This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
optimize SV creation funcs in sv.c
[perl5.git] / sv.c
CommitLineData
a0d0e21e 1/* sv.c
79072805 2 *
1129b882 3 * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
83706693
RGS
4 * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 by Larry Wall
5 * and others
79072805
LW
6 *
7 * You may distribute under the terms of either the GNU General Public
8 * License or the Artistic License, as specified in the README file.
9 *
4ac71550
TC
10 */
11
12/*
13 * 'I wonder what the Entish is for "yes" and "no",' he thought.
14 * --Pippin
15 *
16 * [p.480 of _The Lord of the Rings_, III/iv: "Treebeard"]
17 */
18
19/*
645c22ef
DM
20 *
21 *
5e045b90
AMS
22 * This file contains the code that creates, manipulates and destroys
23 * scalar values (SVs). The other types (AV, HV, GV, etc.) reuse the
24 * structure of an SV, so their creation and destruction is handled
25 * here; higher-level functions are in av.c, hv.c, and so on. Opcode
26 * level functions (eg. substr, split, join) for each of the types are
27 * in the pp*.c files.
79072805
LW
28 */
29
30#include "EXTERN.h"
864dbfa3 31#define PERL_IN_SV_C
79072805 32#include "perl.h"
d2f185dc 33#include "regcomp.h"
9d9a81f0
CB
34#ifdef __VMS
35# include <rms.h>
36#endif
79072805 37
2f8ed50e
OS
38#ifdef __Lynx__
39/* Missing proto on LynxOS */
40 char *gconvert(double, int, int, char *);
41#endif
42
a4eca1d4
JH
43#ifdef USE_QUADMATH
44# define SNPRINTF_G(nv, buffer, size, ndig) \
45 quadmath_snprintf(buffer, size, "%.*Qg", (int)ndig, (NV)(nv))
46#else
47# define SNPRINTF_G(nv, buffer, size, ndig) \
48 PERL_UNUSED_RESULT(Gconvert((NV)(nv), (int)ndig, 0, buffer))
49#endif
50
e8c6a474
YO
51#ifdef PERL_NEW_COPY_ON_WRITE
52# ifndef SV_COW_THRESHOLD
53# define SV_COW_THRESHOLD 0 /* COW iff len > K */
54# endif
55# ifndef SV_COWBUF_THRESHOLD
56# define SV_COWBUF_THRESHOLD 1250 /* COW iff len > K */
57# endif
58# ifndef SV_COW_MAX_WASTE_THRESHOLD
59# define SV_COW_MAX_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
60# endif
61# ifndef SV_COWBUF_WASTE_THRESHOLD
62# define SV_COWBUF_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
63# endif
64# ifndef SV_COW_MAX_WASTE_FACTOR_THRESHOLD
65# define SV_COW_MAX_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
66# endif
67# ifndef SV_COWBUF_WASTE_FACTOR_THRESHOLD
68# define SV_COWBUF_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
69# endif
70#endif
71/* Work around compiler warnings about unsigned >= THRESHOLD when thres-
72 hold is 0. */
73#if SV_COW_THRESHOLD
74# define GE_COW_THRESHOLD(cur) ((cur) >= SV_COW_THRESHOLD)
75#else
76# define GE_COW_THRESHOLD(cur) 1
77#endif
78#if SV_COWBUF_THRESHOLD
79# define GE_COWBUF_THRESHOLD(cur) ((cur) >= SV_COWBUF_THRESHOLD)
80#else
81# define GE_COWBUF_THRESHOLD(cur) 1
82#endif
83#if SV_COW_MAX_WASTE_THRESHOLD
84# define GE_COW_MAX_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COW_MAX_WASTE_THRESHOLD)
85#else
86# define GE_COW_MAX_WASTE_THRESHOLD(cur,len) 1
87#endif
88#if SV_COWBUF_WASTE_THRESHOLD
89# define GE_COWBUF_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COWBUF_WASTE_THRESHOLD)
90#else
91# define GE_COWBUF_WASTE_THRESHOLD(cur,len) 1
92#endif
93#if SV_COW_MAX_WASTE_FACTOR_THRESHOLD
94# define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COW_MAX_WASTE_FACTOR_THRESHOLD * (cur))
95#else
96# define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) 1
97#endif
98#if SV_COWBUF_WASTE_FACTOR_THRESHOLD
99# define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COWBUF_WASTE_FACTOR_THRESHOLD * (cur))
100#else
101# define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) 1
102#endif
103
104#define CHECK_COW_THRESHOLD(cur,len) (\
105 GE_COW_THRESHOLD((cur)) && \
106 GE_COW_MAX_WASTE_THRESHOLD((cur),(len)) && \
107 GE_COW_MAX_WASTE_FACTOR_THRESHOLD((cur),(len)) \
108)
109#define CHECK_COWBUF_THRESHOLD(cur,len) (\
110 GE_COWBUF_THRESHOLD((cur)) && \
111 GE_COWBUF_WASTE_THRESHOLD((cur),(len)) && \
112 GE_COWBUF_WASTE_FACTOR_THRESHOLD((cur),(len)) \
113)
cca0492e 114
e23c8137 115#ifdef PERL_UTF8_CACHE_ASSERT
ab455f60 116/* if adding more checks watch out for the following tests:
e23c8137
JH
117 * t/op/index.t t/op/length.t t/op/pat.t t/op/substr.t
118 * lib/utf8.t lib/Unicode/Collate/t/index.t
119 * --jhi
120 */
6f207bd3 121# define ASSERT_UTF8_CACHE(cache) \
ab455f60
NC
122 STMT_START { if (cache) { assert((cache)[0] <= (cache)[1]); \
123 assert((cache)[2] <= (cache)[3]); \
124 assert((cache)[3] <= (cache)[1]);} \
125 } STMT_END
e23c8137 126#else
6f207bd3 127# define ASSERT_UTF8_CACHE(cache) NOOP
e23c8137
JH
128#endif
129
f8c7b90f 130#ifdef PERL_OLD_COPY_ON_WRITE
765f542d 131#define SV_COW_NEXT_SV(sv) INT2PTR(SV *,SvUVX(sv))
607fa7f2 132#define SV_COW_NEXT_SV_SET(current,next) SvUV_set(current, PTR2UV(next))
765f542d 133#endif
645c22ef
DM
134
135/* ============================================================================
136
137=head1 Allocation and deallocation of SVs.
d2a0f284
JC
138An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
139sv, av, hv...) contains type and reference count information, and for
140many types, a pointer to the body (struct xrv, xpv, xpviv...), which
141contains fields specific to each type. Some types store all they need
142in the head, so don't have a body.
143
486ec47a 144In all but the most memory-paranoid configurations (ex: PURIFY), heads
d2a0f284
JC
145and bodies are allocated out of arenas, which by default are
146approximately 4K chunks of memory parcelled up into N heads or bodies.
93e68bfb
JC
147Sv-bodies are allocated by their sv-type, guaranteeing size
148consistency needed to allocate safely from arrays.
149
d2a0f284
JC
150For SV-heads, the first slot in each arena is reserved, and holds a
151link to the next arena, some flags, and a note of the number of slots.
152Snaked through each arena chain is a linked list of free items; when
153this becomes empty, an extra arena is allocated and divided up into N
154items which are threaded into the free list.
155
156SV-bodies are similar, but they use arena-sets by default, which
157separate the link and info from the arena itself, and reclaim the 1st
158slot in the arena. SV-bodies are further described later.
645c22ef
DM
159
160The following global variables are associated with arenas:
161
7fefc6c1
KW
162 PL_sv_arenaroot pointer to list of SV arenas
163 PL_sv_root pointer to list of free SV structures
645c22ef 164
7fefc6c1
KW
165 PL_body_arenas head of linked-list of body arenas
166 PL_body_roots[] array of pointers to list of free bodies of svtype
167 arrays are indexed by the svtype needed
93e68bfb 168
d2a0f284
JC
169A few special SV heads are not allocated from an arena, but are
170instead directly created in the interpreter structure, eg PL_sv_undef.
93e68bfb
JC
171The size of arenas can be changed from the default by setting
172PERL_ARENA_SIZE appropriately at compile time.
645c22ef
DM
173
174The SV arena serves the secondary purpose of allowing still-live SVs
175to be located and destroyed during final cleanup.
176
177At the lowest level, the macros new_SV() and del_SV() grab and free
178an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
179to return the SV to the free list with error checking.) new_SV() calls
180more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
181SVs in the free list have their SvTYPE field set to all ones.
182
ff276b08 183At the time of very final cleanup, sv_free_arenas() is called from
645c22ef 184perl_destruct() to physically free all the arenas allocated since the
6a93a7e5 185start of the interpreter.
645c22ef 186
645c22ef
DM
187The function visit() scans the SV arenas list, and calls a specified
188function for each SV it finds which is still live - ie which has an SvTYPE
189other than all 1's, and a non-zero SvREFCNT. visit() is used by the
190following functions (specified as [function that calls visit()] / [function
191called by visit() for each SV]):
192
193 sv_report_used() / do_report_used()
f2524eef 194 dump all remaining SVs (debugging aid)
645c22ef 195
e4487e9b 196 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
e76981f9 197 do_clean_named_io_objs(),do_curse()
645c22ef 198 Attempt to free all objects pointed to by RVs,
e76981f9
FC
199 try to do the same for all objects indir-
200 ectly referenced by typeglobs too, and
201 then do a final sweep, cursing any
202 objects that remain. Called once from
645c22ef
DM
203 perl_destruct(), prior to calling sv_clean_all()
204 below.
205
206 sv_clean_all() / do_clean_all()
207 SvREFCNT_dec(sv) each remaining SV, possibly
208 triggering an sv_free(). It also sets the
209 SVf_BREAK flag on the SV to indicate that the
210 refcnt has been artificially lowered, and thus
211 stopping sv_free() from giving spurious warnings
212 about SVs which unexpectedly have a refcnt
213 of zero. called repeatedly from perl_destruct()
214 until there are no SVs left.
215
93e68bfb 216=head2 Arena allocator API Summary
645c22ef
DM
217
218Private API to rest of sv.c
219
220 new_SV(), del_SV(),
221
df0f0429 222 new_XPVNV(), del_XPVGV(),
645c22ef
DM
223 etc
224
225Public API:
226
8cf8f3d1 227 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
645c22ef 228
645c22ef
DM
229=cut
230
3e8320cc 231 * ========================================================================= */
645c22ef 232
4561caa4
CS
233/*
234 * "A time to plant, and a time to uproot what was planted..."
235 */
236
d7a2c63c
MHM
237#ifdef PERL_MEM_LOG
238# define MEM_LOG_NEW_SV(sv, file, line, func) \
239 Perl_mem_log_new_sv(sv, file, line, func)
240# define MEM_LOG_DEL_SV(sv, file, line, func) \
241 Perl_mem_log_del_sv(sv, file, line, func)
242#else
243# define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
244# define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
245#endif
246
fd0854ff 247#ifdef DEBUG_LEAKING_SCALARS
484e6108
FC
248# define FREE_SV_DEBUG_FILE(sv) STMT_START { \
249 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
250 } STMT_END
d7a2c63c
MHM
251# define DEBUG_SV_SERIAL(sv) \
252 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
253 PTR2UV(sv), (long)(sv)->sv_debug_serial))
fd0854ff
DM
254#else
255# define FREE_SV_DEBUG_FILE(sv)
d7a2c63c 256# define DEBUG_SV_SERIAL(sv) NOOP
fd0854ff
DM
257#endif
258
48614a46
NC
259#ifdef PERL_POISON
260# define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
daba3364 261# define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
48614a46
NC
262/* Whilst I'd love to do this, it seems that things like to check on
263 unreferenced scalars
7e337ee0 264# define POSION_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
48614a46 265*/
7e337ee0
JH
266# define POSION_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
267 PoisonNew(&SvREFCNT(sv), 1, U32)
48614a46
NC
268#else
269# define SvARENA_CHAIN(sv) SvANY(sv)
3eef1deb 270# define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
48614a46
NC
271# define POSION_SV_HEAD(sv)
272#endif
273
990198f0
DM
274/* Mark an SV head as unused, and add to free list.
275 *
276 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
277 * its refcount artificially decremented during global destruction, so
278 * there may be dangling pointers to it. The last thing we want in that
279 * case is for it to be reused. */
280
053fc874
GS
281#define plant_SV(p) \
282 STMT_START { \
990198f0 283 const U32 old_flags = SvFLAGS(p); \
d7a2c63c
MHM
284 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
285 DEBUG_SV_SERIAL(p); \
fd0854ff 286 FREE_SV_DEBUG_FILE(p); \
48614a46 287 POSION_SV_HEAD(p); \
053fc874 288 SvFLAGS(p) = SVTYPEMASK; \
990198f0 289 if (!(old_flags & SVf_BREAK)) { \
3eef1deb 290 SvARENA_CHAIN_SET(p, PL_sv_root); \
990198f0
DM
291 PL_sv_root = (p); \
292 } \
053fc874
GS
293 --PL_sv_count; \
294 } STMT_END
a0d0e21e 295
053fc874
GS
296#define uproot_SV(p) \
297 STMT_START { \
298 (p) = PL_sv_root; \
daba3364 299 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
053fc874
GS
300 ++PL_sv_count; \
301 } STMT_END
302
645c22ef 303
cac9b346
NC
304/* make some more SVs by adding another arena */
305
cac9b346
NC
306STATIC SV*
307S_more_sv(pTHX)
308{
309 SV* sv;
9a87bd09
NC
310 char *chunk; /* must use New here to match call to */
311 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
312 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
cac9b346
NC
313 uproot_SV(sv);
314 return sv;
315}
316
645c22ef
DM
317/* new_SV(): return a new, empty SV head */
318
eba0f806
DM
319#ifdef DEBUG_LEAKING_SCALARS
320/* provide a real function for a debugger to play with */
321STATIC SV*
d7a2c63c 322S_new_SV(pTHX_ const char *file, int line, const char *func)
eba0f806
DM
323{
324 SV* sv;
325
eba0f806
DM
326 if (PL_sv_root)
327 uproot_SV(sv);
328 else
cac9b346 329 sv = S_more_sv(aTHX);
eba0f806
DM
330 SvANY(sv) = 0;
331 SvREFCNT(sv) = 1;
332 SvFLAGS(sv) = 0;
fd0854ff 333 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
e385c3bf
DM
334 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
335 ? PL_parser->copline
336 : PL_curcop
f24aceb1
DM
337 ? CopLINE(PL_curcop)
338 : 0
e385c3bf 339 );
fd0854ff 340 sv->sv_debug_inpad = 0;
cd676548 341 sv->sv_debug_parent = NULL;
484e6108 342 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
d7a2c63c
MHM
343
344 sv->sv_debug_serial = PL_sv_serial++;
345
346 MEM_LOG_NEW_SV(sv, file, line, func);
347 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
348 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
349
eba0f806
DM
350 return sv;
351}
d7a2c63c 352# define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
eba0f806
DM
353
354#else
355# define new_SV(p) \
053fc874 356 STMT_START { \
053fc874
GS
357 if (PL_sv_root) \
358 uproot_SV(p); \
359 else \
cac9b346 360 (p) = S_more_sv(aTHX); \
053fc874
GS
361 SvANY(p) = 0; \
362 SvREFCNT(p) = 1; \
363 SvFLAGS(p) = 0; \
d7a2c63c 364 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
053fc874 365 } STMT_END
eba0f806 366#endif
463ee0b2 367
645c22ef
DM
368
369/* del_SV(): return an empty SV head to the free list */
370
a0d0e21e 371#ifdef DEBUGGING
4561caa4 372
053fc874
GS
373#define del_SV(p) \
374 STMT_START { \
aea4f609 375 if (DEBUG_D_TEST) \
053fc874
GS
376 del_sv(p); \
377 else \
378 plant_SV(p); \
053fc874 379 } STMT_END
a0d0e21e 380
76e3520e 381STATIC void
cea2e8a9 382S_del_sv(pTHX_ SV *p)
463ee0b2 383{
7918f24d
NC
384 PERL_ARGS_ASSERT_DEL_SV;
385
aea4f609 386 if (DEBUG_D_TEST) {
4633a7c4 387 SV* sva;
a3b680e6 388 bool ok = 0;
daba3364 389 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
53c1dcc0
AL
390 const SV * const sv = sva + 1;
391 const SV * const svend = &sva[SvREFCNT(sva)];
c0ff570e 392 if (p >= sv && p < svend) {
a0d0e21e 393 ok = 1;
c0ff570e
NC
394 break;
395 }
a0d0e21e
LW
396 }
397 if (!ok) {
9b387841
NC
398 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
399 "Attempt to free non-arena SV: 0x%"UVxf
400 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
a0d0e21e
LW
401 return;
402 }
403 }
4561caa4 404 plant_SV(p);
463ee0b2 405}
a0d0e21e 406
4561caa4
CS
407#else /* ! DEBUGGING */
408
409#define del_SV(p) plant_SV(p)
410
411#endif /* DEBUGGING */
463ee0b2 412
645c22ef
DM
413
414/*
ccfc67b7
JH
415=head1 SV Manipulation Functions
416
645c22ef
DM
417=for apidoc sv_add_arena
418
419Given a chunk of memory, link it to the head of the list of arenas,
420and split it into a list of free SVs.
