PATCH: [perl #122460] Regression with (?{})(?=...) and pos

The commit that really broke this was
a0dd42312a1f26356d2fdf49656e45b77c2cefb5.  The blamed commit fixed some
typos that kept the earlier bad commit from having an effect.

Zero-length assertions, such as (?=...) obviously can match an empty
string.  The bad commit wrongly removed a setting of a flag to that
effect.  I didn't realize the implications at the time, thinking that
the newly-corrected AND logical operations would correctly set things
up, and that the previous code had been a workaround for the buggy AND.
The problem is that there is an overriding context such that whatever
the AND result, since this is in a 0-length assertion, the result will
match the empty string.  Thus it now differs from the pre-bad-commit
code which set to match the empty string only sometimes.  I'm pretty
sure it should be 'always'.  If I'm wrong, it doesn't mean a bug, but
only that an optimisation that could take place won't.

1 files changed, 6 insertions, 0 deletions

diff --git a/t/re/pat_advanced.t b/t/re/pat_advanced.t
index 82f0917655..986eb87de4 100644
--- a/t/re/pat_advanced.t
+++ b/t/re/pat_advanced.t
@@ -2346,6 +2346,12 @@ EOP
         is($word, 'раб', "Handles UTF8 trie correctly");
     }
 
+    { # [perl #122460]
+        my $a = "rdvark";
+        $a =~ /(?{})(?=[A-Za-z0-9_])a*?/g;
+        is (pos $a, 0, "optimizer correctly thinks (?=...) is 0-length");
+    }
+
     #
     # Keep the following tests last -- they may crash perl
     #