1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
//! TeX's mouth: bytes to tokens.
//!
//! The mouth is a three-state machine per line (`tex.web` §303): `N` at the
//! start of a line, `M` in the middle, `S` while skipping blanks. The states are
//! not decoration — they are what makes a blank line a `\par`, what collapses a
//! run of spaces to one, and what makes a space after a control word vanish.
//! Getting them wrong changes what documents mean, so they are modelled
//! directly rather than approximated with trimming.
use crate::catcode::{Cat, CatTable};
use crate::token::Token;
#[derive(Clone, Copy, PartialEq)]
enum State {
NewLine,
MidLine,
SkipBlanks,
}
/// Tokens read ahead of the expander, and the table revision they assume.
struct Ahead {
/// Generation of the catcode table these were lexed under.
generation: u32,
/// Tokens with the character index each ended at, in source order.
toks: Vec<crate::parallel::Placed>,
/// How many have been handed out.
at: usize,
}
pub struct Lexer {
chars: Vec<char>,
pos: usize,
/// The last answer [`Lexer::line`] gave, as `(position, line)`.
///
/// Counting newlines from the start of the file on every call is O(n) per
/// token and so O(n²) per document — which the scaling benchmark caught:
/// quadrupling the input cost 27x. Counting only what has been consumed
/// SINCE the last answer makes the whole walk amortized O(n), and a rewind
/// (the scanner pushes tokens back) simply recounts from the start.
line_cache: std::cell::Cell<(usize, u32)>,
/// Tokens lexed ahead of the expander, with the generation of the catcode
/// table they were produced under.
///
/// The mouth is the largest single cost of compiling a document and it is
/// the one stage that parallelises: its states are per-line, so line-
/// aligned slices lex independently. Reading ahead is a GUESS that the
/// table will not move, and `\catcode` is allowed to move it at any point,
/// so the guess carries the generation it was made under and is dropped the
/// moment that stops matching. Dropping it costs only the work already
/// done, never correctness.
ahead: Option<Ahead>,
/// Set once read-ahead has been abandoned often enough that it is not
/// paying: a document rewriting catcodes throughout would otherwise re-lex
/// its tail on every change.
ahead_disabled: bool,
/// How many times a stale read-ahead has been thrown away.
ahead_misses: u32,
/// Tokens to scan normally before speculating again.
///
/// A document opens by setting catcodes -- `\catcode`\{=1` and friends --
/// and each one invalidates whatever was read ahead. Speculating from the
/// first token therefore lexes the whole file once per catcode in the
/// preamble and throws all of it away. Waiting until the table has been
/// still for a while costs a few hundred tokens of ordinary scanning and
/// saves re-lexing the document.
ahead_cooldown: usize,
state: State,
/// Pushed-back tokens (`\expandafter` and macro expansion feed these).
pub pending: Vec<Token>,
}
impl Lexer {
/// A mouth over an already-decoded slice of characters.
///
/// The parallel pre-lexer hands each worker one line-aligned slice of the
/// document's characters; decoding UTF-8 once for the whole file and
/// slicing it is what keeps the split from paying for the decode N times.
///
/// Read-ahead is OFF here, and that is load-bearing rather than an
/// optimisation: this constructor is what the workers themselves run, so a
/// worker that read ahead would fan out again, and again, each level
/// re-chunking at different offsets. That nesting is what made the split
/// slower than the scan it replaced AND made it disagree with the mouth --
/// a worker's own sub-chunks split at boundaries the outer level never
/// accounted for, so state that should have been `S` began as `N` and
/// invented a `\par`.
pub fn from_chars(chars: Vec<char>) -> Self {
Self {
chars,
pos: 0,
state: State::NewLine,
pending: Vec::new(),
line_cache: std::cell::Cell::new((0, 1)),
ahead: None,
ahead_disabled: true,
ahead_misses: 0,
ahead_cooldown: Self::AHEAD_COOLDOWN,
}
}
/// How far the mouth has read, as an index into the characters.
