home | about | talks | publications

BYO Standard: An Explorer’s Guide to Compiler Plugins

July, 2018

The Rust programming language provides powerful guarantees around memory and thread safety. It also exposes all the knobs required for implementing custom rules, enabling a project to make additional guarantees and enforce opinions on best practice. Embedded standards are very opinionated about software practices—like using floating point values as loop counters or the number of possible exit points of a function—and Rust’s defaults don’t prevent every runtime panic or potential to crash (for example, recursion that goes too deep and overflows the stack).

For PolySync, a runtime panic means the potential for an unsafe situation on the road, and with that in mind, we’ve explored ways to restrict that potential. Of course, we aren’t the only ones thinking about ways to improve the quality of code at compile time by enforcing the right rules for the job. Active projects like rust-clippy are working to do that too by providing lints to supplement the rustc defaults.

In this post we’ll explore how to enforce a rule by prohibiting a practice we’ve formed an opinion about, the indexing of a vector or an array. Preventing indexing represents the potential to eliminate runtime panics from stepping out of bounds.

Lets get started.

Choosing Your Lint Level

Lint levels allow you to configure the severity of a rule violation with the help of rustc.

One size doesn’t fit all and these decisions should be tailored to your project’s specific requirements.

A quick warning about “forbid”: it should be used sparingly. Forbidding everything can prove troublesome, for instance, forbidding calls to unwrap. While it may be feasible to disallow it in your own code, it’s not uncommon for macros exposed by external crates to hide calls to unwrap. Rather than finding another crate, or rewriting it internally to avoid all calls to unwrap, it might be acceptable to implement warning or deny policies that require an explicit allow.

Exploring the AST

There are at least two options for exploring a Rust program’s AST. The -Z ast-json flag passed to rustc helps with getting a feel for the general look and structure of the AST. Using it looks like this:

rustc src/main -Z ast-json

This post uses print statements as they provide context about when the data we’re looking for is accessible (i.e. is EarlyLintPass or LateLintPass right for me?). Because the syntax::ast types implement Debug, they also provided type information that can be a little more parsable.

First, make sure you’re using the nightly toolchain:

rust default nightly

Now setup a clean linting crate, here named customlints:

cargo init customlints --lib

Next, we’ll do some setup in the customlints/src/lib.rs file.

#![feature(plugin_registrar)]
#![feature(box_syntax, rustc_private)]
#![feature(macro_vis_matcher)]

#[macro_use]
extern crate rustc;
extern crate rustc_plugin;
extern crate syntax;
extern crate syntax_pos;

use rustc::hir;
use rustc::lint::{
    EarlyContext, EarlyLintPassObject, 
    LateContext, LateLintPassObject,
    LintArray, LintContext, LintPass,
};
use rustc_plugin::Registry;
use syntax::ast;

struct EarlyPass;
struct LatePass;

impl LintPass for LatePass {
    fn get_lints(&self) -> LintArray {
        // We'll get to this later, kind of...
        lint_array!() 
    }
}

impl LintPass for EarlyPass {
    fn get_lints(&self) -> LintArray {
        // We'll definitely get to this later!
        lint_array!() 
    }
}

#[plugin_registrar]
pub fn register_plugins(reg: &mut Registry) {
    reg.register_early_lint_pass(
        box EarlyPass as EarlyLintPassObject);
    reg.register_late_lint_pass(
        box LatePass as LateLintPassObject);
}

Now, we can implement rustc::lint::EarlyLintPass and rustc::lint::LateLintPass for some preliminary examination.

impl rustc::lint::EarlyLintPass 
for EarlyPass {
   fn check_expr(
       &mut self, 
       cx: &EarlyContext, 
       expr: &ast::Expr) {
       println!("Early pass, expression: {:?}", expr);
   }
}

impl<'a, 'tcx> rustc::lint::LateLintPass<'a, 'tcx> 
for LatePass {
   fn check_expr(
       &mut self, 
       cx: &LateContext<'a, 'tcx>, 
       expr: &'tcx hir::Expr) {
       println!("Late pass, expression: {:?}", expr);
   }
}

