Skip to main content

Tips for Reading Code

I’ve been trying to get better at reading code. It’s an important skill. Too often people neglect this skill, struggle to read other people’s code, and then accuse the code of being poorly written. Sometimes the code is poorly written, but, as Joel Spolsky points out, sometimes the problem is just inherently messy. Not to mention, the source code is the most up to date documentation. If you’re having trouble using a library, reading the source code can clear up a lot of issues.

I’m not amazing at reading code, but here’s my tips:

  • Use a good IDE

Yes, yes, your emacs config is amazing. I know. I’m typing this in emacs right this second. But the truth is that a good IDE is way more functional than an editor, unless that editor is VSCode. You could spend hours setting up your config to use language servers and file navigators and whatever, but an IDE gets you all of the way without any work.

From there, learn to use the editor. Learn the keyboard shortcuts for goto definition, and go back from goto definition. Learn to search in a file and search across directories. If you can, watch someone who is experienced with the IDE.

A good IDE can really help with chasing down a particular execution path. If I’m reading the TypeScript compiler codebase and I come across getContextuallyTypedParameterType that in turn calls isContextSensitiveFunctionOrObjectLiteralMethod (yeesh), I can goto the definition and see that isContextSensitiveFunctionOrObjectLiteralMethod calls isInJSFile, isFunctionDeclaration, isFunctionExpressionOrArrowFunction, etc.

  • Embrace grep/ripgrep and find/fd

While an IDE is great, sometimes it’s better to use a simpler tool. Two of the tools that I use the most are ripgrep and fd. They’re both modern replacements for old school command line tools, ripgrep replaces grep and fd replaces find. I find the newer tools a lot more friendly but the old school ones are great too. Ripgrep searches for a string or regular expression in your files and fd searches for a string or regex in your file names.

These are really great for when you need to dive into a big codebase and get to some really specific code. For instance, let’s say I’m trying to figure out how Istio loads WASM files. I don’t know Istio and I don’t really know Go, the implementation langauge. However, I can load up the Istio codebase and use fd to see if there are any files with “wasm” in their name:

pilot/pkg/model/wasm.go
pilot/pkg/networking/core/v1alpha3/extension/wasmplugin.go
pilot/pkg/networking/core/v1alpha3/extension/wasmplugin_test.go
pkg/test/datasets/validation/dataset/extensions-v1alpha1-WasmPlugin-invalid.yaml
pkg/test/datasets/validation/dataset/extensions-v1alpha1-WasmPlugin-valid.yaml
pkg/wasm
releasenotes/notes/reliable-wasm-remote-load.yaml
releasenotes/notes/wasm-extension-dashboard.yaml
releasenotes/notes/wasmplugin.yaml
tests/integration/telemetry/stats/prometheus/api/stats_wasm_filter_test.go
tests/integration/telemetry/stats/prometheus/api/stats_wasm_tcp_filter_test.go
tests/integration/telemetry/stats/prometheus/wasm
tests/integration/telemetry/stats/prometheus/wasm/bad_wasm_filter_test.go
tests/integration/telemetry/stats/prometheus/wasm/stats_wasm_filter_test.go
tests/integration/telemetry/stats/prometheus/wasm/stats_wasm_tcp_filter_test.go

Okay there’s a few but not too many. It seems like the files in releasenotes/ and tests/ are not that important. Likewise the yaml files don’t see that important. Let’s remove those:

pilot/pkg/model/wasm.go
pilot/pkg/networking/core/v1alpha3/extension/wasmplugin.go
pilot/pkg/networking/core/v1alpha3/extension/wasmplugin_test.go
pkg/wasm

This leaves us three files and one directory. At this point I’d probably figure out what pilot/ and pkg/ mean by reading the documentation and the code. My intuition is that pkg/wasm will have the meat of it because it seems to be an actual package dedicated to WASM.

Of course you could do this in an IDE, but I happen to like using these tools.

  • Find the meat of the code

If you’re diving into a new codebase, sometimes it can be hard to figure out where the action is happening. Too often you’re confronted with a bunch of folders, most of which are rarely touched or hold configuration, documentation, testing, etc. These files can create a lot of noise and mess with the aforementioned tools. For instance, in the prior example we had to filter out all the releasenotes/ and tests/ directory.