421
422=cut
423*/
424
d2bd4e7f
NC
425static void
426S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
463ee0b2 427{
daba3364 428 SV *const sva = MUTABLE_SV(ptr);
eb578fdb
KW
429 SV* sv;
430 SV* svend;
4633a7c4 431
7918f24d
NC
432 PERL_ARGS_ASSERT_SV_ADD_ARENA;
433
4633a7c4 434 /* The first SV in an arena isn't an SV. */
3280af22 435 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
4633a7c4
LW
436 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
437 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
438
3280af22
NIS
439 PL_sv_arenaroot = sva;
440 PL_sv_root = sva + 1;
4633a7c4
LW
441
442 svend = &sva[SvREFCNT(sva) - 1];
443 sv = sva + 1;
463ee0b2 444 while (sv < svend) {
3eef1deb 445 SvARENA_CHAIN_SET(sv, (sv + 1));
03e36789 446#ifdef DEBUGGING
978b032e 447 SvREFCNT(sv) = 0;
03e36789 448#endif
4b69cbe3 449 /* Must always set typemask because it's always checked in on cleanup
03e36789 450 when the arenas are walked looking for objects. */
8990e307 451 SvFLAGS(sv) = SVTYPEMASK;
463ee0b2
LW
452 sv++;
453 }
3eef1deb 454 SvARENA_CHAIN_SET(sv, 0);
03e36789
NC
455#ifdef DEBUGGING
456 SvREFCNT(sv) = 0;
457#endif
4633a7c4
LW
458 SvFLAGS(sv) = SVTYPEMASK;
459}
460
055972dc
DM
461/* visit(): call the named function for each non-free SV in the arenas
462 * whose flags field matches the flags/mask args. */
645c22ef 463
5226ed68 464STATIC I32
de37a194 465S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
8990e307 466{
4633a7c4 467 SV* sva;
5226ed68 468 I32 visited = 0;
8990e307 469
7918f24d
NC
470 PERL_ARGS_ASSERT_VISIT;
471
daba3364 472 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
eb578fdb
KW
473 const SV * const svend = &sva[SvREFCNT(sva)];
474 SV* sv;
4561caa4 475 for (sv = sva + 1; sv < svend; ++sv) {
e4787c0c 476 if (SvTYPE(sv) != (svtype)SVTYPEMASK
055972dc
DM
477 && (sv->sv_flags & mask) == flags
478 && SvREFCNT(sv))
479 {
942481a7 480 (*f)(aTHX_ sv);
5226ed68
JH
481 ++visited;
482 }
8990e307
LW
483 }
484 }
5226ed68 485 return visited;
8990e307
LW
486}
487
758a08c3
JH
488#ifdef DEBUGGING
489
645c22ef
DM
490/* called by sv_report_used() for each live SV */
491
492static void
5fa45a31 493do_report_used(pTHX_ SV *const sv)
645c22ef 494{
e4787c0c 495 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
645c22ef
DM
496 PerlIO_printf(Perl_debug_log, "****\n");
497 sv_dump(sv);
498 }
499}
758a08c3 500#endif
645c22ef
DM
501
502/*
503=for apidoc sv_report_used
504
fde67290 505Dump the contents of all SVs not yet freed (debugging aid).
645c22ef
DM
506
507=cut
508*/
509
8990e307 510void
864dbfa3 511Perl_sv_report_used(pTHX)
4561caa4 512{
ff270d3a 513#ifdef DEBUGGING
055972dc 514 visit(do_report_used, 0, 0);
96a5add6
AL
515#else
516 PERL_UNUSED_CONTEXT;
ff270d3a 517#endif
4561caa4
CS
518}
519
645c22ef
DM
520/* called by sv_clean_objs() for each live SV */
521
522static void
de37a194 523do_clean_objs(pTHX_ SV *const ref)
645c22ef 524{
ea724faa
NC
525 assert (SvROK(ref));
526 {
823a54a3
AL
527 SV * const target = SvRV(ref);
528 if (SvOBJECT(target)) {
529 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
530 if (SvWEAKREF(ref)) {
531 sv_del_backref(target, ref);
532 SvWEAKREF_off(ref);
533 SvRV_set(ref, NULL);
534 } else {
535 SvROK_off(ref);
536 SvRV_set(ref, NULL);
fc2b2dca 537 SvREFCNT_dec_NN(target);
823a54a3 538 }
645c22ef
DM
539 }
540 }
645c22ef
DM
541}
542
645c22ef 543
e4487e9b
DM
544/* clear any slots in a GV which hold objects - except IO;
545 * called by sv_clean_objs() for each live GV */
546
645c22ef 547static void
f30de749 548do_clean_named_objs(pTHX_ SV *const sv)
645c22ef 549{
57ef47cc 550 SV *obj;
ea724faa 551 assert(SvTYPE(sv) == SVt_PVGV);
d011219a 552 assert(isGV_with_GP(sv));
57ef47cc
DM
553 if (!GvGP(sv))
554 return;
555
556 /* freeing GP entries may indirectly free the current GV;
557 * hold onto it while we mess with the GP slots */
558 SvREFCNT_inc(sv);
559
560 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
561 DEBUG_D((PerlIO_printf(Perl_debug_log,
562 "Cleaning named glob SV object:\n "), sv_dump(obj)));
563 GvSV(sv) = NULL;
fc2b2dca 564 SvREFCNT_dec_NN(obj);
57ef47cc
DM
565 }
566 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
567 DEBUG_D((PerlIO_printf(Perl_debug_log,
568 "Cleaning named glob AV object:\n "), sv_dump(obj)));
569 GvAV(sv) = NULL;
fc2b2dca 570 SvREFCNT_dec_NN(obj);
57ef47cc
DM
571 }
572 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
573 DEBUG_D((PerlIO_printf(Perl_debug_log,
574 "Cleaning named glob HV object:\n "), sv_dump(obj)));
575 GvHV(sv) = NULL;
fc2b2dca 576 SvREFCNT_dec_NN(obj);
57ef47cc
DM
577 }
578 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
579 DEBUG_D((PerlIO_printf(Perl_debug_log,
580 "Cleaning named glob CV object:\n "), sv_dump(obj)));
c43ae56f 581 GvCV_set(sv, NULL);
fc2b2dca 582 SvREFCNT_dec_NN(obj);
57ef47cc 583 }
fc2b2dca 584 SvREFCNT_dec_NN(sv); /* undo the inc above */
e4487e9b
DM
585}
586
68b590d9 587/* clear any IO slots in a GV which hold objects (except stderr, defout);
e4487e9b
DM
588 * called by sv_clean_objs() for each live GV */
589
590static void
591do_clean_named_io_objs(pTHX_ SV *const sv)
592{
e4487e9b
DM
593 SV *obj;
594 assert(SvTYPE(sv) == SVt_PVGV);
595 assert(isGV_with_GP(sv));
68b590d9 596 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
e4487e9b
DM
597 return;
598
599 SvREFCNT_inc(sv);
57ef47cc
DM
600 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
601 DEBUG_D((PerlIO_printf(Perl_debug_log,
602 "Cleaning named glob IO object:\n "), sv_dump(obj)));
603 GvIOp(sv) = NULL;
fc2b2dca 604 SvREFCNT_dec_NN(obj);
645c22ef 605 }
fc2b2dca 606 SvREFCNT_dec_NN(sv); /* undo the inc above */
645c22ef 607}
645c22ef 608
4155e4fe
FC
609/* Void wrapper to pass to visit() */
610static void
611do_curse(pTHX_ SV * const sv) {
c2910e6c
FC
612 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
613 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
4155e4fe
FC
614 return;
615 (void)curse(sv, 0);
616}
617
645c22ef
DM
618/*
619=for apidoc sv_clean_objs
620
fde67290 621Attempt to destroy all objects not yet freed.
645c22ef
DM
622
623=cut
624*/
625
4561caa4 626void
864dbfa3 627Perl_sv_clean_objs(pTHX)
4561caa4 628{
68b590d9 629 GV *olddef, *olderr;
3280af22 630 PL_in_clean_objs = TRUE;
055972dc 631 visit(do_clean_objs, SVf_ROK, SVf_ROK);
e4487e9b
DM
632 /* Some barnacles may yet remain, clinging to typeglobs.
633 * Run the non-IO destructors first: they may want to output
634 * error messages, close files etc */
d011219a 635 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
e4487e9b 636 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
4155e4fe
FC
637 /* And if there are some very tenacious barnacles clinging to arrays,
638 closures, or what have you.... */
639 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
68b590d9
DM
640 olddef = PL_defoutgv;
641 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
642 if (olddef && isGV_with_GP(olddef))
643 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
644 olderr = PL_stderrgv;
645 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
646 if (olderr && isGV_with_GP(olderr))
647 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
648 SvREFCNT_dec(olddef);
3280af22 649 PL_in_clean_objs = FALSE;
4561caa4
CS
650}
651
645c22ef
DM
652/* called by sv_clean_all() for each live SV */
653
654static void
de37a194 655do_clean_all(pTHX_ SV *const sv)
645c22ef 656{
daba3364 657 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
cddfcddc 658 /* don't clean pid table and strtab */
d17ea597 659 return;
cddfcddc 660 }
645c22ef
DM
661 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
662 SvFLAGS(sv) |= SVf_BREAK;
fc2b2dca 663 SvREFCNT_dec_NN(sv);
645c22ef
DM
664}
665
666/*
667=for apidoc sv_clean_all
668
669Decrement the refcnt of each remaining SV, possibly triggering a
fde67290 670cleanup. This function may have to be called multiple times to free
ff276b08 671SVs which are in complex self-referential hierarchies.
645c22ef
DM
672
673=cut
674*/
675
5226ed68 676I32
864dbfa3 677Perl_sv_clean_all(pTHX)
8990e307 678{
5226ed68 679 I32 cleaned;
3280af22 680 PL_in_clean_all = TRUE;
055972dc 681 cleaned = visit(do_clean_all, 0,0);
5226ed68 682 return cleaned;
8990e307 683}
463ee0b2 684
5e258f8c
JC
685/*
686 ARENASETS: a meta-arena implementation which separates arena-info
687 into struct arena_set, which contains an array of struct
688 arena_descs, each holding info for a single arena. By separating
689 the meta-info from the arena, we recover the 1st slot, formerly
690 borrowed for list management. The arena_set is about the size of an
39244528 691 arena, avoiding the needless malloc overhead of a naive linked-list.
5e258f8c
JC
692
693 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
694 memory in the last arena-set (1/2 on average). In trade, we get
695 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
d2a0f284 696 smaller types). The recovery of the wasted space allows use of
e15dad31
JC
697 small arenas for large, rare body types, by changing array* fields
698 in body_details_by_type[] below.
5e258f8c 699*/
5e258f8c 700struct arena_desc {
398c677b
NC
701 char *arena; /* the raw storage, allocated aligned */
702 size_t size; /* its size ~4k typ */
e5973ed5 703 svtype utype; /* bodytype stored in arena */
5e258f8c
JC
704};
705
e6148039
NC
706struct arena_set;
707
708/* Get the maximum number of elements in set[] such that struct arena_set
e15dad31 709 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
e6148039
NC
710 therefore likely to be 1 aligned memory page. */
711
712#define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
713 - 2 * sizeof(int)) / sizeof (struct arena_desc))
5e258f8c
JC
714
715struct arena_set {
716 struct arena_set* next;
0a848332
NC
717 unsigned int set_size; /* ie ARENAS_PER_SET */
718 unsigned int curr; /* index of next available arena-desc */
5e258f8c
JC
719 struct arena_desc set[ARENAS_PER_SET];
720};
721
645c22ef
DM
722/*
723=for apidoc sv_free_arenas
724
fde67290 725Deallocate the memory used by all arenas. Note that all the individual SV
645c22ef
DM
726heads and bodies within the arenas must already have been freed.
727
728=cut
7fefc6c1 729
645c22ef 730*/
4633a7c4 731void
864dbfa3 732Perl_sv_free_arenas(pTHX)
4633a7c4
LW
733{
734 SV* sva;
735 SV* svanext;
0a848332 736 unsigned int i;
4633a7c4
LW
737
738 /* Free arenas here, but be careful about fake ones. (We assume
739 contiguity of the fake ones with the corresponding real ones.) */
740
3280af22 741 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
daba3364 742 svanext = MUTABLE_SV(SvANY(sva));
4633a7c4 743 while (svanext && SvFAKE(svanext))
daba3364 744 svanext = MUTABLE_SV(SvANY(svanext));
4633a7c4
LW
745
746 if (!SvFAKE(sva))
1df70142 747 Safefree(sva);
4633a7c4 748 }
93e68bfb 749
5e258f8c 750 {
0a848332
NC
751 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
752
753 while (aroot) {
754 struct arena_set *current = aroot;
755 i = aroot->curr;
756 while (i--) {
5e258f8c
JC
757 assert(aroot->set[i].arena);
758 Safefree(aroot->set[i].arena);
759 }
0a848332
NC
760 aroot = aroot->next;
761 Safefree(current);
5e258f8c
JC
762 }
763 }
dc8220bf 764 PL_body_arenas = 0;
fdda85ca 765
0a848332
NC
766 i = PERL_ARENA_ROOTS_SIZE;
767 while (i--)
93e68bfb 768 PL_body_roots[i] = 0;
93e68bfb 769
3280af22
NIS
770 PL_sv_arenaroot = 0;
771 PL_sv_root = 0;
4633a7c4
LW
772}
773
bd81e77b
NC
774/*
775 Here are mid-level routines that manage the allocation of bodies out
776 of the various arenas. There are 5 kinds of arenas:
29489e7c 777
bd81e77b
NC
778 1. SV-head arenas, which are discussed and handled above
779 2. regular body arenas
780 3. arenas for reduced-size bodies
781 4. Hash-Entry arenas
29489e7c 782
bd81e77b
NC
783 Arena types 2 & 3 are chained by body-type off an array of
784 arena-root pointers, which is indexed by svtype. Some of the
785 larger/less used body types are malloced singly, since a large
786 unused block of them is wasteful. Also, several svtypes dont have
787 bodies; the data fits into the sv-head itself. The arena-root
788 pointer thus has a few unused root-pointers (which may be hijacked
789 later for arena types 4,5)
29489e7c 790
bd81e77b
NC
791 3 differs from 2 as an optimization; some body types have several
792 unused fields in the front of the structure (which are kept in-place
793 for consistency). These bodies can be allocated in smaller chunks,
794 because the leading fields arent accessed. Pointers to such bodies
795 are decremented to point at the unused 'ghost' memory, knowing that
796 the pointers are used with offsets to the real memory.
29489e7c 797
d2a0f284
JC
798
799=head1 SV-Body Allocation
800
f554dfc5
MH
801=cut
802
d2a0f284
JC
803Allocation of SV-bodies is similar to SV-heads, differing as follows;
804the allocation mechanism is used for many body types, so is somewhat
805more complicated, it uses arena-sets, and has no need for still-live
806SV detection.
807
808At the outermost level, (new|del)_X*V macros return bodies of the
809appropriate type. These macros call either (new|del)_body_type or
810(new|del)_body_allocated macro pairs, depending on specifics of the
811type. Most body types use the former pair, the latter pair is used to
812allocate body types with "ghost fields".
813
814"ghost fields" are fields that are unused in certain types, and
69ba284b 815consequently don't need to actually exist. They are declared because
d2a0f284
JC
816they're part of a "base type", which allows use of functions as
817methods. The simplest examples are AVs and HVs, 2 aggregate types
818which don't use the fields which support SCALAR semantics.
819
69ba284b 820For these types, the arenas are carved up into appropriately sized
d2a0f284
JC
821chunks, we thus avoid wasted memory for those unaccessed members.
822When bodies are allocated, we adjust the pointer back in memory by the
69ba284b 823size of the part not allocated, so it's as if we allocated the full
d2a0f284
JC
824structure. (But things will all go boom if you write to the part that
825is "not there", because you'll be overwriting the last members of the
826preceding structure in memory.)
827
69ba284b 828We calculate the correction using the STRUCT_OFFSET macro on the first
a05ea1cf 829member present. If the allocated structure is smaller (no initial NV
69ba284b
NC
830actually allocated) then the net effect is to subtract the size of the NV
831from the pointer, to return a new pointer as if an initial NV were actually
a05ea1cf 832allocated. (We were using structures named *_allocated for this, but
69ba284b
NC
833this turned out to be a subtle bug, because a structure without an NV
834could have a lower alignment constraint, but the compiler is allowed to
835optimised accesses based on the alignment constraint of the actual pointer
836to the full structure, for example, using a single 64 bit load instruction
837because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
d2a0f284 838
a05ea1cf 839This is the same trick as was used for NV and IV bodies. Ironically it
d2a0f284 840doesn't need to be used for NV bodies any more, because NV is now at
5b306eef
DD
841the start of the structure. IV bodies, and also in some builds NV bodies,
842don't need it either, because they are no longer allocated.
d2a0f284
JC
843
844In turn, the new_body_* allocators call S_new_body(), which invokes
845new_body_inline macro, which takes a lock, and takes a body off the
1e30fcd5 846linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
d2a0f284
JC
847necessary to refresh an empty list. Then the lock is released, and
848the body is returned.
849
99816f8d 850Perl_more_bodies allocates a new arena, and carves it up into an array of N
d2a0f284
JC
851bodies, which it strings into a linked list. It looks up arena-size
852and body-size from the body_details table described below, thus
853supporting the multiple body-types.
854
855If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
856the (new|del)_X*V macros are mapped directly to malloc/free.
857
d2a0f284
JC
858For each sv-type, struct body_details bodies_by_type[] carries
859parameters which control these aspects of SV handling:
860
861Arena_size determines whether arenas are used for this body type, and if
862so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
863zero, forcing individual mallocs and frees.
864
865Body_size determines how big a body is, and therefore how many fit into
866each arena. Offset carries the body-pointer adjustment needed for
69ba284b 867"ghost fields", and is used in *_allocated macros.
d2a0f284
JC
868
869But its main purpose is to parameterize info needed in
870Perl_sv_upgrade(). The info here dramatically simplifies the function
69ba284b 871vs the implementation in 5.8.8, making it table-driven. All fields
d2a0f284
JC
872are used for this, except for arena_size.
873
874For the sv-types that have no bodies, arenas are not used, so those
875PL_body_roots[sv_type] are unused, and can be overloaded. In
876something of a special case, SVt_NULL is borrowed for HE arenas;
c6f8b1d0 877PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
d2a0f284 878bodies_by_type[SVt_NULL] slot is not used, as the table is not
c6f8b1d0 879available in hv.c.
d2a0f284 880
29489e7c
DM
881*/
882
bd81e77b 883struct body_details {
0fb58b32 884 U8 body_size; /* Size to allocate */
10666ae3 885 U8 copy; /* Size of structure to copy (may be shorter) */
5b306eef
DD
886 U8 offset; /* Size of unalloced ghost fields to first alloced field*/
887 PERL_BITFIELD8 type : 4; /* We have space for a sanity check. */
888 PERL_BITFIELD8 cant_upgrade : 1;/* Cannot upgrade this type */
889 PERL_BITFIELD8 zero_nv : 1; /* zero the NV when upgrading from this */
890 PERL_BITFIELD8 arena : 1; /* Allocated from an arena */
891 U32 arena_size; /* Size of arena to allocate */
bd81e77b 892};
29489e7c 893
bd81e77b
NC
894#define HADNV FALSE
895#define NONV TRUE
29489e7c 896
d2a0f284 897
bd81e77b
NC
898#ifdef PURIFY
899/* With -DPURFIY we allocate everything directly, and don't use arenas.
