/* sv.c
*
* Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
- * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
+ * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 by Larry Wall
+ * and others
*
* You may distribute under the terms of either the GNU General Public
* License or the Artistic License, as specified in the README file.
use the Encode extension for that.
=cut
+
+The grow version is currently not externally documented. It adds a parameter,
+extra, which is the number of unused bytes the string of 'sv' is guaranteed to
+have free after it upon return. This allows the caller to reserve extra space
+that it intends to fill, to avoid extra grows.
+
+Also externally undocumented for the moment is the flag SV_FORCE_UTF8_UPGRADE,
+which can be used to tell this function to not first check to see if there are
+any characters that are different in UTF-8 (variant characters) which would
+force it to allocate a new string to sv, but to assume there are. Typically
+this flag is used by a routine that has already parsed the string to find that
+there are such characters, and passes this information on so that the work
+doesn't have to be repeated.
+
+(One might think that the calling routine could pass in the position of the
+first such variant, so it wouldn't have to be found again. But that is not the
+case, because typically when the caller is likely to use this flag, it won't be
+calling this routine unless it finds something that won't fit into a byte.
+Otherwise it tries to not upgrade and just use bytes. But some things that
+do fit into a byte are variants in utf8, and the caller may not have been
+keeping track of these.)
+
+If the routine itself changes the string, it adds a trailing NUL. Such a NUL
+isn't guaranteed due to having other routines do the work in some input cases,
+or if the input is already flagged as being in utf8.
+
+The speed of this could perhaps be improved for many cases if someone wanted to
+write a fast function that counts the number of variant characters in a string,
+especially if it could return the position of the first one.
+
*/
STRLEN
-Perl_sv_utf8_upgrade_flags(pTHX_ register SV *const sv, const I32 flags)
+Perl_sv_utf8_upgrade_flags_grow(pTHX_ register SV *const sv, const I32 flags, STRLEN extra)
{
dVAR;
- PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS;
+ PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
if (sv == &PL_sv_undef)
return 0;
STRLEN len = 0;
if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
(void) sv_2pv_flags(sv,&len, flags);
- if (SvUTF8(sv))
+ if (SvUTF8(sv)) {
+ if (extra) SvGROW(sv, SvCUR(sv) + extra);
return len;
+ }
} else {
(void) SvPV_force(sv,len);
}
}
if (SvUTF8(sv)) {
+ if (extra) SvGROW(sv, SvCUR(sv) + extra);
return SvCUR(sv);
}
sv_force_normal_flags(sv, 0);
}
- if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING))
+ if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING)) {
sv_recode_to_utf8(sv, PL_encoding);
- else { /* Assume Latin-1/EBCDIC */
+ if (extra) SvGROW(sv, SvCUR(sv) + extra);
+ return SvCUR(sv);
+ }
+
+ if (SvCUR(sv) > 0) { /* Assume Latin-1/EBCDIC */
/* This function could be much more efficient if we
* had a FLAG in SVs to signal if there are any variant
* chars in the PV. Given that there isn't such a flag
- * make the loop as fast as possible. */
- const U8 * const s = (U8 *) SvPVX_const(sv);
- const U8 * const e = (U8 *) SvEND(sv);
- const U8 *t = s;
+ * make the loop as fast as possible (although there are certainly ways
+ * to speed this up, eg. through vectorization) */
+ U8 * s = (U8 *) SvPVX_const(sv);
+ U8 * e = (U8 *) SvEND(sv);
+ U8 *t = s;
+ STRLEN two_byte_count = 0;
+ if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
+
+ /* See if really will need to convert to utf8. We mustn't rely on our
+ * incoming SV being well formed and having a trailing '\0', as certain
+ * code in pp_formline can send us partially built SVs. */
+
while (t < e) {
const U8 ch = *t++;
- /* Check for variant */
- if (!NATIVE_IS_INVARIANT(ch)) {
- STRLEN len = SvCUR(sv);
- /* *Currently* bytes_to_utf8() adds a '\0' after every string
- it converts. This isn't documented. It's not clear if it's
- a bad thing to be doing, and should be changed to do exactly
- what the documentation says. If so, this code will have to
- be changed.
