SciLex

A small, header-only C++20 contextual lexer built on REAL.

• ReDoS-safe by construction (via REAL): no rule backtracks, nothing is exponential. Linear on every rule a DFA takes and on grammars whose rules stop scanning near their tokens; quadratic in the worst case, through a rule left on Pike (see Performance).
• Modes — contextual lexing: the same byte lexes differently by context (f-strings, XML tag/content, YAML block/flow).
• Layout Awareness — mode-aware indentation (NEWLINE / INDENT / DEDENT).
• Eager tokenize or lazy scan; positioned errors with a context snippet.
• C++20 header-only + abi3 Python binding (CPython 3.11+).
• Zero dependencies beyond REAL headers.

Define an ordered set of token rules — each a (kind, regex, skip) triple — and SciLex tokenizes by maximal munch: the longest anchored match wins, with rule order breaking ties. A rule can also opt into modes (contextual lexing), so the same byte lexes differently by context. Because it is a thin layer over REAL, every rule match is linear in what it scans and ReDoS-safe by construction; tokenizing is linear wherever the rules run on the DFA, and quadratic in the worst case only through a rule left on Pike (see Performance).

What that covers today: significant indentation, plus contexts like f-strings, YAML flow collections, and bracket continuation (modes + Layout Awareness Level A). Cases that need a deeper lexing↔indentation coupling — YAML block scalars | / >, heredocs — are Level B: documented, not in this version.

This follows the same design principles as REAL: purity, simplicity, and measured optimality.

Capabilities

• Ordered token rules: (kind, real::regex, skip)
• Maximal-munch matching (longest match wins, rule order for ties)
• Contextual lexing (modes) — per-rule in_mode + a push / pop / set mode stack
• DFA fast path (automatic) — every mode whose DFA reproduces the per-rule munch is accelerated with one real::dfa pass (5.7–18× on the example grammars, every one wholly on it); the decision is exact (Pike is the floor), the token stream identical; dfa_policy::requested restricts it to dfa_modes
• Layout Awareness — mode-aware indentation (NEWLINE / INDENT / DEDENT)
• Source positions (byte offset, line, column counted in bytes, code points or UTF-16 units); each token carries its start and its mode, and lexer::end_of(source, token) gives where it ends
• Eager (tokenize) and lazy (scan) APIs
• Optional END_OF_INPUT token
• Positioned errors with a context snippet
• ReDoS-safe (via REAL); linear wherever the rules run on the DFA, quadratic in the worst case only through a rule left on Pike
• Nine example grammars — three of them modal (f-strings, XML, YAML)

The three modal grammars differ in shape and each documents its own scope; modes resolve the contexts above, but the one contextual case still outside the model — lexing steered by indentation (block scalars, heredocs) — is Level B.

Not yet: block scalars / heredocs (Layout Awareness Level B), a compile-time static_lexer (a baked DFA — the Phase-0 spike found this wants build-time codegen, not constexpr).

See the guided tour for details.

C++ API

#include <scilex/scilex.hpp>

std::vector<scilex::rule> rules {
  {.kind = 0, .pattern = real::regex(R"(\s+)"), .skip = true}, // whitespace, skipped
  {.kind = 1, .pattern = real::regex("if")},                   // keyword: listed before the identifier
  {.kind = 2, .pattern = real::regex("[a-z_][a-z0-9_]*")},     // identifier
  {.kind = 3, .pattern = real::regex("[0-9]+")},               // number
  {.kind = 4, .pattern = real::regex(R"([-+*/=])")},           // operator
};
const scilex::lexer lexer {std::move(rules)};

// Lazy: one token per step (tokenize() returns them all at once).
for (const scilex::token& tok : lexer.scan("if x + 42")) {
  std::printf("%d %.*s\n", tok.kind, static_cast<int>(tok.lexeme.size()), tok.lexeme.data());
}

See docs/design.dox for the complete C++ API (lexer, token, position, layout, lex_error).

Python binding

An abi3 CPython extension (CPython 3.11+, Limited API).

import scilex

lx = scilex.Lexer([
    (0, r"\s+", True),                 # whitespace (skip)
    (1, r"[0-9]+", False),             # number
    (2, r"[A-Za-z_][A-Za-z0-9_]*", False),
])
# Every mode whose DFA is exact is accelerated (5.7–18×): lx.dfa_modes_active names them;
# scilex.Lexer([...], dfa="requested") keeps the per-rule path.

