regexx/README.md

# regexx

A single-file C regular expression interpreter that reproduces the observable
behavior of Python's `re` module, including its exact identifier names
(`Pattern`, `Match`, `re_compile`, `re_sub`, `IGNORECASE`, and so on; see
Section 9 of `concept.md`), and that additionally supports binary data
(arbitrary byte streams, including embedded NUL) and UTF-8 text alongside
plain ASCII.

The design rationale, the algorithmic trade-offs, and a full accounting of
what is and is not carried over from Python's `re` and from POSIX's native
`regex.h` are recorded in [`concept.md`](concept.md). This file documents the
implementation that exists today at a glance; [`docs/API.md`](docs/API.md) is
the exhaustive reference (every type, every flag, every function's exact
return-value and memory-ownership convention, checked against a real CPython
interpreter, not against memory).

## Implementation status

This is a v1 implementation. It is a complete, tested engine for the pattern
syntax and operations listed below, executed by a single recursive
backtracking engine (`concept.md` Section 7.3) over a fully materialized copy
of the input.

**It does not yet implement the streaming, bounded-memory regular engine of
`concept.md` Section 7.2.** `Input` (the abstraction over "a source of
chunks", `concept.md` 9.2) is implemented, and `Input_from_file` maps or
reads a whole file into memory before matching (see "Memory footprint"
below for exactly how, and for the measured numbers this and other fixes
were checked against). Every public function signature is already exactly
what the streaming design in `concept.md` specifies, so the non-streaming
implementation underneath a given call can be replaced later without
changing any caller. Concretely, today:

- `Pattern_match`/`Pattern_fullmatch` (a single anchored attempt at a fixed
  position) run in time proportional to the length of that attempt, and, for
  the common case of a single character, class, or `.` repeated by a
  quantifier, in *O(1)* recursion depth regardless of input size (the
  `OP_REPEAT1` fast path). A repeated *compound* sub-pattern (for example
  `(ab)*`) still recurses once per repetition, bounded by a configurable
  depth limit (`MAX_DEPTH` in `regexx.c`, currently 60000, not exposed
  through the public API, so changing it means editing `regexx.c` and
  rebuilding); past that limit, matching fails with a reported error rather
  than a stack overflow or a wrong answer.
- **`Pattern_search`/`Pattern_finditer`/`Pattern_split`/`Pattern_sub` run in
  linear time for the common case: a pattern built only from simple,
  non-backreference constructs where every quantifier's body is a single
  character, class, or `.`** (`a*b`, `\s*\d+`, `.*"`, and the great majority
  of patterns actually written by hand). This was not always true of this
  build; measuring it directly is what caught that it previously was not
  (an earlier revision of this file claimed linear time without having
  re-verified it, then had to correct that claim, then fixed the underlying
  defect; the corrected numbers are below). Two independent techniques make
  it true now, neither of them the streaming engine of `concept.md` Section
  7.2, which remains unimplemented:
  1. **Memoized backtracking.** For any pattern with no `OP_BACKREF`
     anywhere, whether execution starting at a given (instruction, text
     position) pair can ever reach a match is a fact that never changes
     once computed, so `run_memo` (`regexx.c`) caches every *failure* (never
     a success, so it cannot change which match is found, only skip
     re-deriving failures already known) and consults the cache before
     redoing that work. This is a published technique, not a house
     invention: a memoization table that records only failures, because
     "a matching success ... immediately propagates to the success of the
     whole problem," is exactly the scheme described in recent work on
     backtracking regex matchers ("Selective Memoization for Efficient
     Backtracking Regular Expression Matching", and, for the
     lookaround/atomic case specifically, "Efficient Matching with
     Memoization for Regexes with Look-around and Atomic Grouping", both
     linked below).
  2. **Precomputed run lengths and skip-ahead tables for `OP_REPEAT1`.**
     Memoizing individual `(instruction, position)` pairs does not help
     when each one is already `O(1)` work, which is exactly `a*b`'s
     situation: counting how many `a`s follow a position, and then trying
     every possible split point against the trailing `b` one at a time,
     are each individually cheap but happen `O(remaining length)` times per
     start position tried. `compute_maxrun` precomputes, once per
     `Pattern_search`/`finditer`/`split` call, how many characters each
     repeat can consume from every position in one backward pass, and
     `compute_next_prevmatch` precomputes, for a repeat immediately
     followed by a single literal/class/`.`, the rightmost position at or
     before any given point where that next atom can match, so the
     backtrack loop jumps directly to candidates worth trying instead of
     visiting every position in between. This is the same idea production
     engines call a literal prefilter (RE2's and Rust's `regex` crate's
     `memchr`/`memmem`/Teddy prefilters skip positions that provably cannot
     match before ever invoking the full engine); those use SIMD-accelerated
     library primitives operating directly on bytes, this build uses a
     precomputed array, which is slower per lookup but the same
     algorithmic idea, and appropriate for `concept.md`'s stated priority of
     convenience over performance.

