In wcmatch, a Python glob and fnmatch library, a star directly before an extended star group, as in a**(b), was not read as * followed by the group *(b). I fixed it in PR #246, merged about half an hour after opening.

When it happens

from wcmatch import fnmatch, glob
fnmatch.fnmatch('abb', 'a**(b)', flags=fnmatch.EXTMATCH)
# False   want True
glob.globmatch('b', '**(a|b)b', flags=glob.GLOBSTAR | glob.EXTGLOB)
# False   want True

Bash reads **(b) as * followed by the group *(b). Release 11.1 returned False for both, and a**(b) matched only names ending in the literal text (b).

How wcmatch reads a run of stars

WcParse in wcmatch/_wcparse.py reads a pattern in order. With GLOBSTAR, a ** at the start of a segment may be a globstar; with GLOBSTARLONG, a third star makes ***. A loop then collapses the remaining stars into one, because ** and * match the same text.

Why the gap existed

All three places took the star that opens the group as part of the run. Before 11.0 the loop ran only at the start of a segment, so a**(b) worked up to 10.2.1. No file under tests held a **( pattern.

The fix

When extended groups are on, each place now stops at a star followed by (, so that star stays for the group. The wcmatch/_wcparse.py diff, without context lines:

@@ -1439 +1439 @@ class WcParse(Generic[AnyStr]):
-                if c != '*':
+                if c != '*' or (self.extend and i.match(RE_EXT_GROUP, False)):
@@ -1448 +1448 @@ class WcParse(Generic[AnyStr]):
-                    if c != '*':
+                    if c != '*' or (self.extend and i.match(RE_EXT_GROUP, False)):
@@ -1494 +1494 @@ class WcParse(Generic[AnyStr]):
-            # Consume duplicate starts
+            # Consume duplicate starts, but leave a star that opens an extended group
@@ -1497 +1497 @@ class WcParse(Generic[AnyStr]):
-                while c == '*':
+                while c == '*' and not (self.extend and i.match(RE_EXT_GROUP, False)):

**(x) now compiles like * followed by *(x), so **(a|aa)c can backtrack on a long non-matching name, as *+(a|aa)c already does.

The test

Eight rows joined existing tables: three in TestFnMatch.cases and five in TestGlobMatch.cases, two of them with glob.GL, such as ***(a|b)b. Without the fix, pytest tests/test_fnmatch.py tests/test_globmatch.py gave 8 failed and 866 passed; with it, 874 passed.

What the maintainer did

@facelessuser wrote “Another good find.” and squash-merged it. The diff did not change.

The lesson

  • A loop that collapses repeated characters must stop where the next character changes meaning. A * before ( opens a group; it is not a repeat.
  • For a regression, name the last release that worked and what the next one widened. Here a loop limited to the start of a segment before 11.0 now runs elsewhere.
  • A fix that compiles a pattern like a known one inherits its costs. Say so: **(a|aa)c can now backtrack as *+(a|aa)c does.

If a parser of yours collapses repeated characters, reply on X.