Usually there’s one or two directories that hold the meat of the project. In smaller projects that’ll often be a src/ directory. In bigger projects you may have a packages or lib directory that stores the libraries that make up the project. Often this will depend on the language. In Rust it would not be too weird to have a crates/ directory, for instance. Another way to find the meat of the code is to look at previous pull requests and see where the majority of the changes occur.

  • Don’t get bogged down in the weeds

If you’re tackling a big codebase, it can be tempting to just start reading. Maybe that works for some people but it doesn’t work for me. Realistically you’re not going to be able to read a few thousand line codebase and understand it in time to complete your task. You’re gonna have to rely on imperfect knowledge. That’s why I recommend using searching techniques to drill down into the relevant parts of the codebase. If you are going to read the code, try to find the central entrypoint. This isn’t always true, but usually there is some function or object that is the main do-er.

For instance, let’s say I’m trying to read the Rust name resolver codebase. I start in the lib.rs and start reading from the top:

enum Weak {
    Yes,
    No,
}

#[derive(Copy, Clone, PartialEq, Debug)]
pub enum Determinacy {
    Determined,
    Undetermined,
}

impl Determinacy {
    fn determined(determined: bool) -> Determinacy {
        if determined { Determinacy::Determined } else { Determinacy::Undetermined }
    }
}

/// A specific scope in which a name can be looked up.
/// This enum is currently used only for early resolution (imports and macros),
/// but not for late resolution yet.
#[derive(Clone, Copy)]
enum Scope<'a> {
    DeriveHelpers(LocalExpnId),
    DeriveHelpersCompat,
    MacroRules(MacroRulesScopeRef<'a>),
    CrateRoot,
    // The node ID is for reporting the `PROC_MACRO_DERIVE_RESOLUTION_FALLBACK`
    // lint if it should be reported.
    Module(Module<'a>, Option<NodeId>),
    RegisteredAttrs,
    MacroUsePrelude,
    BuiltinAttrs,
    ExternPrelude,
    ToolPrelude,
    StdLibPrelude,
    BuiltinTypes,
}

/// Names from different contexts may want to visit different subsets of all specific scopes
/// with different restrictions when looking up the resolution.
/// This enum is currently used only for early resolution (imports and macros),
/// but not for late resolution yet.
#[derive(Clone, Copy)]
enum ScopeSet<'a> {
    /// All scopes with the given namespace.
    All(Namespace, /*is_import*/ bool),
    /// Crate root, then extern prelude (used for mixed 2015-2018 mode in macros).
    AbsolutePath(Namespace),
    /// All scopes with macro namespace and the given macro kind restriction.
    Macro(MacroKind),
    /// All scopes with the given namespace, used for partially performing late resolution.
    /// The node id enables lints and is used for reporting them.
    Late(Namespace, Module<'a>, Option<NodeId>),
}

Bored and confused already? I am. I have some idea of what this code does, just by virtue of having worked on compilers, but it’s still very overwhelming. Instead, lets skip down, scanning for something that seems important…

/// The main resolver class.
///
/// This is the visitor that walks the whole crate.
pub struct Resolver<'a> {
    session: &'a Session,

    definitions: Definitions,

    graph_root: Module<'a>,

    prelude: Option<Module<'a>>,
    extern_prelude: FxHashMap<Ident, ExternPreludeEntry<'a>>,

    /// N.B., this is used only for better diagnostics, not name resolution itself.
    has_self: FxHashSet<DefId>,

    /// Names of fields of an item `DefId` accessible with dot syntax.
    /// Used for hints during error reporting.
    field_names: FxHashMap<DefId, Vec<Spanned<Symbol>>>,

    /// All imports known to succeed or fail.
    determined_imports: Vec<&'a Import<'a>>,

    /// All non-determined imports.
    indeterminate_imports: Vec<&'a Import<'a>>,

    /// FIXME: Refactor things so that these fields are passed through arguments and not resolver.
    /// We are resolving a last import segment during import validation.
    last_import_segment: bool,
    /// This binding should be ignored during in-module resolution, so that we don't get
    /// "self-confirming" import resolutions during import validation.
    unusable_binding: Option<&'a NameBinding<'a>>,
    

