GeezerVison

Manual backport of #154867, cherry-picked from the 3.12 backport (#155166), with one 3.10-only change in the second commit.

Possessive quantifiers were added in 3.11, so (?:(?P=quote){2}|(?!(?P=quote)).)*+ does not compile here. The body is matched as an unrolled loop instead; since the quote character is a backreference, the "not a quote" atom needs a lookahead:

other = r'(?:(?!(?P=quote)).)*'
body = r'%s(?:(?P=quote){2}%s)*' % (other, other)

Every position matches only one of the two alternatives, and each iteration of the outer loop consumes the two quote characters, so it does not backtrack.

Checked against the 3.11 version of the file: no difference in the sniffed dialect on 90000 samples, and the same timings — sniffing '"abcdefghijklmnopqrstuvwxyz"\n' * n with delimiters=',:|\t' takes 0.147, 0.298 and 0.602 s for n = 2500, 5000 and 10000, against 0.142, 0.286 and 0.569 s on 3.11.

• Issue: csv.Sniffer().sniff() Hangs on Single-Column Data #98820