This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Cannot do much if putc fails in debug output.
[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
9f53080a 51#ifndef SV_COW_THRESHOLD
e8c6a474 52# define SV_COW_THRESHOLD 0 /* COW iff len > K */
9f53080a
FC
53#endif
54#ifndef SV_COWBUF_THRESHOLD
e8c6a474 55# define SV_COWBUF_THRESHOLD 1250 /* COW iff len > K */
9f53080a
FC
56#endif
57#ifndef SV_COW_MAX_WASTE_THRESHOLD
e8c6a474 58# define SV_COW_MAX_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
9f53080a
FC
59#endif
60#ifndef SV_COWBUF_WASTE_THRESHOLD
e8c6a474 61# define SV_COWBUF_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
9f53080a
FC
62#endif
63#ifndef SV_COW_MAX_WASTE_FACTOR_THRESHOLD
e8c6a474 64# define SV_COW_MAX_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
9f53080a
FC
65#endif
66#ifndef SV_COWBUF_WASTE_FACTOR_THRESHOLD
e8c6a474 67# define SV_COWBUF_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
e8c6a474
YO
68#endif
69/* Work around compiler warnings about unsigned >= THRESHOLD when thres-
70 hold is 0. */
71#if SV_COW_THRESHOLD
72# define GE_COW_THRESHOLD(cur) ((cur) >= SV_COW_THRESHOLD)
73#else
74# define GE_COW_THRESHOLD(cur) 1
75#endif
76#if SV_COWBUF_THRESHOLD
77# define GE_COWBUF_THRESHOLD(cur) ((cur) >= SV_COWBUF_THRESHOLD)
78#else
79# define GE_COWBUF_THRESHOLD(cur) 1
80#endif
81#if SV_COW_MAX_WASTE_THRESHOLD
82# define GE_COW_MAX_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COW_MAX_WASTE_THRESHOLD)
83#else
84# define GE_COW_MAX_WASTE_THRESHOLD(cur,len) 1
85#endif
86#if SV_COWBUF_WASTE_THRESHOLD
87# define GE_COWBUF_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COWBUF_WASTE_THRESHOLD)
88#else
89# define GE_COWBUF_WASTE_THRESHOLD(cur,len) 1
90#endif
91#if SV_COW_MAX_WASTE_FACTOR_THRESHOLD
92# define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COW_MAX_WASTE_FACTOR_THRESHOLD * (cur))
93#else
94# define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) 1
95#endif
96#if SV_COWBUF_WASTE_FACTOR_THRESHOLD
97# define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COWBUF_WASTE_FACTOR_THRESHOLD * (cur))
98#else
99# define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) 1
100#endif
101
102#define CHECK_COW_THRESHOLD(cur,len) (\
103 GE_COW_THRESHOLD((cur)) && \
104 GE_COW_MAX_WASTE_THRESHOLD((cur),(len)) && \
105 GE_COW_MAX_WASTE_FACTOR_THRESHOLD((cur),(len)) \
106)
107#define CHECK_COWBUF_THRESHOLD(cur,len) (\
108 GE_COWBUF_THRESHOLD((cur)) && \
109 GE_COWBUF_WASTE_THRESHOLD((cur),(len)) && \
110 GE_COWBUF_WASTE_FACTOR_THRESHOLD((cur),(len)) \
111)
cca0492e 112
e23c8137 113#ifdef PERL_UTF8_CACHE_ASSERT
ab455f60 114/* if adding more checks watch out for the following tests:
e23c8137
JH
115 * t/op/index.t t/op/length.t t/op/pat.t t/op/substr.t
116 * lib/utf8.t lib/Unicode/Collate/t/index.t
117 * --jhi
118 */
6f207bd3 119# define ASSERT_UTF8_CACHE(cache) \
ab455f60
NC
120 STMT_START { if (cache) { assert((cache)[0] <= (cache)[1]); \
121 assert((cache)[2] <= (cache)[3]); \
122 assert((cache)[3] <= (cache)[1]);} \
123 } STMT_END
e23c8137 124#else
6f207bd3 125# define ASSERT_UTF8_CACHE(cache) NOOP
e23c8137
JH
126#endif
127
f8c7b90f 128#ifdef PERL_OLD_COPY_ON_WRITE
765f542d 129#define SV_COW_NEXT_SV(sv) INT2PTR(SV *,SvUVX(sv))
607fa7f2 130#define SV_COW_NEXT_SV_SET(current,next) SvUV_set(current, PTR2UV(next))
765f542d 131#endif
645c22ef
DM
132
133/* ============================================================================
134
135=head1 Allocation and deallocation of SVs.
d2a0f284
JC
136An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
137sv, av, hv...) contains type and reference count information, and for
138many types, a pointer to the body (struct xrv, xpv, xpviv...), which
139contains fields specific to each type. Some types store all they need
140in the head, so don't have a body.
141
486ec47a 142In all but the most memory-paranoid configurations (ex: PURIFY), heads
d2a0f284
JC
143and bodies are allocated out of arenas, which by default are
144approximately 4K chunks of memory parcelled up into N heads or bodies.
93e68bfb
JC
145Sv-bodies are allocated by their sv-type, guaranteeing size
146consistency needed to allocate safely from arrays.
147
d2a0f284
JC
148For SV-heads, the first slot in each arena is reserved, and holds a
149link to the next arena, some flags, and a note of the number of slots.
150Snaked through each arena chain is a linked list of free items; when
151this becomes empty, an extra arena is allocated and divided up into N
152items which are threaded into the free list.
153
154SV-bodies are similar, but they use arena-sets by default, which
155separate the link and info from the arena itself, and reclaim the 1st
156slot in the arena. SV-bodies are further described later.
645c22ef
DM
157
158The following global variables are associated with arenas:
159
7fefc6c1
KW
160 PL_sv_arenaroot pointer to list of SV arenas
161 PL_sv_root pointer to list of free SV structures
645c22ef 162
7fefc6c1
KW
163 PL_body_arenas head of linked-list of body arenas
164 PL_body_roots[] array of pointers to list of free bodies of svtype
165 arrays are indexed by the svtype needed
93e68bfb 166
d2a0f284
JC
167A few special SV heads are not allocated from an arena, but are
168instead directly created in the interpreter structure, eg PL_sv_undef.
93e68bfb
JC
169The size of arenas can be changed from the default by setting
170PERL_ARENA_SIZE appropriately at compile time.
645c22ef
DM
171
172The SV arena serves the secondary purpose of allowing still-live SVs
173to be located and destroyed during final cleanup.
174
175At the lowest level, the macros new_SV() and del_SV() grab and free
176an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
177to return the SV to the free list with error checking.) new_SV() calls
178more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
179SVs in the free list have their SvTYPE field set to all ones.
180
ff276b08 181At the time of very final cleanup, sv_free_arenas() is called from
645c22ef 182perl_destruct() to physically free all the arenas allocated since the
6a93a7e5 183start of the interpreter.
645c22ef 184
645c22ef
DM
185The function visit() scans the SV arenas list, and calls a specified
186function for each SV it finds which is still live - ie which has an SvTYPE
187other than all 1's, and a non-zero SvREFCNT. visit() is used by the
188following functions (specified as [function that calls visit()] / [function
189called by visit() for each SV]):
190
191 sv_report_used() / do_report_used()
f2524eef 192 dump all remaining SVs (debugging aid)
645c22ef 193
e4487e9b 194 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
e76981f9 195 do_clean_named_io_objs(),do_curse()
645c22ef 196 Attempt to free all objects pointed to by RVs,
e76981f9
FC
197 try to do the same for all objects indir-
198 ectly referenced by typeglobs too, and
199 then do a final sweep, cursing any
200 objects that remain. Called once from
645c22ef
DM
201 perl_destruct(), prior to calling sv_clean_all()
202 below.
