+
+ } /* End of loop through literal characters */
+
+ /* Here we have either exhausted the input or ran out of room in
+ * the node. (If we encountered a character that can't be in the
+ * node, transfer is made directly to <loopdone>, and so we
+ * wouldn't have fallen off the end of the loop.) In the latter
+ * case, we artificially have to split the node into two, because
+ * we just don't have enough space to hold everything. This
+ * creates a problem if the final character participates in a
+ * multi-character fold in the non-final position, as a match that
+ * should have occurred won't, due to the way nodes are matched,
+ * and our artificial boundary. So back off until we find a non-
+ * problematic character -- one that isn't at the beginning or
+ * middle of such a fold. (Either it doesn't participate in any
+ * folds, or appears only in the final position of all the folds it
+ * does participate in.) A better solution with far fewer false
+ * positives, and that would fill the nodes more completely, would
+ * be to actually have available all the multi-character folds to
+ * test against, and to back-off only far enough to be sure that
+ * this node isn't ending with a partial one. <upper_parse> is set
+ * further below (if we need to reparse the node) to include just
+ * up through that final non-problematic character that this code
+ * identifies, so when it is set to less than the full node, we can
+ * skip the rest of this */
+ if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
+
+ const STRLEN full_len = len;
+
+ assert(len >= MAX_NODE_STRING_SIZE);
+
+ /* Here, <s> points to the final byte of the final character.
+ * Look backwards through the string until find a non-
+ * problematic character */
+
+ if (! UTF) {
+
+ /* These two have no multi-char folds to non-UTF characters
+ */
+ if (ASCII_FOLD_RESTRICTED || LOC) {
+ goto loopdone;
+ }
+
+ while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
+ len = s - s0 + 1;
+ }
+ else {
+ if (! PL_NonL1NonFinalFold) {
+ PL_NonL1NonFinalFold = _new_invlist_C_array(
+ NonL1_Perl_Non_Final_Folds_invlist);
+ }
+
+ /* Point to the first byte of the final character */
+ s = (char *) utf8_hop((U8 *) s, -1);
+
+ while (s >= s0) { /* Search backwards until find
+ non-problematic char */
+ if (UTF8_IS_INVARIANT(*s)) {
+
+ /* There are no ascii characters that participate
+ * in multi-char folds under /aa. In EBCDIC, the
+ * non-ascii invariants are all control characters,
+ * so don't ever participate in any folds. */
+ if (ASCII_FOLD_RESTRICTED
+ || ! IS_NON_FINAL_FOLD(*s))
+ {
+ break;
+ }
+ }
+ else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
+
+ /* No Latin1 characters participate in multi-char
+ * folds under /l */
+ if (LOC
+ || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
+ *s, *(s+1))))
+ {
+ break;
+ }
+ }
+ else if (! _invlist_contains_cp(
+ PL_NonL1NonFinalFold,
+ valid_utf8_to_uvchr((U8 *) s, NULL)))
+ {
+ break;
+ }
+
+ /* Here, the current character is problematic in that
+ * it does occur in the non-final position of some
+ * fold, so try the character before it, but have to
+ * special case the very first byte in the string, so
+ * we don't read outside the string */
+ s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
+ } /* End of loop backwards through the string */
+
+ /* If there were only problematic characters in the string,
+ * <s> will point to before s0, in which case the length
+ * should be 0, otherwise include the length of the
+ * non-problematic character just found */
+ len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
+ }
+
+ /* Here, have found the final character, if any, that is
+ * non-problematic as far as ending the node without splitting
+ * it across a potential multi-char fold. <len> contains the
+ * number of bytes in the node up-to and including that
+ * character, or is 0 if there is no such character, meaning
+ * the whole node contains only problematic characters. In
+ * this case, give up and just take the node as-is. We can't
+ * do any better */
+ if (len == 0) {
+ len = full_len;
+ } else {
+
+ /* Here, the node does contain some characters that aren't
+ * problematic. If one such is the final character in the
+ * node, we are done */
+ if (len == full_len) {
+ goto loopdone;
+ }
+ else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
+
+ /* If the final character is problematic, but the
+ * penultimate is not, back-off that last character to
+ * later start a new node with it */
+ p = oldp;
+ goto loopdone;
+ }
+
+ /* Here, the final non-problematic character is earlier
+ * in the input than the penultimate character. What we do
+ * is reparse from the beginning, going up only as far as
+ * this final ok one, thus guaranteeing that the node ends
+ * in an acceptable character. The reason we reparse is
+ * that we know how far in the character is, but we don't
+ * know how to correlate its position with the input parse.
+ * An alternate implementation would be to build that
+ * correlation as we go along during the original parse,
+ * but that would entail extra work for every node, whereas
+ * this code gets executed only when the string is too
+ * large for the node, and the final two characters are
+ * problematic, an infrequent occurrence. Yet another
+ * possible strategy would be to save the tail of the
+ * string, and the next time regatom is called, initialize
+ * with that. The problem with this is that unless you
+ * back off one more character, you won't be guaranteed
+ * regatom will get called again, unless regbranch,
+ * regpiece ... are also changed. If you do back off that
+ * extra character, so that there is input guaranteed to
+ * force calling regatom, you can't handle the case where
+ * just the first character in the node is acceptable. I
+ * (khw) decided to try this method which doesn't have that
+ * pitfall; if performance issues are found, we can do a
+ * combination of the current approach plus that one */
+ upper_parse = len;
+ len = 0;
+ s = s0;
+ goto reparse;
+ }
+ } /* End of verifying node ends with an appropriate char */
+
https://perl5.git.perl.org / perl5.git / blobdiff