1
Fork 0

Rollup merge of #100398 - nnethercote:improve-Zhir-stats, r=michaelwoerister

Improve `-Zhir-stats`

Add testing, improve coverage, avoid some double counting, and add more detail.

r? `@michaelwoerister`
This commit is contained in:
Dylan DPC 2022-08-11 22:47:06 +05:30 committed by GitHub
commit 7ecc8923db
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 447 additions and 54 deletions

View file

@ -21,21 +21,55 @@ enum Id {
None,
}
struct NodeData {
struct NodeStats {
count: usize,
size: usize,
}
impl NodeStats {
fn new() -> NodeStats {
NodeStats { count: 0, size: 0 }
}
}
struct Node {
stats: NodeStats,
subnodes: FxHashMap<&'static str, NodeStats>,
}
impl Node {
fn new() -> Node {
Node { stats: NodeStats::new(), subnodes: FxHashMap::default() }
}
}
/// This type measures the size of AST and HIR nodes, by implementing the AST
/// and HIR `Visitor` traits. But we don't measure every visited type because
/// that could cause double counting.
///
/// For example, `ast::Visitor` has `visit_ident`, but `Ident`s are always
/// stored inline within other AST nodes, so we don't implement `visit_ident`
/// here. In constrast, we do implement `visit_expr` because `ast::Expr` is
/// always stored as `P<ast::Expr>`, and every such expression should be
/// measured separately.
///
/// In general, a `visit_foo` method should be implemented here if the
/// corresponding `Foo` type is always stored on its own, e.g.: `P<Foo>`,
/// `Box<Foo>`, `Vec<Foo>`, `Box<[Foo]>`.
///
/// There are some types in the AST and HIR tree that the visitors do not have
/// a `visit_*` method for, and so we cannot measure these, which is
/// unfortunate.
struct StatCollector<'k> {
krate: Option<Map<'k>>,
data: FxHashMap<&'static str, NodeData>,
nodes: FxHashMap<&'static str, Node>,
seen: FxHashSet<Id>,
}
pub fn print_hir_stats(tcx: TyCtxt<'_>) {
let mut collector = StatCollector {
krate: Some(tcx.hir()),
data: FxHashMap::default(),
nodes: FxHashMap::default(),
seen: FxHashSet::default(),
};
tcx.hir().walk_toplevel_module(&mut collector);
@ -44,49 +78,88 @@ pub fn print_hir_stats(tcx: TyCtxt<'_>) {
}
pub fn print_ast_stats(krate: &ast::Crate, title: &str) {
use rustc_ast::visit::Visitor;
let mut collector =
StatCollector { krate: None, data: FxHashMap::default(), seen: FxHashSet::default() };
ast_visit::walk_crate(&mut collector, krate);
StatCollector { krate: None, nodes: FxHashMap::default(), seen: FxHashSet::default() };
collector.visit_crate(krate);
collector.print(title);
}
impl<'k> StatCollector<'k> {
fn record<T>(&mut self, label: &'static str, id: Id, node: &T) {
// Record a top-level node.
fn record<T>(&mut self, label: &'static str, id: Id, val: &T) {
self.record_inner(label, None, id, val);
}
// Record a two-level entry, with a top-level enum type and a variant.
fn record_variant<T>(&mut self, label1: &'static str, label2: &'static str, id: Id, val: &T) {
self.record_inner(label1, Some(label2), id, val);
}
fn record_inner<T>(
&mut self,
label1: &'static str,
label2: Option<&'static str>,
id: Id,
val: &T,
) {
if id != Id::None && !self.seen.insert(id) {
return;
}
let entry = self.data.entry(label).or_insert(NodeData { count: 0, size: 0 });
let node = self.nodes.entry(label1).or_insert(Node::new());
node.stats.count += 1;
node.stats.size = std::mem::size_of_val(val);
entry.count += 1;
entry.size = std::mem::size_of_val(node);
if let Some(label2) = label2 {
let subnode = node.subnodes.entry(label2).or_insert(NodeStats::new());
subnode.count += 1;
subnode.size = std::mem::size_of_val(val);
}
}
fn print(&self, title: &str) {
let mut stats: Vec<_> = self.data.iter().collect();
let mut nodes: Vec<_> = self.nodes.iter().collect();
nodes.sort_by_key(|&(_, ref node)| node.stats.count * node.stats.size);
stats.sort_by_key(|&(_, ref d)| d.count * d.size);
let mut total_size = 0;
let total_size = nodes.iter().map(|(_, node)| node.stats.count * node.stats.size).sum();
eprintln!("\n{}\n", title);
eprintln!("{:<18}{:>18}{:>14}{:>14}", "Name", "Accumulated Size", "Count", "Item Size");
