A real POSIX ERE matcher: a hand-written parser, a Thompson automaton run as a parallel simulation for leftmost-longest, and a backtracker for leftmost-first, over one shared syntax tree — so the same pattern string can be read two ways, matched two ways, and the four answers put side by side.
Try any pattern: https://dev48.infy.uk/solve/day77-posix-ere-matcher.html
The most-repeated piece of regex advice
Everyone knows the first rule of alternation: branches are tried in order, so put the longer one first. /a|ab/ against ab matches a in JavaScript, Python and Perl. In grep, sed and awk it matches ab — and so does /ab|a/.
Over 12,000 swapped pairs:
| branches reordered | |
|---|---|
| POSIX answer changes | 0 times |
| ECMAScript answer changes | 1,860 times — 15.50% |
The single most-repeated piece of regex advice is load-bearing in one family and a no-op in the other, with no error either way.
The finding is positional
a*? is three characters. To ECMAScript it is one lazy quantifier and the answer is the empty string. To POSIX it is two operators, (a*)?, and the answer is the whole run. Opposite extremes from an identical spelling, decided only by which binary you piped it through.
Inside a bracket expression POSIX has no escape character at all, so [\.] is a class of two members — backslash and dot. grep and sed agree; gawk, shipped by the same project, does not.
Of 40 frontier patterns, five real implementations return the same answer for only 19. Twenty split two ways, and ^* splits three:
^* sed: syntax error
grep: matches
gawk: no match
The divergence is an identity, not chaos
Over 16,800 generated cases the two policies:
- always agree on whether there is a match
- always start at the same index
- and the ECMAScript match is a prefix of the POSIX match — 16,800 out of 16,800
They differ only in how much, on 1,758 of them — 10.46%. That is a far more useful thing to know than "the engines differ": you can predict the direction every time.
Verifier 524,305 asserts, page self-check 18,971, 0 failures.