203
204 sv_clean_all() / do_clean_all()
205 SvREFCNT_dec(sv) each remaining SV, possibly
206 triggering an sv_free(). It also sets the
207 SVf_BREAK flag on the SV to indicate that the
208 refcnt has been artificially lowered, and thus
209 stopping sv_free() from giving spurious warnings
210 about SVs which unexpectedly have a refcnt
211 of zero. called repeatedly from perl_destruct()
212 until there are no SVs left.
213
93e68bfb 214=head2 Arena allocator API Summary
645c22ef
DM
215
216Private API to rest of sv.c
217
218 new_SV(), del_SV(),
219
df0f0429 220 new_XPVNV(), del_XPVGV(),
645c22ef
DM
221 etc
222
223Public API:
224
8cf8f3d1 225 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
645c22ef 226
645c22ef
DM
227=cut
228
3e8320cc 229 * ========================================================================= */
645c22ef 230
4561caa4
CS
231/*
232 * "A time to plant, and a time to uproot what was planted..."
233 */
234
d7a2c63c
MHM
235#ifdef PERL_MEM_LOG
236# define MEM_LOG_NEW_SV(sv, file, line, func) \
237 Perl_mem_log_new_sv(sv, file, line, func)
238# define MEM_LOG_DEL_SV(sv, file, line, func) \
239 Perl_mem_log_del_sv(sv, file, line, func)
240#else
241# define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
242# define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
243#endif
244
fd0854ff 245#ifdef DEBUG_LEAKING_SCALARS
484e6108
FC
246# define FREE_SV_DEBUG_FILE(sv) STMT_START { \
247 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
248 } STMT_END
d7a2c63c
MHM
249# define DEBUG_SV_SERIAL(sv) \
250 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
251 PTR2UV(sv), (long)(sv)->sv_debug_serial))
fd0854ff
DM
252#else
253# define FREE_SV_DEBUG_FILE(sv)
d7a2c63c 254# define DEBUG_SV_SERIAL(sv) NOOP
fd0854ff
DM
255#endif
256
48614a46
NC
257#ifdef PERL_POISON
258# define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
daba3364 259# define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
48614a46
NC
260/* Whilst I'd love to do this, it seems that things like to check on
261 unreferenced scalars
ce5dbf61 262# define POISON_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
48614a46 263*/
ce5dbf61 264# define POISON_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
7e337ee0 265 PoisonNew(&SvREFCNT(sv), 1, U32)
48614a46
NC
266#else
267# define SvARENA_CHAIN(sv) SvANY(sv)
3eef1deb 268# define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
ce5dbf61 269# define POISON_SV_HEAD(sv)
48614a46
NC
270#endif
271
990198f0
DM
272/* Mark an SV head as unused, and add to free list.
273 *
274 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
275 * its refcount artificially decremented during global destruction, so
276 * there may be dangling pointers to it. The last thing we want in that
277 * case is for it to be reused. */
278
053fc874
GS
279#define plant_SV(p) \
280 STMT_START { \
990198f0 281 const U32 old_flags = SvFLAGS(p); \
d7a2c63c
MHM
282 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
283 DEBUG_SV_SERIAL(p); \
fd0854ff 284 FREE_SV_DEBUG_FILE(p); \
ce5dbf61 285 POISON_SV_HEAD(p); \
053fc874 286 SvFLAGS(p) = SVTYPEMASK; \
990198f0 287 if (!(old_flags & SVf_BREAK)) { \
3eef1deb 288 SvARENA_CHAIN_SET(p, PL_sv_root); \
990198f0
DM
289 PL_sv_root = (p); \
290 } \
053fc874
GS
291 --PL_sv_count; \
292 } STMT_END
a0d0e21e 293
053fc874
GS
294#define uproot_SV(p) \
295 STMT_START { \
296 (p) = PL_sv_root; \
daba3364 297 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
053fc874
GS
298 ++PL_sv_count; \
299 } STMT_END
300
645c22ef 301
cac9b346
NC
302/* make some more SVs by adding another arena */
303
cac9b346
NC
304STATIC SV*
305S_more_sv(pTHX)
306{
307 SV* sv;
9a87bd09
NC
308 char *chunk; /* must use New here to match call to */
309 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
310 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
cac9b346
NC
311 uproot_SV(sv);
312 return sv;
313}
314
645c22ef
DM
315/* new_SV(): return a new, empty SV head */
316
eba0f806
DM
317#ifdef DEBUG_LEAKING_SCALARS
318/* provide a real function for a debugger to play with */
319STATIC SV*
d7a2c63c 320S_new_SV(pTHX_ const char *file, int line, const char *func)
eba0f806
DM
321{
322 SV* sv;
323
eba0f806
DM
324 if (PL_sv_root)
325 uproot_SV(sv);
326 else
cac9b346 327 sv = S_more_sv(aTHX);
eba0f806
DM
328 SvANY(sv) = 0;
329 SvREFCNT(sv) = 1;
330 SvFLAGS(sv) = 0;
fd0854ff 331 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
e385c3bf
DM
332 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
333 ? PL_parser->copline
334 : PL_curcop
f24aceb1
DM
335 ? CopLINE(PL_curcop)
336 : 0
e385c3bf 337 );
fd0854ff 338 sv->sv_debug_inpad = 0;
cd676548 339 sv->sv_debug_parent = NULL;
484e6108 340 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
d7a2c63c
MHM
341
342 sv->sv_debug_serial = PL_sv_serial++;
343
344 MEM_LOG_NEW_SV(sv, file, line, func);
345 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
346 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
347
eba0f806
DM
348 return sv;
349}
d7a2c63c 350# define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
eba0f806
DM
351
352#else
353# define new_SV(p) \
053fc874 354 STMT_START { \
053fc874
GS
355 if (PL_sv_root) \
356 uproot_SV(p); \
357 else \
cac9b346 358 (p) = S_more_sv(aTHX); \
053fc874
GS
359 SvANY(p) = 0; \
360 SvREFCNT(p) = 1; \
361 SvFLAGS(p) = 0; \
d7a2c63c 362 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
053fc874 363 } STMT_END
eba0f806 364#endif
463ee0b2 365
645c22ef
DM
366
367/* del_SV(): return an empty SV head to the free list */
368
a0d0e21e 369#ifdef DEBUGGING
4561caa4 370
053fc874
GS
371#define del_SV(p) \
372 STMT_START { \
aea4f609 373 if (DEBUG_D_TEST) \
053fc874
GS
374 del_sv(p); \
375 else \
376 plant_SV(p); \
053fc874 377 } STMT_END
a0d0e21e 378
76e3520e 379STATIC void
cea2e8a9 380S_del_sv(pTHX_ SV *p)
463ee0b2 381{
7918f24d
NC
382 PERL_ARGS_ASSERT_DEL_SV;
383
aea4f609 384 if (DEBUG_D_TEST) {
4633a7c4 385 SV* sva;
a3b680e6 386 bool ok = 0;
daba3364 387 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
53c1dcc0
AL
388 const SV * const sv = sva + 1;
389 const SV * const svend = &sva[SvREFCNT(sva)];
c0ff570e 390 if (p >= sv && p < svend) {
a0d0e21e 391 ok = 1;
c0ff570e
NC
392 break;
393 }
a0d0e21e
LW
394 }
395 if (!ok) {
9b387841
NC
396 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
397 "Attempt to free non-arena SV: 0x%"UVxf
398 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
a0d0e21e
LW
399 return;
400 }
401 }
4561caa4 402 plant_SV(p);
463ee0b2 403}
a0d0e21e 404
4561caa4
CS
405#else /* ! DEBUGGING */
406
407#define del_SV(p) plant_SV(p)
408
409#endif /* DEBUGGING */
463ee0b2 410
dc6369ef
EH
411/*
412 * Bodyless IVs and NVs!