     Measured directly: `a*b` searched over *n* bytes of `a` with no `b`
     anywhere took 19.2s at *n* = 80,000 before these two techniques and
     0.0016s after, with time now scaling linearly (2x per doubling of *n*)
     rather than quadratically (4x per doubling).

     Cost: these tables use `O(k * n)` memory, `k` being the number of
     `OP_REPEAT1` instructions actually present in the compiled pattern,
     `n` the input length; `Pattern_match`/`Pattern_fullmatch` never
     allocate them (a single anchored attempt gets no benefit from them).
- **A pattern with a backreference can, like CPython's own `_sre`, still
  take worst-case exponential time on an adversarial input** (`concept.md`
  Section 5, 13.2; memoization above is unsound and therefore disabled
  whenever a pattern contains `OP_BACKREF` anywhere, since a backreference's
  outcome depends on capture history, not on position alone). This is the
  same catastrophic backtracking (ReDoS) behavior CPython itself exhibits on
  such patterns, not a regression specific to this engine; a pattern author
  who needs to rule it out for a specific pattern can use an atomic group
  (`(?>...)`) or a possessive quantifier around the ambiguous repetition,
  exactly as they would for CPython's `re`, and exactly as general ReDoS
  mitigation guidance recommends (linked below).
- **A backreference-free pattern shaped like nested, overlapping
  quantifiers (`(a+)+b`, the textbook ReDoS shape) is no longer
  exponential, because it has no backreference and so is memoized, but is
  not necessarily linear either.** Measured directly: `(a+)+b` against *n*
  characters of `a` with no trailing `b` took over a minute already at
  *n* = 40 before memoization (exponential; this build's own stress test
  keeps that case at *n* = 24 for exactly this reason) and 3.2s at
  *n* = 32,000 after (empirically quadratic: roughly 4x per doubling),
  because the inner `a+`'s `OP_REPEAT1` is followed by the group's closing
  save, not by a simple atom, so the skip-ahead technique above does not
  apply to it, only the memoization does. Exponential to quadratic is a
  qualitative change (the *n* = 40 case alone went from longer than this
  project could wait for, to instant), not a complete fix; a pattern
  actually meant to run unattended against adversarial input should still
  avoid this shape, or wrap the inner repetition in an atomic group,
  regardless of the improvement.

Further reading on the techniques above: Russ Cox, ["Regular Expression
Matching: the Virtual Machine
Approach"](https://swtch.com/~rsc/regexp/regexp2.html) (why prepending
`.*?` gives linear-time unanchored search only in a Thompson/Pike VM, not in
a backtracking engine, which is why this build needed a different fix);
["Selective Memoization for Efficient Backtracking Regular Expression
Matching"](https://arxiv.org/html/2606.26678) and ["Efficient Matching with
Memoization for Regexes with Look-around and Atomic
Grouping"](https://arxiv.org/pdf/2401.12639) (the failure-only memoization
scheme this build's `run_memo` implements, and a more memory-efficient
selective variant, memoizing only at loop "feedback nodes" rather than every
instruction, that this build does not implement but could); the [Snyk
writeup on ReDoS and catastrophic
backtracking](https://snyk.io/blog/redos-and-catastrophic-backtracking/) for
the general phenomenon and mitigation guidance.

### Memory footprint

Measured directly with Valgrind/Massif (a real 10MB search) and by watching
`VmRSS`/`VmHWM` on real 200MB-1GB files, not estimated from reading the
code. `Pattern_search`/`finditer`/`split`/`sub` (the operations that try
more than one start position) currently use, at peak, about **8.4x** the
input length in memory for a pattern using `OP_REPEAT1`
(`concept.md` 7.5's `compute_maxrun` and `compute_next_prevmatch` tables,
`int32_t`-per-input-position each, are the entire remaining cost: 47.75%
each in the Massif profile, `alloc_memo`'s bitset a further 4.5%).
`Pattern_match`/`fullmatch` (a single attempt, no search tables) use
proportionally less.

