Optimise index so that if the big string is ISO-8859-1 but the little

string is UTF-8, it tries to downgrade the little string, rather than
upgrade the big string. For half-meg big strings this is a fourfold
speed gain.

p4raw-id: //depot/perl@27113

diff --git a/pp.c b/pp.c
index 96d5ef6..e9f159b 100644 (file)
--- a/pp.c
+++ b/pp.c
@@ -3117,8 +3117,8 @@ PP(pp_index)
     const char *tmps2;
     STRLEN biglen;
     const I32 arybase = PL_curcop->cop_arybase;
-    int big_utf8;
-    int little_utf8;
+    bool big_utf8;
+    bool little_utf8;
 
     if (MAXARG < 3)
        offset = 0;
@@ -3130,22 +3130,43 @@ PP(pp_index)
     little_utf8 = DO_UTF8(little);
     if (big_utf8 ^ little_utf8) {
        /* One needs to be upgraded.  */
-       SV * const bytes = little_utf8 ? big : little;
-       STRLEN len;
-       const char * const p = SvPV_const(bytes, len);
-
-       temp = newSVpvn(p, len);
+       if (little_utf8 && !PL_encoding) {
+           /* Well, maybe instead we might be able to downgrade the small
+              string?  */
+           STRLEN little_len;
+           const U8 * const little_pv = (U8*) SvPV_const(little, little_len);
+           char * const pv = (char*)bytes_from_utf8(little_pv, &little_len,
+                                                    &little_utf8);
+           if (little_utf8) {
+               /* If the large string is ISO-8859-1, and it's not possible to
+                  convert the small string to ISO-8859-1, then there is no
+                  way that it could be found anywhere by index.  */
+               retval = -1;
+               goto fail;
+           }
 
-       if (PL_encoding) {
-           sv_recode_to_utf8(temp, PL_encoding);
-       } else {
-           sv_utf8_upgrade(temp);
-       }
-       if (little_utf8) {
-           big = temp;
-           big_utf8 = TRUE;
+           /* At this point, pv is a malloc()ed string. So donate it to temp
+              to ensure it will get free()d  */
+           little = temp = newSV(0);
+           sv_usepvn(temp, pv, little_len);
        } else {
-           little = temp;
+           SV * const bytes = little_utf8 ? big : little;
+           STRLEN len;
+           const char * const p = SvPV_const(bytes, len);
+
+           temp = newSVpvn(p, len);
+
+           if (PL_encoding) {
+               sv_recode_to_utf8(temp, PL_encoding);
+           } else {
+               sv_utf8_upgrade(temp);
+           }
+           if (little_utf8) {
+               big = temp;
+               big_utf8 = TRUE;
+           } else {
+               little = temp;
+           }
        }
     }
     if (big_utf8 && offset > 0)
@@ -3158,12 +3179,14 @@ PP(pp_index)
     if (!(tmps2 = fbm_instr((unsigned char*)tmps + offset,
       (unsigned char*)tmps + biglen, little, 0)))
        retval = -1;
-    else
+    else {
        retval = tmps2 - tmps;
-    if (retval > 0 && big_utf8)
-       sv_pos_b2u(big, &retval);
+       if (big_utf8)
+           sv_pos_b2u(big, &retval);
+    }
     if (temp)
        SvREFCNT_dec(temp);
+ fail:
     PUSHi(retval + arybase);
     RETURN;
 }