# Eager
tokens = lx.tokenize("foo 42", eof=True)

# Lazy (generator)
for tok in lx.scan("foo 42"):
    print(tok.kind, tok.lexeme, tok.position)

# Text in pieces: exactly tokenize's tokens, each once no text to come can change it
stream = lx.stream()
for chunk in ("fo", "o 4", "2"):
    for tok in stream.feed(chunk):
        print(tok.lexeme)
tail = stream.finish()

# Errors with context
try:
    lx.tokenize("foo @")
except scilex.error as e:
    e.position
    e.context

For significant indentation:

laid = scilex.Layout().apply(lx.tokenize(src, eof=True))

pip install scilex (wheels + sdist). Use scilex.get_include() to compile C++ code against the installed headers.

Build locally: make python && make python-test.

Contextual lexing — modes

A flat rule list can't separate contexts where the same byte means different things — { opens a Python f-string interpolation but a dict elsewhere; < opens an XML tag in content but is just a character inside CDATA. SciLex handles this with an opt-in mode stack: a rule may be restricted to named modes (in_mode) and may push / pop / set the mode when it wins. The engine is unchanged — maximal munch and the exact first-byte dispatch simply run per mode.

This unlocks, with no engine change:

• f-strings — f"sum={a+b}": code ↔ string body ↔ interpolation, nesting through the stack;
• XML — content ↔ tag (a shallow two-mode flip; CDATA and comments are single regex tokens, so an inner < is literal);
• YAML — block ↔ flow (significant indentation plus flow collections).

using op = scilex::mode_action::op;
scilex::rule open {.kind = OPEN, .pattern = real::regex("f\"")};
open.in_mode = {"default", "interp"};                      // active in code
open.action  = {.operation = op::push, .target = "fstr"};  // enters the f-string body
// "{" pushes "interp"; the closing quote pops "fstr"; the stack tracks nesting.

NAME, OPEN, TEXT, LB, RB, CLOSE = range(6)
fstr = scilex.Lexer([
    (NAME, r"[a-z]+", False, ["default", "interp"]),               # code, shared
    (OPEN, r'f"', False, ["default", "interp"], ("push", "fstr")),
    (TEXT, r'[^{}"]+', False, ["fstr"]),
    (LB, r"\{", False, ["fstr"], ("push", "interp")),         # "{" opens it from the body
    (CLOSE, r'"', False, ["fstr"], ("pop",)),
    (RB, r"\}", False, ["interp"], ("pop",)),
])
[t.kind for t in fstr.tokenize(r'f"hi {name}"')]   # OPEN TEXT LB NAME RB CLOSE

An action is None | ("push", mode) | ("set", mode) | ("pop",); a plain (kind, pattern, skip) rule needs neither field, so existing grammars are unaffected. See examples/python.hpp, examples/xml.hpp, examples/yaml.hpp for the three modal profiles in full. The stack is bounded: a push past scilex::max_mode_depth (65 536 frames, ~2 MiB) is a lexical error under either error policy, so an input made only of openers cannot grow it without end.

DFA fast path (automatic)

Every mode is accelerated by a real::dfa where that is exact: instead of trying each candidate rule at every position, one DFA pass recognizes the winning rule — the same maximal munch, with the order tie-break baked into the automaton. On a mode where many rules share leading bytes that is 5.7–18× the regular path on the full token path of the example grammars.

scilex::lexer lexer(std::move(rules));   // dfa_policy::automatic: every mode is tried
lexer.dfa_modes_active();                // the modes actually accelerated
// Only some modes, or none: dfa_policy::requested with the names (empty = the per-rule path).
scilex::lexer pike(std::move(other_rules), {}, {}, scilex::error_policy::raise,
                   scilex::column_unit::bytes, scilex::dfa_policy::requested);

It is best-effort and invisible: a rule that needs a zero-width assertion no DFA can represent, or whose DFA would change an answer, silently stays on the regular Pike engine beside its mode's DFA (pike_rules(mode) names it). A DFA takes each rule's longest match while Pike takes the match the rule's priority order prefers, and which rules keep the two equal is not visible in the syntax: as|assert stops at as on "assert" and so stays on Pike, while the lazy x*?y agrees on every input and keeps it. The constructor decides this for every rule with real::dfa_faithful — exactly, not by sampling — so the token stream is byte identical either way (Pike is the floor) and layout is unchanged. The DFA is built once, in the constructor, and that is its cost: measured 2026-09-27 (arm64, -O2, minimum of 7, REAL 2026.9.9), building the example grammars' lexers takes 0.06–2.9 ms instead of 0.01–0.12 ms, and the Python grammar's five modes ~13.5 ms (~26 ms against REAL 2026.9.8 and ~140 ms against 2026.9.6, whose DFA constructions were slower). A caller that builds many short-lived lexers can pass dfa_policy::requested. From Python: Lexer(..., dfa="requested").