Two real issues were found and fixed getting to that number, in order:

1. **`build_matbuf` widened every byte to a 4-byte `uint32_t`, even in
   `BINARY`/`ASCII` mode, where a byte never exceeds 255 and the widening
   bought nothing.** This cost as much extra memory as the input itself,
   four times over, unconditionally, on top of the search tables above.
   Fixed: `BINARY`/`ASCII` mode now reads the input's own bytes directly
   (`MatBuf`/`MCtx`'s `text8` field, `text_at()`/`buf_at()` in `regexx.c`);
   only `UTF8` mode still widens, because it actually needs code points up
   to `0x10FFFF`, which do not fit in a byte. This dropped the measured
   10MB-search peak from 140.3MB (13.4x) to 87.8MB (8.4x).
2. **`Input_from_file` read every file into a fresh, private, `malloc`'d
   copy, even though the OS's page cache already holds the file's bytes.**
   For a regular, seekable, non-empty file this now uses `mmap()`
   (`PROT_READ`, `MAP_PRIVATE`) instead: the mapped pages are backed
   directly by the file and stay clean (never written), so the kernel can
   reclaim them under memory pressure and re-fault them in from disk later,
   rather than them being pinned for the whole match attempt the way a
   `malloc`'d copy is; it also removes one whole redundant copy of the
   file's bytes. Falls back to the previous `read()`-based incremental
   copy for anything `mmap` does not apply to (a pipe, a FIFO, process
   substitution, stdin, an empty file, or an `mmap()` call that itself
   fails).

A **third fix attempt was tried, measured, and reverted** specifically
because "prevent OOM, keep the footprint small" turned out to have a
sharp edge worth recording: capping `compute_maxrun`/`compute_next_prevmatch`
above a size budget and falling back to the plain scan already used when
either table is `NULL` seemed like an obvious bounded-memory safety valve.
Measured directly against a real 200MB non-matching search, it was worse
than doing nothing: both tables are needed together to keep this pattern
shape (`x*y`-style, unbounded quantifier followed by a required literal
that never occurs) at linear time; disabling either one alone reintroduces
the `O(n^2)` behavior they exist to fix, and `O(n^2)` at `n` in the hundreds
of millions does not finish in any practical amount of time. A fast,
diagnosable allocation failure (see below) is a better failure mode than a
silent, effectively-unbounded hang, so the cap was removed; these two
tables are allocated unconditionally again. There is no way to get both
bounded memory and linear time out of this technique for this pattern
shape; only `concept.md` Section 7.2's actual streaming automaton (still
unimplemented) gets both at once, by construction, which is why it remains
the correct long-term fix for this axis specifically.

**Failing safely.** Every allocation on the input-proportional paths above
(`build_matbuf`, `compute_maxrun`, `compute_next_prevmatch`,
`Input_from_file`, the UTF-8 decode arrays) is now checked; a failure
returns a `PatternError`/`-1` through the ordinary error path instead of
crashing on a `NULL` dereference, which several of them did before this
was audited (found by deliberately reasoning through "what happens when
this specific `malloc` fails on a huge request", not by a tool). This does
not prevent an out-of-memory condition on a genuinely memory-constrained
machine; the operating system's OOM killer can still end the process for
an allocation this library made in good faith (`malloc`/`mmap` returning
`NULL`/`MAP_FAILED` is the case this library can catch; being killed by
the kernel before that happens is not something a userspace library can
intercept). Measured concretely on the machine this was developed on: a
200MB file search that previously crashed via the OOM killer now completes
successfully in about 6 seconds at roughly 1.7GB peak RSS; a 1GB file on
the same machine still exceeded what was available at the time. Both
numbers are specific to that machine's available memory at the time, not
a hard property of the library; the 8.4x multiplier above is what actually
determines the practical ceiling on a given machine (roughly
`available memory / 8.4` for search-family operations on a pattern using
`OP_REPEAT1`, more forgiving for `match`/`fullmatch` or for patterns
without a simple-atom quantifier at all).