413 *
414 * Since 5.9.2, we can avoid allocating a body for SVt_IV-type SVs.
415 * Since the larger IV-holding variants of SVs store their integer
416 * values in their respective bodies, the family of SvIV() accessor
417 * macros would naively have to branch on the SV type to find the
418 * integer value either in the HEAD or BODY. In order to avoid this
419 * expensive branch, a clever soul has deployed a great hack:
420 * We set up the SvANY pointer such that instead of pointing to a
421 * real body, it points into the memory before the location of the
422 * head. We compute this pointer such that the location of
423 * the integer member of the hypothetical body struct happens to
424 * be the same as the location of the integer member of the bodyless
425 * SV head. This now means that the SvIV() family of accessors can
426 * always read from the (hypothetical or real) body via SvANY.
427 *
428 * Since the 5.21 dev series, we employ the same trick for NVs
429 * if the architecture can support it (NVSIZE <= IVSIZE).
430 */
431
432/* The following two macros compute the necessary offsets for the above
433 * trick and store them in SvANY for SvIV() (and friends) to use. */
434#define SET_SVANY_FOR_BODYLESS_IV(sv) \
435 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv))
436
437#define SET_SVANY_FOR_BODYLESS_NV(sv) \
438 SvANY(sv) = (XPVNV*)((char*)&(sv->sv_u.svu_nv) - STRUCT_OFFSET(XPVNV, xnv_u.xnv_nv))
645c22ef
DM
439
440/*
ccfc67b7
JH
441=head1 SV Manipulation Functions
442
645c22ef
DM
443=for apidoc sv_add_arena
444
445Given a chunk of memory, link it to the head of the list of arenas,
446and split it into a list of free SVs.
447
448=cut
449*/
450
d2bd4e7f
NC
451static void
452S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
463ee0b2 453{
daba3364 454 SV *const sva = MUTABLE_SV(ptr);
eb578fdb
KW
455 SV* sv;
456 SV* svend;
4633a7c4 457
7918f24d
NC
458 PERL_ARGS_ASSERT_SV_ADD_ARENA;
459
4633a7c4 460 /* The first SV in an arena isn't an SV. */
3280af22 461 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
4633a7c4
LW
462 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
463 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
464
3280af22
NIS
465 PL_sv_arenaroot = sva;
466 PL_sv_root = sva + 1;
4633a7c4
LW
467
468 svend = &sva[SvREFCNT(sva) - 1];
469 sv = sva + 1;
463ee0b2 470 while (sv < svend) {
3eef1deb 471 SvARENA_CHAIN_SET(sv, (sv + 1));
03e36789 472#ifdef DEBUGGING
978b032e 473 SvREFCNT(sv) = 0;
03e36789 474#endif
4b69cbe3 475 /* Must always set typemask because it's always checked in on cleanup
03e36789 476 when the arenas are walked looking for objects. */
8990e307 477 SvFLAGS(sv) = SVTYPEMASK;
463ee0b2
LW
478 sv++;
479 }
3eef1deb 480 SvARENA_CHAIN_SET(sv, 0);
03e36789
NC
481#ifdef DEBUGGING
482 SvREFCNT(sv) = 0;
483#endif
4633a7c4
LW
484 SvFLAGS(sv) = SVTYPEMASK;
485}
486
055972dc
DM
487/* visit(): call the named function for each non-free SV in the arenas
488 * whose flags field matches the flags/mask args. */
645c22ef 489
5226ed68 490STATIC I32
de37a194 491S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
8990e307 492{
4633a7c4 493 SV* sva;
5226ed68 494 I32 visited = 0;
8990e307 495
7918f24d
NC
496 PERL_ARGS_ASSERT_VISIT;
497
daba3364 498 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
eb578fdb
KW
499 const SV * const svend = &sva[SvREFCNT(sva)];
500 SV* sv;
4561caa4 501 for (sv = sva + 1; sv < svend; ++sv) {
e4787c0c 502 if (SvTYPE(sv) != (svtype)SVTYPEMASK
055972dc
DM
503 && (sv->sv_flags & mask) == flags
504 && SvREFCNT(sv))
505 {
942481a7 506 (*f)(aTHX_ sv);
5226ed68
JH
507 ++visited;
508 }
8990e307
LW
509 }
510 }
5226ed68 511 return visited;
8990e307
LW
512}
513
758a08c3
JH
514#ifdef DEBUGGING
515
645c22ef
DM
516/* called by sv_report_used() for each live SV */
517
518static void
5fa45a31 519do_report_used(pTHX_ SV *const sv)
645c22ef 520{
e4787c0c 521 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
645c22ef
DM
522 PerlIO_printf(Perl_debug_log, "****\n");
523 sv_dump(sv);
524 }
525}
758a08c3 526#endif
645c22ef
DM
527
528/*
529=for apidoc sv_report_used
530
fde67290 531Dump the contents of all SVs not yet freed (debugging aid).
