Remove MatcherPosHandle.

This type was a small performance win for `html5ever`, which uses a
macro with hundreds of very simple rules that don't contain any
metavariables. But this type is complicated (extra lifetimes) and
perf-neutral for macros that do have metavariables.

This commit removes `MatcherPosHandle`, simplifying things a lot. This
increases the allocation rate for `html5ever` and similar cases a bit,
but makes things easier for follow-up changes that will improve
performance more than what we lost here.
This commit is contained in:
Nicholas Nethercote 2022-03-19 08:56:24 +11:00
parent 10644e0789
commit cedb787f6e
2 changed files with 23 additions and 88 deletions

View File

@ -1,5 +1,6 @@
#![feature(associated_type_bounds)]
#![feature(associated_type_defaults)]
#![feature(box_syntax)]
#![feature(crate_visibility_modifier)]
#![feature(decl_macro)]
#![feature(if_let_guard)]

View File

@ -89,7 +89,6 @@ use rustc_span::symbol::Ident;
use std::borrow::Cow;
use std::collections::hash_map::Entry::{Occupied, Vacant};
use std::mem;
use std::ops::{Deref, DerefMut};
// To avoid costly uniqueness checks, we require that `MatchSeq` always has a nonempty body.
@ -136,24 +135,8 @@ type NamedMatchVec = SmallVec<[NamedMatch; 4]>;
/// Represents a single "position" (aka "matcher position", aka "item"), as
/// described in the module documentation.
///
/// Here:
///
/// - `'root` represents the lifetime of the stack slot that holds the root
/// `MatcherPos`. As described in `MatcherPosHandle`, the root `MatcherPos`
/// structure is stored on the stack, but subsequent instances are put into
/// the heap.
/// - `'tt` represents the lifetime of the token trees that this matcher
/// position refers to.
///
/// It is important to distinguish these two lifetimes because we have a
/// `SmallVec<TokenTreeOrTokenTreeSlice<'tt>>` below, and the destructor of
/// that is considered to possibly access the data from its elements (it lacks
/// a `#[may_dangle]` attribute). As a result, the compiler needs to know that
/// all the elements in that `SmallVec` strictly outlive the root stack slot
/// lifetime. By separating `'tt` from `'root`, we can show that.
#[derive(Clone)]
struct MatcherPos<'root, 'tt> {
struct MatcherPos<'tt> {
/// The token or slice of tokens that make up the matcher. `elts` is short for "elements".
top_elts: TokenTreeOrTokenTreeSlice<'tt>,
@ -185,7 +168,7 @@ struct MatcherPos<'root, 'tt> {
match_hi: usize,
/// This field is only used if we are matching a repetition.
repetition: Option<MatcherPosRepetition<'root, 'tt>>,
repetition: Option<MatcherPosRepetition<'tt>>,
/// Specifically used to "unzip" token trees. By "unzip", we mean to unwrap the delimiters from
/// a delimited token tree (e.g., something wrapped in `(` `)`) or to get the contents of a doc
@ -200,9 +183,9 @@ struct MatcherPos<'root, 'tt> {
// This type is used a lot. Make sure it doesn't unintentionally get bigger.
#[cfg(all(target_arch = "x86_64", target_pointer_width = "64"))]
rustc_data_structures::static_assert_size!(MatcherPos<'_, '_>, 240);
rustc_data_structures::static_assert_size!(MatcherPos<'_>, 232);
impl<'root, 'tt> MatcherPos<'root, 'tt> {
impl<'tt> MatcherPos<'tt> {
/// `len` `Vec`s (initially shared and empty) that will store matches of metavars.
fn create_matches(len: usize) -> Box<[Lrc<NamedMatchVec>]> {
if len == 0 {
@ -241,11 +224,7 @@ impl<'root, 'tt> MatcherPos<'root, 'tt> {
}
}
fn repetition(
up: MatcherPosHandle<'root, 'tt>,
sp: DelimSpan,
seq: Lrc<SequenceRepetition>,
) -> Self {
fn repetition(up: Box<MatcherPos<'tt>>, sp: DelimSpan, seq: Lrc<SequenceRepetition>) -> Self {
MatcherPos {
stack: smallvec![],
idx: 0,
@ -270,7 +249,7 @@ impl<'root, 'tt> MatcherPos<'root, 'tt> {
}
#[derive(Clone)]
struct MatcherPosRepetition<'root, 'tt> {
struct MatcherPosRepetition<'tt> {
/// The KleeneOp of this sequence.
seq_op: mbe::KleeneOp,
@ -279,55 +258,12 @@ struct MatcherPosRepetition<'root, 'tt> {