Unicode identifiers vs DFA speed — the grammar author's choice

A real trade-off worth stating plainly. Write an identifier rule as \w+ (or [^\W\d]\w*) with the default flags and it reads Unicode identifiers — café, 変数 — the faithful behaviour for a language like Python 3. But a Unicode \w expands into more UTF-8 byte transitions than a DFA is built from, and \b is a zero-width assertion no DFA represents, so a rule holding either stays on the general engine beside its mode's DFA (same tokens, visible via pike_rules(mode)). The narrower Unicode \d and \s expand and stay on the DFA. Concretely the general engine runs at ~7–14 MB/s while every shipped grammar runs wholly on the DFA at 5.7–18× that on arm64 (8.6–27.5× on x86-64, where Pike reads lower); the python-unicode grammar, whose identifier rule stays on Pike, runs at 34.7 MB/s against 10.4 MB/s for Pike alone (3.3×; 2026-09-24, arm64, -O2, 256 KiB, BENCHMARKS.md) — the Unicode identifier costs part of the DFA.

So: if your identifiers are ASCII by specification (JSON, SQL, C), pin (?a) inline in the pattern (or pass real::flags::ascii) to keep \w \d \s \b ASCII, small, and DFA-representable — what the examples/ grammars do. If you want Unicode identifiers, write \w+ and accept that its rule runs on the general engine. The two tokenize ASCII input identically; they differ only on non-ASCII input and on whether the mode can be a DFA. The python-unicode example (scilex --example python-unicode) is the faithful-Python-3 variant of python, identical but for that one rule.

Thread safety

A lexer is immutable once built. Its DFAs are built in the constructor, and every tokenize call and every scan range keeps its own mode stack and walk memos, so one const lexer can be shared by any number of threads, each lexing its own text — checked under ThreadSanitizer with eight threads over four grammars, the hybrid ones included. An iterator from scan and a stream from stream() are cursors: drive each from a single thread. Rules left on Pike (pike_rules(mode)) call real::regex, whose lazy DFAs each thread leases from the regex's own pool, so they take no lock. In Python, scan and a stream's feed hold the GIL for each call while tokenize releases it around inputs of 4 KB or more.

Layout Awareness (Level A)

The layout pass is positional, and by default mode-blind. Layout Awareness Level A lets a mode be marked insignificant (Lexer(insignificant_modes=…)), so its tokens pass through without shaping indentation — and every token carries its mode (Token.mode) for the pass to read.

That lifts two real cases a decoupled positional pass otherwise gets wrong:

• YAML multi-line flow — [\n 1,\n 2\n] adds no spurious INDENT/DEDENT;
• Python implicit continuation — a call/list/dict wrapped across lines inside () [] {} reads as continuation, not a new block.

laid = lexer.layout(lexer.tokenize(src, eof=True))   # uses the lexer's own policy

Two invariants hold: with no insignificant mode the result is byte-for-byte the positional pass (zero cost); and the mode is the single source of the policy (no per-rule flag).

Honest scope. Level A covers multi-line flow and implicit continuation. Block scalars (| / >) and heredocs need a reference indent carried in the mode frame — that is Level B, a designed next step, not yet built. The bundled grammars demonstrate the features; each examples/<lang>.hpp header documents its own scope.

CLI

scilex is a command-line lexer — make cli builds it, make install puts it on your PATH (PREFIX=/BINDIR= to choose where). It has two input modes.

Built-in grammars — a showcase over the nine example languages (JSON, Python, C++, SQL, CSS, Lisp, math, XML, YAML):

$ scilex --list                       # the built-in grammars
$ scilex --version                    # SciLex's version and the REAL it was built with
$ scilex --example json file.json     # lex a file …
$ scilex --example python --layout    # … or its bundled sample, with INDENT/DEDENT

Your own grammar — the universal mode: bring a .lex file and lex anything. A grammar is one rule per line — name, a tab, regex, then an optional tab and space-separated options: skip, in=m1,m2 (the modes the rule is active in), and one of push=m, set=m, pop (# comments and blank lines are ignored):

$ cat my.lex
WS	\s+	skip
NUMBER	[0-9]+(\.[0-9]+)?
IDENT	[A-Za-z_][A-Za-z0-9_]*
OP	<=|>=|==|!=|[-+*/%=<>]

$ echo 'x = 41 + 1' | scilex my.lex        # stdin when no file is given
IDENT	x	1:1
OP	=	1:3
NUMBER	41	1:5
OP	+	1:8
NUMBER	1	1:10

Output is one token per line — the kind, a tab, the lexeme, a tab, then line:col; --layout adds the indentation tokens. A malformed grammar is reported with a clear, positioned error (my.lex:3: invalid regex: …) — never a crash. See examples/sample.lex for a worked file.

