Rust build scripts vs. Meson

- meson, rust

One of the pain points in trying to make the Meson build system work with Rust and Cargo is Cargo's use of build scripts, i.e. the build.rs that many Rust programs use for doing things before the main build. This post is about my exploration of what build.rs does.

Thanks to Nirbheek Chauhan for his comments and additions to a draft of this article!

TL;DR: build.rs is pretty ad-hoc and somewhat primitive, when compared to Meson's very nice, high-level patterns for build-time things.

I have the intuition that giving names to the things that are usually done in build.rs scripts, and creating abstractions for them, can make it easier later to implement those abstractions in terms of Meson. Maybe we can eliminate build.rs in most cases? Maybe Cargo can acquire higher-level concepts that plug well to Meson?

(That is... I think we can refactor our way out of this mess.)

What does build.rs do?

The first paragraph in the documentation for Cargo build scripts tells us this:

Some packages need to compile third-party non-Rust code, for example C libraries. Other packages need to link to C libraries which can either be located on the system or possibly need to be built from source. Others still need facilities for functionality such as code generation before building (think parser generators).

That is,

  • Compiling third-party non-Rust code. For example, maybe there is a C sub-library that the Rust crate needs.

  • Link to C libraries... located on the system... or built from source. For example, in gtk-rs, the sys crates link to libgtk-3.so, libcairo.so, etc. and need to find a way to locate those libraries with pkg-config.

  • Code generation. In the C world this could be generating a parser with yacc; in the Rust world there are many utilities to generate code that is later used in your actual program.

In the next sections I'll look briefly at each of these cases, but in a different order.

Code generation

Here is an example, in how librsvg generates code for a couple of things that get autogenerated before compiling the main library:

  • A perfect hash function (PHF) of attributes and CSS property names.
  • A pair of lookup tables for SRGB linearization and un-linearization.

For example, this is main() in build.rs:

fn main() {
    generate_phf_of_svg_attributes();
    generate_srgb_tables();
}

And this is the first few lines of of the first function:

fn generate_phf_of_svg_attributes() {
    let path = Path::new(&env::var("OUT_DIR").unwrap()).join("attributes-codegen.rs");
    let mut file = BufWriter::new(File::create(&path).unwrap());

    writeln!(&mut file, "#[repr(C)]").unwrap();

    // ... etc
}

Generate a path like $OUT_DIR/attributes-codegen.rs, create a file with that name, a BufWriter for the file, and start outputting code to it.

Similarly, the second function:

fn generate_srgb_tables() {
    let linearize_table = compute_table(linearize);
    let unlinearize_table = compute_table(unlinearize);

    let path = Path::new(&env::var("OUT_DIR").unwrap()).join("srgb-codegen.rs");
    let mut file = BufWriter::new(File::create(&path).unwrap());

    // ...

    print_table(&mut file, "LINEARIZE", &linearize_table);
    print_table(&mut file, "UNLINEARIZE", &unlinearize_table);
}

Compute two lookup tables, create a file named $OUT_DIR/srgb-codegen.rs, and write the lookup tables to the file.

Later in the actual librsvg code, the generated files get included into the source code using the include! macro. For example, here is where attributes-codegen.rs gets included:

// attributes.rs

extern crate phf;  // crate for perfect hash function

// the generated file has the declaration for enum Attribute
include!(concat!(env!("OUT_DIR"), "/attributes-codegen.rs"));

One thing to note here is that the generated source files (attributes-codegen.rs, srgb-codegen.rs) get put in $OUT_DIR, a directory that Cargo creates for the compilation artifacts. The files do not get put into the original source directories with the rest of the library's code; the idea is to keep the source directories pristine.

At least in those terms, Meson and Cargo agree that source directories should be kept clean of autogenerated files.

The Code Generation section of Cargo's documentation agrees:

In general, build scripts should not modify any files outside of OUT_DIR. It may seem fine on the first blush, but it does cause problems when you use such crate as a dependency, because there's an implicit invariant that sources in .cargo/registry should be immutable. cargo won't allow such scripts when packaging.

Now, some things to note here:

  • Both the build.rs program and the actual library sources look at the $OUT_DIR environment variable for the location of the generated sources.

  • The Cargo docs say that if the code generator needs input files, it can look for them based on its current directory, which will be the toplevel of your source package (i.e. your toplevel Cargo.toml).

Meson hates this scheme of things. In particular, Meson is very systematic about where it finds input files and sources, and where things like code generators are allowed to place their output.

The way Meson communicates these paths to code generators is via command-line arguments to "custom targets". Here is an example that is easier to read than the documentation:

gen = find_program('generator.py')

outputs = custom_target('generated',
  output : ['foo.h', 'foo.c'],
  command : [gen, '@OUTDIR@'],
  ...
)

This defines a target named 'generated', which will use the generator.py program to output two files, foo.h and foo.c. That Python program will get called with @OUTDIR@ as a command-line argument; in effect, meson will call /full/path/to/generator.py @OUTDIR@ explicitly, without any magic passed through environment variables.

If this looks similar to what Cargo does above with build.rs, it's because it is similar. It's just that Meson gives a name to the concept of generating code at build time (Meson's name for this is a custom target), and provides a mechanism to say which program is the generator, which files it is expected to generate, and how to call the program with appropriate arguments to put files in the right place.

In contrast, Cargo assumes that all of that information can be inferred from an environment variable.

In addition, if the custom target takes other files as input (say, so it can call yacc my-grammar.y), the custom_target() command can take an input: argument. This way, Meson can add a dependency on those input files, so that the appropriate things will be rebuilt if the input files change.