/// The "parent" matcher position. That is, the matcher position just before we enter the
/// sequence.
up: MatcherPosHandle<'root, 'tt>,
up: Box<MatcherPos<'tt>>,
}
// Lots of MatcherPos instances are created at runtime. Allocating them on the
// heap is slow. Furthermore, using SmallVec<MatcherPos> to allocate them all
// on the stack is also slow, because MatcherPos is quite a large type and
// instances get moved around a lot between vectors, which requires lots of
// slow memcpy calls.
//
// Therefore, the initial MatcherPos is always allocated on the stack,
// subsequent ones (of which there aren't that many) are allocated on the heap,
// and this type is used to encapsulate both cases.
enum MatcherPosHandle<'root, 'tt> {
Ref(&'root mut MatcherPos<'root, 'tt>),
Box(Box<MatcherPos<'root, 'tt>>),
}
impl<'root, 'tt> Clone for MatcherPosHandle<'root, 'tt> {
// This always produces a new Box.
fn clone(&self) -> Self {
MatcherPosHandle::Box(match *self {
MatcherPosHandle::Ref(ref r) => Box::new((**r).clone()),
MatcherPosHandle::Box(ref b) => b.clone(),
})
}
}
impl<'root, 'tt> Deref for MatcherPosHandle<'root, 'tt> {
type Target = MatcherPos<'root, 'tt>;
fn deref(&self) -> &Self::Target {
match *self {
MatcherPosHandle::Ref(ref r) => r,
MatcherPosHandle::Box(ref b) => b,
}
}
}
impl<'root, 'tt> DerefMut for MatcherPosHandle<'root, 'tt> {
fn deref_mut(&mut self) -> &mut MatcherPos<'root, 'tt> {
match *self {
MatcherPosHandle::Ref(ref mut r) => r,
MatcherPosHandle::Box(ref mut b) => b,
}
}
}
enum EofItems<'root, 'tt> {
enum EofItems<'tt> {
None,
One(MatcherPosHandle<'root, 'tt>),
One(Box<MatcherPos<'tt>>),
Multiple,
}
@ -494,6 +430,10 @@ fn token_name_eq(t1: &Token, t2: &Token) -> bool {
pub struct TtParser {
macro_name: Ident,
cur_items: Vec<Box<MatcherPos<'tt>>>,
next_items: Vec<Box<MatcherPos<'tt>>>,
bb_items: Vec<Box<MatcherPos<'tt>>>,
}
impl TtParser {
@ -520,13 +460,13 @@ impl TtParser {
///
/// `Some(result)` if everything is finished, `None` otherwise. Note that matches are kept
/// track of through the items generated.
fn parse_tt_inner<'root, 'tt>(
fn parse_tt_inner<'tt>(
&self,
sess: &ParseSess,
ms: &[TokenTree],
cur_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
next_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
bb_items: &mut SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
cur_items: &mut SmallVec<[Box<MatcherPos<'tt>>; 1]>,
next_items: &mut SmallVec<[Box<MatcherPos<'tt>>; 1]>,
bb_items: &mut SmallVec<[Box<MatcherPos<'tt>>; 1]>,
token: &Token,
) -> Option<NamedParseResult> {
// Matcher positions that would be valid if the macro invocation was over now. Only
@ -570,9 +510,7 @@ impl TtParser {
}
// Allow for the possibility of one or more matches of this sequence.
cur_items.push(MatcherPosHandle::Box(Box::new(MatcherPos::repetition(
item, sp, seq,
))));
cur_items.push(box MatcherPos::repetition(item, sp, seq));
}
TokenTree::MetaVarDecl(span, _, None) => {
@ -706,11 +644,7 @@ impl TtParser {
// `parse_tt_inner` then processes all of these possible matcher positions and produces
// possible next positions into `next_items`. After some post-processing, the contents of
// `next_items` replenish `cur_items` and we start over again.
//
// This MatcherPos instance is allocated on the stack. All others -- and there are
// frequently *no* others! -- are allocated on the heap.
let mut initial = MatcherPos::new(ms);
let mut cur_items = smallvec![MatcherPosHandle::Ref(&mut initial)];
let mut cur_items = smallvec![box MatcherPos::new(ms)];
loop {
let mut next_items = SmallVec::new();
@ -793,10 +727,10 @@ impl TtParser {
}
}
fn ambiguity_error<'root, 'tt>(
fn ambiguity_error<'tt>(
&self,
next_items: SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
bb_items: SmallVec<[MatcherPosHandle<'root, 'tt>; 1]>,
next_items: SmallVec<[Box<MatcherPos<'tt>>; 1]>,
bb_items: SmallVec<[Box<MatcherPos<'tt>>; 1]>,
token_span: rustc_span::Span,
) -> NamedParseResult {
let nts = bb_items