### Pattern syntax supported

Literals; `.` (with `DOTALL`); character classes with ranges, negation, and
`\d \D \w \W \s \S`; `\b \B`; anchors `^ $ \A \Z` (with `MULTILINE`);
quantifiers `* + ? {m,n} {m,} {,n} {m}`, greedy and lazy; possessive
quantifiers `*+ ++ ?+ {m,n}+`; groups `(...) (?:...) (?P<name>...)`;
alternation `|`; backreferences `\1`-`\99`, `(?P=name)`, `\g<name>`,
`\g<N>`; lookahead `(?=...) (?!...)`; fixed-width lookbehind
`(?<=...) (?<!...)`; atomic groups `(?>...)`; comments `(?#...)`; global
inline flags `(?aiLmsux)` at the start of a pattern; escapes
`\n \r \t \f \v \a`, octal `\0`-prefixed escapes, `\xhh`, `\uxxxx`,
`\Uxxxxxxxx`; flags `IGNORECASE`, `MULTILINE`, `DOTALL`, `VERBOSE`, `ASCII`
(all with observable effect; see `docs/API.md` Section 2 for exactly what
each one does), plus `UNICODE`, `LOCALE`, and `DEBUG` (accepted for source
compatibility with Python, currently no-ops in this build: `LOCALE` because
this build commits to the "C" locale only, under which `\w`/`\b`/`\B`
classify only ASCII letters and digits regardless of the flag, verified
directly against both the C standard and a real CPython interpreter).

Rejected at compile time with a clear `PatternError`, rather than
mis-parsed: conditional groups `(?(id)yes|no)`, scoped inline flags
`(?flags:...)`, `\N{NAME}` named code points, and POSIX bracket classes
`[:alpha:]` (which are not part of Python `re` at all, `concept.md` 14.3).
Variable-width lookbehind is also rejected at compile time, matching
CPython.

### Operations supported

`Pattern_match/fullmatch/search/finditer/findall/split/sub/subn/free`,
`Pattern_groupindex_lookup`,
`Match_group/start/end/span/start_byte/end_byte/span_byte/free`,
`re_compile/match/fullmatch/search/finditer/findall/split/sub/subn/escape/purge`,
`PatternError_free`. See `regexx.h` for exact signatures, `docs/API.md` for
the full reference (return values, memory ownership, exact Python
correspondence for each one), and `concept.md` Section 9 for the naming
convention they follow.

### Known deviations from `concept.md` and from CPython, beyond the items above

- `\w`, `\s`, `IGNORECASE` case folding, and `\d` in `UTF8` mode are backed
  by glibc's `wctype.h` functions under the `C.utf8` locale, not by a
  hand-generated Unicode table (`concept.md` 13.3 anticipated a reduced
  static table; using the C library's own tables turned out to be simpler
  and more complete, at the cost of depending on the platform's Unicode
  version rather than a pinned one). Concretely verified, not just
  theoretical: U+00A0 (NO-BREAK SPACE) is in Unicode's `White_Space`
  property, so CPython's `\s` matches it, but glibc's `iswspace()` under
  `C.utf8` does not, so this build's `\s` does not either. Found by the
  large combinatorial test expansion (`tests/cases.py` Category F,
  `tests/TEST_PLAN.md`), not anticipated in advance; recorded here rather
  than patched, since hand-patching individual code points would start
  down the path of maintaining an ad hoc table this design deliberately
  avoided by delegating to `wctype.h` in the first place. A second,
  same-class instance was found by the later Category I expansion:
  fullwidth digits (`U+FF10`-`U+FF19`, Unicode category `Nd`) match
  CPython's `\d` but not glibc's `iswdigit()` under `C.utf8` either.
- `lastindex`/`lastgroup` report the highest-numbered capturing group that
  participated in the match, which coincides with CPython's "most recently
  closed group" rule for straightforward patterns but can differ from it in
  pathological cases (nested alternation re-executing a lower-numbered group
  after a higher one). Not exercised by the test suite; documented here
  rather than silently accepted.
- `Match_free`, `PatternError_free`, `Input_from_buffer`, `Input_from_file`,
  and `Input_free` have no Python counterpart and are not mentioned in
  `concept.md`'s API surface; they exist because C has no garbage collector.
  `Pattern_sub`/`Pattern_subn` take the replacement template and the
  callback as two separate parameters rather than one polymorphic argument,
  for the same reason (`concept.md` 9.4 already anticipates and justifies
  this one).
- Python's `Match.start(group)`/`.end(group)` raise `IndexError` for an
  invalid group number and return `-1` only for a valid group that did not
  participate; `Match_start`/`Match_end` return `-1` for both cases, since C
  has no exception to raise. `Match_group` does distinguish them (`-1` for
  no such group, `0` for an unparticipated one), see `docs/API.md` Section
  3.23.
- `Pattern.groupindex` has no enumeration function in this build, only
  `Pattern_groupindex_lookup(pattern, name)`; there is no way to list every
  name a compiled pattern defines without already knowing what to look for.