645c22ef
DM
532
533=cut
534*/
535
8990e307 536void
864dbfa3 537Perl_sv_report_used(pTHX)
4561caa4 538{
ff270d3a 539#ifdef DEBUGGING
055972dc 540 visit(do_report_used, 0, 0);
96a5add6
AL
541#else
542 PERL_UNUSED_CONTEXT;
ff270d3a 543#endif
4561caa4
CS
544}
545
645c22ef
DM
546/* called by sv_clean_objs() for each live SV */
547
548static void
de37a194 549do_clean_objs(pTHX_ SV *const ref)
645c22ef 550{
ea724faa
NC
551 assert (SvROK(ref));
552 {
823a54a3
AL
553 SV * const target = SvRV(ref);
554 if (SvOBJECT(target)) {
555 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
556 if (SvWEAKREF(ref)) {
557 sv_del_backref(target, ref);
558 SvWEAKREF_off(ref);
559 SvRV_set(ref, NULL);
560 } else {
561 SvROK_off(ref);
562 SvRV_set(ref, NULL);
fc2b2dca 563 SvREFCNT_dec_NN(target);
823a54a3 564 }
645c22ef
DM
565 }
566 }
645c22ef
DM
567}
568
645c22ef 569
e4487e9b
DM
570/* clear any slots in a GV which hold objects - except IO;
571 * called by sv_clean_objs() for each live GV */
572
645c22ef 573static void
f30de749 574do_clean_named_objs(pTHX_ SV *const sv)
645c22ef 575{
57ef47cc 576 SV *obj;
ea724faa 577 assert(SvTYPE(sv) == SVt_PVGV);
d011219a 578 assert(isGV_with_GP(sv));
57ef47cc
DM
579 if (!GvGP(sv))
580 return;
581
582 /* freeing GP entries may indirectly free the current GV;
583 * hold onto it while we mess with the GP slots */
584 SvREFCNT_inc(sv);
585
586 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
587 DEBUG_D((PerlIO_printf(Perl_debug_log,
588 "Cleaning named glob SV object:\n "), sv_dump(obj)));
589 GvSV(sv) = NULL;
fc2b2dca 590 SvREFCNT_dec_NN(obj);
57ef47cc
DM
591 }
592 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
593 DEBUG_D((PerlIO_printf(Perl_debug_log,
594 "Cleaning named glob AV object:\n "), sv_dump(obj)));
595 GvAV(sv) = NULL;
fc2b2dca 596 SvREFCNT_dec_NN(obj);
57ef47cc
DM
597 }
598 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
599 DEBUG_D((PerlIO_printf(Perl_debug_log,
600 "Cleaning named glob HV object:\n "), sv_dump(obj)));
601 GvHV(sv) = NULL;
fc2b2dca 602 SvREFCNT_dec_NN(obj);
57ef47cc
DM
603 }
604 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
605 DEBUG_D((PerlIO_printf(Perl_debug_log,
606 "Cleaning named glob CV object:\n "), sv_dump(obj)));
c43ae56f 607 GvCV_set(sv, NULL);
fc2b2dca 608 SvREFCNT_dec_NN(obj);
57ef47cc 609 }
fc2b2dca 610 SvREFCNT_dec_NN(sv); /* undo the inc above */
e4487e9b
DM
611}
612
68b590d9 613/* clear any IO slots in a GV which hold objects (except stderr, defout);
e4487e9b
DM
614 * called by sv_clean_objs() for each live GV */
615
616static void
617do_clean_named_io_objs(pTHX_ SV *const sv)
618{
e4487e9b
DM
619 SV *obj;
620 assert(SvTYPE(sv) == SVt_PVGV);
621 assert(isGV_with_GP(sv));
68b590d9 622 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
e4487e9b
DM
623 return;
624
625 SvREFCNT_inc(sv);
57ef47cc
DM
626 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
627 DEBUG_D((PerlIO_printf(Perl_debug_log,
628 "Cleaning named glob IO object:\n "), sv_dump(obj)));
629 GvIOp(sv) = NULL;
fc2b2dca 630 SvREFCNT_dec_NN(obj);
645c22ef 631 }
fc2b2dca 632 SvREFCNT_dec_NN(sv); /* undo the inc above */
645c22ef 633}
645c22ef 634
4155e4fe
FC
635/* Void wrapper to pass to visit() */
636static void
637do_curse(pTHX_ SV * const sv) {
c2910e6c
FC
638 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
639 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
4155e4fe
FC
640 return;
641 (void)curse(sv, 0);
642}
643
645c22ef
DM
644/*
645=for apidoc sv_clean_objs
646
fde67290 647Attempt to destroy all objects not yet freed.
645c22ef
DM
648
649=cut
650*/
651
4561caa4 652void
864dbfa3 653Perl_sv_clean_objs(pTHX)
4561caa4 654{
68b590d9 655 GV *olddef, *olderr;
3280af22 656 PL_in_clean_objs = TRUE;
055972dc 657 visit(do_clean_objs, SVf_ROK, SVf_ROK);
e4487e9b
DM
658 /* Some barnacles may yet remain, clinging to typeglobs.
659 * Run the non-IO destructors first: they may want to output
660 * error messages, close files etc */
d011219a 661 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
e4487e9b 662 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
4155e4fe
FC
663 /* And if there are some very tenacious barnacles clinging to arrays,
664 closures, or what have you.... */
665 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
68b590d9
DM
666 olddef = PL_defoutgv;
667 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
668 if (olddef && isGV_with_GP(olddef))
669 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
670 olderr = PL_stderrgv;
671 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
672 if (olderr && isGV_with_GP(olderr))
673 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
674 SvREFCNT_dec(olddef);
3280af22 675 PL_in_clean_objs = FALSE;
4561caa4
CS
676}
677
645c22ef
DM
678/* called by sv_clean_all() for each live SV */
679
680static void
de37a194 681do_clean_all(pTHX_ SV *const sv)
645c22ef 682{
daba3364 683 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
cddfcddc 684 /* don't clean pid table and strtab */
d17ea597 685 return;
cddfcddc 686 }
645c22ef
DM
687 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
688 SvFLAGS(sv) |= SVf_BREAK;
fc2b2dca 689 SvREFCNT_dec_NN(sv);
645c22ef
DM
690}
691
692/*
693=for apidoc sv_clean_all
694
695Decrement the refcnt of each remaining SV, possibly triggering a
fde67290 696cleanup. This function may have to be called multiple times to free
ff276b08 697SVs which are in complex self-referential hierarchies.
645c22ef
DM
698
699=cut
700*/
701
5226ed68 702I32
864dbfa3 703Perl_sv_clean_all(pTHX)
8990e307 704{
5226ed68 705 I32 cleaned;
3280af22 706 PL_in_clean_all = TRUE;
055972dc 707 cleaned = visit(do_clean_all, 0,0);
5226ed68 708 return cleaned;
8990e307 709}
463ee0b2 710
5e258f8c
JC
711/*
712 ARENASETS: a meta-arena implementation which separates arena-info
713 into struct arena_set, which contains an array of struct
714 arena_descs, each holding info for a single arena. By separating
715 the meta-info from the arena, we recover the 1st slot, formerly
716 borrowed for list management. The arena_set is about the size of an
39244528 717 arena, avoiding the needless malloc overhead of a naive linked-list.
5e258f8c
JC
718
719 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
720 memory in the last arena-set (1/2 on average). In trade, we get
721 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
d2a0f284 722 smaller types). The recovery of the wasted space allows use of
e15dad31
JC
723 small arenas for large, rare body types, by changing array* fields
724 in body_details_by_type[] below.
5e258f8c 725*/
5e258f8c 726struct arena_desc {
398c677b
NC
727 char *arena; /* the raw storage, allocated aligned */
728 size_t size; /* its size ~4k typ */
e5973ed5 729 svtype utype; /* bodytype stored in arena */
5e258f8c
JC
730};
731
e6148039
NC
732struct arena_set;
733
734/* Get the maximum number of elements in set[] such that struct arena_set
e15dad31 735 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
e6148039
NC
736 therefore likely to be 1 aligned memory page. */
737
738#define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
739 - 2 * sizeof(int)) / sizeof (struct arena_desc))
5e258f8c
JC
740
741struct arena_set {
742 struct arena_set* next;
0a848332
NC
743 unsigned int set_size; /* ie ARENAS_PER_SET */
744 unsigned int curr; /* index of next available arena-desc */
5e258f8c
JC
745 struct arena_desc set[ARENAS_PER_SET];
746};
747
645c22ef
DM
748/*
749=for apidoc sv_free_arenas
750
fde67290 751Deallocate the memory used by all arenas. Note that all the individual SV
645c22ef
DM
752heads and bodies within the arenas must already have been freed.