900 This seems a rather elegant way to simplify some of the code below. */
901#define HASARENA FALSE
902#else
903#define HASARENA TRUE
904#endif
905#define NOARENA FALSE
29489e7c 906
d2a0f284
JC
907/* Size the arenas to exactly fit a given number of bodies. A count
908 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
909 simplifying the default. If count > 0, the arena is sized to fit
910 only that many bodies, allowing arenas to be used for large, rare
911 bodies (XPVFM, XPVIO) without undue waste. The arena size is
912 limited by PERL_ARENA_SIZE, so we can safely oversize the
913 declarations.
914 */
95db5f15
MB
915#define FIT_ARENA0(body_size) \
916 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
917#define FIT_ARENAn(count,body_size) \
918 ( count * body_size <= PERL_ARENA_SIZE) \
919 ? count * body_size \
920 : FIT_ARENA0 (body_size)
921#define FIT_ARENA(count,body_size) \
cd1dc8e2 922 (U32)(count \
95db5f15 923 ? FIT_ARENAn (count, body_size) \
cd1dc8e2 924 : FIT_ARENA0 (body_size))
d2a0f284 925
bd81e77b
NC
926/* Calculate the length to copy. Specifically work out the length less any
927 final padding the compiler needed to add. See the comment in sv_upgrade
928 for why copying the padding proved to be a bug. */
29489e7c 929
bd81e77b
NC
930#define copy_length(type, last_member) \
931 STRUCT_OFFSET(type, last_member) \
daba3364 932 + sizeof (((type*)SvANY((const SV *)0))->last_member)
29489e7c 933
bd81e77b 934static const struct body_details bodies_by_type[] = {
829cd18a
NC
935 /* HEs use this offset for their arena. */
936 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
d2a0f284 937
db93c0c4
NC
938 /* IVs are in the head, so the allocation size is 0. */
939 { 0,
d2a0f284 940 sizeof(IV), /* This is used to copy out the IV body. */
10666ae3 941 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
db93c0c4 942 NOARENA /* IVS don't need an arena */, 0
d2a0f284
JC
943 },
944
5b306eef
DD
945#if NVSIZE <= IVSIZE
946 { 0, sizeof(NV),
947 STRUCT_OFFSET(XPVNV, xnv_u),
948 SVt_NV, FALSE, HADNV, NOARENA, 0 },
949#else
6e128786
NC
950 { sizeof(NV), sizeof(NV),
951 STRUCT_OFFSET(XPVNV, xnv_u),
952 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
5b306eef 953#endif
d2a0f284 954
bc337e5c 955 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
956 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
957 + STRUCT_OFFSET(XPV, xpv_cur),
69ba284b 958 SVt_PV, FALSE, NONV, HASARENA,
889d28b2 959 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 960
d361b004
KW
961 { sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur),
962 copy_length(XINVLIST, is_offset) - STRUCT_OFFSET(XPV, xpv_cur),
963 + STRUCT_OFFSET(XPV, xpv_cur),
964 SVt_INVLIST, TRUE, NONV, HASARENA,
965 FIT_ARENA(0, sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur)) },
e94d9b54 966
bc337e5c 967 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
968 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
969 + STRUCT_OFFSET(XPV, xpv_cur),
970 SVt_PVIV, FALSE, NONV, HASARENA,
971 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 972
bc337e5c 973 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
974 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
975 + STRUCT_OFFSET(XPV, xpv_cur),
976 SVt_PVNV, FALSE, HADNV, HASARENA,
977 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 978
6e128786 979 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
d2a0f284 980 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
4df7f6af 981
601dfd0a
NC
982 { sizeof(regexp),
983 sizeof(regexp),
984 0,
ecff11eb 985 SVt_REGEXP, TRUE, NONV, HASARENA,
eaeb1e7f 986 FIT_ARENA(0, sizeof(regexp))
5c35adbb 987 },
4df7f6af 988
10666ae3 989 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
d2a0f284
JC
990 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
991
10666ae3 992 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
d2a0f284
JC
993 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
994
601dfd0a 995 { sizeof(XPVAV),
4f7003f5 996 copy_length(XPVAV, xav_alloc),
601dfd0a 997 0,
69ba284b 998 SVt_PVAV, TRUE, NONV, HASARENA,
601dfd0a 999 FIT_ARENA(0, sizeof(XPVAV)) },
d2a0f284 1000
601dfd0a 1001 { sizeof(XPVHV),
359164a0 1002 copy_length(XPVHV, xhv_max),
601dfd0a 1003 0,
69ba284b 1004 SVt_PVHV, TRUE, NONV, HASARENA,
601dfd0a 1005 FIT_ARENA(0, sizeof(XPVHV)) },
d2a0f284 1006
601dfd0a
NC
1007 { sizeof(XPVCV),
1008 sizeof(XPVCV),
1009 0,
69ba284b 1010 SVt_PVCV, TRUE, NONV, HASARENA,
601dfd0a 1011 FIT_ARENA(0, sizeof(XPVCV)) },
69ba284b 1012
601dfd0a
NC
1013 { sizeof(XPVFM),
1014 sizeof(XPVFM),
1015 0,
69ba284b 1016 SVt_PVFM, TRUE, NONV, NOARENA,
601dfd0a 1017 FIT_ARENA(20, sizeof(XPVFM)) },
d2a0f284 1018
601dfd0a
NC
1019 { sizeof(XPVIO),
1020 sizeof(XPVIO),
1021 0,
b6f60916 1022 SVt_PVIO, TRUE, NONV, HASARENA,
601dfd0a 1023 FIT_ARENA(24, sizeof(XPVIO)) },
bd81e77b 1024};
29489e7c 1025
bd81e77b 1026#define new_body_allocated(sv_type) \
d2a0f284 1027 (void *)((char *)S_new_body(aTHX_ sv_type) \
bd81e77b 1028 - bodies_by_type[sv_type].offset)
29489e7c 1029
26359cfa
NC
1030/* return a thing to the free list */
1031
1032#define del_body(thing, root) \
1033 STMT_START { \
1034 void ** const thing_copy = (void **)thing; \
1035 *thing_copy = *root; \
1036 *root = (void*)thing_copy; \
1037 } STMT_END
29489e7c 1038
bd81e77b 1039#ifdef PURIFY
5b306eef
DD
1040#if !(NVSIZE <= IVSIZE)
1041# define new_XNV() safemalloc(sizeof(XPVNV))
1042#endif
beeec492
NC
1043#define new_XPVNV() safemalloc(sizeof(XPVNV))
1044#define new_XPVMG() safemalloc(sizeof(XPVMG))
29489e7c 1045
beeec492 1046#define del_XPVGV(p) safefree(p)
29489e7c 1047
bd81e77b 1048#else /* !PURIFY */
29489e7c 1049
5b306eef
DD
1050#if !(NVSIZE <= IVSIZE)
1051# define new_XNV() new_body_allocated(SVt_NV)
1052#endif
65ac1738 1053#define new_XPVNV() new_body_allocated(SVt_PVNV)
65ac1738 1054#define new_XPVMG() new_body_allocated(SVt_PVMG)
645c22ef 1055
26359cfa
NC
1056#define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1057 &PL_body_roots[SVt_PVGV])
1d7c1841 1058
bd81e77b 1059#endif /* PURIFY */
93e68bfb 1060
bd81e77b 1061/* no arena for you! */
93e68bfb 1062
bd81e77b 1063#define new_NOARENA(details) \
beeec492 1064 safemalloc((details)->body_size + (details)->offset)
bd81e77b 1065#define new_NOARENAZ(details) \
beeec492 1066 safecalloc((details)->body_size + (details)->offset, 1)
d2a0f284 1067
1e30fcd5
NC
1068void *
1069Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1070 const size_t arena_size)
d2a0f284 1071{
d2a0f284 1072 void ** const root = &PL_body_roots[sv_type];
99816f8d
NC
1073 struct arena_desc *adesc;
1074 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1075 unsigned int curr;
d2a0f284
JC
1076 char *start;
1077 const char *end;
02982131 1078 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
20b7effb
JH
1079#if defined(DEBUGGING) && defined(PERL_GLOBAL_STRUCT)
1080 dVAR;
1081#endif
0b2d3faa 1082#if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
23e9d66c
NC
1083 static bool done_sanity_check;
1084
0b2d3faa
JH
1085 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1086 * variables like done_sanity_check. */
10666ae3 1087 if (!done_sanity_check) {
ea471437 1088 unsigned int i = SVt_LAST;
10666ae3
NC
1089
1090 done_sanity_check = TRUE;
1091
1092 while (i--)
1093 assert (bodies_by_type[i].type == i);
1094 }
1095#endif
1096
02982131 1097 assert(arena_size);
23e9d66c 1098
99816f8d
NC
1099 /* may need new arena-set to hold new arena */
1100 if (!aroot || aroot->curr >= aroot->set_size) {
1101 struct arena_set *newroot;
1102 Newxz(newroot, 1, struct arena_set);
1103 newroot->set_size = ARENAS_PER_SET;
1104 newroot->next = aroot;
1105 aroot = newroot;
1106 PL_body_arenas = (void *) newroot;
1107 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1108 }
1109
1110 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1111 curr = aroot->curr++;
1112 adesc = &(aroot->set[curr]);
1113 assert(!adesc->arena);
1114
1115 Newx(adesc->arena, good_arena_size, char);
1116 adesc->size = good_arena_size;
1117 adesc->utype = sv_type;
1118 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %"UVuf"\n",
1119 curr, (void*)adesc->arena, (UV)good_arena_size));
1120
1121 start = (char *) adesc->arena;
d2a0f284 1122
29657bb6
NC
1123 /* Get the address of the byte after the end of the last body we can fit.
1124 Remember, this is integer division: */
02982131 1125 end = start + good_arena_size / body_size * body_size;
d2a0f284 1126
486ec47a 1127 /* computed count doesn't reflect the 1st slot reservation */
d8fca402
NC
1128#if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1129 DEBUG_m(PerlIO_printf(Perl_debug_log,
1130 "arena %p end %p arena-size %d (from %d) type %d "
1131 "size %d ct %d\n",
02982131
NC
1132 (void*)start, (void*)end, (int)good_arena_size,
1133 (int)arena_size, sv_type, (int)body_size,
1134 (int)good_arena_size / (int)body_size));
d8fca402 1135#else
d2a0f284
JC
1136 DEBUG_m(PerlIO_printf(Perl_debug_log,
1137 "arena %p end %p arena-size %d type %d size %d ct %d\n",
6c9570dc 1138 (void*)start, (void*)end,
02982131
NC
1139 (int)arena_size, sv_type, (int)body_size,
1140 (int)good_arena_size / (int)body_size));
d8fca402 1141#endif
d2a0f284
JC
1142 *root = (void *)start;
1143
29657bb6
NC
1144 while (1) {
1145 /* Where the next body would start: */
d2a0f284 1146 char * const next = start + body_size;
29657bb6
NC
1147
1148 if (next >= end) {
1149 /* This is the last body: */
1150 assert(next == end);
1151
1152 *(void **)start = 0;
1153 return *root;
1154 }
1155
d2a0f284
JC
1156 *(void**) start = (void *)next;
1157 start = next;
1158 }
d2a0f284
JC
1159}
1160
1161/* grab a new thing from the free list, allocating more if necessary.
1162 The inline version is used for speed in hot routines, and the
1163 function using it serves the rest (unless PURIFY).
1164*/
1165#define new_body_inline(xpv, sv_type) \
1166 STMT_START { \
1167 void ** const r3wt = &PL_body_roots[sv_type]; \
11b79775 1168 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1e30fcd5 1169 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
02982131
NC
1170 bodies_by_type[sv_type].body_size,\
1171 bodies_by_type[sv_type].arena_size)); \
d2a0f284 1172 *(r3wt) = *(void**)(xpv); \
d2a0f284
JC
1173 } STMT_END
1174
1175#ifndef PURIFY
1176
1177STATIC void *
de37a194 1178S_new_body(pTHX_ const svtype sv_type)
d2a0f284 1179{
d2a0f284
JC
1180 void *xpv;
1181 new_body_inline(xpv, sv_type);
1182 return xpv;
1183}
1184
1185#endif
93e68bfb 1186
238b27b3
NC
1187static const struct body_details fake_rv =
1188 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1189
bd81e77b
NC
1190/*
1191=for apidoc sv_upgrade
93e68bfb 1192
bd81e77b
NC
1193Upgrade an SV to a more complex form. Generally adds a new body type to the
1194SV, then copies across as much information as possible from the old body.
9521ca61
FC
1195It croaks if the SV is already in a more complex form than requested. You
1196generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1197before calling C<sv_upgrade>, and hence does not croak. See also
1198C<svtype>.
93e68bfb 1199
bd81e77b 1200=cut
93e68bfb 1201*/
93e68bfb 1202
bd81e77b 1203void
5aaab254 1204Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
cac9b346 1205{
bd81e77b
NC
1206 void* old_body;
1207 void* new_body;
42d0e0b7 1208 const svtype old_type = SvTYPE(sv);
d2a0f284 1209 const struct body_details *new_type_details;
238b27b3 1210 const struct body_details *old_type_details
bd81e77b 1211 = bodies_by_type + old_type;
4df7f6af 1212 SV *referant = NULL;
cac9b346 1213
7918f24d
NC
1214 PERL_ARGS_ASSERT_SV_UPGRADE;
1215
1776cbe8
NC
1216 if (old_type == new_type)
1217 return;
1218
1219 /* This clause was purposefully added ahead of the early return above to
1220 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1221 inference by Nick I-S that it would fix other troublesome cases. See
1222 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1223
1224 Given that shared hash key scalars are no longer PVIV, but PV, there is
1225 no longer need to unshare so as to free up the IVX slot for its proper
1226 purpose. So it's safe to move the early return earlier. */
1227
093085a8 1228 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
bd81e77b
NC
1229 sv_force_normal_flags(sv, 0);
1230 }
cac9b346 1231
bd81e77b 1232 old_body = SvANY(sv);
de042e1d 1233
bd81e77b
NC
1234 /* Copying structures onto other structures that have been neatly zeroed
1235 has a subtle gotcha. Consider XPVMG
cac9b346 1236
bd81e77b
NC
1237 +------+------+------+------+------+-------+-------+
1238 | NV | CUR | LEN | IV | MAGIC | STASH |
1239 +------+------+------+------+------+-------+-------+
1240 0 4 8 12 16 20 24 28
645c22ef 1241
bd81e77b
NC
1242 where NVs are aligned to 8 bytes, so that sizeof that structure is
1243 actually 32 bytes long, with 4 bytes of padding at the end:
08742458 1244
bd81e77b
NC
1245 +------+------+------+------+------+-------+-------+------+
1246 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1247 +------+------+------+------+------+-------+-------+------+
1248 0 4 8 12 16 20 24 28 32
08742458 1249
bd81e77b 1250 so what happens if you allocate memory for this structure:
30f9da9e 1251
bd81e77b
NC
1252 +------+------+------+------+------+-------+-------+------+------+...
1253 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1254 +------+------+------+------+------+-------+-------+------+------+...
1255 0 4 8 12 16 20 24 28 32 36
bfc44f79 1256
bd81e77b
NC
1257 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1258 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1259 started out as zero once, but it's quite possible that it isn't. So now,
1260 rather than a nicely zeroed GP, you have it pointing somewhere random.
1261 Bugs ensue.
bfc44f79 1262
bd81e77b
NC
1263 (In fact, GP ends up pointing at a previous GP structure, because the
1264 principle cause of the padding in XPVMG getting garbage is a copy of
6c9e42f7
NC
1265 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1266 this happens to be moot because XPVGV has been re-ordered, with GP
1267 no longer after STASH)
30f9da9e 1268
bd81e77b
NC
1269 So we are careful and work out the size of used parts of all the
1270 structures. */
bfc44f79 1271
bd81e77b
NC
1272 switch (old_type) {
1273 case SVt_NULL:
1274 break;
1275 case SVt_IV:
4df7f6af
NC
1276 if (SvROK(sv)) {
1277 referant = SvRV(sv);
238b27b3
NC
1278 old_type_details = &fake_rv;
1279 if (new_type == SVt_NV)
1280 new_type = SVt_PVNV;
4df7f6af
NC
1281 } else {
1282 if (new_type < SVt_PVIV) {
1283 new_type = (new_type == SVt_NV)
1284 ? SVt_PVNV : SVt_PVIV;
1285 }
bd81e77b
NC
1286 }
1287 break;
1288 case SVt_NV:
1289 if (new_type < SVt_PVNV) {
1290 new_type = SVt_PVNV;
bd81e77b
NC
1291 }
1292 break;
bd81e77b
NC
1293 case SVt_PV:
1294 assert(new_type > SVt_PV);
1295 assert(SVt_IV < SVt_PV);
1296 assert(SVt_NV < SVt_PV);
1297 break;
1298 case SVt_PVIV:
1299 break;
1300 case SVt_PVNV:
1301 break;
1302 case SVt_PVMG:
1303 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1304 there's no way that it can be safely upgraded, because perl.c
1305 expects to Safefree(SvANY(PL_mess_sv)) */
1306 assert(sv != PL_mess_sv);
1307 /* This flag bit is used to mean other things in other scalar types.