- As is, we mustn't rely on our incoming SV being well formed
- and having a trailing '\0', as certain code in pp_formline
- can send us partially built SVs. */
- U8 * const recoded = bytes_to_utf8((U8*)s, &len);
-
- SvPV_free(sv); /* No longer using what was there before. */
- SvPV_set(sv, (char*)recoded);
- SvCUR_set(sv, len);
- SvLEN_set(sv, len + 1); /* No longer know the real size. */
- break;
- }
+ if (NATIVE_IS_INVARIANT(ch)) continue;
+
+ t--; /* t already incremented; re-point to first variant */
+ two_byte_count = 1;
+ goto must_be_utf8;
}
- /* Mark as UTF-8 even if no variant - saves scanning loop */
+
+ /* utf8 conversion not needed because all are invariants. Mark as
+ * UTF-8 even if no variant - saves scanning loop */
SvUTF8_on(sv);
+ return SvCUR(sv);
+
+must_be_utf8:
+
+ /* Here, the string should be converted to utf8, either because of an
+ * input flag (two_byte_count = 0), or because a character that
+ * requires 2 bytes was found (two_byte_count = 1). t points either to
+ * the beginning of the string (if we didn't examine anything), or to
+ * the first variant. In either case, everything from s to t - 1 will
+ * occupy only 1 byte each on output.
+ *
+ * There are two main ways to convert. One is to create a new string
+ * and go through the input starting from the beginning, appending each
+ * converted value onto the new string as we go along. It's probably
+ * best to allocate enough space in the string for the worst possible
+ * case rather than possibly running out of space and having to
+ * reallocate and then copy what we've done so far. Since everything
+ * from s to t - 1 is invariant, the destination can be initialized
+ * with these using a fast memory copy
+ *
+ * The other way is to figure out exactly how big the string should be
+ * by parsing the entire input. Then you don't have to make it big
+ * enough to handle the worst possible case, and more importantly, if
+ * the string you already have is large enough, you don't have to
+ * allocate a new string, you can copy the last character in the input
+ * string to the final position(s) that will be occupied by the
+ * converted string and go backwards, stopping at t, since everything
+ * before that is invariant.
+ *
+ * There are advantages and disadvantages to each method.
+ *
+ * In the first method, we can allocate a new string, do the memory
+ * copy from the s to t - 1, and then proceed through the rest of the
+ * string byte-by-byte.
+ *
+ * In the second method, we proceed through the rest of the input
+ * string just calculating how big the converted string will be. Then
+ * there are two cases:
+ * 1) if the string has enough extra space to handle the converted
+ * value. We go backwards through the string, converting until we
+ * get to the position we are at now, and then stop. If this
+ * position is far enough along in the string, this method is
+ * faster than the other method. If the memory copy were the same
+ * speed as the byte-by-byte loop, that position would be about
+ * half-way, as at the half-way mark, parsing to the end and back
+ * is one complete string's parse, the same amount as starting
+ * over and going all the way through. Actually, it would be
+ * somewhat less than half-way, as it's faster to just count bytes
+ * than to also copy, and we don't have the overhead of allocating
+ * a new string, changing the scalar to use it, and freeing the
+ * existing one. But if the memory copy is fast, the break-even
+ * point is somewhere after half way. The counting loop could be
+ * sped up by vectorization, etc, to move the break-even point
+ * further towards the beginning.
+ * 2) if the string doesn't have enough space to handle the converted
+ * value. A new string will have to be allocated, and one might
+ * as well, given that, start from the beginning doing the first
+ * method. We've spent extra time parsing the string and in
+ * exchange all we've gotten is that we know precisely how big to
+ * make the new one. Perl is more optimized for time than space,
+ * so this case is a loser.
