Skip to content

Fix lint perf regressions #105485

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
Dec 12, 2022
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
149 changes: 97 additions & 52 deletions compiler/rustc_lint/src/early.rs
Original file line number Diff line number Diff line change
Expand Up @@ -20,23 +20,23 @@ use rustc_ast::ptr::P;
use rustc_ast::visit::{self as ast_visit, Visitor};
use rustc_ast::{self as ast, walk_list, HasAttrs};
use rustc_middle::ty::RegisteredTools;
use rustc_session::lint::{BufferedEarlyLint, LintBuffer};
use rustc_session::lint::{BufferedEarlyLint, LintBuffer, LintPass};
use rustc_session::Session;
use rustc_span::symbol::Ident;
use rustc_span::Span;

macro_rules! run_early_passes { ($cx:expr, $f:ident, $($args:expr),*) => ({
for pass in $cx.passes.iter_mut() {
pass.$f(&$cx.context, $($args),*);
}
macro_rules! lint_callback { ($cx:expr, $f:ident, $($args:expr),*) => ({
$cx.pass.$f(&$cx.context, $($args),*);
}) }

pub struct EarlyContextAndPasses<'a> {
/// Implements the AST traversal for early lint passes. `T` provides the the
/// `check_*` methods.
pub struct EarlyContextAndPass<'a, T: EarlyLintPass> {
context: EarlyContext<'a>,
passes: Vec<EarlyLintPassObject>,
pass: T,
}

impl<'a> EarlyContextAndPasses<'a> {
impl<'a, T: EarlyLintPass> EarlyContextAndPass<'a, T> {
// This always-inlined function is for the hot call site.
#[inline(always)]
fn inlined_check_id(&mut self, id: ast::NodeId) {
Expand Down Expand Up @@ -70,27 +70,27 @@ impl<'a> EarlyContextAndPasses<'a> {

self.inlined_check_id(id);
debug!("early context: enter_attrs({:?})", attrs);
run_early_passes!(self, enter_lint_attrs, attrs);
lint_callback!(self, enter_lint_attrs, attrs);
f(self);
debug!("early context: exit_attrs({:?})", attrs);
run_early_passes!(self, exit_lint_attrs, attrs);
lint_callback!(self, exit_lint_attrs, attrs);
self.context.builder.pop(push);
}
}

impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
impl<'a, T: EarlyLintPass> ast_visit::Visitor<'a> for EarlyContextAndPass<'a, T> {
fn visit_param(&mut self, param: &'a ast::Param) {
self.with_lint_attrs(param.id, &param.attrs, |cx| {
run_early_passes!(cx, check_param, param);
lint_callback!(cx, check_param, param);
ast_visit::walk_param(cx, param);
});
}

fn visit_item(&mut self, it: &'a ast::Item) {
self.with_lint_attrs(it.id, &it.attrs, |cx| {
run_early_passes!(cx, check_item, it);
lint_callback!(cx, check_item, it);
ast_visit::walk_item(cx, it);
run_early_passes!(cx, check_item_post, it);
lint_callback!(cx, check_item_post, it);
})
}

Expand All @@ -101,10 +101,10 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
}

fn visit_pat(&mut self, p: &'a ast::Pat) {
run_early_passes!(self, check_pat, p);
lint_callback!(self, check_pat, p);
self.check_id(p.id);
ast_visit::walk_pat(self, p);
run_early_passes!(self, check_pat_post, p);
lint_callback!(self, check_pat_post, p);
}

fn visit_pat_field(&mut self, field: &'a ast::PatField) {
Expand All @@ -120,7 +120,7 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {

fn visit_expr(&mut self, e: &'a ast::Expr) {
self.with_lint_attrs(e.id, &e.attrs, |cx| {
run_early_passes!(cx, check_expr, e);
lint_callback!(cx, check_expr, e);
ast_visit::walk_expr(cx, e);
})
}
Expand All @@ -141,7 +141,7 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
// Note that statements get their attributes from
// the AST struct that they wrap (e.g. an item)
self.with_lint_attrs(s.id, s.attrs(), |cx| {
run_early_passes!(cx, check_stmt, s);
lint_callback!(cx, check_stmt, s);
cx.check_id(s.id);
});
// The visitor for the AST struct wrapped
Expand All @@ -152,7 +152,7 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
}

fn visit_fn(&mut self, fk: ast_visit::FnKind<'a>, span: Span, id: ast::NodeId) {
run_early_passes!(self, check_fn, fk, span, id);
lint_callback!(self, check_fn, fk, span, id);
self.check_id(id);
ast_visit::walk_fn(self, fk);