753
754=cut
7fefc6c1 755
645c22ef 756*/
4633a7c4 757void
864dbfa3 758Perl_sv_free_arenas(pTHX)
4633a7c4
LW
759{
760 SV* sva;
761 SV* svanext;
0a848332 762 unsigned int i;
4633a7c4
LW
763
764 /* Free arenas here, but be careful about fake ones. (We assume
765 contiguity of the fake ones with the corresponding real ones.) */
766
3280af22 767 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
daba3364 768 svanext = MUTABLE_SV(SvANY(sva));
4633a7c4 769 while (svanext && SvFAKE(svanext))
daba3364 770 svanext = MUTABLE_SV(SvANY(svanext));
4633a7c4
LW
771
772 if (!SvFAKE(sva))
1df70142 773 Safefree(sva);
4633a7c4 774 }
93e68bfb 775
5e258f8c 776 {
0a848332
NC
777 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
778
779 while (aroot) {
780 struct arena_set *current = aroot;
781 i = aroot->curr;
782 while (i--) {
5e258f8c
JC
783 assert(aroot->set[i].arena);
784 Safefree(aroot->set[i].arena);
785 }
0a848332
NC
786 aroot = aroot->next;
787 Safefree(current);
5e258f8c
JC
788 }
789 }
dc8220bf 790 PL_body_arenas = 0;
fdda85ca 791
0a848332
NC
792 i = PERL_ARENA_ROOTS_SIZE;
793 while (i--)
93e68bfb 794 PL_body_roots[i] = 0;
93e68bfb 795
3280af22
NIS
796 PL_sv_arenaroot = 0;
797 PL_sv_root = 0;
4633a7c4
LW
798}
799
bd81e77b
NC
800/*
801 Here are mid-level routines that manage the allocation of bodies out
802 of the various arenas. There are 5 kinds of arenas:
29489e7c 803
bd81e77b
NC
804 1. SV-head arenas, which are discussed and handled above
805 2. regular body arenas
806 3. arenas for reduced-size bodies
807 4. Hash-Entry arenas
29489e7c 808
bd81e77b
NC
809 Arena types 2 & 3 are chained by body-type off an array of
810 arena-root pointers, which is indexed by svtype. Some of the
811 larger/less used body types are malloced singly, since a large
812 unused block of them is wasteful. Also, several svtypes dont have
813 bodies; the data fits into the sv-head itself. The arena-root
814 pointer thus has a few unused root-pointers (which may be hijacked
815 later for arena types 4,5)
29489e7c 816
bd81e77b
NC
817 3 differs from 2 as an optimization; some body types have several
818 unused fields in the front of the structure (which are kept in-place
819 for consistency). These bodies can be allocated in smaller chunks,
820 because the leading fields arent accessed. Pointers to such bodies
821 are decremented to point at the unused 'ghost' memory, knowing that
822 the pointers are used with offsets to the real memory.
29489e7c 823
d2a0f284
JC
824
825=head1 SV-Body Allocation
826
f554dfc5
MH
827=cut
828
d2a0f284
JC
829Allocation of SV-bodies is similar to SV-heads, differing as follows;
830the allocation mechanism is used for many body types, so is somewhat
831more complicated, it uses arena-sets, and has no need for still-live
832SV detection.
833
834At the outermost level, (new|del)_X*V macros return bodies of the
835appropriate type. These macros call either (new|del)_body_type or
836(new|del)_body_allocated macro pairs, depending on specifics of the
837type. Most body types use the former pair, the latter pair is used to
838allocate body types with "ghost fields".
839
840"ghost fields" are fields that are unused in certain types, and
69ba284b 841consequently don't need to actually exist. They are declared because
d2a0f284
JC
842they're part of a "base type", which allows use of functions as
843methods. The simplest examples are AVs and HVs, 2 aggregate types
844which don't use the fields which support SCALAR semantics.
845
69ba284b 846For these types, the arenas are carved up into appropriately sized
d2a0f284
JC
847chunks, we thus avoid wasted memory for those unaccessed members.
848When bodies are allocated, we adjust the pointer back in memory by the
69ba284b 849size of the part not allocated, so it's as if we allocated the full
d2a0f284
JC
850structure. (But things will all go boom if you write to the part that
851is "not there", because you'll be overwriting the last members of the
852preceding structure in memory.)
853
69ba284b 854We calculate the correction using the STRUCT_OFFSET macro on the first
a05ea1cf 855member present. If the allocated structure is smaller (no initial NV
69ba284b
NC
856actually allocated) then the net effect is to subtract the size of the NV
857from the pointer, to return a new pointer as if an initial NV were actually
a05ea1cf 858allocated. (We were using structures named *_allocated for this, but
69ba284b
NC
859this turned out to be a subtle bug, because a structure without an NV
860could have a lower alignment constraint, but the compiler is allowed to
861optimised accesses based on the alignment constraint of the actual pointer
862to the full structure, for example, using a single 64 bit load instruction
863because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
d2a0f284 864
a05ea1cf 865This is the same trick as was used for NV and IV bodies. Ironically it
d2a0f284 866doesn't need to be used for NV bodies any more, because NV is now at
5b306eef
DD
867the start of the structure. IV bodies, and also in some builds NV bodies,
868don't need it either, because they are no longer allocated.
d2a0f284
JC
869
870In turn, the new_body_* allocators call S_new_body(), which invokes
871new_body_inline macro, which takes a lock, and takes a body off the
1e30fcd5 872linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
d2a0f284
JC
873necessary to refresh an empty list. Then the lock is released, and
874the body is returned.
875
99816f8d 876Perl_more_bodies allocates a new arena, and carves it up into an array of N
d2a0f284
JC
877bodies, which it strings into a linked list. It looks up arena-size
878and body-size from the body_details table described below, thus
879supporting the multiple body-types.
880
881If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
882the (new|del)_X*V macros are mapped directly to malloc/free.
883
d2a0f284
JC
884For each sv-type, struct body_details bodies_by_type[] carries
885parameters which control these aspects of SV handling:
886
887Arena_size determines whether arenas are used for this body type, and if
888so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
889zero, forcing individual mallocs and frees.
890
891Body_size determines how big a body is, and therefore how many fit into
892each arena. Offset carries the body-pointer adjustment needed for
69ba284b 893"ghost fields", and is used in *_allocated macros.
d2a0f284
JC
894
895But its main purpose is to parameterize info needed in
896Perl_sv_upgrade(). The info here dramatically simplifies the function
69ba284b 897vs the implementation in 5.8.8, making it table-driven. All fields
d2a0f284
JC
898are used for this, except for arena_size.
899
900For the sv-types that have no bodies, arenas are not used, so those
901PL_body_roots[sv_type] are unused, and can be overloaded. In
902something of a special case, SVt_NULL is borrowed for HE arenas;
c6f8b1d0 903PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
d2a0f284 904bodies_by_type[SVt_NULL] slot is not used, as the table is not
c6f8b1d0 905available in hv.c.
d2a0f284 906
29489e7c
DM
907*/
908
bd81e77b 909struct body_details {
0fb58b32 910 U8 body_size; /* Size to allocate */
10666ae3 911 U8 copy; /* Size of structure to copy (may be shorter) */
5b306eef
DD
912 U8 offset; /* Size of unalloced ghost fields to first alloced field*/
913 PERL_BITFIELD8 type : 4; /* We have space for a sanity check. */
914 PERL_BITFIELD8 cant_upgrade : 1;/* Cannot upgrade this type */
915 PERL_BITFIELD8 zero_nv : 1; /* zero the NV when upgrading from this */
916 PERL_BITFIELD8 arena : 1; /* Allocated from an arena */
917 U32 arena_size; /* Size of arena to allocate */
bd81e77b 918};
29489e7c 919
bd81e77b
NC
920#define HADNV FALSE
921#define NONV TRUE
29489e7c 922
d2a0f284 923
bd81e77b
NC
924#ifdef PURIFY
925/* With -DPURFIY we allocate everything directly, and don't use arenas.
926 This seems a rather elegant way to simplify some of the code below. */
927#define HASARENA FALSE
928#else
929#define HASARENA TRUE
930#endif
931#define NOARENA FALSE
29489e7c 932
d2a0f284
JC
933/* Size the arenas to exactly fit a given number of bodies. A count
934 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
935 simplifying the default. If count > 0, the arena is sized to fit
936 only that many bodies, allowing arenas to be used for large, rare
937 bodies (XPVFM, XPVIO) without undue waste. The arena size is
938 limited by PERL_ARENA_SIZE, so we can safely oversize the
939 declarations.
