45001: Fix zero-length matches with ${...%...}

As the shortest match is preferred, zero-length matches beat any other.

ChangeLog

@@ -1,3 +1,9 @@
+2019-12-12  Peter Stephenson  <p.stephenson@samsung.com>
+
+	* 45001: Src/glob.c, Test/D04parameter.ztst: ${...%...} didn't
+	test for zero-length matches, which should be preferred where
+	possible.
+
 2019-12-11  Martijn Dekker  <martijn@inlv.org>
 
 	* 45004: Src/builtin.c, Src/compat.c, Src/exec.c, Src/glob.c,

Src/glob.c

@@ -2909,6 +2909,12 @@ igetmatch(char **sp, Patprog p, int fl, int n, char *replstr,
 	     */
 	    mb_charinit();
 	    tmatch = NULL;
+	    set_pat_start(p, l);
+	    if (pattrylen(p, send, 0, 0, &patstralloc, umltot) &&
+		!--n) {
+		*sp = get_match_ret(&imd, umltot, umltot);
+		return 1;
+	    }
 	    for (ioff = 0, t = s, umlen = umltot; t < send; ioff++) {
 		set_pat_start(p, t-s);
 		if (pattrylen(p, t, umlen, 0, &patstralloc, ioff))

Test/D04parameter.ztst

@@ -2534,3 +2534,26 @@ F:behavior, see http://austingroupbugs.net/view.php?id=888
 0:Global variables are not trashed by "foo=bar builtin" (regression test)
 >function-value
 >global-value
+
+ foo=pws
+ print ${foo%*}
+0:Smallest match at end can match zero-length string
+>pws
+
+ foo=pws
+ print ${foo%?}
+0:Smallest match at end with a character always matches one
+>pw
+
+ setopt extendedglob
+ foo=pws
+ print ${foo%s#}
+ print ${foo%%s#}
+0:Smallest / largest match with non-trivial closure
+>pws
+>pw
+
+ foo=pws
+ print ${foo%%*}
+0:Largest match at end matches entire string
+>