+ * So what I've decided to do is not use the 2nd method unless it is
+ * guaranteed that a new string won't have to be allocated, assuming
+ * the worst case. I also decided not to put any more conditions on it
+ * than this, for now. It seems likely that, since the worst case is
+ * twice as big as the unknown portion of the string (plus 1), we won't
+ * be guaranteed enough space, causing us to go to the first method,
+ * unless the string is short, or the first variant character is near
+ * the end of it. In either of these cases, it seems best to use the
+ * 2nd method. The only circumstance I can think of where this would
+ * be really slower is if the string had once had much more data in it
+ * than it does now, but there is still a substantial amount in it */
+
+ {
+ STRLEN invariant_head = t - s;
+ STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
+ if (SvLEN(sv) < size) {
+
+ /* Here, have decided to allocate a new string */
+
+ U8 *dst;
+ U8 *d;
+
+ Newx(dst, size, U8);
+
+ /* If no known invariants at the beginning of the input string,
+ * set so starts from there. Otherwise, can use memory copy to
+ * get up to where we are now, and then start from here */
+
+ if (invariant_head <= 0) {
+ d = dst;
+ } else {
+ Copy(s, dst, invariant_head, char);
+ d = dst + invariant_head;
+ }
+
+ while (t < e) {
+ const UV uv = NATIVE8_TO_UNI(*t++);
+ if (UNI_IS_INVARIANT(uv))
+ *d++ = (U8)UNI_TO_NATIVE(uv);
+ else {
+ *d++ = (U8)UTF8_EIGHT_BIT_HI(uv);
+ *d++ = (U8)UTF8_EIGHT_BIT_LO(uv);
+ }
+ }
+ *d = '\0';
+ SvPV_free(sv); /* No longer using pre-existing string */
+ SvPV_set(sv, (char*)dst);
+ SvCUR_set(sv, d - dst);
+ SvLEN_set(sv, size);
+ } else {
+
+ /* Here, have decided to get the exact size of the string.
+ * Currently this happens only when we know that there is
+ * guaranteed enough space to fit the converted string, so
+ * don't have to worry about growing. If two_byte_count is 0,
+ * then t points to the first byte of the string which hasn't
+ * been examined yet. Otherwise two_byte_count is 1, and t
+ * points to the first byte in the string that will expand to
+ * two. Depending on this, start examining at t or 1 after t.
+ * */
+
+ U8 *d = t + two_byte_count;
+
+
+ /* Count up the remaining bytes that expand to two */
+
+ while (d < e) {
+ const U8 chr = *d++;
+ if (! NATIVE_IS_INVARIANT(chr)) two_byte_count++;
+ }
+
+ /* The string will expand by just the number of bytes that
+ * occupy two positions. But we are one afterwards because of
+ * the increment just above. This is the place to put the
+ * trailing NUL, and to set the length before we decrement */
+
+ d += two_byte_count;
+ SvCUR_set(sv, d - s);
+ *d-- = '\0';
+
+
+ /* Having decremented d, it points to the position to put the
+ * very last byte of the expanded string. Go backwards through
+ * the string, copying and expanding as we go, stopping when we
+ * get to the part that is invariant the rest of the way down */
+
+ e--;
+ while (e >= t) {
+ const U8 ch = NATIVE8_TO_UNI(*e--);
+ if (UNI_IS_INVARIANT(ch)) {
+ *d-- = UNI_TO_NATIVE(ch);
+ } else {
+ *d-- = (U8)UTF8_EIGHT_BIT_LO(ch);
+ *d-- = (U8)UTF8_EIGHT_BIT_HI(ch);
+ }
+ }
+ }
+ }
}
+
+ /* Mark as UTF-8 even if no variant - saves scanning loop */
+ SvUTF8_on(sv);
return SvCUR(sv);
}
IoBOTTOM_NAME(dstr) = SAVEPV(IoBOTTOM_NAME(dstr));
break;
case SVt_PVAV:
- if (AvARRAY((const AV *)sstr)) {
+ /* avoid cloning an empty array */
+ if (AvARRAY((const AV *)sstr) && AvFILLp((const AV *)sstr) >= 0) {
SV **dst_ary, **src_ary;
SSize_t items = AvFILLp((const AV *)sstr) + 1;
else {
AvARRAY(MUTABLE_AV(dstr)) = NULL;
AvALLOC((const AV *)dstr) = (SV**)NULL;
+ AvMAX( (const AV *)dstr) = -1;
+ AvFILLp((const AV *)dstr) = -1;
}
break;
case SVt_PVHV:
PTR2UV(PL_watchok));
}
+ PL_registered_mros = hv_dup_inc(proto_perl->Iregistered_mros, param);
+
if (!(flags & CLONEf_KEEP_PTR_TABLE)) {
ptr_table_free(PL_ptr_table);
PL_ptr_table = NULL;
https://perl5.git.perl.org / perl5.git / blobdiff