## Building

Requires a C11 compiler and, for the test suite, Python 3 (used only to
generate ground truth from CPython's own `re` module, `concept.md` Section
11; the library itself has no runtime dependency beyond the C standard
library and `libc`'s `wctype.h`/`locale.h`).

```sh
make            # builds libregexx.a and the rxgrep example
make test       # regenerates tests/generated_tests.c from Python `re`
                # ground truth and runs the full suite
make check      # same, under AddressSanitizer + UndefinedBehaviorSanitizer
make clean
```

`make install` installs `libregexx.a` and `regexx.h` under `PREFIX`
(default `/usr/local`).

## Using the library

```c
#include "regexx.h"
#include <string.h>

const char *pattern = "(\\w+)@(\\w+)";
Pattern *pat = re_compile(pattern, strlen(pattern), UTF8, NULL);
Input *in = Input_from_buffer((const uint8_t *)"user@host", strlen("user@host"));

Match m;
if (Pattern_search(pat, in, 0, -1, &m) == 1) {
    const char *g; size_t glen;
    Match_group(&m, NULL, 1, &g, &glen);   /* g/glen -> "user" */
    Match_free(&m);
}

char *out; size_t outlen;
Pattern_sub(pat, in, "\\2@\\1", NULL, NULL, 0, &out, &outlen); /* "host@user" */
free(out);

Pattern_free(pat);
Input_free(in);
```

`flags` to `re_compile` combine a data mode, exactly one of `BINARY`,
`ASCII`, or `UTF8` (`concept.md` 9.3), with any of the Python-named flags
(`IGNORECASE`, `MULTILINE`, `DOTALL`, `VERBOSE`, `ASCII` as a flag also
forces ASCII-only `\w`/`\s`/`\d` inside `UTF8` mode, `LOCALE`, `DEBUG`).

## Example: rxgrep

`examples/rxgrep.c` is a small grep-like program built on the library,
demonstrating all three data modes and both the matching and substitution
API:

```sh
./rxgrep -in 'hello' file.txt        # case-insensitive, line numbers
./rxgrep -m utf8 -o '\w+' file.txt   # print every UTF-8 word, one per line
./rxgrep -c 'error' log.txt          # count matching lines
./rxgrep -m binary 'a.c' data.bin    # match raw bytes, embedded NUL included
./rxgrep -m utf8 --sub 'REDACTED' '\d{3}-\d{4}' file.txt
```

Run `./rxgrep --help` for the full option list.

## Testing

`tests/cases.py` lists pattern/subject/operation triples, both hand-written
and, for most of the file, generated programmatically from combinations of
quantifiers, groups, backreferences, lookaround, flags, and encoding modes
(`tests/TEST_PLAN.md` records the exact category breakdown and why each one
exists). `tests/gen.py` computes every case's expected result with
CPython's own `re` module and writes `tests/generated_tests.c`, which is
then compiled against `regexx.c` and checked. This is a direct
implementation of the strategy `concept.md` Section 11 describes:
conformance is measured against what CPython actually does, not against a
re-derived reading of its documentation, at a scale (3,252 cases as of this
writing, `make test` reports the current exact count) large enough that it
has already found real defects this way, not only confirmed the absence of
ones anyone thought to write by hand (`tests/TEST_PLAN.md` "Result" names
all four: a C trigraph bug in the test generator itself, a real, previously
undocumented `Pattern_finditer`/`Pattern_split` empty-match defect, a wrong
ground truth for `ASCII` mode in the generator itself, and a real,
verified-wrong implementation of the `LOCALE` flag). `make check`
additionally runs the suite under AddressSanitizer and
UndefinedBehaviorSanitizer.

## License

MIT. See [`LICENSE`](LICENSE).