In order to incorporate this exploratory plugin, we can create another crate, a [[bin]] this time:

cargo init example --bin

Then point to our linter in that project’s Cargo.toml. (Note the optional = true , we’ll revisit that later.):

[dependencies.customlints]
path = "/path/to/customlints"
optional = true

Next, we can fill in the main function:

#![cfg_attr(feature="customlints", feature(plugin))]
#![cfg_attr(feature="customlints", plugin(customlints))]

fn main() {
   // Initialize a vector containing a single element.
   let x = vec![0;1];
   // Attempt to access the vector's 10th element.
   // This is what we want to prohibit!
   let _a = x[10];
}

Time to build. Note that the dependency on the customlints crate is optional, so it needs to be enabled in the build command:

cargo build --features "customlints"

Looking closely, we can see what looks like our indexing operation in our output. Here’s one:

Early pass, expression: expr(13: x[10])

And later, another:

Late pass, expression: expr(13: x[10])

Now, we can modify our print statements in order to unpack the hir::Expr a bit more:

println!("Early pass, expression node: {:?}", expr.node);

To minimize noise, we can comment out our late pass output:

// println!("Late pass, expression node: {:?}", expr.node);

After building our [[bin]] project again, there’s some information that looks promising, the following Index type that contains x and 10:

Early pass, expression node: Index(expr(11: x), expr(12: 10))

Enforcing a Lint

Using the compiler plugin docs and the docs for syntax::ast let’s lint the Index type. First, we’ll need a few modifications. We’ll begin by declaring the lint with a Deny qualification. This means a program that indexes will fail to compile, but if necessary, can be allowed with #![allow(indexing_lint)].

declare_lint!(INDEXING_LINT, Deny, "Deny indexing operations.");

Then, we’ll populate that empty lint_array in the impl LintPass for EarlyPass that we mentioned earlier.

impl LintPass for EarlyPass {
   fn get_lints(&self) -> LintArray {
       lint_array!(INDEXING_LINT)
   }
}

Finally, we can define when our lint occurs (any occurance of the Index type) and the report it provides. Replace that EarlyPass print statement as follows (feel free to remove the references to LatePass, we’re only going to use EarlyPass from here on out):

impl rustc::lint::EarlyLintPass 
for EarlyPass {
   fn check_expr(
       &mut self, 
       cx: &EarlyContext, 
       expr: &ast::Expr) {
       if let ast::ExprKind::Index(_ , _) = expr.node {
           cx.span_lint(
               INDEXING_LINT, 
               expr.span, 
               "Indexing operations disallowed!");
       }
   }
}

After building our [[bin]] project again, tada!

error: Indexing operations disallowed!
--> src/main.rs:6:14
  |
6 |     let _a = x[10];
  |              ^^^^^
  |
 = note: #[deny(indexing_lint)] on by default
error: aborting due to previous error

We’ve implemented a lint.

Fine Tuning

Now, in an attempt to eliminate unintended side effects, we need to stress the edges (or even just look for them). For instance, maybe since let _b = &x[..] won’t cause a runtime panic we decide to allow its behavior. Let’s add that to our [[bin]] project and build it.

Sure enough, we’re denying behavior we’ve decided we don’t want to. By putting our print statement back we can take another look at some debugging output (informed by the syntax::ast::ExprKind).

impl rustc::lint::EarlyLintPass 
for EarlyPass {
   fn check_expr(
       &mut self, 
       cx: &EarlyContext, 
       expr: &ast::Expr) {
       if let ast::ExprKind::Index(_, ref e) = 
       expr.node {
           println!(
               "Early pass, expression node: {:?}", 
               e.node);
           // cx.span_lint(
           //    INDEXING_LINT, 
           //    expr.span, 
           //    "Indexing operations disallowed!");
       }
   }
}

That provides the following:

Early pass, expression node: Lit(Spanned { node: Int(10, Unsuffixed), \
span: Span { lo: BytePos(167), hi: BytePos(169), ctxt: #0 } })

Early pass, expression node: Range(None, None, HalfOpen)

But what do we do with that? One option is experimenting with our [[bin]] project to see if we can get some more context. This can be achieved with a few more indexing operations.