940 */
95db5f15
MB
941#define FIT_ARENA0(body_size) \
942 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
943#define FIT_ARENAn(count,body_size) \
944 ( count * body_size <= PERL_ARENA_SIZE) \
945 ? count * body_size \
946 : FIT_ARENA0 (body_size)
947#define FIT_ARENA(count,body_size) \
cd1dc8e2 948 (U32)(count \
95db5f15 949 ? FIT_ARENAn (count, body_size) \
cd1dc8e2 950 : FIT_ARENA0 (body_size))
d2a0f284 951
bd81e77b
NC
952/* Calculate the length to copy. Specifically work out the length less any
953 final padding the compiler needed to add. See the comment in sv_upgrade
954 for why copying the padding proved to be a bug. */
29489e7c 955
bd81e77b
NC
956#define copy_length(type, last_member) \
957 STRUCT_OFFSET(type, last_member) \
daba3364 958 + sizeof (((type*)SvANY((const SV *)0))->last_member)
29489e7c 959
bd81e77b 960static const struct body_details bodies_by_type[] = {
829cd18a
NC
961 /* HEs use this offset for their arena. */
962 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
d2a0f284 963
db93c0c4
NC
964 /* IVs are in the head, so the allocation size is 0. */
965 { 0,
d2a0f284 966 sizeof(IV), /* This is used to copy out the IV body. */
10666ae3 967 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
db93c0c4 968 NOARENA /* IVS don't need an arena */, 0
d2a0f284
JC
969 },
970
5b306eef
DD
971#if NVSIZE <= IVSIZE
972 { 0, sizeof(NV),
973 STRUCT_OFFSET(XPVNV, xnv_u),
974 SVt_NV, FALSE, HADNV, NOARENA, 0 },
975#else
6e128786
NC
976 { sizeof(NV), sizeof(NV),
977 STRUCT_OFFSET(XPVNV, xnv_u),
978 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
5b306eef 979#endif
d2a0f284 980
bc337e5c 981 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
982 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
983 + STRUCT_OFFSET(XPV, xpv_cur),
69ba284b 984 SVt_PV, FALSE, NONV, HASARENA,
889d28b2 985 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 986
d361b004
KW
987 { sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur),
988 copy_length(XINVLIST, is_offset) - STRUCT_OFFSET(XPV, xpv_cur),
989 + STRUCT_OFFSET(XPV, xpv_cur),
990 SVt_INVLIST, TRUE, NONV, HASARENA,
991 FIT_ARENA(0, sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur)) },
e94d9b54 992
bc337e5c 993 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
994 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
995 + STRUCT_OFFSET(XPV, xpv_cur),
996 SVt_PVIV, FALSE, NONV, HASARENA,
997 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 998
bc337e5c 999 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
1000 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
1001 + STRUCT_OFFSET(XPV, xpv_cur),
1002 SVt_PVNV, FALSE, HADNV, HASARENA,
1003 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 1004
6e128786 1005 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
d2a0f284 1006 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
4df7f6af 1007
601dfd0a
NC
1008 { sizeof(regexp),
1009 sizeof(regexp),
1010 0,
ecff11eb 1011 SVt_REGEXP, TRUE, NONV, HASARENA,
eaeb1e7f 1012 FIT_ARENA(0, sizeof(regexp))
5c35adbb 1013 },
4df7f6af 1014
10666ae3 1015 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
d2a0f284
JC
1016 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
1017
10666ae3 1018 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
d2a0f284
JC
1019 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
1020
601dfd0a 1021 { sizeof(XPVAV),
4f7003f5 1022 copy_length(XPVAV, xav_alloc),
601dfd0a 1023 0,
69ba284b 1024 SVt_PVAV, TRUE, NONV, HASARENA,
601dfd0a 1025 FIT_ARENA(0, sizeof(XPVAV)) },
d2a0f284 1026
601dfd0a 1027 { sizeof(XPVHV),
359164a0 1028 copy_length(XPVHV, xhv_max),
601dfd0a 1029 0,
69ba284b 1030 SVt_PVHV, TRUE, NONV, HASARENA,
601dfd0a 1031 FIT_ARENA(0, sizeof(XPVHV)) },
d2a0f284 1032
601dfd0a
NC
1033 { sizeof(XPVCV),
1034 sizeof(XPVCV),
1035 0,
69ba284b 1036 SVt_PVCV, TRUE, NONV, HASARENA,
601dfd0a 1037 FIT_ARENA(0, sizeof(XPVCV)) },
69ba284b 1038
601dfd0a
NC
1039 { sizeof(XPVFM),
1040 sizeof(XPVFM),
1041 0,
69ba284b 1042 SVt_PVFM, TRUE, NONV, NOARENA,
601dfd0a 1043 FIT_ARENA(20, sizeof(XPVFM)) },
d2a0f284 1044
601dfd0a
NC
1045 { sizeof(XPVIO),
1046 sizeof(XPVIO),
1047 0,
b6f60916 1048 SVt_PVIO, TRUE, NONV, HASARENA,
601dfd0a 1049 FIT_ARENA(24, sizeof(XPVIO)) },
bd81e77b 1050};
29489e7c 1051
bd81e77b 1052#define new_body_allocated(sv_type) \
d2a0f284 1053 (void *)((char *)S_new_body(aTHX_ sv_type) \
bd81e77b 1054 - bodies_by_type[sv_type].offset)
29489e7c 1055
26359cfa
NC
1056/* return a thing to the free list */
1057
1058#define del_body(thing, root) \
1059 STMT_START { \
1060 void ** const thing_copy = (void **)thing; \
1061 *thing_copy = *root; \
1062 *root = (void*)thing_copy; \
1063 } STMT_END
29489e7c 1064
bd81e77b 1065#ifdef PURIFY
5b306eef
DD
1066#if !(NVSIZE <= IVSIZE)
1067# define new_XNV() safemalloc(sizeof(XPVNV))
1068#endif
beeec492
NC
1069#define new_XPVNV() safemalloc(sizeof(XPVNV))
1070#define new_XPVMG() safemalloc(sizeof(XPVMG))
29489e7c 1071
beeec492 1072#define del_XPVGV(p) safefree(p)
29489e7c 1073
bd81e77b 1074#else /* !PURIFY */
29489e7c 1075
5b306eef
DD
1076#if !(NVSIZE <= IVSIZE)
1077# define new_XNV() new_body_allocated(SVt_NV)
1078#endif
65ac1738 1079#define new_XPVNV() new_body_allocated(SVt_PVNV)
65ac1738 1080#define new_XPVMG() new_body_allocated(SVt_PVMG)
645c22ef 1081
26359cfa
NC
1082#define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1083 &PL_body_roots[SVt_PVGV])
1d7c1841 1084
bd81e77b 1085#endif /* PURIFY */
93e68bfb 1086
bd81e77b 1087/* no arena for you! */
93e68bfb 1088
bd81e77b 1089#define new_NOARENA(details) \
beeec492 1090 safemalloc((details)->body_size + (details)->offset)
bd81e77b 1091#define new_NOARENAZ(details) \
beeec492 1092 safecalloc((details)->body_size + (details)->offset, 1)
d2a0f284 1093
1e30fcd5
NC
1094void *
1095Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1096 const size_t arena_size)
d2a0f284 1097{
d2a0f284 1098 void ** const root = &PL_body_roots[sv_type];
99816f8d
NC
1099 struct arena_desc *adesc;
1100 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1101 unsigned int curr;
d2a0f284
JC
1102 char *start;
1103 const char *end;
02982131 1104 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
20b7effb
JH
1105#if defined(DEBUGGING) && defined(PERL_GLOBAL_STRUCT)
1106 dVAR;
1107#endif
0b2d3faa 1108#if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
23e9d66c
NC
1109 static bool done_sanity_check;
1110
0b2d3faa
JH
1111 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1112 * variables like done_sanity_check. */
10666ae3 1113 if (!done_sanity_check) {
ea471437 1114 unsigned int i = SVt_LAST;
10666ae3
NC
1115
1116 done_sanity_check = TRUE;
1117
1118 while (i--)
1119 assert (bodies_by_type[i].type == i);
1120 }
1121#endif
1122
02982131 1123 assert(arena_size);
23e9d66c 1124
99816f8d
NC
1125 /* may need new arena-set to hold new arena */
1126 if (!aroot || aroot->curr >= aroot->set_size) {
1127 struct arena_set *newroot;
1128 Newxz(newroot, 1, struct arena_set);
1129 newroot->set_size = ARENAS_PER_SET;
1130 newroot->next = aroot;
1131 aroot = newroot;
1132 PL_body_arenas = (void *) newroot;
1133 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1134 }
1135
1136 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1137 curr = aroot->curr++;
1138 adesc = &(aroot->set[curr]);
1139 assert(!adesc->arena);
1140
1141 Newx(adesc->arena, good_arena_size, char);
1142 adesc->size = good_arena_size;
1143 adesc->utype = sv_type;
1144 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %"UVuf"\n",
1145 curr, (void*)adesc->arena, (UV)good_arena_size));
1146
1147 start = (char *) adesc->arena;
d2a0f284 1148
29657bb6
NC
1149 /* Get the address of the byte after the end of the last body we can fit.