eprintln!("----------------------------------------------------------------");
for (label, data) in stats {
eprintln!(
"{:<18}{:>18}{:>14}{:>14}",
label,
to_readable_str(data.count * data.size),
to_readable_str(data.count),
to_readable_str(data.size)
);
let percent = |m, n| (m * 100) as f64 / n as f64;
total_size += data.count * data.size;
for (label, node) in nodes {
let size = node.stats.count * node.stats.size;
eprintln!(
"{:<18}{:>10} ({:4.1}%){:>14}{:>14}",
label,
to_readable_str(size),
percent(size, total_size),
to_readable_str(node.stats.count),
to_readable_str(node.stats.size)
);
if !node.subnodes.is_empty() {
let mut subnodes: Vec<_> = node.subnodes.iter().collect();
subnodes.sort_by_key(|&(_, ref subnode)| subnode.count * subnode.size);
for (label, subnode) in subnodes {
let size = subnode.count * subnode.size;
eprintln!(
"- {:<18}{:>10} ({:4.1}%){:>14}",
label,
to_readable_str(size),
percent(size, total_size),
to_readable_str(subnode.count),
);
}
}
}
eprintln!("----------------------------------------------------------------");
eprintln!("{:<18}{:>18}\n", "Total", to_readable_str(total_size));
eprintln!("{:<18}{:>10}\n", "Total", to_readable_str(total_size));
}
}
@ -228,6 +301,10 @@ impl<'v> hir_visit::Visitor<'v> for StatCollector<'v> {
hir_visit::walk_path(self, path)
}
// `PathSegment` has one inline use (in `ast::ExprKind::MethodCall`) and
// one non-inline use (in `Path::segments`). The latter case is more common
// than the former case, so we implement this visitor and tolerate the
// double counting in the former case.
fn visit_path_segment(&mut self, path_span: Span, path_segment: &'v hir::PathSegment<'v>) {
self.record("PathSegment", Id::None, path_segment);
hir_visit::walk_path_segment(self, path_span, path_segment)
@ -243,14 +320,54 @@ impl<'v> hir_visit::Visitor<'v> for StatCollector<'v> {
}
}
// Used to avoid boilerplate for types with many variants.
macro_rules! record_variants {
(
($self:ident, $val:expr, $kind:expr, $ty:ty, $tykind:ident), // mandatory pieces
[$($variant:ident),*]
) => {
match $kind {
$(
ast::$tykind::$variant { .. } => {
$self.record_variant(stringify!($ty), stringify!($variant), Id::None, $val)
}
)*
}
};
}
impl<'v> ast_visit::Visitor<'v> for StatCollector<'v> {
fn visit_foreign_item(&mut self, i: &'v ast::ForeignItem) {
self.record("ForeignItem", Id::None, i);
record_variants!(
(self, i, i.kind, ForeignItem, ForeignItemKind),
[Static, Fn, TyAlias, MacCall]
);
ast_visit::walk_foreign_item(self, i)
}
fn visit_item(&mut self, i: &'v ast::Item) {
self.record("Item", Id::None, i);
record_variants!(
(self, i, i.kind, Item, ItemKind),
[
ExternCrate,
Use,
Static,
Const,
Fn,
Mod,
ForeignMod,
GlobalAsm,
TyAlias,
Enum,
Struct,
Union,
Trait,
TraitAlias,
Impl,
MacCall,
MacroDef
]
);
ast_visit::walk_item(self, i)
}
@ -265,47 +382,116 @@ impl<'v> ast_visit::Visitor<'v> for StatCollector<'v> {
}
fn visit_stmt(&mut self, s: &'v ast::Stmt) {
self.record("Stmt", Id::None, s);
record_variants!(
(self, s, s.kind, Stmt, StmtKind),
[Local, Item, Expr, Semi, Empty, MacCall]
);
ast_visit::walk_stmt(self, s)
}
fn visit_param(&mut self, p: &'v ast::Param) {
self.record("Param", Id::None, p);
ast_visit::walk_param(self, p)
}
fn visit_arm(&mut self, a: &'v ast::Arm) {
self.record("Arm", Id::None, a);
ast_visit::walk_arm(self, a)
}
fn visit_pat(&mut self, p: &'v ast::Pat) {
self.record("Pat", Id::None, p);
record_variants!(
(self, p, p.kind, Pat, PatKind),
[
Wild,
Ident,
Struct,
TupleStruct,
Or,
Path,
Tuple,
Box,
Ref,
Lit,
Range,
Slice,
Rest,
Paren,
MacCall
]
);
ast_visit::walk_pat(self, p)
}
fn visit_expr(&mut self, ex: &'v ast::Expr) {
self.record("Expr", Id::None, ex);
ast_visit::walk_expr(self, ex)
fn visit_expr(&mut self, e: &'v ast::Expr) {
record_variants!(
(self, e, e.kind, Expr, ExprKind),
[
Box, Array, ConstBlock, Call, MethodCall, Tup, Binary, Unary, Lit, Cast, Type, Let,
If, While, ForLoop, Loop, Match, Closure, Block, Async, Await, TryBlock, Assign,
AssignOp, Field, Index, Range, Underscore, Path, AddrOf, Break, Continue, Ret,