Oh hey this seems important! It’s the Resolver struct in rustc_resolve and indeed the comment indicates that it’s the main class. But geez there’s a lot of fields. Let’s skip them too. Now that we know the main class is called Resolver, we can look for its methods by searching for an impl block for Resolver. Oh hey we found it:

impl<'a> Resolver<'a> {
    pub fn new(
        session: &'a Session,
        krate: &Crate,
        crate_name: &str,
        metadata_loader: Box<MetadataLoaderDyn>,
        arenas: &'a ResolverArenas<'a>,
    ) -> Resolver<'a> {
        let root_def_id = CRATE_DEF_ID.to_def_id();
        let mut module_map = FxHashMap::default();
        let graph_root = arenas.new_module(
            None,
            ModuleKind::Def(DefKind::Mod, root_def_id, kw::Empty),
            ExpnId::root(),
            krate.span,
            session.contains_name(&krate.attrs, sym::no_implicit_prelude),
            &mut module_map,
        );
        let empty_module = arenas.new_module(
            None,
            ModuleKind::Def(DefKind::Mod, root_def_id, kw::Empty),
            ExpnId::root(),
            DUMMY_SP,
            true,
            &mut FxHashMap::default(),
        );

Again, lots of stuff here. At this point we have our pick of methods that we can trace. I like the look of resolve_crate:

    /// Entry point to crate resolution.
    pub fn resolve_crate(&mut self, krate: &Crate) {
        self.session.time("resolve_crate", || {
            self.session.time("finalize_imports", || ImportResolver { r: self }.finalize_imports());
            self.session.time("finalize_macro_resolutions", || self.finalize_macro_resolutions());
            self.session.time("late_resolve_crate", || self.late_resolve_crate(krate));
            self.session.time("resolve_main", || self.resolve_main());
            self.session.time("resolve_check_unused", || self.check_unused(krate));
            self.session.time("resolve_report_errors", || self.report_errors(krate));
            self.session.time("resolve_postprocess", || self.crate_loader.postprocess(krate));
        });
    }

The comment indicates it’s an entrypoint and it calls a bunch of stuff. At this point I’d use my IDE to goto definition for each of these functions and start reading. If I encounter stuff that requires understanding the code we skipped before, only then will I go back and read that code.

  • Take notes

Taking notes can help a lot with organizing your brain. I’m not the greatest note taker so I can’t give too much advice. Some people like typing notes on a computer, others enjoy writing in a notebook. I’ve started to enjoy writing in a notebook, albeit more for designing my own code rather than reading others’. You can take notes on which functions do what, on the overall architecture, etc. Just make sure you’re not spending time taking notes just for the sake of taking notes. Too often in school we learn to take notes because our teacher says to take notes. These notes should be genuinely new pieces of knowledge that you want to remember.

  • Build the code and break something

You should try to build the code pretty soon after you start reading the codebase. It’s almost impossible to learn a codebase solely by reading it. You need to execute code and see what happens. You can stick print statements into the codebase and see what they print. You can try to add a feature and see how it breaks.

A good, mature codebase should have tests and static analysis that prevents you from doing anything too stupid. Often times you can learn a lot just by seeing what tests your change breaks. If your codebase doesn’t have sufficient testing, you can still learn a lot by showing your code to someone who does understand the codebase and having them give you feedback. Of course, this person should be tactful and nice in their feedback.

It may feel really uncomfortable contributing to a codebase where you don’t understand a significant portion. That’s completely understandable and truth be told it’s a sign that you’re accurately assessing the risks of modern, large scale software development. It’s hard to make and maintain big code bases. That’s (partially) why we get paid a lot. But you have to do your best. Besides, at a certain point it’s not necessarily your fault if the company lets people contribute bad or broken code to a codebase. Sure, you should always aim to write good code, but at scale, that’s really hard. The lead developers have to set up reasonable precautions and safety measures such as testing, QA, fuzzing, verification, code review, etc. if they want any assurance that the codebase will stay functional and well maintained.

And my final piece of advice:

  • Ask the original author or maintainer

You should do all of the above, but really, asking the person who wrote the code is your best bet. You can use git blame to figure out who wrote the code. Hopefully they’re still at the company, at which point you can schedule a meeting or DM them. If they’re not at the company, they may still respond to a polite email if they’re nice. However, I wouldn’t abuse this option unless you are explicitly being mentored/supervised by this person, they have the bandwidth to attend to you, or you absolutely cannot figure this out on your own. Likely they have their own tasks, so if you can get away with figuring stuff out on your own, that’s probably better for both their productivity and your skills as a programmer. That said, if you politely reach out and they’re mean to you, don’t let that deter you! That’s their problem and not yours. Hopefully the next person will be more accomodating.