regcomp.c: Fix bug in /[A-Z]/i

This also fixes /[a-z]/i.

When not under /i, these two ranges alone in a bracketed character class
can be optimized into qr/[[:upper:]]/a and qr/[[:lower:]]/a respectively.
This optimization saves space in the pattern (as no bitmap is needed),
and I think it executes faster.  But this optimization has to be
foregone under /i (unless /a is also present) because otherwise
certain non-ASCII characters such as the \N{KELVIN SIGN} don't match,
and they should.

1 files changed, 6 insertions, 1 deletions

diff --git a/regcomp.c b/regcomp.c
index ef529279f3..2a75b240e3 100644
--- a/regcomp.c
+++ b/regcomp.c
@@ -14895,7 +14895,11 @@ S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
                     op = POSIXA;
                 }
             }
-            else if (prevvalue == 'A') {
+            else if (AT_LEAST_ASCII_RESTRICTED || ! FOLD) {
+                /* We can optimize A-Z or a-z, but not if they could match
+                 * something like the KELVIN SIGN under /i (/a means they
+                 * can't) */
+            if (prevvalue == 'A') {
                 if (value == 'Z'
 #ifdef EBCDIC
                     && literal_endpoint == 2
@@ -14915,6 +14919,7 @@ S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
                     op = POSIXA;
                 }
             }
+            }
         }
 
         /* Here, we have changed <op> away from its initial value iff we found