///
/// The pre-lexer records this after each token so a cached stream can be
/// abandoned mid-way and scanning resumed at exactly the right character.
pub fn pos(&self) -> usize {
self.pos
}
/// The document's characters, for a caller that wants to slice them.
pub fn chars(&self) -> &[char] {
&self.chars
}
pub fn new(src: &str) -> Self {
Self {
chars: src.chars().collect(),
pos: 0,
line_cache: std::cell::Cell::new((0, 1)),
state: State::NewLine,
pending: Vec::new(),
ahead: None,
ahead_disabled: false,
ahead_misses: 0,
ahead_cooldown: Self::AHEAD_COOLDOWN,
}
}
/// Put a token stream back in front of the input, to be read next.
///
/// Stored reversed so `next` can pop from the end — expansion pushes whole
/// macro bodies here and they must come back out in order.
pub fn push_back(&mut self, toks: &[Token]) {
for t in toks.iter().rev() {
self.pending.push(*t);
}
}
/// The 1-based line the mouth has reached, for a diagnostic that has to
/// point somewhere. Counted from the characters already consumed rather
/// than tracked incrementally: the scanner rewinds and pushes back, and a
/// counter that has to be maintained through both is a counter that drifts.
pub fn line(&self) -> u32 {
let pos = self.pos.min(self.chars.len());
let (from, base) = match self.line_cache.get() {
(cached_pos, line) if cached_pos <= pos => (cached_pos, line),
// The scanner moved backwards: recount.
_ => (0, 1),
};
let line = base + self.chars[from..pos].iter().filter(|c| **c == '\n').count() as u32;
self.line_cache.set((pos, line));
line
}
fn peek(&self) -> Option<char> {
self.chars.get(self.pos).copied()
}
/// `^^X` notation (`tex.web` §352): a superscript character twice, then one
/// character, denotes a control character. `^^M` is a carriage return, which
/// is how a line end gets written inside a macro body.
fn double_superscript(&mut self, cats: &CatTable, c: char) -> Option<char> {
if cats.get(c) != Cat::Superscript {
return None;
}
if self.chars.get(self.pos + 1).copied() != Some(c) {
return None;
}
let third = self.chars.get(self.pos + 2).copied()?;
let code = u32::from(third);
if code >= 128 {
return None;
}
self.pos += 3;
let shifted = match code < 64 {
true => code + 64,
false => code - 64,
};
char::from_u32(shifted)
}
/// Smallest tail worth handing to other cores.
///
/// Below this the threads cost more than the scan they replace; the figure
/// is a slice per core of a few tens of kilobytes, which is where the split
/// starts to pay on the documents in `bench/`.
const AHEAD_MIN_CHARS: usize = 65_536;
/// How many times a stale read-ahead is tolerated before giving up on it.
///
/// A document that rewrites catcodes throughout would otherwise re-lex its
/// tail after every change, turning a speedup into quadratic work. Three
/// strikes is enough to tell a preamble (which settles) from a document
/// that keeps moving the table.
const AHEAD_MAX_MISSES: u32 = 3;
/// Tokens of quiet required before (re)building the read-ahead.
const AHEAD_COOLDOWN: usize = 512;
/// Serve the next token from the speculative read-ahead, filling it first
/// if that is worthwhile.
///
/// Returns `None` when there is nothing cached and nothing worth caching,
/// in which case the caller scans as it always did.
fn next_from_ahead(&mut self, cats: &CatTable) -> Option<Token> {
if self.ahead_disabled {
return None;
}
// OFF by default, and that is a measured decision rather than caution.
//
// Splitting the mouth across cores is correct (tests/parallel_lex.rs
// pins it against the sequential mouth) but it LOSES: the sequential
// mouth streams, never materialising a token list, while reading ahead
// has to copy the characters per worker, build a token-and-position
// vector for the whole tail, and concatenate the pieces. On a 5.6 MB
// document that is tens of megabytes of traffic to save a scan worth
// about 0.03s, and it measured 3x SLOWER end to end than not doing it.
