[`finalize_and_compute_crate_hash`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_middle/hir/map/collector/struct.NodeCollector.html#method.finalize_and_compute_crate_hash

### Crate Disambiguator

The [`CrateDisambiguator`] is a 128-bit hash used to distinguish between
different crates of the same name. It is a hash of all the [`-C metadata`] CLI
options computed in [`compute_crate_disambiguator`]. It is used in a variety
of places, such as symbol name mangling, crate loading, and much more.

By default, all Rust symbols are mangled and incorporate the disambiguator
hash. This allows multiple versions of the same crate to be included together.
Cargo automatically generates `-C metadata` hashes based on a variety of
factors, like the package version, source, and the target kind (a lib and bin
can have the same crate name, so they need to be disambiguated).

[`CrateDisambiguator`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_ast/crate_disambiguator/struct.CrateDisambiguator.html
[`compute_crate_disambiguator`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_interface/util/fn.compute_crate_disambiguator.html
[`-C metadata`]: https://doc.rust-lang.org/rustc/codegen-options/index.html#metadata

## Crate loading

Crate loading can have quite a few subtle complexities. During [name
resolution], when an external crate is referenced (via an `extern crate` or
path), the resolver uses the [`CrateLoader`] which is responsible for finding
the crate libraries and loading the [metadata] for them. After the dependency
is loaded, the `CrateLoader` will provide the information the resolver needs
to perform its job (such as expanding macros, resolving paths, etc.).

To load each external crate, the `CrateLoader` uses a [`CrateLocator`] to
actually find the correct files for one specific crate. There is some great
documentation in the [`locator`] module that goes into detail on how loading
works, and I strongly suggest reading it to get the full picture.

The location of a dependency can come from several different places. Direct
dependencies are usually passed with `--extern` flags, and the loader can look
at those directly. Direct dependencies often have references to their own
dependencies, which need to be loaded, too. These are usually found by
scanning the directories passed with the `-L` flag for any file whose metadata
contains a matching crate name and [SVH](#strict-version-hash). The loader
will also look at the [sysroot] to find dependencies.

As crates are loaded, they are kept in the [`CStore`] with the crate metadata
wrapped in the [`CrateMetadata`] struct. After resolution and expansion, the
`CStore` will make its way into the [`GlobalCtxt`] for the rest of
compilation.

[name resolution]: ../name-resolution.md
[`CrateLoader`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/creader/struct.CrateLoader.html
[`CrateLocator`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/locator/struct.CrateLocator.html
[`locator`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/locator/index.html
[`CStore`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/creader/struct.CStore.html
[`CrateMetadata`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_metadata/rmeta/decoder/struct.CrateMetadata.html
[`GlobalCtxt`]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_middle/ty/struct.GlobalCtxt.html
[sysroot]: ../building/bootstrapping.md#what-is-a-sysroot

## Pipelining

One trick to improve compile times is to start building a crate as soon as the
metadata for its dependencies is available. For a library, there is no need to
wait for the code generation of dependencies to finish. Cargo implements this
technique by telling `rustc` to emit an [`rmeta`](#rmeta) file for each
dependency as well as an [`rlib`](#rlib). As early as it can, `rustc` will
save the `rmeta` file to disk before it continues to the code generation
phase. The compiler sends a JSON message to let the build tool know that it
can start building the next crate if possible.

The [crate loading](#crate-loading) system is smart enough to know when it
sees an `rmeta` file to use that if the `rlib` is not there (or has only been
partially written).

This pipelining isn't possible for binaries, because the linking phase will
require the code generation of all its dependencies. In the future, it may be
possible to further improve this scenario by splitting linking into a separate
command (see [#64191]).

[#64191]: https://github.com/rust-lang/rust/issues/64191

[metadata]: #metadata