Expand Down Expand Up @@ -180,37 +180,37 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {

fn visit_variant(&mut self, v: &'a ast::Variant) {
self.with_lint_attrs(v.id, &v.attrs, |cx| {
run_early_passes!(cx, check_variant, v);
lint_callback!(cx, check_variant, v);
ast_visit::walk_variant(cx, v);
})
}

fn visit_ty(&mut self, t: &'a ast::Ty) {
run_early_passes!(self, check_ty, t);
lint_callback!(self, check_ty, t);
self.check_id(t.id);
ast_visit::walk_ty(self, t);
}

fn visit_ident(&mut self, ident: Ident) {
run_early_passes!(self, check_ident, ident);
lint_callback!(self, check_ident, ident);
}

fn visit_local(&mut self, l: &'a ast::Local) {
self.with_lint_attrs(l.id, &l.attrs, |cx| {
run_early_passes!(cx, check_local, l);
lint_callback!(cx, check_local, l);
ast_visit::walk_local(cx, l);
})
}

fn visit_block(&mut self, b: &'a ast::Block) {
run_early_passes!(self, check_block, b);
lint_callback!(self, check_block, b);
self.check_id(b.id);
ast_visit::walk_block(self, b);
}

fn visit_arm(&mut self, a: &'a ast::Arm) {
self.with_lint_attrs(a.id, &a.attrs, |cx| {
run_early_passes!(cx, check_arm, a);
lint_callback!(cx, check_arm, a);
ast_visit::walk_arm(cx, a);
})
}
Expand All @@ -229,19 +229,19 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
}

fn visit_generic_arg(&mut self, arg: &'a ast::GenericArg) {
run_early_passes!(self, check_generic_arg, arg);
lint_callback!(self, check_generic_arg, arg);
ast_visit::walk_generic_arg(self, arg);
}

fn visit_generic_param(&mut self, param: &'a ast::GenericParam) {
self.with_lint_attrs(param.id, &param.attrs, |cx| {
run_early_passes!(cx, check_generic_param, param);
lint_callback!(cx, check_generic_param, param);
ast_visit::walk_generic_param(cx, param);
});
}

fn visit_generics(&mut self, g: &'a ast::Generics) {
run_early_passes!(self, check_generics, g);
lint_callback!(self, check_generics, g);
ast_visit::walk_generics(self, g);
}

Expand All @@ -250,18 +250,18 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
}

fn visit_poly_trait_ref(&mut self, t: &'a ast::PolyTraitRef) {
run_early_passes!(self, check_poly_trait_ref, t);
lint_callback!(self, check_poly_trait_ref, t);
ast_visit::walk_poly_trait_ref(self, t);
}

fn visit_assoc_item(&mut self, item: &'a ast::AssocItem, ctxt: ast_visit::AssocCtxt) {
self.with_lint_attrs(item.id, &item.attrs, |cx| match ctxt {
ast_visit::AssocCtxt::Trait => {
run_early_passes!(cx, check_trait_item, item);
lint_callback!(cx, check_trait_item, item);
ast_visit::walk_assoc_item(cx, item, ctxt);
}
ast_visit::AssocCtxt::Impl => {
run_early_passes!(cx, check_impl_item, item);
lint_callback!(cx, check_impl_item, item);
ast_visit::walk_assoc_item(cx, item, ctxt);
}
});
Expand All @@ -282,28 +282,57 @@ impl<'a> ast_visit::Visitor<'a> for EarlyContextAndPasses<'a> {
}

fn visit_attribute(&mut self, attr: &'a ast::Attribute) {
run_early_passes!(self, check_attribute, attr);
lint_callback!(self, check_attribute, attr);
}

fn visit_mac_def(&mut self, mac: &'a ast::MacroDef, id: ast::NodeId) {
run_early_passes!(self, check_mac_def, mac);
lint_callback!(self, check_mac_def, mac);
self.check_id(id);
}

fn visit_mac_call(&mut self, mac: &'a ast::MacCall) {
run_early_passes!(self, check_mac, mac);
lint_callback!(self, check_mac, mac);
ast_visit::walk_mac(self, mac);
}
}

// Combines multiple lint passes into a single pass, at runtime. Each
// `check_foo` method in `$methods` within this pass simply calls `check_foo`
// once per `$pass`. Compare with `declare_combined_early_lint_pass`, which is
// similar, but combines lint passes at compile time.
struct RuntimeCombinedEarlyLintPass<'a> {
passes: &'a mut [EarlyLintPassObject],
}
Comment on lines +303 to +305
Copy link
Contributor

Choose a reason for hiding this comment

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

Instezad of having separate branches for the builtin-only vs builtin+others, can something like this work?

Suggested change
struct RuntimeCombinedEarlyLintPass<'a> {
passes: &'a mut [EarlyLintPassObject],
}
struct RuntimeCombinedEarlyLintPass<'a, T: EarlyLintPass> {
builtin: T,
passes: &'a mut [EarlyLintPassObject],
}

This would require changing the macros, but we may have both the static dispatch for builtin lints and limit the complexity.

Copy link
Contributor Author

@nnethercote nnethercote Dec 9, 2022

Choose a reason for hiding this comment

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

That won't help. Under the original structure, we had builtin_lints and passes, but passes is always for empty normal compilation. (It's non-empty when bootstrapping, because rustc has some rustc-specific lints, and it's also non-empty for clippy.)

I then combined builtin_lints into passes. This meant that even in the normal builtins-only cases, the code had to iterate over the passes slice for every check_* call. This is what caused the slowdown. It took me a while to work out because the iterator methods are inlined so it wasn't that obvious.