//
// The lexer is simply not where a texrs run spends its time: it
// compiles to bytecode, so the mouth is walked once and the VM does the
// repeated work. Left in, off, and honest -- `TEXRS_PARALLEL=1` turns
// it on for anyone measuring a document shaped differently.
if std::env::var_os("TEXRS_PARALLEL").is_none() {
self.ahead_disabled = true;
return None;
}
// Anything cached under a superseded table is wrong from here on.
if let Some(a) = &self.ahead {
if a.generation != cats.generation() {
self.ahead = None;
crate::parallel::DROP_GEN.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
self.ahead_misses += 1;
self.ahead_cooldown = Self::AHEAD_COOLDOWN;
if self.ahead_misses >= Self::AHEAD_MAX_MISSES {
self.ahead_disabled = true;
return None;
}
}
}
if self.ahead.is_none() {
if self.ahead_cooldown > 0 {
self.ahead_cooldown -= 1;
return None;
}
if self.chars.len().saturating_sub(self.pos) < Self::AHEAD_MIN_CHARS {
return None;
}
let threads = crate::parallel::default_threads();
if threads < 2 {
return None;
}
let toks = crate::parallel::prelex(&self.chars, self.pos, cats, threads);
if toks.is_empty() {
return None;
}
self.ahead = Some(Ahead {
generation: cats.generation(),
toks,
at: 0,
});
}
let a = self.ahead.as_mut()?;
let (tok, end) = *a.toks.get(a.at)?;
a.at += 1;
// Keep `pos` exact: `line()` reads it, and abandoning the cache has to
// resume scanning at the character this token ended on.
self.pos = end;
if a.at == a.toks.len() {
crate::parallel::DROP_EXHAUST.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
self.ahead = None;
}
Some(tok)
}
/// One token, or `None` at end of input.
pub fn next_token(&mut self, cats: &CatTable) -> Option<Token> {
if let Some(t) = self.pending.pop() {
return Some(t);
}
if let Some(t) = self.next_from_ahead(cats) {
return Some(t);
}
loop {
let c = self.peek()?;
if let Some(decoded) = self.double_superscript(cats, c) {
// Re-read the decoded character in place of the `^^X` triple.
self.chars.splice(self.pos..self.pos, [decoded]);
continue;
}
let cat = cats.get(c);
self.pos += 1;
match cat {
Cat::Escape => return Some(self.control_sequence(cats)),
Cat::Comment => {
// The comment and the line end it eats produce nothing, and
// the next line starts in state N.
while let Some(ch) = self.peek() {
self.pos += 1;
if cats.get(ch) == Cat::EndLine {
break;
}
}
self.state = State::NewLine;
}
Cat::EndLine => {
// §304: a line end is a space in state M, a `\par` in state
// N (the blank-line rule), and nothing while skipping blanks.
let out = match self.state {
State::MidLine => Some(Token::Char(' ', Cat::Space)),
State::NewLine => Some(Token::cs("par")),
State::SkipBlanks => None,
};
self.state = State::NewLine;
if out.is_some() {
return out;
}
}
Cat::Space => {
// A run of spaces is one space, and only in state M.
if self.state == State::MidLine {
self.state = State::SkipBlanks;
return Some(Token::Char(' ', Cat::Space));
}
}
Cat::Ignored => {}
_ => {
self.state = State::MidLine;
return Some(Token::Char(c, cat));
}
}
}
}
/// The escape character has been consumed; read the control sequence.
///
/// A control WORD is a maximal run of letters and swallows the spaces after
/// it (state S); a control SYMBOL is exactly one character and does not.
/// That asymmetry is why `\foo bar` loses its space and `\! bar` keeps it.
fn control_sequence(&mut self, cats: &CatTable) -> Token {
let Some(first) = self.peek() else {
return Token::cs("");
};
self.pos += 1;
if cats.get(first) != Cat::Letter {
// A control space (`\ `) leaves the state mid-line, like any other
// single-character control sequence.
self.state = match cats.get(first) == Cat::Space {
true => State::SkipBlanks,
false => State::MidLine,
};
return Token::cs(first.encode_utf8(&mut [0u8; 4]));
}
let mut name = String::from(first);
while let Some(c) = self.peek() {
if cats.get(c) != Cat::Letter {
break;
}
name.push(c);
self.pos += 1;
}
self.state = State::SkipBlanks;
Token::cs(&name)
}
}