1308 Given that it only has meaning inside the pad, it shouldn't be set
1309 on anything that can get upgraded. */
00b1698f 1310 assert(!SvPAD_TYPED(sv));
bd81e77b
NC
1311 break;
1312 default:
2439e033 1313 if (UNLIKELY(old_type_details->cant_upgrade))
c81225bc
NC
1314 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1315 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
bd81e77b 1316 }
3376de98 1317
2439e033 1318 if (UNLIKELY(old_type > new_type))
3376de98
NC
1319 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1320 (int)old_type, (int)new_type);
1321
2fa1109b 1322 new_type_details = bodies_by_type + new_type;
645c22ef 1323
bd81e77b
NC
1324 SvFLAGS(sv) &= ~SVTYPEMASK;
1325 SvFLAGS(sv) |= new_type;
932e9ff9 1326
ab4416c0
NC
1327 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1328 the return statements above will have triggered. */
1329 assert (new_type != SVt_NULL);
bd81e77b 1330 switch (new_type) {
bd81e77b
NC
1331 case SVt_IV:
1332 assert(old_type == SVt_NULL);
1333 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv));
1334 SvIV_set(sv, 0);
1335 return;
1336 case SVt_NV:
1337 assert(old_type == SVt_NULL);
5b306eef
DD
1338#if NVSIZE <= IVSIZE
1339 SvANY(sv) = (XPVNV*)((char*)&(sv->sv_u.svu_nv) - STRUCT_OFFSET(XPVNV, xnv_u.xnv_nv));
1340#else
bd81e77b 1341 SvANY(sv) = new_XNV();
5b306eef 1342#endif
bd81e77b
NC
1343 SvNV_set(sv, 0);
1344 return;
bd81e77b 1345 case SVt_PVHV:
bd81e77b 1346 case SVt_PVAV:
d2a0f284 1347 assert(new_type_details->body_size);
c1ae03ae
NC
1348
1349#ifndef PURIFY
1350 assert(new_type_details->arena);
d2a0f284 1351 assert(new_type_details->arena_size);
c1ae03ae 1352 /* This points to the start of the allocated area. */
d2a0f284
JC
1353 new_body_inline(new_body, new_type);
1354 Zero(new_body, new_type_details->body_size, char);
c1ae03ae
NC
1355 new_body = ((char *)new_body) - new_type_details->offset;
1356#else
1357 /* We always allocated the full length item with PURIFY. To do this
1358 we fake things so that arena is false for all 16 types.. */
1359 new_body = new_NOARENAZ(new_type_details);
1360#endif
1361 SvANY(sv) = new_body;
1362 if (new_type == SVt_PVAV) {
1363 AvMAX(sv) = -1;
1364 AvFILLp(sv) = -1;
1365 AvREAL_only(sv);
64484faa 1366 if (old_type_details->body_size) {
ac572bf4
NC
1367 AvALLOC(sv) = 0;
1368 } else {
1369 /* It will have been zeroed when the new body was allocated.
1370 Lets not write to it, in case it confuses a write-back
1371 cache. */
1372 }
78ac7dd9
NC
1373 } else {
1374 assert(!SvOK(sv));
1375 SvOK_off(sv);
1376#ifndef NODEFAULT_SHAREKEYS
1377 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1378#endif
586fc6a3
SM
1379 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1380 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
c1ae03ae 1381 }
aeb18a1e 1382
bd81e77b
NC
1383 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1384 The target created by newSVrv also is, and it can have magic.
1385 However, it never has SvPVX set.
1386 */
4df7f6af
NC
1387 if (old_type == SVt_IV) {
1388 assert(!SvROK(sv));
1389 } else if (old_type >= SVt_PV) {
bd81e77b
NC
1390 assert(SvPVX_const(sv) == 0);
1391 }
aeb18a1e 1392
bd81e77b 1393 if (old_type >= SVt_PVMG) {
e736a858 1394 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
bd81e77b 1395 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
797c7171
NC
1396 } else {
1397 sv->sv_u.svu_array = NULL; /* or svu_hash */
bd81e77b
NC
1398 }
1399 break;
93e68bfb 1400
bd81e77b
NC
1401 case SVt_PVIV:
1402 /* XXX Is this still needed? Was it ever needed? Surely as there is
1403 no route from NV to PVIV, NOK can never be true */
1404 assert(!SvNOKp(sv));
1405 assert(!SvNOK(sv));
1406 case SVt_PVIO:
1407 case SVt_PVFM:
bd81e77b
NC
1408 case SVt_PVGV:
1409 case SVt_PVCV:
1410 case SVt_PVLV:
d361b004 1411 case SVt_INVLIST:
12c45b25 1412 case SVt_REGEXP:
bd81e77b
NC
1413 case SVt_PVMG:
1414 case SVt_PVNV:
1415 case SVt_PV:
93e68bfb 1416
d2a0f284 1417 assert(new_type_details->body_size);
bd81e77b
NC
1418 /* We always allocated the full length item with PURIFY. To do this
1419 we fake things so that arena is false for all 16 types.. */
1420 if(new_type_details->arena) {
1421 /* This points to the start of the allocated area. */
d2a0f284
JC
1422 new_body_inline(new_body, new_type);
1423 Zero(new_body, new_type_details->body_size, char);
bd81e77b
NC
1424 new_body = ((char *)new_body) - new_type_details->offset;
1425 } else {
1426 new_body = new_NOARENAZ(new_type_details);
1427 }
1428 SvANY(sv) = new_body;
5e2fc214 1429
bd81e77b 1430 if (old_type_details->copy) {
f9ba3d20
NC
1431 /* There is now the potential for an upgrade from something without
1432 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1433 int offset = old_type_details->offset;
1434 int length = old_type_details->copy;
1435
1436 if (new_type_details->offset > old_type_details->offset) {
d4c19fe8 1437 const int difference
f9ba3d20
NC
1438 = new_type_details->offset - old_type_details->offset;
1439 offset += difference;
1440 length -= difference;
1441 }
1442 assert (length >= 0);
1443
1444 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1445 char);
bd81e77b
NC
1446 }
1447
1448#ifndef NV_ZERO_IS_ALLBITS_ZERO
f2524eef 1449 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
e5ce394c
NC
1450 * correct 0.0 for us. Otherwise, if the old body didn't have an
1451 * NV slot, but the new one does, then we need to initialise the
1452 * freshly created NV slot with whatever the correct bit pattern is
1453 * for 0.0 */
e22a937e
NC
1454 if (old_type_details->zero_nv && !new_type_details->zero_nv
1455 && !isGV_with_GP(sv))
bd81e77b 1456 SvNV_set(sv, 0);
82048762 1457#endif
5e2fc214 1458
2439e033 1459 if (UNLIKELY(new_type == SVt_PVIO)) {
85dca89a 1460 IO * const io = MUTABLE_IO(sv);
d963bf01 1461 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
85dca89a
NC
1462
1463 SvOBJECT_on(io);
1464 /* Clear the stashcache because a new IO could overrule a package
1465 name */
103f5a36 1466 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
85dca89a
NC
1467 hv_clear(PL_stashcache);
1468
85dca89a 1469 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
f2524eef 1470 IoPAGE_LEN(sv) = 60;
85dca89a 1471 }
2439e033 1472 if (UNLIKELY(new_type == SVt_REGEXP))
8d919b0a
FC
1473 sv->sv_u.svu_rx = (regexp *)new_body;
1474 else if (old_type < SVt_PV) {
4df7f6af
NC
1475 /* referant will be NULL unless the old type was SVt_IV emulating
1476 SVt_RV */
1477 sv->sv_u.svu_rv = referant;
1478 }
bd81e77b
NC
1479 break;
1480 default:
afd78fd5
JH
1481 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1482 (unsigned long)new_type);
bd81e77b 1483 }
73171d91 1484
5b306eef
DD
1485 /* if this is zero, this is a body-less SVt_NULL, SVt_IV/SVt_RV,
1486 and sometimes SVt_NV */
1487 if (old_type_details->body_size) {
bd81e77b 1488#ifdef PURIFY
beeec492 1489 safefree(old_body);
bd81e77b 1490#else
bc786448
GG
1491 /* Note that there is an assumption that all bodies of types that
1492 can be upgraded came from arenas. Only the more complex non-
1493 upgradable types are allowed to be directly malloc()ed. */
1494 assert(old_type_details->arena);
bd81e77b
NC
1495 del_body((void*)((char*)old_body + old_type_details->offset),
1496 &PL_body_roots[old_type]);
1497#endif
1498 }
1499}
73171d91 1500
bd81e77b
NC
1501/*
1502=for apidoc sv_backoff
73171d91 1503
fde67290 1504Remove any string offset. You should normally use the C<SvOOK_off> macro
bd81e77b 1505wrapper instead.
73171d91 1506
bd81e77b 1507=cut
73171d91
NC
1508*/
1509
bd81e77b 1510int
ddeaf645 1511Perl_sv_backoff(SV *const sv)
bd81e77b 1512{
69240efd 1513 STRLEN delta;
7a4bba22 1514 const char * const s = SvPVX_const(sv);
7918f24d
NC
1515
1516 PERL_ARGS_ASSERT_SV_BACKOFF;
7918f24d 1517
bd81e77b
NC
1518 assert(SvOOK(sv));
1519 assert(SvTYPE(sv) != SVt_PVHV);
1520 assert(SvTYPE(sv) != SVt_PVAV);
7a4bba22 1521
69240efd
NC
1522 SvOOK_offset(sv, delta);
1523
7a4bba22
NC
1524 SvLEN_set(sv, SvLEN(sv) + delta);
1525 SvPV_set(sv, SvPVX(sv) - delta);
1526 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
bd81e77b
NC
1527 SvFLAGS(sv) &= ~SVf_OOK;
1528 return 0;
1529}
73171d91 1530
bd81e77b
NC
1531/*
1532=for apidoc sv_grow
73171d91 1533
bd81e77b
NC
1534Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1535upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1536Use the C<SvGROW> wrapper instead.
93e68bfb 1537
bd81e77b
NC
1538=cut
1539*/
93e68bfb 1540
e0060e30
FC
1541static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1542
bd81e77b 1543char *
5aaab254 1544Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
bd81e77b 1545{
eb578fdb 1546 char *s;
93e68bfb 1547
7918f24d
NC
1548 PERL_ARGS_ASSERT_SV_GROW;
1549
bd81e77b
NC
1550 if (SvROK(sv))
1551 sv_unref(sv);
1552 if (SvTYPE(sv) < SVt_PV) {
1553 sv_upgrade(sv, SVt_PV);
1554 s = SvPVX_mutable(sv);
1555 }
1556 else if (SvOOK(sv)) { /* pv is offset? */
1557 sv_backoff(sv);
1558 s = SvPVX_mutable(sv);
1559 if (newlen > SvLEN(sv))
1560 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
bd81e77b
NC
1561 }
1562 else
db2c6cb3 1563 {
e0060e30 1564 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
bd81e77b 1565 s = SvPVX_mutable(sv);
db2c6cb3 1566 }
aeb18a1e 1567
cbcb2a16
DM
1568#ifdef PERL_NEW_COPY_ON_WRITE
1569 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
3c239bea 1570 * to store the COW count. So in general, allocate one more byte than
cbcb2a16
DM
1571 * asked for, to make it likely this byte is always spare: and thus
1572 * make more strings COW-able.
1573 * If the new size is a big power of two, don't bother: we assume the
1574 * caller wanted a nice 2^N sized block and will be annoyed at getting
1575 * 2^N+1 */
1576 if (newlen & 0xff)
1577 newlen++;
1578#endif
1579
ce861ea7
YO
1580#if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1581#define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1582#endif
1583
bd81e77b 1584 if (newlen > SvLEN(sv)) { /* need more room? */
f1200559 1585 STRLEN minlen = SvCUR(sv);
3c239bea 1586 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
f1200559
WH
1587 if (newlen < minlen)
1588 newlen = minlen;
ce861ea7 1589#ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
7c641603
KW
1590
1591 /* Don't round up on the first allocation, as odds are pretty good that
1592 * the initial request is accurate as to what is really needed */
ce861ea7 1593 if (SvLEN(sv)) {
e8c6a474 1594 newlen = PERL_STRLEN_ROUNDUP(newlen);
ce861ea7 1595 }
bd81e77b 1596#endif
98653f18 1597 if (SvLEN(sv) && s) {
10edeb5d 1598 s = (char*)saferealloc(s, newlen);
bd81e77b
NC
1599 }
1600 else {
10edeb5d 1601 s = (char*)safemalloc(newlen);
bd81e77b
NC
1602 if (SvPVX_const(sv) && SvCUR(sv)) {
1603 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1604 }
1605 }
1606 SvPV_set(sv, s);
ce861ea7 1607#ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
98653f18
NC
1608 /* Do this here, do it once, do it right, and then we will never get
1609 called back into sv_grow() unless there really is some growing
1610 needed. */
ca7c1a29 1611 SvLEN_set(sv, Perl_safesysmalloc_size(s));
98653f18 1612#else
bd81e77b 1613 SvLEN_set(sv, newlen);
98653f18 1614#endif
bd81e77b
NC
1615 }
1616 return s;
1617}
aeb18a1e 1618
bd81e77b
NC
1619/*
1620=for apidoc sv_setiv
932e9ff9 1621
bd81e77b
NC
1622Copies an integer into the given SV, upgrading first if necessary.
1623Does not handle 'set' magic. See also C<sv_setiv_mg>.
463ee0b2 1624
bd81e77b
NC
1625=cut
1626*/
463ee0b2 1627
bd81e77b 1628void
5aaab254 1629Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
bd81e77b 1630{
7918f24d
NC
1631 PERL_ARGS_ASSERT_SV_SETIV;
1632
bd81e77b
NC
1633 SV_CHECK_THINKFIRST_COW_DROP(sv);
1634 switch (SvTYPE(sv)) {
1635 case SVt_NULL:
bd81e77b 1636 case SVt_NV:
3376de98 1637 sv_upgrade(sv, SVt_IV);
bd81e77b 1638 break;
bd81e77b
NC
1639 case SVt_PV:
1640 sv_upgrade(sv, SVt_PVIV);
1641 break;
463ee0b2 1642
bd81e77b 1643 case SVt_PVGV:
6e592b3a
BM
1644 if (!isGV_with_GP(sv))
1645 break;
bd81e77b
NC
1646 case SVt_PVAV:
1647 case SVt_PVHV:
1648 case SVt_PVCV:
1649 case SVt_PVFM:
1650 case SVt_PVIO:
22e74366 1651 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b
NC
1652 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1653 OP_DESC(PL_op));
42d0e0b7 1654 default: NOOP;
bd81e77b
NC
1655 }
1656 (void)SvIOK_only(sv); /* validate number */
1657 SvIV_set(sv, i);
1658 SvTAINT(sv);
1659}
932e9ff9 1660
bd81e77b
NC
1661/*
1662=for apidoc sv_setiv_mg
d33b2eba 1663
bd81e77b 1664Like C<sv_setiv>, but also handles 'set' magic.
1c846c1f 1665
bd81e77b
NC
1666=cut
1667*/
d33b2eba 1668
bd81e77b 1669void
5aaab254 1670Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
bd81e77b 1671{
7918f24d
NC
1672 PERL_ARGS_ASSERT_SV_SETIV_MG;
1673
bd81e77b
NC
1674 sv_setiv(sv,i);
1675 SvSETMAGIC(sv);
1676}
727879eb 1677
bd81e77b
NC
1678/*
1679=for apidoc sv_setuv
d33b2eba 1680
bd81e77b
NC
1681Copies an unsigned integer into the given SV, upgrading first if necessary.
1682Does not handle 'set' magic. See also C<sv_setuv_mg>.
9b94d1dd 1683
bd81e77b
NC
1684=cut
1685*/
d33b2eba 1686
bd81e77b 1687void
5aaab254 1688Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
bd81e77b 1689{
7918f24d
NC
1690 PERL_ARGS_ASSERT_SV_SETUV;
1691
013abb9b
NC
1692 /* With the if statement to ensure that integers are stored as IVs whenever
1693 possible:
bd81e77b 1694 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
d33b2eba 1695
bd81e77b
NC
1696 without
1697 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1c846c1f 1698
013abb9b
NC
1699 If you wish to remove the following if statement, so that this routine
1700 (and its callers) always return UVs, please benchmark to see what the
1701 effect is. Modern CPUs may be different. Or may not :-)
bd81e77b
NC
1702 */
1703 if (u <= (UV)IV_MAX) {
1704 sv_setiv(sv, (IV)u);
1705 return;
1706 }
1707 sv_setiv(sv, 0);
1708 SvIsUV_on(sv);
1709 SvUV_set(sv, u);
1710}
d33b2eba 1711
bd81e77b
NC
1712/*
1713=for apidoc sv_setuv_mg
727879eb 1714
bd81e77b 1715Like C<sv_setuv>, but also handles 'set' magic.
9b94d1dd 1716
bd81e77b
NC
1717=cut
1718*/
5e2fc214 1719
bd81e77b 1720void
5aaab254 1721Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
bd81e77b 1722{
7918f24d
NC
1723 PERL_ARGS_ASSERT_SV_SETUV_MG;
1724
bd81e77b
NC
1725 sv_setuv(sv,u);
1726 SvSETMAGIC(sv);
1727}
5e2fc214 1728
954c1994 1729/*
bd81e77b 1730=for apidoc sv_setnv
954c1994 1731
bd81e77b
NC
1732Copies a double into the given SV, upgrading first if necessary.
1733Does not handle 'set' magic. See also C<sv_setnv_mg>.
954c1994
GS
1734
1735=cut
1736*/
1737
63f97190 1738void
5aaab254 1739Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
79072805 1740{
7918f24d
NC
1741 PERL_ARGS_ASSERT_SV_SETNV;
1742
bd81e77b
NC
1743 SV_CHECK_THINKFIRST_COW_DROP(sv);
1744 switch (SvTYPE(sv)) {
79072805 1745 case SVt_NULL:
79072805 1746 case SVt_IV:
bd81e77b 1747 sv_upgrade(sv, SVt_NV);
79072805
LW
1748 break;
1749 case SVt_PV:
79072805 1750 case SVt_PVIV:
bd81e77b 1751 sv_upgrade(sv, SVt_PVNV);
79072805 1752 break;
bd4b1eb5 1753
bd4b1eb5 1754 case SVt_PVGV:
6e592b3a
BM
1755 if (!isGV_with_GP(sv))
1756 break;
bd81e77b
NC
1757 case SVt_PVAV:
1758 case SVt_PVHV:
79072805 1759 case SVt_PVCV:
bd81e77b
NC
1760 case SVt_PVFM:
1761 case SVt_PVIO:
22e74366 1762 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b 1763 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
94bbb3f4 1764 OP_DESC(PL_op));
42d0e0b7 1765 default: NOOP;
2068cd4d 1766 }
bd81e77b
NC
1767 SvNV_set(sv, num);
1768 (void)SvNOK_only(sv); /* validate number */
1769 SvTAINT(sv);
79072805
LW
1770}
1771
645c22ef 1772/*
bd81e77b 1773=for apidoc sv_setnv_mg
645c22ef 1774
bd81e77b 1775Like C<sv_setnv>, but also handles 'set' magic.
