1//! Client-side interner used for symbols.
2//!
3//! This is roughly based on the symbol interner from `rustc_span` and the
4//! DroplessArena from `rustc_arena`. It is unfortunately a complete
5//! copy/re-implementation rather than a dependency as it is difficult to depend
6//! on crates from within `proc_macro`, due to it being built at the same time
7//! as `std`.
8//!
9//! If at some point in the future it becomes easier to add dependencies to
10//! proc_macro, this module should probably be removed or simplified.
11
12use std::cell::RefCell;
13use std::num::NonZero;
14
15use super::*;
16
17/// Handle for a symbol string stored within the Interner.
18#[derive(Copy, Clone, PartialEq, Eq, Hash)]
19pub struct Symbol(NonZero<u32>);
20
21impl !Send for Symbol {}
22impl !Sync for Symbol {}
23
24impl Symbol {
25 /// Intern a new `Symbol`
26 pub(crate) fn new(string: &str) -> Self {
27 INTERNER.with_borrow_mut(|i| i.intern(string))
28 }
29
30 /// Creates a new `Symbol` for an identifier.
31 ///
32 /// Validates and normalizes before converting it to a symbol.
33 pub(crate) fn new_ident(string: &str, is_raw: bool) -> Self {
34 // Fast-path: check if this is a valid ASCII identifier
35 if Self::is_valid_ascii_ident(string.as_bytes()) || string == "$crate" {
36 if is_raw && !Self::can_be_raw(string) {
37 panic!("`{}` cannot be a raw identifier", string);
38 }
39 return Self::new(string);
40 }
41
42 // Slow-path: If the string is already ASCII we're done, otherwise ask
43 // our server to do this for us over RPC.
44 // We don't need to check for identifiers which can't be raw here,
45 // because all of them are ASCII.
46 if string.is_ascii() {
47 Err(())
48 } else {
49 client::Methods::symbol_normalize_and_validate_ident(string)
50 }
51 .unwrap_or_else(|_| panic!("`{:?}` is not a valid identifier", string))
52 }
53
54 /// Run a callback with the symbol's string value.
55 pub(crate) fn with<R>(self, f: impl FnOnce(&str) -> R) -> R {
56 INTERNER.with_borrow(|i| f(i.get(self)))
57 }
58
59 /// Clear out the thread-local symbol interner, making all previously
60 /// created symbols invalid such that `with` will panic when called on them.
61 pub(crate) fn invalidate_all() {
62 INTERNER.with_borrow_mut(|i| i.clear());
63 }
64
65 /// Checks if the ident is a valid ASCII identifier.
66 ///
67 /// This is a short-circuit which is cheap to implement within the
68 /// proc-macro client to avoid RPC when creating simple idents, but may
69 /// return `false` for a valid identifier if it contains non-ASCII
70 /// characters.
71 fn is_valid_ascii_ident(bytes: &[u8]) -> bool {
72 matches!(bytes.first(), Some(b'_' | b'a'..=b'z' | b'A'..=b'Z'))
73 && bytes[1..]
74 .iter()
75 .all(|b| matches!(b, b'_' | b'a'..=b'z' | b'A'..=b'Z' | b'0'..=b'9'))
76 }
77
78 // Mimics the behavior of `Symbol::can_be_raw` from `rustc_span`
79 fn can_be_raw(string: &str) -> bool {
80 !matches!(string, "_" | "super" | "self" | "Self" | "crate" | "$crate")
81 }
82}
83
84impl fmt::Debug for Symbol {
85 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
86 self.with(|s| fmt::Debug::fmt(s, f))
87 }
88}
89
90impl fmt::Display for Symbol {
91 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
92 self.with(|s| fmt::Display::fmt(s, f))
93 }
94}
95
96impl<S> Encode<S> for Symbol {
97 fn encode(self, w: &mut Buffer, s: &mut S) {
98 self.with(|sym| sym.encode(w, s))
99 }
100}
101
102impl<S: server::Server> Decode<'_, '_, server::HandleStore<S>> for server::MarkedSymbol<S> {
103 fn decode(r: &mut &[u8], s: &mut server::HandleStore<S>) -> Self {
104 Mark::mark(S::intern_symbol(<&str>::decode(r, s)))
105 }
106}
107
108impl<S: server::Server> Encode<server::HandleStore<S>> for server::MarkedSymbol<S> {
109 fn encode(self, w: &mut Buffer, s: &mut server::HandleStore<S>) {
110 S::with_symbol_string(&self.unmark(), |sym| sym.encode(w, s))
111 }
112}
113
114impl<S> Decode<'_, '_, S> for Symbol {
115 fn decode(r: &mut &[u8], s: &mut S) -> Self {
116 Symbol::new(<&str>::decode(r, s))
117 }
118}
119
120thread_local! {
121 static INTERNER: RefCell<Interner> = RefCell::new(Interner {
122 arena: arena::Arena::new(),
123 names: fxhash::FxHashMap::default(),
124 strings: Vec::new(),
125 // Start with a base of 1 to make sure that `NonZero<u32>` works.
126 sym_base: NonZero::new(1).unwrap(),
127 });
128}
129
130/// Basic interner for a `Symbol`, inspired by the one in `rustc_span`.
131struct Interner {
132 arena: arena::Arena,
133 // SAFETY: These `'static` lifetimes are actually references to data owned
134 // by the Arena. This is safe, as we never return them as static references
135 // from `Interner`.
136 names: fxhash::FxHashMap<&'static str, Symbol>,
137 strings: Vec<&'static str>,
138 // The offset to apply to symbol names stored in the interner. This is used
139 // to ensure that symbol names are not re-used after the interner is
140 // cleared.
141 sym_base: NonZero<u32>,
142}
143
144impl Interner {
145 fn intern(&mut self, string: &str) -> Symbol {
146 if let Some(&name) = self.names.get(string) {
147 return name;
148 }
149
150 let name = Symbol(
151 self.sym_base
152 .checked_add(self.strings.len() as u32)
153 .expect("`proc_macro` symbol name overflow"),
154 );
155
156 let string: &str = self.arena.alloc_str(string);
157
158 // SAFETY: we can extend the arena allocation to `'static` because we
159 // only access these while the arena is still alive.
160 let string: &'static str = unsafe { &*(string as *const str) };
161 self.strings.push(string);
162 self.names.insert(string, name);
163 name
164 }
165
166 /// Reads a symbol's value from the store while it is held.
167 fn get(&self, symbol: Symbol) -> &str {
168 // NOTE: Subtract out the offset which was added to make the symbol
169 // nonzero and prevent symbol name re-use.
170 let name = symbol
171 .0
172 .get()
173 .checked_sub(self.sym_base.get())
174 .expect("use-after-free of `proc_macro` symbol");
175 self.strings[name as usize]
176 }
177
178 /// Clear all symbols from the store, invalidating them such that `get` will
179 /// panic if they are accessed in the future.
180 fn clear(&mut self) {
181 // NOTE: Be careful not to panic here, as we may be called on the client
182 // when a `catch_unwind` isn't installed.
183 self.sym_base = self.sym_base.saturating_add(self.strings.len() as u32);
184 self.names.clear();
185 self.strings.clear();
186
187 // SAFETY: This is cleared after the names and strings tables are
188 // cleared out, so no references into the arena should remain.
189 self.arena = arena::Arena::new();
190 }
191}
192