struct regnode_charclass_class *start_class;
} scan_data_t;
+/* The below is perhaps overboard, but this allows us to save a test at the
+ * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
+ * and 'a' differ by a single bit; the same with the upper and lower case of
+ * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
+ * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
+ * then inverts it to form a mask, with just a single 0, in the bit position
+ * where the upper- and lowercase differ. XXX There are about 40 other
+ * instances in the Perl core where this micro-optimization could be used.
+ * Should decide if maintenance cost is worse, before changing those
+ *
+ * Returns a boolean as to whether or not 'v' is either a lowercase or
+ * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
+ * compile-time constant, the generated code is better than some optimizing
+ * compilers figure out, amounting to a mask and test. The results are
+ * meaningless if 'c' is not one of [A-Za-z] */
+#define isARG2_lower_or_UPPER_ARG1(c, v) \
+ (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
+
/*
* Forward declarations for pregcomp()'s friends.
*/
* have to find at least two characters for a multi-fold */
const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
- /* The below is perhaps overboard, but this allows us to save a
- * test each time through the loop at the expense of a mask. This
- * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
- * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
- * are 64. This uses an exclusive 'or' to find that bit and then
- * inverts it to form a mask, with just a single 0, in the bit
- * position where 'S' and 's' differ. */
- const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
- const U8 s_masked = 's' & S_or_s_mask;
-
while (s < upper) {
int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
if (! len) { /* Not a multi-char fold. */
}
if (len == 2
- && ((*s & S_or_s_mask) == s_masked)
- && ((*(s+1) & S_or_s_mask) == s_masked))
+ && isARG2_lower_or_UPPER_ARG1('s', *s)
+ && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
{
/* EXACTF nodes need to know that the minimum length
defchar: {
STRLEN len = 0;
- UV ender;
+ UV ender = 0;
char *p;
char *s;
#define MAX_NODE_STRING_SIZE 127
char *s0;
U8 upper_parse = MAX_NODE_STRING_SIZE;
STRLEN foldlen;
- U8 node_type;
+ U8 node_type = compute_EXACTish(pRExC_state);
bool next_is_quantifier;
char * oldp = NULL;
+ /* We can convert EXACTF nodes to EXACTFU if they contain only
+ * characters that match identically regardless of the target
+ * string's UTF8ness. The reason to do this is that EXACTF is not
+ * trie-able, EXACTFU is. (We don't need to figure this out until
+ * pass 2) */
+ bool maybe_exactfu = node_type == EXACTF && PASS2;
+
/* If a folding node contains only code points that don't
* participate in folds, it can be changed into an EXACT node,
* which allows the optimizer more things to look for */
bool maybe_exact;
- ender = 0;
- node_type = compute_EXACTish(pRExC_state);
ret = reg_node(pRExC_state, node_type);
/* In pass1, folded, we use a temporary buffer instead of the
/* We do the EXACTFish to EXACT node only if folding, and not if in
* locale, as whether a character folds or not isn't known until
- * runtime */
- maybe_exact = FOLD && ! LOC;
+ * runtime. (And we don't need to figure this out until pass 2) */
+ maybe_exact = FOLD && ! LOC && PASS2;
/* XXX The node can hold up to 255 bytes, yet this only goes to
* 127. I (khw) do not know why. Keeping it somewhat less than
|| (node_type == EXACTFU
&& ender == LATIN_SMALL_LETTER_SHARP_S)))
{
+ if (IS_IN_SOME_FOLD_L1(ender)) {
+ maybe_exact = FALSE;
+
+ /* See if the character's fold differs between /d and
+ * /u. This includes the multi-char fold SHARP S to
+ * 'ss' */
+ if (maybe_exactfu
+ && (PL_fold[ender] != PL_fold_latin1[ender]
+ || ender == LATIN_SMALL_LETTER_SHARP_S
+ || (len > 0
+ && isARG2_lower_or_UPPER_ARG1('s', ender)
+ && isARG2_lower_or_UPPER_ARG1('s', *(s-1)))))
+ {
+ maybe_exactfu = FALSE;
+ }
+ }
*(s++) = (char) ender;
- maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
}
else { /* UTF */
* do any better */
if (len == 0) {
len = full_len;
+
+ /* If the node ends in an 's' we make sure it stays EXACTF,
+ * as if it turns into an EXACTFU, it could later get
+ * joined with another 's' that would then wrongly match
+ * the sharp s */
+ if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
+ {
+ maybe_exactfu = FALSE;
+ }
} else {
/* Here, the node does contain some characters that aren't
if (len == 0) {
OP(ret) = NOTHING;
}
- else{
-
- /* If 'maybe_exact' is still set here, means there are no
- * code points in the node that participate in folds */
- if (FOLD && maybe_exact) {
- OP(ret) = EXACT;
+ else {
+ if (FOLD) {
+ /* If 'maybe_exact' is still set here, means there are no
+ * code points in the node that participate in folds;
+ * similarly for 'maybe_exactfu' and code points that match
+ * differently depending on UTF8ness of the target string
+ * */
+ if (maybe_exact) {
+ OP(ret) = EXACT;
+ }
+ else if (maybe_exactfu) {
+ OP(ret) = EXACTFU;
+ }
}
alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
}
sv_catpvs(sv, "{unicode_all}");
else if (ANYOF_NONBITMAP(o)) {
SV *lv; /* Set if there is something outside the bit map. */
- SV * sw;
bool byte_output = FALSE; /* If something in the bitmap has been
output */
}
/* Get the stuff that wasn't in the bitmap */
- sw = regclass_swash(prog, o, FALSE, &lv, NULL);
+ (void) regclass_swash(prog, o, FALSE, &lv, NULL);
if (lv && lv != &PL_sv_undef) {
char *s = savesvpv(lv);
char * const origs = s;