InlineAsm, MacCall, Struct, Repeat, Paren, Try, Yield, Yeet, Err
]
);
ast_visit::walk_expr(self, e)
}
fn visit_ty(&mut self, t: &'v ast::Ty) {
self.record("Ty", Id::None, t);
record_variants!(
(self, t, t.kind, Ty, TyKind),
[
Slice,
Array,
Ptr,
Rptr,
BareFn,
Never,
Tup,
Path,
TraitObject,
ImplTrait,
Paren,
Typeof,
Infer,
ImplicitSelf,
MacCall,
Err,
CVarArgs
]
);
ast_visit::walk_ty(self, t)
}
fn visit_generic_param(&mut self, g: &'v ast::GenericParam) {
self.record("GenericParam", Id::None, g);
ast_visit::walk_generic_param(self, g)
}
fn visit_where_predicate(&mut self, p: &'v ast::WherePredicate) {
record_variants!(
(self, p, p, WherePredicate, WherePredicate),
[BoundPredicate, RegionPredicate, EqPredicate]
);
ast_visit::walk_where_predicate(self, p)
}
fn visit_fn(&mut self, fk: ast_visit::FnKind<'v>, s: Span, _: NodeId) {
self.record("FnDecl", Id::None, fk.decl());
ast_visit::walk_fn(self, fk, s)
}
fn visit_assoc_item(&mut self, item: &'v ast::AssocItem, ctxt: ast_visit::AssocCtxt) {
let label = match ctxt {
ast_visit::AssocCtxt::Trait => "TraitItem",
ast_visit::AssocCtxt::Impl => "ImplItem",
};
self.record(label, Id::None, item);
ast_visit::walk_assoc_item(self, item, ctxt);
fn visit_assoc_item(&mut self, i: &'v ast::AssocItem, ctxt: ast_visit::AssocCtxt) {
record_variants!(
(self, i, i.kind, AssocItem, AssocItemKind),
[Const, Fn, TyAlias, MacCall]
);
ast_visit::walk_assoc_item(self, i, ctxt);
}
fn visit_param_bound(&mut self, bounds: &'v ast::GenericBound, _ctxt: BoundKind) {
self.record("GenericBound", Id::None, bounds);
ast_visit::walk_param_bound(self, bounds)
fn visit_param_bound(&mut self, b: &'v ast::GenericBound, _ctxt: BoundKind) {
record_variants!((self, b, b, GenericBound, GenericBound), [Trait, Outlives]);
ast_visit::walk_param_bound(self, b)
}
fn visit_field_def(&mut self, s: &'v ast::FieldDef) {
@ -318,27 +504,42 @@ impl<'v> ast_visit::Visitor<'v> for StatCollector<'v> {
ast_visit::walk_variant(self, v)
}
fn visit_lifetime(&mut self, lifetime: &'v ast::Lifetime, _: ast_visit::LifetimeCtxt) {
self.record("Lifetime", Id::None, lifetime);
ast_visit::walk_lifetime(self, lifetime)
}
fn visit_mac_call(&mut self, mac: &'v ast::MacCall) {
self.record("MacCall", Id::None, mac);
ast_visit::walk_mac(self, mac)
}
// `UseTree` has one inline use (in `ast::ItemKind::Use`) and one
// non-inline use (in `ast::UseTreeKind::Nested). The former case is more
// common, so we don't implement `visit_use_tree` and tolerate the missed
// coverage in the latter case.
fn visit_path_segment(&mut self, path_span: Span, path_segment: &'v ast::PathSegment) {
self.record("PathSegment", Id::None, path_segment);
ast_visit::walk_path_segment(self, path_span, path_segment)
}
fn visit_assoc_constraint(&mut self, constraint: &'v ast::AssocConstraint) {
self.record("AssocConstraint", Id::None, constraint);
ast_visit::walk_assoc_constraint(self, constraint)
// `GenericArgs` has one inline use (in `ast::AssocConstraint::gen_args`) and one
// non-inline use (in `ast::PathSegment::args`). The latter case is more
// common, so we implement `visit_generic_args` and tolerate the double
// counting in the former case.
fn visit_generic_args(&mut self, sp: Span, g: &'v ast::GenericArgs) {
record_variants!((self, g, g, GenericArgs, GenericArgs), [AngleBracketed, Parenthesized]);
ast_visit::walk_generic_args(self, sp, g)
}
fn visit_attribute(&mut self, attr: &'v ast::Attribute) {
self.record("Attribute", Id::None, attr);
record_variants!((self, attr, attr.kind, Attribute, AttrKind), [Normal, DocComment]);
ast_visit::walk_attribute(self, attr)
}
fn visit_expr_field(&mut self, f: &'v ast::ExprField) {
self.record("ExprField", Id::None, f);
ast_visit::walk_expr_field(self, f)
}
fn visit_crate(&mut self, krate: &'v ast::Crate) {
self.record("Crate", Id::None, krate);
ast_visit::walk_crate(self, krate)
}
fn visit_inline_asm(&mut self, asm: &'v ast::InlineAsm) {
self.record("InlineAsm", Id::None, asm);
ast_visit::walk_inline_asm(self, asm)
}
}