Here’s the main function.

fn main() {
   let x = vec![0;1];
   let _a = x[10];
   let _b = &x[10..];
   let _c = &x[..10];
   let _d = &x[10..100];
   let _e = &x[..];
}

Now let’s look at the output and see what we can tell.

Because of the Int(10, Unsuffixed), it looks like it corresponds to let _a = x[10];:

Early pass, expression node: Lit(Spanned { node: Int(10, Unsuffixed), \
span: Span { lo: BytePos(167), hi: BytePos(169), ctxt: #0 } })

Assuming the output occurs in the same order as the indexing operations in the source code, &x[..] probably corresponds to:

Early pass, expression node: Range(None, None, HalfOpen)

and &x[10..] to:

Early pass, expression node: Range(Some(expr(23: 10)), None, HalfOpen)

and &x[..10] to:

Early pass, expression node: Range(None, Some(expr(30: 10)), HalfOpen)

and &x[10..100] to:

Early pass, expression node: Range(Some(expr(37: 10)), \
Some(expr(38: 100)), HalfOpen)

The first two values in that Range type correspond to Some if there is a position defined and None if there isn’t. We still need to disallow cases where there is potential for out-of-bounds access (all but the last case), so the next step ends up looking like the following:

impl rustc::lint::EarlyLintPass 
for EarlyPass {
   fn check_expr(
       &mut self, 
       cx: &EarlyContext, 
       expr: &ast::Expr) {
       if let ast::ExprKind::Index(_, ref e) =
       expr.node {
           match e.node {
               ast::ExprKind::Range(
                   None, None, _
                   ) => (),  // allow &[..]
               _ => {
                   cx.span_lint(
                       INDEXING_LINT, 
                       expr.span,
                       "Indexing operations disallowed.");
               }
           }
       }
   }
}

And that did it! All that’s left are the errors for the behavior we decided to deny.

Continuing the Work

This lint introduces the potential for some false positives. If it’s possible to prove at compile time that an index is in range, then our error may be overstepping its utility. Taking this lint to the next level likely means allowing indexing in those cases. It may even be feasible to implement a MIR processing lint that triggers on all reachable panics. If that existed, a lint like ours could be phased out entirely.

How a Stable Product Uses Unstable Lints

Nightly Rust should be backward compatible with the last stable release. That means a stable project’s build success can depend on nightly toolchain linting without having to worry about something unrelated breaking. Revisiting the setup we did for our [[bin]] example, we see that it allows us to opt out of linting on a stable toolchain.

First, the customlints dependency needs be optional in your Cargo.toml:

[dependencies.customlints]
path = "/<path>/<to>/customlints"
optional = true

Then, any project that’s subject to your custom lints can use cfg_attr to ensure the linting plugin isn’t enabled unless the customlints feature is specified:

[#![cfg_attr(feature="customlints", feature(plugin))]
#![cfg_attr(feature="customlints", plugin(customlints))]

Using that set up, a nightly build can be invoked with:

cargo build --features "customlints"

and a stable build with:

cargo build

Additional Reading

It can be helpful to start this kind of exploratory effort by looking at other projects that also implement linting compiler plugins. Some of the places that have a lot to offer in that area are:

There are other powerful applications of compiler plugins too. llogiq maintains a couple of my favorites, mutagen and overflower. Mutagen is the foundation of a mutation testing framework that enables you to evaluate whether your tests work as they should. Overflower lets you decide what to do when integer overflow occurs.

Additionally, there is the possibility that others in the community will think your lints are useful. If you think that’s the case, consider running them by the rust-clippy team.

While implementing lints can be an illuminating window into the compiler and a Rust program’s AST, compile time lints are a powerful approach to restricting program behavior. Because if a program with violations doesn’t compile, it can’t crash.

[A version of this post was originally published at polysync.io/blog.]