645c22ef
DM
1776
1777=cut
1778*/
1779
bd81e77b 1780void
5aaab254 1781Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
79072805 1782{
7918f24d
NC
1783 PERL_ARGS_ASSERT_SV_SETNV_MG;
1784
bd81e77b
NC
1785 sv_setnv(sv,num);
1786 SvSETMAGIC(sv);
79072805
LW
1787}
1788
3f7602fa
TC
1789/* Return a cleaned-up, printable version of sv, for non-numeric, or
1790 * not incrementable warning display.
1791 * Originally part of S_not_a_number().
1792 * The return value may be != tmpbuf.
bd81e77b 1793 */
954c1994 1794
3f7602fa
TC
1795STATIC const char *
1796S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1797 const char *pv;
94463019 1798
3f7602fa 1799 PERL_ARGS_ASSERT_SV_DISPLAY;
7918f24d 1800
94463019 1801 if (DO_UTF8(sv)) {
3f7602fa 1802 SV *dsv = newSVpvs_flags("", SVs_TEMP);
472394e4 1803 pv = sv_uni_display(dsv, sv, 10, UNI_DISPLAY_ISPRINT);
94463019
JH
1804 } else {
1805 char *d = tmpbuf;
3f7602fa 1806 const char * const limit = tmpbuf + tmpbuf_size - 8;
94463019
JH
1807 /* each *s can expand to 4 chars + "...\0",
1808 i.e. need room for 8 chars */
ecdeb87c 1809
00b6aa41
AL
1810 const char *s = SvPVX_const(sv);
1811 const char * const end = s + SvCUR(sv);
1812 for ( ; s < end && d < limit; s++ ) {
94463019 1813 int ch = *s & 0xFF;
bd27cf70 1814 if (! isASCII(ch) && !isPRINT_LC(ch)) {
94463019
JH
1815 *d++ = 'M';
1816 *d++ = '-';
bd27cf70
KW
1817
1818 /* Map to ASCII "equivalent" of Latin1 */
1819 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
94463019
JH
1820 }
1821 if (ch == '\n') {
1822 *d++ = '\\';
1823 *d++ = 'n';
1824 }
1825 else if (ch == '\r') {
1826 *d++ = '\\';
1827 *d++ = 'r';
1828 }
1829 else if (ch == '\f') {
1830 *d++ = '\\';
1831 *d++ = 'f';
1832 }
1833 else if (ch == '\\') {
1834 *d++ = '\\';
1835 *d++ = '\\';
1836 }
1837 else if (ch == '\0') {
1838 *d++ = '\\';
1839 *d++ = '0';
1840 }
1841 else if (isPRINT_LC(ch))
1842 *d++ = ch;
1843 else {
1844 *d++ = '^';
1845 *d++ = toCTRL(ch);
1846 }
1847 }
1848 if (s < end) {
1849 *d++ = '.';
1850 *d++ = '.';
1851 *d++ = '.';
1852 }
1853 *d = '\0';
1854 pv = tmpbuf;
a0d0e21e 1855 }
a0d0e21e 1856
3f7602fa
TC
1857 return pv;
1858}
1859
1860/* Print an "isn't numeric" warning, using a cleaned-up,
1861 * printable version of the offending string
1862 */
1863
1864STATIC void
1865S_not_a_number(pTHX_ SV *const sv)
1866{
3f7602fa
TC
1867 char tmpbuf[64];
1868 const char *pv;
1869
1870 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1871
1872 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1873
533c011a 1874 if (PL_op)
9014280d 1875 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1876 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019
JH
1877 "Argument \"%s\" isn't numeric in %s", pv,
1878 OP_DESC(PL_op));
a0d0e21e 1879 else
9014280d 1880 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1881 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019 1882 "Argument \"%s\" isn't numeric", pv);
a0d0e21e
LW
1883}
1884
3f7602fa
TC
1885STATIC void
1886S_not_incrementable(pTHX_ SV *const sv) {
3f7602fa
TC
1887 char tmpbuf[64];
1888 const char *pv;
1889
1890 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1891
1892 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1893
1894 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1895 "Argument \"%s\" treated as 0 in increment (++)", pv);
1896}
1897
c2988b20
NC
1898/*
1899=for apidoc looks_like_number
1900
645c22ef
DM
1901Test if the content of an SV looks like a number (or is a number).
1902C<Inf> and C<Infinity> are treated as numbers (so will not issue a
f52e41ad
FC
1903non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1904ignored.
c2988b20
NC
1905
1906=cut
1907*/
1908
1909I32
aad570aa 1910Perl_looks_like_number(pTHX_ SV *const sv)
c2988b20 1911{
eb578fdb 1912 const char *sbegin;
c2988b20
NC
1913 STRLEN len;
1914
7918f24d
NC
1915 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1916
f52e41ad
FC
1917 if (SvPOK(sv) || SvPOKp(sv)) {
1918 sbegin = SvPV_nomg_const(sv, len);
c2988b20 1919 }
c2988b20 1920 else
e0ab1c0e 1921 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
c2988b20
NC
1922 return grok_number(sbegin, len, NULL);
1923}
25da4f38 1924
19f6321d
NC
1925STATIC bool
1926S_glob_2number(pTHX_ GV * const gv)
180488f8 1927{
7918f24d
NC
1928 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1929
675c862f
AL
1930 /* We know that all GVs stringify to something that is not-a-number,
1931 so no need to test that. */
1932 if (ckWARN(WARN_NUMERIC))
8e629ff4
FC
1933 {
1934 SV *const buffer = sv_newmortal();
1935 gv_efullname3(buffer, gv, "*");
675c862f 1936 not_a_number(buffer);
8e629ff4 1937 }
675c862f
AL
1938 /* We just want something true to return, so that S_sv_2iuv_common
1939 can tail call us and return true. */
19f6321d 1940 return TRUE;
675c862f
AL
1941}
1942
25da4f38
IZ
1943/* Actually, ISO C leaves conversion of UV to IV undefined, but
1944 until proven guilty, assume that things are not that bad... */
1945
645c22ef
DM
1946/*
1947 NV_PRESERVES_UV:
1948
1949 As 64 bit platforms often have an NV that doesn't preserve all bits of
28e5dec8
JH
1950 an IV (an assumption perl has been based on to date) it becomes necessary
1951 to remove the assumption that the NV always carries enough precision to
1952 recreate the IV whenever needed, and that the NV is the canonical form.
1953 Instead, IV/UV and NV need to be given equal rights. So as to not lose
645c22ef 1954 precision as a side effect of conversion (which would lead to insanity
28e5dec8 1955 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
8a4a3196
KW
1956 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1957 where precision was lost, and IV/UV/NV slots that have a valid conversion
1958 which has lost no precision
645c22ef 1959 2) to ensure that if a numeric conversion to one form is requested that
28e5dec8
JH
1960 would lose precision, the precise conversion (or differently
1961 imprecise conversion) is also performed and cached, to prevent
1962 requests for different numeric formats on the same SV causing
1963 lossy conversion chains. (lossless conversion chains are perfectly
1964 acceptable (still))
1965
1966
1967 flags are used:
1968 SvIOKp is true if the IV slot contains a valid value
1969 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1970 SvNOKp is true if the NV slot contains a valid value
1971 SvNOK is true only if the NV value is accurate
1972
1973 so
645c22ef 1974 while converting from PV to NV, check to see if converting that NV to an
28e5dec8
JH
1975 IV(or UV) would lose accuracy over a direct conversion from PV to
1976 IV(or UV). If it would, cache both conversions, return NV, but mark
1977 SV as IOK NOKp (ie not NOK).
1978
645c22ef 1979 While converting from PV to IV, check to see if converting that IV to an
28e5dec8
JH
1980 NV would lose accuracy over a direct conversion from PV to NV. If it
1981 would, cache both conversions, flag similarly.
1982
1983 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
1984 correctly because if IV & NV were set NV *always* overruled.
645c22ef
DM
1985 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
1986 changes - now IV and NV together means that the two are interchangeable:
28e5dec8 1987 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
d460ef45 1988
645c22ef
DM
1989 The benefit of this is that operations such as pp_add know that if
1990 SvIOK is true for both left and right operands, then integer addition
1991 can be used instead of floating point (for cases where the result won't
1992 overflow). Before, floating point was always used, which could lead to
28e5dec8
JH
1993 loss of precision compared with integer addition.
1994
1995 * making IV and NV equal status should make maths accurate on 64 bit
1996 platforms
1997 * may speed up maths somewhat if pp_add and friends start to use
645c22ef 1998 integers when possible instead of fp. (Hopefully the overhead in
28e5dec8
JH
1999 looking for SvIOK and checking for overflow will not outweigh the
2000 fp to integer speedup)
2001 * will slow down integer operations (callers of SvIV) on "inaccurate"
2002 values, as the change from SvIOK to SvIOKp will cause a call into
2003 sv_2iv each time rather than a macro access direct to the IV slot
2004 * should speed up number->string conversion on integers as IV is
645c22ef 2005 favoured when IV and NV are equally accurate
28e5dec8
JH
2006
2007 ####################################################################
645c22ef
DM
2008 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2009 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2010 On the other hand, SvUOK is true iff UV.
28e5dec8
JH
2011 ####################################################################
2012
645c22ef 2013 Your mileage will vary depending your CPU's relative fp to integer
28e5dec8
JH
2014 performance ratio.
2015*/
2016
2017#ifndef NV_PRESERVES_UV
645c22ef
DM
2018# define IS_NUMBER_UNDERFLOW_IV 1
2019# define IS_NUMBER_UNDERFLOW_UV 2
2020# define IS_NUMBER_IV_AND_UV 2
2021# define IS_NUMBER_OVERFLOW_IV 4
2022# define IS_NUMBER_OVERFLOW_UV 5
2023
2024/* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
28e5dec8
JH
2025
2026/* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2027STATIC int
5aaab254 2028S_sv_2iuv_non_preserve(pTHX_ SV *const sv
47031da6
NC
2029# ifdef DEBUGGING
2030 , I32 numtype
2031# endif
2032 )
28e5dec8 2033{
7918f24d 2034 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
23491f1d 2035 PERL_UNUSED_CONTEXT;
7918f24d 2036
3f7c398e 2037 DEBUG_c(PerlIO_printf(Perl_debug_log,"sv_2iuv_non '%s', IV=0x%"UVxf" NV=%"NVgf" inttype=%"UVXf"\n", SvPVX_const(sv), SvIVX(sv), SvNVX(sv), (UV)numtype));
28e5dec8
JH
2038 if (SvNVX(sv) < (NV)IV_MIN) {
2039 (void)SvIOKp_on(sv);
2040 (void)SvNOK_on(sv);
45977657 2041 SvIV_set(sv, IV_MIN);
28e5dec8
JH
2042 return IS_NUMBER_UNDERFLOW_IV;
2043 }
2044 if (SvNVX(sv) > (NV)UV_MAX) {
2045 (void)SvIOKp_on(sv);
2046 (void)SvNOK_on(sv);
2047 SvIsUV_on(sv);
607fa7f2 2048 SvUV_set(sv, UV_MAX);
28e5dec8
JH
2049 return IS_NUMBER_OVERFLOW_UV;
2050 }
c2988b20
NC
2051 (void)SvIOKp_on(sv);
2052 (void)SvNOK_on(sv);
2053 /* Can't use strtol etc to convert this string. (See truth table in
2054 sv_2iv */
2055 if (SvNVX(sv) <= (UV)IV_MAX) {
45977657 2056 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2057 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2058 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2059 } else {
2060 /* Integer is imprecise. NOK, IOKp */
2061 }
2062 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2063 }
2064 SvIsUV_on(sv);
607fa7f2 2065 SvUV_set(sv, U_V(SvNVX(sv)));
659c4b96 2066 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2067 if (SvUVX(sv) == UV_MAX) {
2068 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2069 possibly be preserved by NV. Hence, it must be overflow.
2070 NOK, IOKp */
2071 return IS_NUMBER_OVERFLOW_UV;
2072 }
2073 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2074 } else {
2075 /* Integer is imprecise. NOK, IOKp */
28e5dec8 2076 }
c2988b20 2077 return IS_NUMBER_OVERFLOW_IV;
28e5dec8 2078}
645c22ef
DM
2079#endif /* !NV_PRESERVES_UV*/
2080
a13f4dff 2081/* If numtype is infnan, set the NV of the sv accordingly.
5564cd7f 2082 * If numtype is anything else, try setting the NV using Atof(PV). */
a13f4dff 2083static void
5564cd7f 2084S_sv_setnv(pTHX_ SV* sv, int numtype)
a13f4dff 2085{
07925c5e 2086 bool pok = cBOOL(SvPOK(sv));
5564cd7f 2087 bool nok = FALSE;
a13f4dff
JH
2088 if ((numtype & IS_NUMBER_INFINITY)) {
2089 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
5564cd7f 2090 nok = TRUE;
a13f4dff
JH
2091 }
2092 else if ((numtype & IS_NUMBER_NAN)) {
2093 SvNV_set(sv, NV_NAN);
d48bd569 2094 nok = TRUE;
a13f4dff 2095 }
d48bd569 2096 else if (pok) {
a13f4dff 2097 SvNV_set(sv, Atof(SvPVX_const(sv)));
d48bd569
JH
2098 /* Purposefully no true nok here, since we don't want to blow
2099 * away the possible IOK/UV of an existing sv. */
2100 }
5564cd7f 2101 if (nok) {
d48bd569 2102 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
5564cd7f
JH
2103 if (pok)
2104 SvPOK_on(sv); /* PV is okay, though. */
2105 }
a13f4dff
JH
2106}
2107
af359546 2108STATIC bool
7918f24d
NC
2109S_sv_2iuv_common(pTHX_ SV *const sv)
2110{
7918f24d
NC
2111 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2112
af359546 2113 if (SvNOKp(sv)) {
28e5dec8
JH
2114 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2115 * without also getting a cached IV/UV from it at the same time
2116 * (ie PV->NV conversion should detect loss of accuracy and cache
af359546
NC
2117 * IV or UV at same time to avoid this. */
2118 /* IV-over-UV optimisation - choose to cache IV if possible */
25da4f38 2119
a83034fc 2120 if (UNLIKELY(Perl_isinfnan(SvNVX(sv))))
0f83c5a4
JH
2121 return FALSE;
2122
25da4f38
IZ
2123 if (SvTYPE(sv) == SVt_NV)
2124 sv_upgrade(sv, SVt_PVNV);
2125
28e5dec8
JH
2126 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2127 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2128 certainly cast into the IV range at IV_MAX, whereas the correct
2129 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2130 cases go to UV */
2131 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
45977657 2132 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2133 if (SvNVX(sv) == (NV) SvIVX(sv)
28e5dec8
JH
2134#ifndef NV_PRESERVES_UV
2135 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2136 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2137 /* Don't flag it as "accurately an integer" if the number
2138 came from a (by definition imprecise) NV operation, and
2139 we're outside the range of NV integer precision */
2140#endif
2141 ) {
a43d94f2
NC
2142 if (SvNOK(sv))
2143 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2144 else {
2145 /* scalar has trailing garbage, eg "42a" */
2146 }
28e5dec8 2147 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2148 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
28e5dec8
JH
2149 PTR2UV(sv),
2150 SvNVX(sv),
2151 SvIVX(sv)));
2152
2153 } else {
2154 /* IV not precise. No need to convert from PV, as NV
2155 conversion would already have cached IV if it detected
2156 that PV->IV would be better than PV->NV->IV
2157 flags already correct - don't set public IOK. */
2158 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2159 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
28e5dec8
JH
2160 PTR2UV(sv),
2161 SvNVX(sv),
2162 SvIVX(sv)));
2163 }
2164 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2165 but the cast (NV)IV_MIN rounds to a the value less (more
2166 negative) than IV_MIN which happens to be equal to SvNVX ??
2167 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2168 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2169 (NV)UVX == NVX are both true, but the values differ. :-(
2170 Hopefully for 2s complement IV_MIN is something like
2171 0x8000000000000000 which will be exact. NWC */
d460ef45 2172 }
25da4f38 2173 else {
607fa7f2 2174 SvUV_set(sv, U_V(SvNVX(sv)));
28e5dec8 2175 if (
659c4b96 2176 (SvNVX(sv) == (NV) SvUVX(sv))
28e5dec8
JH
2177#ifndef NV_PRESERVES_UV
2178 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2179 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2180 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2181 /* Don't flag it as "accurately an integer" if the number
2182 came from a (by definition imprecise) NV operation, and
2183 we're outside the range of NV integer precision */
2184#endif
a43d94f2 2185 && SvNOK(sv)
28e5dec8
JH
2186 )
2187 SvIOK_on(sv);
25da4f38 2188 SvIsUV_on(sv);
1c846c1f 2189 DEBUG_c(PerlIO_printf(Perl_debug_log,
57def98f 2190 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
56431972 2191 PTR2UV(sv),
57def98f
JH
2192 SvUVX(sv),
2193 SvUVX(sv)));
25da4f38 2194 }
748a9306 2195 }
cd84013a 2196 else if (SvPOKp(sv)) {
c2988b20 2197 UV value;
504618e9 2198 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
af359546 2199 /* We want to avoid a possible problem when we cache an IV/ a UV which
25da4f38 2200 may be later translated to an NV, and the resulting NV is not
c2988b20
NC
2201 the same as the direct translation of the initial string
2202 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2203 be careful to ensure that the value with the .456 is around if the
2204 NV value is requested in the future).
1c846c1f 2205
af359546 2206 This means that if we cache such an IV/a UV, we need to cache the
25da4f38 2207 NV as well. Moreover, we trade speed for space, and do not
28e5dec8 2208 cache the NV if we are sure it's not needed.