1150 Remember, this is integer division: */
02982131 1151 end = start + good_arena_size / body_size * body_size;
d2a0f284 1152
486ec47a 1153 /* computed count doesn't reflect the 1st slot reservation */
d8fca402
NC
1154#if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1155 DEBUG_m(PerlIO_printf(Perl_debug_log,
1156 "arena %p end %p arena-size %d (from %d) type %d "
1157 "size %d ct %d\n",
02982131
NC
1158 (void*)start, (void*)end, (int)good_arena_size,
1159 (int)arena_size, sv_type, (int)body_size,
1160 (int)good_arena_size / (int)body_size));
d8fca402 1161#else
d2a0f284
JC
1162 DEBUG_m(PerlIO_printf(Perl_debug_log,
1163 "arena %p end %p arena-size %d type %d size %d ct %d\n",
6c9570dc 1164 (void*)start, (void*)end,
02982131
NC
1165 (int)arena_size, sv_type, (int)body_size,
1166 (int)good_arena_size / (int)body_size));
d8fca402 1167#endif
d2a0f284
JC
1168 *root = (void *)start;
1169
29657bb6
NC
1170 while (1) {
1171 /* Where the next body would start: */
d2a0f284 1172 char * const next = start + body_size;
29657bb6
NC
1173
1174 if (next >= end) {
1175 /* This is the last body: */
1176 assert(next == end);
1177
1178 *(void **)start = 0;
1179 return *root;
1180 }
1181
d2a0f284
JC
1182 *(void**) start = (void *)next;
1183 start = next;
1184 }
d2a0f284
JC
1185}
1186
1187/* grab a new thing from the free list, allocating more if necessary.
1188 The inline version is used for speed in hot routines, and the
1189 function using it serves the rest (unless PURIFY).
1190*/
1191#define new_body_inline(xpv, sv_type) \
1192 STMT_START { \
1193 void ** const r3wt = &PL_body_roots[sv_type]; \
11b79775 1194 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1e30fcd5 1195 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
02982131
NC
1196 bodies_by_type[sv_type].body_size,\
1197 bodies_by_type[sv_type].arena_size)); \
d2a0f284 1198 *(r3wt) = *(void**)(xpv); \
d2a0f284
JC
1199 } STMT_END
1200
1201#ifndef PURIFY
1202
1203STATIC void *
de37a194 1204S_new_body(pTHX_ const svtype sv_type)
d2a0f284 1205{
d2a0f284
JC
1206 void *xpv;
1207 new_body_inline(xpv, sv_type);
1208 return xpv;
1209}
1210
1211#endif
93e68bfb 1212
238b27b3
NC
1213static const struct body_details fake_rv =
1214 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1215
bd81e77b
NC
1216/*
1217=for apidoc sv_upgrade
93e68bfb 1218
bd81e77b
NC
1219Upgrade an SV to a more complex form. Generally adds a new body type to the
1220SV, then copies across as much information as possible from the old body.
9521ca61
FC
1221It croaks if the SV is already in a more complex form than requested. You
1222generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1223before calling C<sv_upgrade>, and hence does not croak. See also
1224C<svtype>.
93e68bfb 1225
bd81e77b 1226=cut
93e68bfb 1227*/
93e68bfb 1228
bd81e77b 1229void
5aaab254 1230Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
cac9b346 1231{
bd81e77b
NC
1232 void* old_body;
1233 void* new_body;
42d0e0b7 1234 const svtype old_type = SvTYPE(sv);
d2a0f284 1235 const struct body_details *new_type_details;
238b27b3 1236 const struct body_details *old_type_details
bd81e77b 1237 = bodies_by_type + old_type;
4df7f6af 1238 SV *referant = NULL;
cac9b346 1239
7918f24d
NC
1240 PERL_ARGS_ASSERT_SV_UPGRADE;
1241
1776cbe8
NC
1242 if (old_type == new_type)
1243 return;
1244
1245 /* This clause was purposefully added ahead of the early return above to
1246 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1247 inference by Nick I-S that it would fix other troublesome cases. See
1248 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1249
1250 Given that shared hash key scalars are no longer PVIV, but PV, there is
1251 no longer need to unshare so as to free up the IVX slot for its proper
1252 purpose. So it's safe to move the early return earlier. */
1253
093085a8 1254 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
bd81e77b
NC
1255 sv_force_normal_flags(sv, 0);
1256 }
cac9b346 1257
bd81e77b 1258 old_body = SvANY(sv);
de042e1d 1259
bd81e77b
NC
1260 /* Copying structures onto other structures that have been neatly zeroed
1261 has a subtle gotcha. Consider XPVMG
cac9b346 1262
bd81e77b
NC
1263 +------+------+------+------+------+-------+-------+
1264 | NV | CUR | LEN | IV | MAGIC | STASH |
1265 +------+------+------+------+------+-------+-------+
1266 0 4 8 12 16 20 24 28
645c22ef 1267
bd81e77b
NC
1268 where NVs are aligned to 8 bytes, so that sizeof that structure is
1269 actually 32 bytes long, with 4 bytes of padding at the end:
08742458 1270
bd81e77b
NC
1271 +------+------+------+------+------+-------+-------+------+
1272 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1273 +------+------+------+------+------+-------+-------+------+
1274 0 4 8 12 16 20 24 28 32
08742458 1275
bd81e77b 1276 so what happens if you allocate memory for this structure:
30f9da9e 1277
bd81e77b
NC
1278 +------+------+------+------+------+-------+-------+------+------+...
1279 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1280 +------+------+------+------+------+-------+-------+------+------+...
1281 0 4 8 12 16 20 24 28 32 36
bfc44f79 1282
bd81e77b
NC
1283 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1284 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1285 started out as zero once, but it's quite possible that it isn't. So now,
1286 rather than a nicely zeroed GP, you have it pointing somewhere random.