A modal grammar reads the same way — a string mode entered by a double quote and left by the next one:

WS	\s+	skip
STRING	"	push=str
TEXT	[^"\\]+	in=str
ESCAPE	\\.	in=str
END	"	in=str pop
IDENT	[A-Za-z_]\w*

The format has one parser, in the optional header scilex/grammar.hpp (scilex::parse_grammar(text, origin), scilex::load_grammar(path), errors as scilex::grammar_error with a line and column); scilex.hpp does not include it, so the lexer itself stays plain C++ rule lists (std::vector<scilex::rule>). Python reaches the same parser: scilex.parse_grammar(text) and scilex.load_grammar(path) return a Grammar whose .rules feed Lexer, .names name each kind and .lexer(...) builds one; a malformed grammar raises scilex.GrammarError with .line, .column and .cause.

Dependencies

SciLex is header-only and depends only on REAL's headers (the package real-regex on PyPI / https://github.com/RECHE23/real-regex).

By default the build looks for them in a sibling checkout:

~/Projects/
├── real-regex/   # REAL (https://github.com/RECHE23/real-regex)
└── scilex/       # SciLex  (uses ../real-regex/include by default)

Point the build elsewhere with REAL_INCLUDE (Makefile) or -DSCILEX_REAL_INCLUDE=... (CMake) — for instance at the path printed by python -c "import real; print(real.get_include())" when REAL is installed via pip.

For CI or a reproducible build — where no on-disk layout can be assumed — fetch REAL with CMake FetchContent instead (make build FETCH=1, or -DSCILEX_FETCH_DEPS=ON); point it at a remote and pin a tag with -DSCILEX_REAL_REPO=https://… -DSCILEX_REAL_TAG=v2026.9.6.

Development

make test        # build and run the test suite
make coverage    # line-coverage summary + HTML report
make sanitize    # tests under AddressSanitizer + UndefinedBehaviorSanitizer
make lint        # clang-tidy
make format      # uncrustify, in place
make doc         # API reference (Doxygen) with embedded coverage
make fuzz        # libFuzzer + ASan/UBSan over the property oracle (FUZZ_TIME seconds)
make full-local-gate  # every gate, cheap first; the pre-push check
make sabotage-help    # the sabotage harness: break one thing, run one check, put it back

The API reference is published at https://reche23.github.io/scilex/.

Override the compiler with make test CXX=g++-14.

Coverage bar. SciLex holds the SciLang-stack gate — 100% on all four dimensions (lines, functions, regions and branches) of include/, enforced by make coverage-gate: locally by make full-local-gate (Apple clang), and in CI under Linux clang 18 on every push.

scilex::scilex is the CMake target — add_subdirectory, FetchContent, or an installed config package. The config calls find_dependency(real), so installing REAL's config package alongside (on the same prefix) makes the whole chain resolve from one find_package:

# With REAL and SciLex installed under <prefix>:
find_package(scilex CONFIG REQUIRED)   # pulls in real:: transitively
target_link_libraries(app PRIVATE scilex::scilex)

Releasing

make release computes the next calendar version YYYY.M.PATCH (the patch resets each month; PEP 440 drops leading zeros). The pushed tag drives the release workflow — wheels + sdist + the API-reference tarball + a GitHub Release, published via Trusted Publishing — while docs.yml deploys the reference to GitHub Pages.

Design

A guided tour of how SciLex works (maximal munch, REAL foundation, layout, C++/Python API, current scope) lives in docs/design.dox (also rendered by make doc).

Performance

See BENCHMARKS.md. In C++ the example grammars lex at 66–134 MB/s on their DFAs (arm64, -O2); through the Python binding SciLex is 1.3–1.9× faster than re on the benign case measured there, and ahead of Pygments and tree-sitter on the two corpora of the cross-tool table; on a ReDoS pattern SciLex stays linear while re explodes. flex, a code generator, remains ~5–7× faster.

Linear on the DFA; the worst case is quadratic, and only on Pike. Every rule match is linear in the text it scans, but at each token start every candidate rule is tried, and a rule may scan far past the token that finally wins: a*b and a on aaa… scan to the end looking for b at every position, then lose to a. On the DFA each walk is memoized over the whole source (Reps, 1998): a state a walk proved leads to no accept stops every later walk that reaches it, so the rules on the DFA cost O(n × states) in total. A rule the DFA cannot take (see DFA fast path) keeps the per-position scan, and the worst case with it. Measured on 2026-09-24 (arm64, Apple clang 16, -O2): a*b and a on aaa…, both on the DFA, lex 256 KiB in 8.8 ms and double with the input; the same pair kept on Pike (dfa_policy::requested) still quadruples per doubling, 266 ms at 4 000 bytes and 16.9 s at 32 000 (2026-09-23). The shipped grammars stay linear (flat MB/s in BENCHMARKS.md); a grammar fed by users (.lex files) reaches the worst case only through a rule left on Pike (pike_rules(mode) names them). See docs/spec.dox.

License

MIT — see LICENSE.

Author

René Chenard