25da4f38 2209 */
16b7a9a4 2210
c2988b20
NC
2211 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2212 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2213 == IS_NUMBER_IN_UV) {
5e045b90 2214 /* It's definitely an integer, only upgrade to PVIV */
28e5dec8
JH
2215 if (SvTYPE(sv) < SVt_PVIV)
2216 sv_upgrade(sv, SVt_PVIV);
f7bbb42a 2217 (void)SvIOK_on(sv);
c2988b20
NC
2218 } else if (SvTYPE(sv) < SVt_PVNV)
2219 sv_upgrade(sv, SVt_PVNV);
28e5dec8 2220
a13f4dff 2221 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
5564cd7f 2222 S_sv_setnv(aTHX_ sv, numtype);
a13f4dff
JH
2223 return FALSE;
2224 }
2225
f2524eef 2226 /* If NVs preserve UVs then we only use the UV value if we know that
c2988b20
NC
2227 we aren't going to call atof() below. If NVs don't preserve UVs
2228 then the value returned may have more precision than atof() will
2229 return, even though value isn't perfectly accurate. */
2230 if ((numtype & (IS_NUMBER_IN_UV
2231#ifdef NV_PRESERVES_UV
2232 | IS_NUMBER_NOT_INT
2233#endif
2234 )) == IS_NUMBER_IN_UV) {
2235 /* This won't turn off the public IOK flag if it was set above */
2236 (void)SvIOKp_on(sv);
2237
2238 if (!(numtype & IS_NUMBER_NEG)) {
2239 /* positive */;
2240 if (value <= (UV)IV_MAX) {
45977657 2241 SvIV_set(sv, (IV)value);
c2988b20 2242 } else {
af359546 2243 /* it didn't overflow, and it was positive. */
607fa7f2 2244 SvUV_set(sv, value);
c2988b20
NC
2245 SvIsUV_on(sv);
2246 }
2247 } else {
2248 /* 2s complement assumption */
2249 if (value <= (UV)IV_MIN) {
45977657 2250 SvIV_set(sv, -(IV)value);
c2988b20
NC
2251 } else {
2252 /* Too negative for an IV. This is a double upgrade, but
d1be9408 2253 I'm assuming it will be rare. */
c2988b20
NC
2254 if (SvTYPE(sv) < SVt_PVNV)
2255 sv_upgrade(sv, SVt_PVNV);
2256 SvNOK_on(sv);
2257 SvIOK_off(sv);
2258 SvIOKp_on(sv);
9d6ce603 2259 SvNV_set(sv, -(NV)value);
45977657 2260 SvIV_set(sv, IV_MIN);
c2988b20
NC
2261 }
2262 }
2263 }
2264 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2265 will be in the previous block to set the IV slot, and the next
2266 block to set the NV slot. So no else here. */
2267
2268 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2269 != IS_NUMBER_IN_UV) {
2270 /* It wasn't an (integer that doesn't overflow the UV). */
5564cd7f 2271 S_sv_setnv(aTHX_ sv, numtype);
28e5dec8 2272
c2988b20
NC
2273 if (! numtype && ckWARN(WARN_NUMERIC))
2274 not_a_number(sv);
28e5dec8 2275
88cb8500 2276 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
c2988b20 2277 PTR2UV(sv), SvNVX(sv)));
28e5dec8 2278
28e5dec8 2279#ifdef NV_PRESERVES_UV
af359546
NC
2280 (void)SvIOKp_on(sv);
2281 (void)SvNOK_on(sv);
2282 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2283 SvIV_set(sv, I_V(SvNVX(sv)));
2284 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2285 SvIOK_on(sv);
2286 } else {
6f207bd3 2287 NOOP; /* Integer is imprecise. NOK, IOKp */
af359546
NC
2288 }
2289 /* UV will not work better than IV */
2290 } else {
2291 if (SvNVX(sv) > (NV)UV_MAX) {
2292 SvIsUV_on(sv);
2293 /* Integer is inaccurate. NOK, IOKp, is UV */
2294 SvUV_set(sv, UV_MAX);
af359546
NC
2295 } else {
2296 SvUV_set(sv, U_V(SvNVX(sv)));
2297 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2298 NV preservse UV so can do correct comparison. */
2299 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2300 SvIOK_on(sv);
af359546 2301 } else {
6f207bd3 2302 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
af359546
NC
2303 }
2304 }
4b0c9573 2305 SvIsUV_on(sv);
af359546 2306 }
28e5dec8 2307#else /* NV_PRESERVES_UV */
c2988b20
NC
2308 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2309 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
af359546 2310 /* The IV/UV slot will have been set from value returned by
c2988b20
NC
2311 grok_number above. The NV slot has just been set using
2312 Atof. */
560b0c46 2313 SvNOK_on(sv);
c2988b20
NC
2314 assert (SvIOKp(sv));
2315 } else {
2316 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2317 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2318 /* Small enough to preserve all bits. */
2319 (void)SvIOKp_on(sv);
2320 SvNOK_on(sv);
45977657 2321 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2322 if ((NV)(SvIVX(sv)) == SvNVX(sv))
c2988b20
NC
2323 SvIOK_on(sv);
2324 /* Assumption: first non-preserved integer is < IV_MAX,
2325 this NV is in the preserved range, therefore: */
2326 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2327 < (UV)IV_MAX)) {
32fdb065 2328 Perl_croak(aTHX_ "sv_2iv assumed (U_V(fabs((double)SvNVX(sv))) < (UV)IV_MAX) but SvNVX(sv)=%"NVgf" U_V is 0x%"UVxf", IV_MAX is 0x%"UVxf"\n", SvNVX(sv), U_V(SvNVX(sv)), (UV)IV_MAX);
c2988b20
NC
2329 }
2330 } else {
2331 /* IN_UV NOT_INT
2332 0 0 already failed to read UV.
2333 0 1 already failed to read UV.
2334 1 0 you won't get here in this case. IV/UV
2335 slot set, public IOK, Atof() unneeded.
2336 1 1 already read UV.
2337 so there's no point in sv_2iuv_non_preserve() attempting
2338 to use atol, strtol, strtoul etc. */
47031da6 2339# ifdef DEBUGGING
40a17c4c 2340 sv_2iuv_non_preserve (sv, numtype);
47031da6
NC
2341# else
2342 sv_2iuv_non_preserve (sv);
2343# endif
c2988b20
NC
2344 }
2345 }
28e5dec8 2346#endif /* NV_PRESERVES_UV */
a43d94f2
NC
2347 /* It might be more code efficient to go through the entire logic above
2348 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2349 gets complex and potentially buggy, so more programmer efficient
2350 to do it this way, by turning off the public flags: */
2351 if (!numtype)
2352 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
25da4f38 2353 }
af359546
NC
2354 }
2355 else {
675c862f 2356 if (isGV_with_GP(sv))
159b6efe 2357 return glob_2number(MUTABLE_GV(sv));
180488f8 2358
4f62cd62 2359 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
af359546 2360 report_uninit(sv);
25da4f38
IZ
2361 if (SvTYPE(sv) < SVt_IV)
2362 /* Typically the caller expects that sv_any is not NULL now. */
2363 sv_upgrade(sv, SVt_IV);
af359546
NC
2364 /* Return 0 from the caller. */
2365 return TRUE;
2366 }
2367 return FALSE;
2368}
2369
2370/*
2371=for apidoc sv_2iv_flags
2372
2373Return the integer value of an SV, doing any necessary string
2374conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2375Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2376
2377=cut
2378*/
2379
2380IV
5aaab254 2381Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
af359546 2382{
1061065f 2383 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
4bac9ae4 2384
217f6fa3
FC
2385 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2386 && SvTYPE(sv) != SVt_PVFM);
2387
4bac9ae4
CS
2388 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2389 mg_get(sv);
2390
a83034fc 2391 if (SvNOK(sv) && UNLIKELY(Perl_isinfnan(SvNVX(sv))))
0f83c5a4
JH
2392 return 0; /* So wrong but what can we do. */
2393
4bac9ae4
CS
2394 if (SvROK(sv)) {
2395 if (SvAMAGIC(sv)) {
2396 SV * tmpstr;
2397 if (flags & SV_SKIP_OVERLOAD)
2398 return 0;
2399 tmpstr = AMG_CALLunary(sv, numer_amg);
2400 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2401 return SvIV(tmpstr);
2402 }
2403 }
2404 return PTR2IV(SvRV(sv));
2405 }
2406
8d919b0a 2407 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d
NC
2408 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2409 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2410 In practice they are extremely unlikely to actually get anywhere
2411 accessible by user Perl code - the only way that I'm aware of is when
2412 a constant subroutine which is used as the second argument to index.
cd84013a
FC
2413
2414 Regexps have no SvIVX and SvNVX fields.
2b2b6d6d 2415 */
8d919b0a 2416 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2417 {
71c558c3 2418 UV value;
8d919b0a
FC
2419 const char * const ptr =
2420 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
0f83c5a4
JH
2421 const int numtype = grok_number(ptr, SvCUR(sv), &value);
2422
2423 assert((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN)) == 0);
71c558c3
NC
2424
2425 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2426 == IS_NUMBER_IN_UV) {
2427 /* It's definitely an integer */
2428 if (numtype & IS_NUMBER_NEG) {
2429 if (value < (UV)IV_MIN)
2430 return -(IV)value;
2431 } else {
2432 if (value < (UV)IV_MAX)
2433 return (IV)value;
2434 }
2435 }
058b8ae2 2436
71c558c3
NC
2437 if (!numtype) {
2438 if (ckWARN(WARN_NUMERIC))
2439 not_a_number(sv);
2440 }
8d919b0a 2441 return I_V(Atof(ptr));
e20b6c3b 2442 }
4bac9ae4
CS
2443 }
2444
2445 if (SvTHINKFIRST(sv)) {
73381561 2446#ifdef PERL_OLD_COPY_ON_WRITE
af359546
NC
2447 if (SvIsCOW(sv)) {
2448 sv_force_normal_flags(sv, 0);
2449 }
73381561 2450#endif
af359546
NC
2451 if (SvREADONLY(sv) && !SvOK(sv)) {
2452 if (ckWARN(WARN_UNINITIALIZED))
2453 report_uninit(sv);
2454 return 0;
2455 }
2456 }
4bac9ae4 2457
af359546
NC
2458 if (!SvIOKp(sv)) {
2459 if (S_sv_2iuv_common(aTHX_ sv))
2460 return 0;
79072805 2461 }
4bac9ae4 2462
1d7c1841
GS
2463 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2464 PTR2UV(sv),SvIVX(sv)));
25da4f38 2465 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
79072805
LW
2466}
2467
645c22ef 2468/*
891f9566 2469=for apidoc sv_2uv_flags
645c22ef
DM
2470
2471Return the unsigned integer value of an SV, doing any necessary string
891f9566
YST
2472conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2473Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
645c22ef
DM
2474
2475=cut
2476*/
2477
ff68c719 2478UV
5aaab254 2479Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
ff68c719 2480{
1061065f 2481 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
4bac9ae4
CS
2482
2483 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2484 mg_get(sv);
2485
a83034fc 2486 if (SvNOK(sv) && UNLIKELY(Perl_isinfnan(SvNVX(sv))))
0f83c5a4
JH
2487 return 0; /* So wrong but what can we do. */
2488
4bac9ae4
CS
2489 if (SvROK(sv)) {
2490 if (SvAMAGIC(sv)) {
2491 SV *tmpstr;
2492 if (flags & SV_SKIP_OVERLOAD)
2493 return 0;
2494 tmpstr = AMG_CALLunary(sv, numer_amg);
2495 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2496 return SvUV(tmpstr);
2497 }
2498 }
2499 return PTR2UV(SvRV(sv));
2500 }
2501
8d919b0a 2502 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2503 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2504 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2505 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2506 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2507 {
71c558c3 2508 UV value;
8d919b0a
FC
2509 const char * const ptr =
2510 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
0f83c5a4
JH
2511 const int numtype = grok_number(ptr, SvCUR(sv), &value);
2512
2513 assert((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN)) == 0);
71c558c3
NC
2514
2515 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2516 == IS_NUMBER_IN_UV) {
2517 /* It's definitely an integer */
2518 if (!(numtype & IS_NUMBER_NEG))
2519 return value;
2520 }
058b8ae2 2521
71c558c3
NC
2522 if (!numtype) {
2523 if (ckWARN(WARN_NUMERIC))
2524 not_a_number(sv);
2525 }
8d919b0a 2526 return U_V(Atof(ptr));
e20b6c3b 2527 }
4bac9ae4
CS
2528 }
2529
2530 if (SvTHINKFIRST(sv)) {
73381561 2531#ifdef PERL_OLD_COPY_ON_WRITE
765f542d
NC
2532 if (SvIsCOW(sv)) {
2533 sv_force_normal_flags(sv, 0);
8a818333 2534 }
73381561 2535#endif
0336b60e 2536 if (SvREADONLY(sv) && !SvOK(sv)) {
0336b60e 2537 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2538 report_uninit(sv);
ff68c719
PP
2539 return 0;
2540 }
2541 }
4bac9ae4 2542
af359546
NC
2543 if (!SvIOKp(sv)) {
2544 if (S_sv_2iuv_common(aTHX_ sv))
2545 return 0;
ff68c719 2546 }
25da4f38 2547
1d7c1841
GS
2548 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2549 PTR2UV(sv),SvUVX(sv)));
25da4f38 2550 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
ff68c719
PP
2551}
2552
645c22ef 2553/*
196007d1 2554=for apidoc sv_2nv_flags
645c22ef
DM
2555
2556Return the num value of an SV, doing any necessary string or integer
fde67290 2557conversion. If flags includes SV_GMAGIC, does an mg_get() first.
39d5de13 2558Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
645c22ef
DM
2559
2560=cut
2561*/
2562
65202027 2563NV
5aaab254 2564Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
79072805 2565{
1061065f
DD
2566 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2567
217f6fa3
FC
2568 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2569 && SvTYPE(sv) != SVt_PVFM);
8d919b0a 2570 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2571 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2572 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2573 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2574 const char *ptr;
39d5de13
DM
2575 if (flags & SV_GMAGIC)
2576 mg_get(sv);
463ee0b2
LW
2577 if (SvNOKp(sv))
2578 return SvNVX(sv);
cd84013a 2579 if (SvPOKp(sv) && !SvIOKp(sv)) {
8d919b0a
FC
2580 ptr = SvPVX_const(sv);
2581 grokpv:
041457d9 2582 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
8d919b0a 2583 !grok_number(ptr, SvCUR(sv), NULL))
a0d0e21e 2584 not_a_number(sv);
8d919b0a 2585 return Atof(ptr);
a0d0e21e 2586 }
25da4f38 2587 if (SvIOKp(sv)) {
1c846c1f 2588 if (SvIsUV(sv))
65202027 2589 return (NV)SvUVX(sv);
25da4f38 2590 else
65202027 2591 return (NV)SvIVX(sv);
47a72cb8
NC
2592 }
2593 if (SvROK(sv)) {
2594 goto return_rok;
2595 }
8d919b0a
FC
2596 if (isREGEXP(sv)) {
2597 ptr = RX_WRAPPED((REGEXP *)sv);
2598 goto grokpv;
2599 }
47a72cb8
NC
2600 assert(SvTYPE(sv) >= SVt_PVMG);
2601 /* This falls through to the report_uninit near the end of the
2602 function. */
2603 } else if (SvTHINKFIRST(sv)) {
a0d0e21e 2604 if (SvROK(sv)) {
47a72cb8 2605 return_rok:
deb46114 2606 if (SvAMAGIC(sv)) {
aee036bb
DM
2607 SV *tmpstr;
2608 if (flags & SV_SKIP_OVERLOAD)
2609 return 0;
31d632c3 2610 tmpstr = AMG_CALLunary(sv, numer_amg);
deb46114
NC
2611 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2612 return SvNV(tmpstr);
2613 }
2614 }
2615 return PTR2NV(SvRV(sv));
a0d0e21e 2616 }
73381561 2617#ifdef PERL_OLD_COPY_ON_WRITE
765f542d
NC
2618 if (SvIsCOW(sv)) {
2619 sv_force_normal_flags(sv, 0);
8a818333 2620 }
73381561 2621#endif
0336b60e 2622 if (SvREADONLY(sv) && !SvOK(sv)) {
599cee73 2623 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2624 report_uninit(sv);
ed6116ce
LW
2625 return 0.0;
2626 }
79072805
LW
2627 }
2628 if (SvTYPE(sv) < SVt_NV) {
7e25a7e9
NC
2629 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2630 sv_upgrade(sv, SVt_NV);
097ee67d 2631 DEBUG_c({
f93f4e46 2632 STORE_NUMERIC_LOCAL_SET_STANDARD();
1d7c1841 2633 PerlIO_printf(Perl_debug_log,
88cb8500 2634 "0x%"UVxf" num(%" NVgf ")\n",
1d7c1841 2635 PTR2UV(sv), SvNVX(sv));
097ee67d
JH
2636 RESTORE_NUMERIC_LOCAL();
2637 });
79072805
LW
2638 }
2639 else if (SvTYPE(sv) < SVt_PVNV)
2640 sv_upgrade(sv, SVt_PVNV);
59d8ce62
NC
2641 if (SvNOKp(sv)) {
2642 return SvNVX(sv);
61604483 2643 }
59d8ce62 2644 if (SvIOKp(sv)) {
9d6ce603 2645 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
28e5dec8 2646#ifdef NV_PRESERVES_UV
a43d94f2
NC
2647 if (SvIOK(sv))
2648 SvNOK_on(sv);
2649 else
2650 SvNOKp_on(sv);
28e5dec8
JH
2651#else
2652 /* Only set the public NV OK flag if this NV preserves the IV */
2653 /* Check it's not 0xFFFFFFFFFFFFFFFF */
a43d94f2
NC
2654 if (SvIOK(sv) &&
2655 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
28e5dec8
JH
2656 : (SvIVX(sv) == I_V(SvNVX(sv))))
2657 SvNOK_on(sv);
2658 else
2659 SvNOKp_on(sv);
2660#endif
93a17b20 2661 }
cd84013a 2662 else if (SvPOKp(sv)) {
c2988b20 2663 UV value;
3f7c398e 2664 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
041457d9 2665 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
a0d0e21e 2666 not_a_number(sv);
28e5dec8 2667#ifdef NV_PRESERVES_UV
c2988b20
NC
2668 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2669 == IS_NUMBER_IN_UV) {
5e045b90 2670 /* It's definitely an integer */
9d6ce603 2671 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
66d83377 2672 } else {
5564cd7f 2673 S_sv_setnv(aTHX_ sv, numtype);
66d83377 2674 }
a43d94f2
NC
2675 if (numtype)
2676 SvNOK_on(sv);
2677 else
2678 SvNOKp_on(sv);
28e5dec8 2679#else
0f83c5a4
JH
2680 if ((numtype & IS_NUMBER_INFINITY)) {
2681 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
2682 SvNOK_on(sv);
2683 } else if ((numtype & IS_NUMBER_NAN)) {
2684 SvNV_set(sv, NV_NAN);
c2988b20
NC
2685 SvNOK_on(sv);
2686 } else {
0f83c5a4
JH
2687 SvNV_set(sv, Atof(SvPVX_const(sv)));
2688 /* Only set the public NV OK flag if this NV preserves the value in
2689 the PV at least as well as an IV/UV would.