Now, Cargo could very well provide a small utility crate that build scripts could use to figure out all that information. Meson would tell Cargo to use its scheme of things, and pass it down to build scripts via that utility crate. I.e. to have

// build.rs

extern crate cargo_high_level;

let output = Path::new(cargo_high_level::get_output_path()).join("codegen.rs");
//                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ this, instead of:

let output = Path::new(&env::var("OUT_DIR").unwrap()).join("codegen.rs");

// let the build system know about generated dependencies
cargo_high_level::add_output(output);

A similar mechanism could be used for the way Meson likes to pass command-line arguments to the programs that deal with custom targets.

Linking to C libraries on the system

Some Rust crates need to link to lower-level C libraries that actually do the work. For example, in gtk-rs, there are high-level binding crates called gtk, gdk, cairo, etc. These use low-level crates called gtk-sys, gdk-sys, cairo-sys. Those -sys crates are just direct wrappers on top of the C functions of the respective system libraries: gtk-sys makes almost every function in libgtk-3.so available as a Rust-callable function.

System libraries sometimes live in a well-known part of the filesystem (/usr/lib64, for example); other times, like in Windows and MacOS, they could be anywhere. To find that location plus other related metadata (include paths for C header files, library version), many system libraries use pkg-config. At the highest level, one can run pkg-config on the command line, or from build scripts, to query some things about libraries. For example:

# what's the system's installed version of GTK?
$ pkg-config --modversion gtk+-3.0
3.24.4

# what compiler flags would a C compiler need for GTK?
$ pkg-config --cflags gtk+-3.0
-pthread -I/usr/include/gtk-3.0 -I/usr/include/at-spi2-atk/2.0 -I/usr/include/at-spi-2.0 -I/usr/include/dbus-1.0 -I/usr/lib64/dbus-1.0/include -I/usr/include/gtk-3.0 -I/usr/include/gio-unix-2.0/ -I/usr/include/libxkbcommon -I/usr/include/wayland -I/usr/include/cairo -I/usr/include/pango-1.0 -I/usr/include/harfbuzz -I/usr/include/pango-1.0 -I/usr/include/fribidi -I/usr/include/atk-1.0 -I/usr/include/cairo -I/usr/include/pixman-1 -I/usr/include/freetype2 -I/usr/include/libdrm -I/usr/include/libpng16 -I/usr/include/gdk-pixbuf-2.0 -I/usr/include/libmount -I/usr/include/blkid -I/usr/include/uuid -I/usr/include/glib-2.0 -I/usr/lib64/glib-2.0/include

# and which libraries?
$ pkg-config --libs gtk+-3.0
-lgtk-3 -lgdk-3 -lpangocairo-1.0 -lpango-1.0 -latk-1.0 -lcairo-gobject -lcairo -lgdk_pixbuf-2.0 -lgio-2.0 -lgobject-2.0 -lglib-2.0

There is a pkg-config crate which build.rs can use to call this, and communicate that information to Cargo. The example in the crate's documentation is for asking pkg-config for the foo package, with version at least 1.2.3:

extern crate pkg_config;

fn main() {
    pkg_config::Config::new().atleast_version("1.2.3").probe("foo").unwrap();
}

And the documentation says,

After running pkg-config all appropriate Cargo metadata will be printed on stdout if the search was successful.

Wait, what?

Indeed, printing specially-formated stuff on stdout is how build.rs scripts communicate back to Cargo about their findings. To quote Cargo's docs on build scripts; the following is talking about the stdout of build.rs:

Any line that starts with cargo: is interpreted directly by Cargo. This line must be of the form cargo:key=value, like the examples below:

# specially recognized by Cargo
cargo:rustc-link-lib=static=foo
cargo:rustc-link-search=native=/path/to/foo
cargo:rustc-cfg=foo
cargo:rustc-env=FOO=bar
# arbitrary user-defined metadata
cargo:root=/path/to/foo
cargo:libdir=/path/to/foo/lib
cargo:include=/path/to/foo/include

One can use the stdout of a build.rs program to add additional command-line options for rustc, or set environment variables for it, or add library paths, or specific libraries.

Meson hates this scheme of things. I suppose it would prefer to do the pkg-config calls itself, and then pass that information down to Cargo, you guessed it, via command-line options or something well-defined like that. Again, the example cargo_high_level crate I proposed above could be used to communicate this information from Meson to Cargo scripts. Meson also doesn't like this because it would prefer to know about pkg-config-based libraries in a declarative fashion, without having to run a random script like build.rs.

Building C code from Rust

Finally, some Rust crates build a bit of C code and then link that into the compiled Rust code. I have no experience with that, but the respective build scripts generally use the cc crate to call a C compiler and pass options to it conveniently. I suppose Meson would prefer to do this instead, or at least to have a high-level way of passing down information to Cargo.

In effect, Meson has to be in charge of picking the C compiler. Having the thing-to-be-built pick on its own has caused big problems in the past: GObject-Introspection made the same mistake years ago when it decided to use distutils to detect the C compiler; gtk-doc did as well. When those tools are used, we still deal with problems with cross-compilation and when the system has more than one C compiler in it.

Snarky comments about the Unix philosophy

If part of the Unix philosophy is that shit can be glued together with environment variables and stringly-typed stdout... it's a pretty bad philosophy. All the cases above boil down to having a well-defined, more or less strongly-typed way to pass information between programs instead of shaking proverbial tree of the filesystem and the environment and seeing if something usable falls down.

Would we really have to modify all build.rs scripts for this?

Probably. Why not? Meson already has a lot of very well-structured knowledge of how to deal with multi-platform compilation and installation. Re-creating this knowledge in ad-hoc ways in build.rs is not very pleasant or maintainable.

Related work