With this new design, which is very similar to the original design, if passes is empty then we avoid all iteration over the slice. This gives the speed wins.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Pseudocode may make it clearer. Simplifying greatly, it's basically the difference between this:

for node in ast {
    builtin_lints.check_node(node);
}

and this:

let passes = vec![builtin_lints]
for node in ast {
    for pass in passes {
        pass.check_node(node);
    }
}


#[allow(rustc::lint_pass_impl_without_macro)]
impl LintPass for RuntimeCombinedEarlyLintPass<'_> {
fn name(&self) -> &'static str {
panic!()
}
}

macro_rules! impl_early_lint_pass {
([], [$($(#[$attr:meta])* fn $f:ident($($param:ident: $arg:ty),*);)*]) => (
impl EarlyLintPass for RuntimeCombinedEarlyLintPass<'_> {
$(fn $f(&mut self, context: &EarlyContext<'_>, $($param: $arg),*) {
for pass in self.passes.iter_mut() {
pass.$f(context, $($param),*);
}
})*
}
)
}
Copy link
Contributor

Choose a reason for hiding this comment

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

Do you have an explanation why this commit has an effect on performance? Is this a matter of code size vs indirection vs inlining?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

As above: the last commit is the one that regains the performance, because it separates the builtin_lints handling from the passes handling, avoiding all slice iteration when passes is empty. The second last commit is preparatory work.

(And I forgot to mention that the first five commits are repeated from #105416, sorry.)


crate::early_lint_methods!(impl_early_lint_pass, []);

/// Early lints work on different nodes - either on the crate root, or on freshly loaded modules.
/// This trait generalizes over those nodes.
pub trait EarlyCheckNode<'a>: Copy {
fn id(self) -> ast::NodeId;
fn attrs<'b>(self) -> &'b [ast::Attribute]
where
'a: 'b;
fn check<'b>(self, cx: &mut EarlyContextAndPasses<'b>)
fn check<'b, T: EarlyLintPass>(self, cx: &mut EarlyContextAndPass<'b, T>)
where
'a: 'b;
}
Expand All @@ -318,13 +347,13 @@ impl<'a> EarlyCheckNode<'a> for &'a ast::Crate {
{
&self.attrs
}
fn check<'b>(self, cx: &mut EarlyContextAndPasses<'b>)
fn check<'b, T: EarlyLintPass>(self, cx: &mut EarlyContextAndPass<'b, T>)
where
'a: 'b,
{
run_early_passes!(cx, check_crate, self);
lint_callback!(cx, check_crate, self);
ast_visit::walk_crate(cx, self);
run_early_passes!(cx, check_crate_post, self);
lint_callback!(cx, check_crate_post, self);
}
}

Expand All @@ -338,7 +367,7 @@ impl<'a> EarlyCheckNode<'a> for (ast::NodeId, &'a [ast::Attribute], &'a [P<ast::
{
self.1
}
fn check<'b>(self, cx: &mut EarlyContextAndPasses<'b>)
fn check<'b, T: EarlyLintPass>(self, cx: &mut EarlyContextAndPass<'b, T>)
where
'a: 'b,
{
Expand All @@ -356,21 +385,37 @@ pub fn check_ast_node<'a>(
builtin_lints: impl EarlyLintPass + 'static,
check_node: impl EarlyCheckNode<'a>,
) {
let context = EarlyContext::new(
sess,
!pre_expansion,
lint_store,
registered_tools,
lint_buffer.unwrap_or_default(),
);

// Note: `passes` is often empty. In that case, it's faster to run
// `builtin_lints` directly rather than bundling it up into the
// `RuntimeCombinedEarlyLintPass`.
let passes =
if pre_expansion { &lint_store.pre_expansion_passes } else { &lint_store.early_passes };
let mut passes: Vec<EarlyLintPassObject> = passes.iter().map(|p| (p)()).collect();
passes.push(Box::new(builtin_lints));

let mut cx = EarlyContextAndPasses {
context: EarlyContext::new(
sess,
!pre_expansion,
lint_store,
registered_tools,
lint_buffer.unwrap_or_default(),
),
passes,
};
if passes.is_empty() {
check_ast_node_inner(sess, check_node, context, builtin_lints);
} else {
let mut passes: Vec<_> = passes.iter().map(|mk_pass| (mk_pass)()).collect();
passes.push(Box::new(builtin_lints));
let pass = RuntimeCombinedEarlyLintPass { passes: &mut passes[..] };
check_ast_node_inner(sess, check_node, context, pass);
}
}

pub fn check_ast_node_inner<'a, T: EarlyLintPass>(
sess: &Session,
check_node: impl EarlyCheckNode<'a>,
context: EarlyContext<'_>,
pass: T,
) {
let mut cx = EarlyContextAndPass { context, pass };

cx.with_lint_attrs(check_node.id(), check_node.attrs(), |cx| check_node.check(cx));

// All of the buffered lints should have been emitted at this point.
Expand Down
Loading