2690 Not sure how to do this 100% reliably. */
2691 /* if that shift count is out of range then Configure's test is
2692 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2693 UV_BITS */
2694 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2695 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2696 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2697 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2698 /* Can't use strtol etc to convert this string, so don't try.
2699 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2700 SvNOK_on(sv);
c2988b20 2701 } else {
0f83c5a4
JH
2702 /* value has been set. It may not be precise. */
2703 if ((numtype & IS_NUMBER_NEG) && (value > (UV)IV_MIN)) {
2704 /* 2s complement assumption for (UV)IV_MIN */
2705 SvNOK_on(sv); /* Integer is too negative. */
c2988b20 2706 } else {
0f83c5a4
JH
2707 SvNOKp_on(sv);
2708 SvIOKp_on(sv);
6fa402ec 2709
0f83c5a4
JH
2710 if (numtype & IS_NUMBER_NEG) {
2711 SvIV_set(sv, -(IV)value);
2712 } else if (value <= (UV)IV_MAX) {
2713 SvIV_set(sv, (IV)value);
2714 } else {
2715 SvUV_set(sv, value);
2716 SvIsUV_on(sv);
2717 }
c2988b20 2718
0f83c5a4
JH
2719 if (numtype & IS_NUMBER_NOT_INT) {
2720 /* I believe that even if the original PV had decimals,
2721 they are lost beyond the limit of the FP precision.
2722 However, neither is canonical, so both only get p
2723 flags. NWC, 2000/11/25 */
2724 /* Both already have p flags, so do nothing */
2725 } else {
2726 const NV nv = SvNVX(sv);
2727 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2728 if (SvIVX(sv) == I_V(nv)) {
c2988b20 2729 SvNOK_on(sv);
0f83c5a4
JH
2730 } else {
2731 /* It had no "." so it must be integer. */
2732 }
2733 SvIOK_on(sv);
2734 } else {
2735 /* between IV_MAX and NV(UV_MAX).
2736 Could be slightly > UV_MAX */
2737
2738 if (numtype & IS_NUMBER_NOT_INT) {
2739 /* UV and NV both imprecise. */
2740 } else {
2741 const UV nv_as_uv = U_V(nv);
2742
2743 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2744 SvNOK_on(sv);
2745 }
2746 SvIOK_on(sv);
c2988b20
NC
2747 }
2748 }
2749 }
2750 }
2751 }
0f83c5a4
JH
2752 /* It might be more code efficient to go through the entire logic above
2753 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2754 gets complex and potentially buggy, so more programmer efficient
a43d94f2 2755 to do it this way, by turning off the public flags: */
0f83c5a4
JH
2756 if (!numtype)
2757 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2758 }
28e5dec8 2759#endif /* NV_PRESERVES_UV */
93a17b20 2760 }
79072805 2761 else {
0f83c5a4
JH
2762 if (isGV_with_GP(sv)) {
2763 glob_2number(MUTABLE_GV(sv));
2764 return 0.0;
2765 }
180488f8 2766
0f83c5a4
JH
2767 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2768 report_uninit(sv);
2769 assert (SvTYPE(sv) >= SVt_NV);
2770 /* Typically the caller expects that sv_any is not NULL now. */
2771 /* XXX Ilya implies that this is a bug in callers that assume this
2772 and ideally should be fixed. */
2773 return 0.0;
79072805 2774 }
097ee67d 2775 DEBUG_c({
0f83c5a4
JH
2776 STORE_NUMERIC_LOCAL_SET_STANDARD();
2777 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" NVgf ")\n",
2778 PTR2UV(sv), SvNVX(sv));
2779 RESTORE_NUMERIC_LOCAL();
2780 });
463ee0b2 2781 return SvNVX(sv);
79072805
LW
2782}
2783
800401ee
JH
2784/*
2785=for apidoc sv_2num
2786
2787Return an SV with the numeric value of the source SV, doing any necessary
a196a5fa
JH
2788reference or overload conversion. You must use the C<SvNUM(sv)> macro to
2789access this function.
800401ee
JH
2790
2791=cut
2792*/
2793
2794SV *
5aaab254 2795Perl_sv_2num(pTHX_ SV *const sv)
800401ee 2796{
7918f24d
NC
2797 PERL_ARGS_ASSERT_SV_2NUM;
2798
b9ee0594
RGS
2799 if (!SvROK(sv))
2800 return sv;
800401ee 2801 if (SvAMAGIC(sv)) {
31d632c3 2802 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
a02ec77a 2803 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
800401ee
JH
2804 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2805 return sv_2num(tmpsv);
2806 }
2807 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2808}
2809
645c22ef
DM
2810/* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2811 * UV as a string towards the end of buf, and return pointers to start and
2812 * end of it.
2813 *
2814 * We assume that buf is at least TYPE_CHARS(UV) long.
2815 */
2816
864dbfa3 2817static char *
5de3775c 2818S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
25da4f38 2819{
25da4f38 2820 char *ptr = buf + TYPE_CHARS(UV);
823a54a3 2821 char * const ebuf = ptr;
25da4f38 2822 int sign;
25da4f38 2823
7918f24d
NC
2824 PERL_ARGS_ASSERT_UIV_2BUF;
2825
25da4f38
IZ
2826 if (is_uv)
2827 sign = 0;
2828 else if (iv >= 0) {
2829 uv = iv;
2830 sign = 0;
2831 } else {
2832 uv = -iv;
2833 sign = 1;
2834 }
2835 do {
eb160463 2836 *--ptr = '0' + (char)(uv % 10);
25da4f38
IZ
2837 } while (uv /= 10);
2838 if (sign)
2839 *--ptr = '-';
2840 *peob = ebuf;
2841 return ptr;
2842}
2843
0c7e610f 2844/* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
fde71869
JH
2845 * infinity or a not-a-number, writes the appropriate strings to the
2846 * buffer, including a zero byte. On success returns the written length,
5e85836e 2847 * excluding the zero byte, on failure (not an infinity, not a nan, or the
fb8cdbc5
JH
2848 * maxlen too small) returns zero.
2849 *
2850 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2851 * shared string constants we point to, instead of generating a new
2852 * string for each instance. */
0c7e610f 2853STATIC size_t
5e85836e 2854S_infnan_2pv(NV nv, char* buffer, size_t maxlen) {
fb8cdbc5 2855 assert(maxlen >= 4);
5e85836e 2856 if (maxlen < 4) /* "Inf\0", "NaN\0" */
0c7e610f
JH
2857 return 0;
2858 else {
2859 char* s = buffer;
3fd158ea 2860 if (Perl_isinf(nv)) {
128eeacb
DD
2861 if (nv < 0) {
2862 if (maxlen < 5) /* "-Inf\0" */
2863 return 0;
2864 *s++ = '-';
2865 }
2866 *s++ = 'I';
2867 *s++ = 'n';
2868 *s++ = 'f';
3fd158ea
JH
2869 } else if (Perl_isnan(nv)) {
2870 *s++ = 'N';
2871 *s++ = 'a';
2872 *s++ = 'N';
2873 /* XXX optionally output the payload mantissa bits as
2874 * "(unsigned)" (to match the nan("...") C99 function,
2875 * or maybe as "(0xhhh...)" would make more sense...
2876 * provide a format string so that the user can decide?
2877 * NOTE: would affect the maxlen and assert() logic.*/
128eeacb
DD
2878 }
2879
0c7e610f
JH
2880 else
2881 return 0;
5e85836e 2882 assert((s == buffer + 3) || (s == buffer + 4));
0c7e610f 2883 *s++ = 0;
5e85836e 2884 return s - buffer - 1; /* -1: excluding the zero byte */
0c7e610f
JH
2885 }
2886}
2887
645c22ef
DM
2888/*
2889=for apidoc sv_2pv_flags
2890
ff276b08 2891Returns a pointer to the string value of an SV, and sets *lp to its length.
fde67290
FC
2892If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2893string if necessary. Normally invoked via the C<SvPV_flags> macro.
2894C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
645c22ef
DM
2895
2896=cut
2897*/
2898
8d6d96c1 2899char *
5aaab254 2900Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
8d6d96c1 2901{
eb578fdb 2902 char *s;
79072805 2903
1061065f
DD
2904 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2905
217f6fa3
FC
2906 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2907 && SvTYPE(sv) != SVt_PVFM);
4bac9ae4
CS
2908 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2909 mg_get(sv);
2910 if (SvROK(sv)) {
2911 if (SvAMAGIC(sv)) {
2912 SV *tmpstr;
2913 if (flags & SV_SKIP_OVERLOAD)
2914 return NULL;
2915 tmpstr = AMG_CALLunary(sv, string_amg);
2916 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2917 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2918 /* Unwrap this: */
2919 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2920 */
75dfc8ec 2921
4bac9ae4
CS
2922 char *pv;
2923 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2924 if (flags & SV_CONST_RETURN) {
2925 pv = (char *) SvPVX_const(tmpstr);
50adf7d2 2926 } else {
4bac9ae4
CS
2927 pv = (flags & SV_MUTABLE_RETURN)
2928 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
50adf7d2 2929 }
4bac9ae4
CS
2930 if (lp)
2931 *lp = SvCUR(tmpstr);
2932 } else {
2933 pv = sv_2pv_flags(tmpstr, lp, flags);
50adf7d2 2934 }
4bac9ae4
CS
2935 if (SvUTF8(tmpstr))
2936 SvUTF8_on(sv);
2937 else
2938 SvUTF8_off(sv);
2939 return pv;
deb46114 2940 }
4bac9ae4
CS
2941 }
2942 {
2943 STRLEN len;
2944 char *retval;
2945 char *buffer;
2946 SV *const referent = SvRV(sv);
2947
2948 if (!referent) {
2949 len = 7;
2950 retval = buffer = savepvn("NULLREF", len);
2951 } else if (SvTYPE(referent) == SVt_REGEXP &&
2952 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
2953 amagic_is_enabled(string_amg))) {
2954 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2955
2956 assert(re);
67d2d14d 2957
4bac9ae4
CS
2958 /* If the regex is UTF-8 we want the containing scalar to
2959 have an UTF-8 flag too */
2960 if (RX_UTF8(re))
2961 SvUTF8_on(sv);
2962 else
2963 SvUTF8_off(sv);
67d2d14d 2964
4bac9ae4
CS
2965 if (lp)
2966 *lp = RX_WRAPLEN(re);
67d2d14d 2967
4bac9ae4
CS
2968 return RX_WRAPPED(re);
2969 } else {
2970 const char *const typestr = sv_reftype(referent, 0);
2971 const STRLEN typelen = strlen(typestr);
2972 UV addr = PTR2UV(referent);
2973 const char *stashname = NULL;
2974 STRLEN stashnamelen = 0; /* hush, gcc */
2975 const char *buffer_end;
2976
2977 if (SvOBJECT(referent)) {
2978 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
2979
2980 if (name) {
2981 stashname = HEK_KEY(name);
2982 stashnamelen = HEK_LEN(name);
2983
2984 if (HEK_UTF8(name)) {
2985 SvUTF8_on(sv);
fafee734 2986 } else {
4bac9ae4 2987 SvUTF8_off(sv);
fafee734 2988 }
fafee734 2989 } else {
4bac9ae4
CS
2990 stashname = "__ANON__";
2991 stashnamelen = 8;
fafee734 2992 }
4bac9ae4
CS
2993 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
2994 + 2 * sizeof(UV) + 2 /* )\0 */;
2995 } else {
2996 len = typelen + 3 /* (0x */
2997 + 2 * sizeof(UV) + 2 /* )\0 */;
2998 }
fafee734 2999
4bac9ae4
CS
3000 Newx(buffer, len, char);
3001 buffer_end = retval = buffer + len;
3002
3003 /* Working backwards */
3004 *--retval = '\0';
3005 *--retval = ')';
3006 do {
3007 *--retval = PL_hexdigit[addr & 15];
3008 } while (addr >>= 4);
3009 *--retval = 'x';
3010 *--retval = '0';
3011 *--retval = '(';
3012
3013 retval -= typelen;
3014 memcpy(retval, typestr, typelen);
3015
3016 if (stashname) {
3017 *--retval = '=';
3018 retval -= stashnamelen;
3019 memcpy(retval, stashname, stashnamelen);
c080367d 3020 }
4bac9ae4
CS
3021 /* retval may not necessarily have reached the start of the
3022 buffer here. */
3023 assert (retval >= buffer);
3024
3025 len = buffer_end - retval - 1; /* -1 for that \0 */
463ee0b2 3026 }
cdb061a3 3027 if (lp)
4bac9ae4
CS
3028 *lp = len;
3029 SAVEFREEPV(buffer);
3030 return retval;
79072805 3031 }
79072805 3032 }
4bac9ae4
CS
3033
3034 if (SvPOKp(sv)) {
3035 if (lp)
3036 *lp = SvCUR(sv);
3037 if (flags & SV_MUTABLE_RETURN)
3038 return SvPVX_mutable(sv);
3039 if (flags & SV_CONST_RETURN)
3040 return (char *)SvPVX_const(sv);
3041 return SvPVX(sv);
3042 }
3043
3044 if (SvIOK(sv)) {
28e5dec8
JH
3045 /* I'm assuming that if both IV and NV are equally valid then
3046 converting the IV is going to be more efficient */
e1ec3a88 3047 const U32 isUIOK = SvIsUV(sv);
28e5dec8
JH
3048 char buf[TYPE_CHARS(UV)];
3049 char *ebuf, *ptr;
97a130b8 3050 STRLEN len;
28e5dec8
JH
3051
3052 if (SvTYPE(sv) < SVt_PVIV)
3053 sv_upgrade(sv, SVt_PVIV);
4ea1d550 3054 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
97a130b8 3055 len = ebuf - ptr;
5902b6a9 3056 /* inlined from sv_setpvn */
97a130b8
NC
3057 s = SvGROW_mutable(sv, len + 1);
3058 Move(ptr, s, len, char);
3059 s += len;
28e5dec8 3060 *s = '\0';
b127e37e 3061 SvPOK_on(sv);
28e5dec8 3062 }
4bac9ae4 3063 else if (SvNOK(sv)) {
79072805
LW
3064 if (SvTYPE(sv) < SVt_PVNV)
3065 sv_upgrade(sv, SVt_PVNV);
128eeacb
DD
3066 if (SvNVX(sv) == 0.0
3067#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
0f83c5a4 3068 /* XXX Create SvNVXeq(sv, x)? Or just SvNVXzero(sv)? */
128eeacb
DD
3069 && !Perl_isnan(SvNVX(sv))
3070#endif
3071 ) {
29912d93
NC
3072 s = SvGROW_mutable(sv, 2);
3073 *s++ = '0';
3074 *s = '\0';
3075 } else {
5e85836e 3076 STRLEN len;
fb8cdbc5 3077 STRLEN size = 5; /* "-Inf\0" */
0c7e610f 3078
fb8cdbc5 3079 s = SvGROW_mutable(sv, size);
5e85836e 3080 len = S_infnan_2pv(SvNVX(sv), s, size);
fb8cdbc5 3081 if (len > 0) {
0c7e610f 3082 s += len;
fb8cdbc5
JH
3083 SvPOK_on(sv);
3084 }
0c7e610f 3085 else {
0c7e610f 3086 /* some Xenix systems wipe out errno here */
fb8cdbc5
JH
3087 dSAVE_ERRNO;
3088
3840bff0
JH
3089 size =
3090 1 + /* sign */
3091 1 + /* "." */
3092 NV_DIG +
3093 1 + /* "e" */
3094 1 + /* sign */
3095 5 + /* exponent digits */
3096 1 + /* \0 */
3097 2; /* paranoia */
b127e37e 3098
fb8cdbc5 3099 s = SvGROW_mutable(sv, size);
b127e37e 3100#ifndef USE_LOCALE_NUMERIC
a4eca1d4
JH
3101 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3102
0c7e610f
JH
3103 SvPOK_on(sv);
3104#else
28acfe03 3105 {
3840bff0 3106 bool local_radix;
0c7e610f 3107 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
3840bff0
JH
3108
3109 local_radix =
3110 PL_numeric_local &&
3111 PL_numeric_radix_sv &&
3112 SvUTF8(PL_numeric_radix_sv);
3113 if (local_radix && SvLEN(PL_numeric_radix_sv) > 1) {
3114 size += SvLEN(PL_numeric_radix_sv) - 1;
3115 s = SvGROW_mutable(sv, size);
3116 }
3117
a4eca1d4 3118 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
0c7e610f
JH
3119
3120 /* If the radix character is UTF-8, and actually is in the
3121 * output, turn on the UTF-8 flag for the scalar */
3840bff0
JH
3122 if (local_radix &&
3123 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3124 SvUTF8_on(sv);
3125 }
3126
0c7e610f 3127 RESTORE_LC_NUMERIC();
28acfe03 3128 }
68e8f474 3129
0c7e610f
JH
3130 /* We don't call SvPOK_on(), because it may come to
3131 * pass that the locale changes so that the
3132 * stringification we just did is no longer correct. We
3133 * will have to re-stringify every time it is needed */
b127e37e 3134#endif
0c7e610f
JH
3135 RESTORE_ERRNO;
3136 }
3137 while (*s) s++;
bbce6d69 3138 }
79072805 3139 }
4bac9ae4
CS
3140 else if (isGV_with_GP(sv)) {
3141 GV *const gv = MUTABLE_GV(sv);
3142 SV *const buffer = sv_newmortal();
8d1c3e26 3143
4bac9ae4 3144 gv_efullname3(buffer, gv, "*");
180488f8 3145
4bac9ae4
CS
3146 assert(SvPOK(buffer));
3147 if (SvUTF8(buffer))
3148 SvUTF8_on(sv);
3149 if (lp)
3150 *lp = SvCUR(buffer);
3151 return SvPVX(buffer);
3152 }
8d919b0a
FC
3153 else if (isREGEXP(sv)) {
3154 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3155 return RX_WRAPPED((REGEXP *)sv);
3156 }
4bac9ae4 3157 else {
cdb061a3 3158 if (lp)
00b6aa41 3159 *lp = 0;
9f621bb0
NC
3160 if (flags & SV_UNDEF_RETURNS_NULL)
3161 return NULL;
4f62cd62 3162 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
9f621bb0 3163 report_uninit(sv);
4bac9ae4
CS
3164 /* Typically the caller expects that sv_any is not NULL now. */
3165 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
25da4f38 3166 sv_upgrade(sv, SVt_PV);
73d840c0 3167 return (char *)"";
79072805 3168 }
4bac9ae4 3169
cdb061a3 3170 {
823a54a3 3171 const STRLEN len = s - SvPVX_const(sv);
cdb061a3
NC
3172 if (lp)
3173 *lp = len;
3174 SvCUR_set(sv, len);
3175 }
1d7c1841 3176 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3f7c398e 3177 PTR2UV(sv),SvPVX_const(sv)));
4d84ee25
NC
3178 if (flags & SV_CONST_RETURN)
3179 return (char *)SvPVX_const(sv);
10516c54
NC
3180 if (flags & SV_MUTABLE_RETURN)
3181 return SvPVX_mutable(sv);
463ee0b2
LW
3182 return SvPVX(sv);
3183}
3184
645c22ef 3185/*
6050d10e
JP
3186=for apidoc sv_copypv
3187
3188Copies a stringified representation of the source SV into the
3189destination SV. Automatically performs any necessary mg_get and
54f0641b 3190coercion of numeric values into strings. Guaranteed to preserve
2575c402 3191UTF8 flag even from overloaded objects. Similar in nature to
54f0641b
NIS
3192sv_2pv[_flags] but operates directly on an SV instead of just the
3193string. Mostly uses sv_2pv_flags to do its work, except when that
6050d10e
JP
3194would lose the UTF-8'ness of the PV.
