#llvm #safe-wrapper #pass #safe #wrapper

macro llvm-plugin-macros

Proc macros for llvm-plugin

2 unstable releases

0.2.0 Oct 1, 2022
0.1.0 Sep 7, 2022

#62 in #pass


Used in llvm-plugin

Apache-2.0

14KB
76 lines

llvm-plugin-rs

version doc linux windows macos

This crate gives the ability to safely implement passes for the new LLVM pass manager, by leveraging the strongly typed interface provided by Inkwell.

If you have never developed LLVM passes before, you can take a look at the available examples. They will (hopefully) give you a better idea of how to use this crate.

If you want a deeper understanding of the many concepts surrounding the new LLVM pass manager, you should read the official LLVM documentation.

Usage

When importing this crate in your Cargo.toml, you will need to specify the LLVM version to use with a corresponding feature flag:

[dependencies]
llvm-plugin = { version = "0.2", features = ["llvm10-0"] }

Supported versions:

LLVM Version Cargo Feature Flag Linux Windows MacOS
10.0.x llvm10-0
11.0.x llvm11-0
12.0.x llvm12-0
13.0.x llvm13-0
14.0.x llvm14-0

Getting Started

An LLVM plugin is merely a dylib that is given a PassBuilder by the LLVM tool (e.g. opt, lld) loading it. Therefore, you must add the following line in your Cargo.toml:

[lib]
crate-type = ["cdylib"]

A PassBuilder allows registering callbacks on specific actions being performed by the LLVM tool.

For instance, the --passes parameter of opt allows specifying a custom pass pipeline to be run on a given IR module. A plugin could therefore register a callback for parsing an element of the given pipeline (e.g. a pass name), in order to insert a custom pass to run by opt.

The following code illustrates the idea:

use llvm_plugin::inkwell::module::Module;
use llvm_plugin::{
    LlvmModulePass, ModuleAnalysisManager, PassBuilder, PipelineParsing, PreservedAnalyses,
};

// A name and version is required.
#[llvm_plugin::plugin(name = "plugin_name", version = "0.1")]
fn plugin_registrar(builder: &mut PassBuilder) {
    // Add a callback to parse a name from the textual representation of
    // the pipeline to be run.
    builder.add_module_pipeline_parsing_callback(|name, manager| {
        if name == "custom-pass" {
            // the input pipeline contains the name "custom-pass",
            // so we add our custom pass to the pass manager
            manager.add_pass(CustomPass);

            // we notify the caller that we were able to parse
            // the given name
            PipelineParsing::Parsed
        } else {
            // in any other cases, we notify the caller that our
            // callback wasn't able to parse the given name
            PipelineParsing::NotParsed
        }
    });
}

struct CustomPass;
impl LlvmModulePass for CustomPass {
    fn run_pass(
        &self,
        module: &mut Module,
        manager: &ModuleAnalysisManager
    ) -> PreservedAnalyses {
        // transform the IR
        todo!()
    }
}

Now, executing this command would run our custom pass on some input module.bc:

opt --load-pass-plugin=libplugin.so --passes=custom-pass module.bc -disable-output

However, executing this command would not (custom-pass2 cannot be parsed by our plugin):

opt --load-pass-plugin=libplugin.so --passes=custom-pass2 module.bc -disable-output

More callbacks are available, read the documentation for more details.

To learn more about how to sequentially apply more than one pass, read this opt guide.

Linux & MacOS Requirements

Your LLVM toolchain should dynamically link the LLVM library. Fortunately, this is the case for toolchains distributed on apt and homebrew registeries.

If you are not in this case, you have to compile LLVM from sources by specifying the LLVM_LINK_LLVM_DYLIB=ON cmake flag.

Compile LLVM-14
$ wget https://github.com/llvm/llvm-project/releases/download/llvmorg-14.0.0/llvm-14.0.0.src.tar.xz
$ tar xf llvm-14.0.0.src.tar.xz && cd llvm-14.0.0.src
$ mkdir build && cd build
$ cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_INSTALL_PREFIX="$HOME/llvm" \
    -DLLVM_LINK_LLVM_DYLIB=ON \
    -G Ninja
$ ninja install

Windows Requirements

You have to compile LLVM from sources in any case, because you need to apply some patches to the LLVM code base before compiling. Then, you need to specify the LLVM_EXPORT_SYMBOLS_FOR_PLUGINS=ON cmake flag while leaving the LLVM_TARGETS_TO_BUILD flag to its default value.

Compile LLVM-14
$ wget https://github.com/llvm/llvm-project/releases/download/llvmorg-14.0.0/llvm-14.0.0.src.tar.xz
$ tar xf llvm-14.0.0.src.tar.xz && cd llvm-14.0.0.src
$ cat ../ci/windows/llvm-14.patch | patch -p1
$ mkdir build && cd build
$ cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_INSTALL_PREFIX="C:\\llvm" \
    -DLLVM_EXPORT_SYMBOLS_FOR_PLUGINS=ON \
    -G Ninja
$ ninja install
$ cp lib/opt.lib /c/llvm/lib

Don't forget to update your PATH environment variable with your LLVM installation path.

Missing Features

  • Support for loop passes (Inkwell doesn't currently provide safe wrappers)
  • Support for CGSCC passes (Inkwell doesn't currently provide safe wrappers)
  • FFI over the full manager proxy API (only a subset is currently implemented)
  • FFI over the full analysis invalidation API (only a subset is currently implemented)
  • FFI over builtin LLVM analyses (e.g. dominator tree)

Contributions are very welcome, make sure to check out the Contributing Guide first!

Dependencies

~1.5MB
~34K SLoC