3195
4bac9ae4
CS
3196=for apidoc sv_copypv_nomg
3197
3198Like sv_copypv, but doesn't invoke get magic first.
3199
3200=for apidoc sv_copypv_flags
3201
3202Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3203include SV_GMAGIC.
3204
6050d10e
JP
3205=cut
3206*/
3207
3208void
5aaab254 3209Perl_sv_copypv(pTHX_ SV *const dsv, SV *const ssv)
6050d10e 3210{
4bac9ae4
CS
3211 PERL_ARGS_ASSERT_SV_COPYPV;
3212
3213 sv_copypv_flags(dsv, ssv, 0);
3214}
3215
3216void
5aaab254 3217Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
4bac9ae4 3218{
446eaa42 3219 STRLEN len;
4bac9ae4 3220 const char *s;
7918f24d 3221
4bac9ae4 3222 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
7918f24d 3223
c77ed9ca 3224 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
cb50f42d 3225 sv_setpvn(dsv,s,len);
446eaa42 3226 if (SvUTF8(ssv))
cb50f42d 3227 SvUTF8_on(dsv);
446eaa42 3228 else
cb50f42d 3229 SvUTF8_off(dsv);
6050d10e
JP
3230}
3231
3232/*
645c22ef
DM
3233=for apidoc sv_2pvbyte
3234
3235Return a pointer to the byte-encoded representation of the SV, and set *lp
1e54db1a 3236to its length. May cause the SV to be downgraded from UTF-8 as a
645c22ef
DM
3237side-effect.
3238
3239Usually accessed via the C<SvPVbyte> macro.
3240
3241=cut
3242*/
3243
7340a771 3244char *
5aaab254 3245Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3246{
7918f24d
NC
3247 PERL_ARGS_ASSERT_SV_2PVBYTE;
3248
48120f8f 3249 SvGETMAGIC(sv);
4499db73
FC
3250 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3251 || isGV_with_GP(sv) || SvROK(sv)) {
a901b181 3252 SV *sv2 = sv_newmortal();
48120f8f 3253 sv_copypv_nomg(sv2,sv);
a901b181
FC
3254 sv = sv2;
3255 }
0875d2fe 3256 sv_utf8_downgrade(sv,0);
71eb6d8c 3257 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771
GS
3258}
3259
645c22ef 3260/*
035cbb0e
RGS
3261=for apidoc sv_2pvutf8
3262
3263Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3264to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3265
3266Usually accessed via the C<SvPVutf8> macro.
3267
3268=cut
3269*/
645c22ef 3270
7340a771 3271char *
5aaab254 3272Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3273{
7918f24d
NC
3274 PERL_ARGS_ASSERT_SV_2PVUTF8;
3275
4499db73
FC
3276 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3277 || isGV_with_GP(sv) || SvROK(sv))
fe46cbda 3278 sv = sv_mortalcopy(sv);
4bac9ae4
CS
3279 else
3280 SvGETMAGIC(sv);
3281 sv_utf8_upgrade_nomg(sv);
c3ec315f 3282 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771 3283}
1c846c1f 3284
7ee2227d 3285
645c22ef
DM
3286/*
3287=for apidoc sv_2bool
3288
06c841cf
FC
3289This macro is only used by sv_true() or its macro equivalent, and only if
3290the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3291It calls sv_2bool_flags with the SV_GMAGIC flag.
3292
3293=for apidoc sv_2bool_flags
3294
3295This function is only used by sv_true() and friends, and only if
fde67290 3296the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
06c841cf
FC
3297contain SV_GMAGIC, then it does an mg_get() first.
3298
645c22ef
DM
3299
3300=cut
3301*/
3302
463ee0b2 3303bool
9d176cd8 3304Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
463ee0b2 3305{
06c841cf 3306 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
7918f24d 3307
9d176cd8 3308 restart:
06c841cf 3309 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
463ee0b2 3310
a0d0e21e
LW
3311 if (!SvOK(sv))
3312 return 0;
3313 if (SvROK(sv)) {
fabdb6c0 3314 if (SvAMAGIC(sv)) {
31d632c3 3315 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
9d176cd8
DD
3316 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3317 bool svb;
3318 sv = tmpsv;
3319 if(SvGMAGICAL(sv)) {
3320 flags = SV_GMAGIC;
3321 goto restart; /* call sv_2bool */
3322 }
3323 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3324 else if(!SvOK(sv)) {
3325 svb = 0;
3326 }
3327 else if(SvPOK(sv)) {
3328 svb = SvPVXtrue(sv);
3329 }
3330 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3331 svb = (SvIOK(sv) && SvIVX(sv) != 0)
659c4b96 3332 || (SvNOK(sv) && SvNVX(sv) != 0.0);
9d176cd8
DD
3333 }
3334 else {
3335 flags = 0;
3336 goto restart; /* call sv_2bool_nomg */
3337 }
3338 return cBOOL(svb);
3339 }
fabdb6c0
AL
3340 }
3341 return SvRV(sv) != 0;
a0d0e21e 3342 }
85b7d9b3
FC
3343 if (isREGEXP(sv))
3344 return
3345 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
4bac9ae4 3346 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
79072805
LW
3347}
3348
c461cf8f
JH
3349/*
3350=for apidoc sv_utf8_upgrade
3351
78ea37eb 3352Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3353Forces the SV to string form if it is not already.
2bbc8d55 3354Will C<mg_get> on C<sv> if appropriate.
4411f3b6 3355Always sets the SvUTF8 flag to avoid future validity checks even
2bbc8d55
SP
3356if the whole string is the same in UTF-8 as not.
3357Returns the number of bytes in the converted string
c461cf8f 3358
0efd0472 3359This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3360use the Encode extension for that.
3361
fe749c9a
KW
3362=for apidoc sv_utf8_upgrade_nomg
3363
fde67290 3364Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
fe749c9a 3365
8d6d96c1
HS
3366=for apidoc sv_utf8_upgrade_flags
3367
78ea37eb 3368Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3369Forces the SV to string form if it is not already.
8d6d96c1 3370Always sets the SvUTF8 flag to avoid future validity checks even
960b0271
FC
3371if all the bytes are invariant in UTF-8.
3372If C<flags> has C<SV_GMAGIC> bit set,
2bbc8d55 3373will C<mg_get> on C<sv> if appropriate, else not.
2a590426
KW
3374
3375If C<flags> has SV_FORCE_UTF8_UPGRADE set, this function assumes that the PV
3376will expand when converted to UTF-8, and skips the extra work of checking for
3377that. Typically this flag is used by a routine that has already parsed the
3378string and found such characters, and passes this information on so that the
3379work doesn't have to be repeated.
3380
3381Returns the number of bytes in the converted string.
8d6d96c1 3382
0efd0472 3383This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3384use the Encode extension for that.
3385
2a590426 3386=for apidoc sv_utf8_upgrade_flags_grow
b3ab6785 3387
2a590426
KW
3388Like sv_utf8_upgrade_flags, but has an additional parameter C<extra>, which is
3389the number of unused bytes the string of 'sv' is guaranteed to have free after
3390it upon return. This allows the caller to reserve extra space that it intends
3391to fill, to avoid extra grows.
b3ab6785 3392
2a590426
KW
3393C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3394are implemented in terms of this function.
3395
3396Returns the number of bytes in the converted string (not including the spares).
3397
3398=cut
b3ab6785
KW
3399
3400(One might think that the calling routine could pass in the position of the
2a590426
KW
3401first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3402have to be found again. But that is not the case, because typically when the
3403caller is likely to use this flag, it won't be calling this routine unless it
3404finds something that won't fit into a byte. Otherwise it tries to not upgrade
3405and just use bytes. But some things that do fit into a byte are variants in
3406utf8, and the caller may not have been keeping track of these.)
b3ab6785 3407
6602b933
KW
3408If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3409C<NUL> isn't guaranteed due to having other routines do the work in some input
3410cases, or if the input is already flagged as being in utf8.
b3ab6785
KW
3411
3412The speed of this could perhaps be improved for many cases if someone wanted to
3413write a fast function that counts the number of variant characters in a string,
3414especially if it could return the position of the first one.
3415
8d6d96c1
HS
3416*/
3417
3418STRLEN
5aaab254 3419Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
8d6d96c1 3420{
b3ab6785 3421 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
7918f24d 3422
808c356f
RGS
3423 if (sv == &PL_sv_undef)
3424 return 0;
892f9127 3425 if (!SvPOK_nog(sv)) {
e0e62c2a 3426 STRLEN len = 0;
d52b7888
NC
3427 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3428 (void) sv_2pv_flags(sv,&len, flags);
b3ab6785
KW
3429 if (SvUTF8(sv)) {
3430 if (extra) SvGROW(sv, SvCUR(sv) + extra);
d52b7888 3431 return len;
b3ab6785 3432 }
d52b7888 3433 } else {
33fb6f35 3434 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
d52b7888 3435 }
e0e62c2a 3436 }
4411f3b6 3437
f5cee72b 3438 if (SvUTF8(sv)) {
b3ab6785 3439 if (extra) SvGROW(sv, SvCUR(sv) + extra);
5fec3b1d 3440 return SvCUR(sv);
f5cee72b 3441 }
5fec3b1d 3442
765f542d 3443 if (SvIsCOW(sv)) {
c56ed9f6 3444 S_sv_uncow(aTHX_ sv, 0);
db42d148
NIS
3445 }
3446
b3ab6785 3447 if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING)) {
799ef3cb 3448 sv_recode_to_utf8(sv, PL_encoding);
b3ab6785
KW
3449 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3450 return SvCUR(sv);
3451 }
3452
4e93345f
KW
3453 if (SvCUR(sv) == 0) {
3454 if (extra) SvGROW(sv, extra);
3455 } else { /* Assume Latin-1/EBCDIC */
c4e7c712 3456 /* This function could be much more efficient if we
2bbc8d55 3457 * had a FLAG in SVs to signal if there are any variant
c4e7c712 3458 * chars in the PV. Given that there isn't such a flag
b3ab6785
KW
3459 * make the loop as fast as possible (although there are certainly ways
3460 * to speed this up, eg. through vectorization) */
3461 U8 * s = (U8 *) SvPVX_const(sv);
3462 U8 * e = (U8 *) SvEND(sv);
3463 U8 *t = s;
3464 STRLEN two_byte_count = 0;
c4e7c712 3465
b3ab6785
KW
3466 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3467
3468 /* See if really will need to convert to utf8. We mustn't rely on our
3469 * incoming SV being well formed and having a trailing '\0', as certain
3470 * code in pp_formline can send us partially built SVs. */
3471
c4e7c712 3472 while (t < e) {
53c1dcc0 3473 const U8 ch = *t++;
6f2d5cbc 3474 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
b3ab6785
KW
3475
3476 t--; /* t already incremented; re-point to first variant */
3477 two_byte_count = 1;
3478 goto must_be_utf8;
c4e7c712 3479 }
b3ab6785
KW
3480
3481 /* utf8 conversion not needed because all are invariants. Mark as
3482 * UTF-8 even if no variant - saves scanning loop */
c4e7c712 3483 SvUTF8_on(sv);
7f0bfbea 3484 if (extra) SvGROW(sv, SvCUR(sv) + extra);
b3ab6785
KW
3485 return SvCUR(sv);
3486
3487must_be_utf8:
3488
3489 /* Here, the string should be converted to utf8, either because of an
3490 * input flag (two_byte_count = 0), or because a character that
3491 * requires 2 bytes was found (two_byte_count = 1). t points either to
3492 * the beginning of the string (if we didn't examine anything), or to
3493 * the first variant. In either case, everything from s to t - 1 will
3494 * occupy only 1 byte each on output.
3495 *
3496 * There are two main ways to convert. One is to create a new string
3497 * and go through the input starting from the beginning, appending each
3498 * converted value onto the new string as we go along. It's probably
3499 * best to allocate enough space in the string for the worst possible
3500 * case rather than possibly running out of space and having to
3501 * reallocate and then copy what we've done so far. Since everything
3502 * from s to t - 1 is invariant, the destination can be initialized
3503 * with these using a fast memory copy
3504 *
3505 * The other way is to figure out exactly how big the string should be
3506 * by parsing the entire input. Then you don't have to make it big
3507 * enough to handle the worst possible case, and more importantly, if
3508 * the string you already have is large enough, you don't have to
3509 * allocate a new string, you can copy the last character in the input
3510 * string to the final position(s) that will be occupied by the
3511 * converted string and go backwards, stopping at t, since everything
3512 * before that is invariant.
3513 *
3514 * There are advantages and disadvantages to each method.
3515 *
3516 * In the first method, we can allocate a new string, do the memory
3517 * copy from the s to t - 1, and then proceed through the rest of the
3518 * string byte-by-byte.
3519 *
3520 * In the second method, we proceed through the rest of the input
3521 * string just calculating how big the converted string will be. Then
3522 * there are two cases:
3523 * 1) if the string has enough extra space to handle the converted
3524 * value. We go backwards through the string, converting until we
3525 * get to the position we are at now, and then stop. If this
3526 * position is far enough along in the string, this method is
3527 * faster than the other method. If the memory copy were the same
3528 * speed as the byte-by-byte loop, that position would be about
3529 * half-way, as at the half-way mark, parsing to the end and back
3530 * is one complete string's parse, the same amount as starting
3531 * over and going all the way through. Actually, it would be
3532 * somewhat less than half-way, as it's faster to just count bytes
3533 * than to also copy, and we don't have the overhead of allocating
3534 * a new string, changing the scalar to use it, and freeing the
3535 * existing one. But if the memory copy is fast, the break-even
3536 * point is somewhere after half way. The counting loop could be
3537 * sped up by vectorization, etc, to move the break-even point
3538 * further towards the beginning.
3539 * 2) if the string doesn't have enough space to handle the converted
3540 * value. A new string will have to be allocated, and one might
3541 * as well, given that, start from the beginning doing the first
3542 * method. We've spent extra time parsing the string and in
3543 * exchange all we've gotten is that we know precisely how big to
3544 * make the new one. Perl is more optimized for time than space,
3545 * so this case is a loser.
3546 * So what I've decided to do is not use the 2nd method unless it is
3547 * guaranteed that a new string won't have to be allocated, assuming
3548 * the worst case. I also decided not to put any more conditions on it
3549 * than this, for now. It seems likely that, since the worst case is
3550 * twice as big as the unknown portion of the string (plus 1), we won't
3551 * be guaranteed enough space, causing us to go to the first method,
3552 * unless the string is short, or the first variant character is near
3553 * the end of it. In either of these cases, it seems best to use the
3554 * 2nd method. The only circumstance I can think of where this would
3555 * be really slower is if the string had once had much more data in it
3556 * than it does now, but there is still a substantial amount in it */
3557
3558 {
3559 STRLEN invariant_head = t - s;
3560 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3561 if (SvLEN(sv) < size) {
3562
3563 /* Here, have decided to allocate a new string */
3564
3565 U8 *dst;
3566 U8 *d;
3567
3568 Newx(dst, size, U8);
3569
3570 /* If no known invariants at the beginning of the input string,
3571 * set so starts from there. Otherwise, can use memory copy to
3572 * get up to where we are now, and then start from here */
3573
5b26a7b3 3574 if (invariant_head == 0) {
b3ab6785
KW
3575 d = dst;
3576 } else {
3577 Copy(s, dst, invariant_head, char);
3578 d = dst + invariant_head;
3579 }
3580
3581 while (t < e) {
55d09dc8
KW
3582 append_utf8_from_native_byte(*t, &d);
3583 t++;
b3ab6785
KW
3584 }
3585 *d = '\0';
3586 SvPV_free(sv); /* No longer using pre-existing string */
3587 SvPV_set(sv, (char*)dst);
3588 SvCUR_set(sv, d - dst);
3589 SvLEN_set(sv, size);
3590 } else {
3591
3592 /* Here, have decided to get the exact size of the string.
3593 * Currently this happens only when we know that there is
3594 * guaranteed enough space to fit the converted string, so
3595 * don't have to worry about growing. If two_byte_count is 0,
3596 * then t points to the first byte of the string which hasn't
3597 * been examined yet. Otherwise two_byte_count is 1, and t
3598 * points to the first byte in the string that will expand to
3599 * two. Depending on this, start examining at t or 1 after t.
3600 * */
3601
3602 U8 *d = t + two_byte_count;
3603
3604
3605 /* Count up the remaining bytes that expand to two */
3606
3607 while (d < e) {
3608 const U8 chr = *d++;
6f2d5cbc 3609 if (! NATIVE_BYTE_IS_