1287 Bugs ensue.
bfc44f79 1288
bd81e77b
NC
1289 (In fact, GP ends up pointing at a previous GP structure, because the
1290 principle cause of the padding in XPVMG getting garbage is a copy of
6c9e42f7
NC
1291 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1292 this happens to be moot because XPVGV has been re-ordered, with GP
1293 no longer after STASH)
30f9da9e 1294
bd81e77b
NC
1295 So we are careful and work out the size of used parts of all the
1296 structures. */
bfc44f79 1297
bd81e77b
NC
1298 switch (old_type) {
1299 case SVt_NULL:
1300 break;
1301 case SVt_IV:
4df7f6af
NC
1302 if (SvROK(sv)) {
1303 referant = SvRV(sv);
238b27b3
NC
1304 old_type_details = &fake_rv;
1305 if (new_type == SVt_NV)
1306 new_type = SVt_PVNV;
4df7f6af
NC
1307 } else {
1308 if (new_type < SVt_PVIV) {
1309 new_type = (new_type == SVt_NV)
1310 ? SVt_PVNV : SVt_PVIV;
1311 }
bd81e77b
NC
1312 }
1313 break;
1314 case SVt_NV:
1315 if (new_type < SVt_PVNV) {
1316 new_type = SVt_PVNV;
bd81e77b
NC
1317 }
1318 break;
bd81e77b
NC
1319 case SVt_PV:
1320 assert(new_type > SVt_PV);
6d59e610
LM
1321 STATIC_ASSERT_STMT(SVt_IV < SVt_PV);
1322 STATIC_ASSERT_STMT(SVt_NV < SVt_PV);
bd81e77b
NC
1323 break;
1324 case SVt_PVIV:
1325 break;
1326 case SVt_PVNV:
1327 break;
1328 case SVt_PVMG:
1329 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1330 there's no way that it can be safely upgraded, because perl.c
1331 expects to Safefree(SvANY(PL_mess_sv)) */
1332 assert(sv != PL_mess_sv);
bd81e77b
NC
1333 break;
1334 default:
2439e033 1335 if (UNLIKELY(old_type_details->cant_upgrade))
c81225bc
NC
1336 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1337 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
bd81e77b 1338 }
3376de98 1339
2439e033 1340 if (UNLIKELY(old_type > new_type))
3376de98
NC
1341 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1342 (int)old_type, (int)new_type);
1343
2fa1109b 1344 new_type_details = bodies_by_type + new_type;
645c22ef 1345
bd81e77b
NC
1346 SvFLAGS(sv) &= ~SVTYPEMASK;
1347 SvFLAGS(sv) |= new_type;
932e9ff9 1348
ab4416c0
NC
1349 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1350 the return statements above will have triggered. */
1351 assert (new_type != SVt_NULL);
bd81e77b 1352 switch (new_type) {
bd81e77b
NC
1353 case SVt_IV:
1354 assert(old_type == SVt_NULL);
dc6369ef 1355 SET_SVANY_FOR_BODYLESS_IV(sv);
bd81e77b
NC
1356 SvIV_set(sv, 0);
1357 return;
1358 case SVt_NV:
1359 assert(old_type == SVt_NULL);
5b306eef 1360#if NVSIZE <= IVSIZE
dc6369ef 1361 SET_SVANY_FOR_BODYLESS_NV(sv);
5b306eef 1362#else
bd81e77b 1363 SvANY(sv) = new_XNV();
5b306eef 1364#endif
bd81e77b
NC
1365 SvNV_set(sv, 0);
1366 return;
bd81e77b 1367 case SVt_PVHV:
bd81e77b 1368 case SVt_PVAV:
d2a0f284 1369 assert(new_type_details->body_size);
c1ae03ae
NC
1370
1371#ifndef PURIFY
1372 assert(new_type_details->arena);
d2a0f284 1373 assert(new_type_details->arena_size);
c1ae03ae 1374 /* This points to the start of the allocated area. */
d2a0f284
JC
1375 new_body_inline(new_body, new_type);
1376 Zero(new_body, new_type_details->body_size, char);
c1ae03ae
NC
1377 new_body = ((char *)new_body) - new_type_details->offset;
1378#else
1379 /* We always allocated the full length item with PURIFY. To do this
1380 we fake things so that arena is false for all 16 types.. */
1381 new_body = new_NOARENAZ(new_type_details);
1382#endif
1383 SvANY(sv) = new_body;
1384 if (new_type == SVt_PVAV) {
1385 AvMAX(sv) = -1;
1386 AvFILLp(sv) = -1;
1387 AvREAL_only(sv);
64484faa 1388 if (old_type_details->body_size) {
ac572bf4
NC
1389 AvALLOC(sv) = 0;
1390 } else {
1391 /* It will have been zeroed when the new body was allocated.
1392 Lets not write to it, in case it confuses a write-back
1393 cache. */
1394 }
78ac7dd9
NC
1395 } else {
1396 assert(!SvOK(sv));
1397 SvOK_off(sv);
1398#ifndef NODEFAULT_SHAREKEYS
1399 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1400#endif
586fc6a3
SM
1401 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1402 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
c1ae03ae 1403 }
aeb18a1e 1404
bd81e77b
NC
1405 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1406 The target created by newSVrv also is, and it can have magic.
1407 However, it never has SvPVX set.
1408 */
4df7f6af
NC
1409 if (old_type == SVt_IV) {
1410 assert(!SvROK(sv));
1411 } else if (old_type >= SVt_PV) {
bd81e77b
NC
1412 assert(SvPVX_const(sv) == 0);
1413 }
aeb18a1e 1414
bd81e77b 1415 if (old_type >= SVt_PVMG) {
e736a858 1416 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
bd81e77b 1417 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
797c7171
NC
1418 } else {
1419 sv->sv_u.svu_array = NULL; /* or svu_hash */
bd81e77b
NC
1420 }
1421 break;
93e68bfb 1422
bd81e77b
NC
1423 case SVt_PVIV:
1424 /* XXX Is this still needed? Was it ever needed? Surely as there is
1425 no route from NV to PVIV, NOK can never be true */
1426 assert(!SvNOKp(sv));
1427 assert(!SvNOK(sv));
2b5060ae 1428 /* FALLTHROUGH */
bd81e77b
NC
1429 case SVt_PVIO:
1430 case SVt_PVFM:
bd81e77b
NC
1431 case SVt_PVGV:
1432 case SVt_PVCV:
1433 case SVt_PVLV:
d361b004 1434 case SVt_INVLIST:
12c45b25 1435 case SVt_REGEXP:
bd81e77b
NC
1436 case SVt_PVMG:
1437 case SVt_PVNV:
1438 case SVt_PV:
93e68bfb 1439
d2a0f284 1440 assert(new_type_details->body_size);
bd81e77b
NC
1441 /* We always allocated the full length item with PURIFY. To do this
1442 we fake things so that arena is false for all 16 types.. */
1443 if(new_type_details->arena) {
1444 /* This points to the start of the allocated area. */
d2a0f284
JC
1445 new_body_inline(new_body, new_type);
1446 Zero(new_body, new_type_details->body_size, char);
bd81e77b
NC
1447 new_body = ((char *)new_body) - new_type_details->offset;
1448 } else {
1449 new_body = new_NOARENAZ(new_type_details);
1450 }
1451 SvANY(sv) = new_body;
5e2fc214 1452
bd81e77b 1453 if (old_type_details->copy) {
f9ba3d20
NC
1454 /* There is now the potential for an upgrade from something without
1455 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1456 int offset = old_type_details->offset;
1457 int length = old_type_details->copy;
1458
1459 if (new_type_details->offset > old_type_details->offset) {
d4c19fe8 1460 const int difference
f9ba3d20
NC
1461 = new_type_details->offset - old_type_details->offset;
1462 offset += difference;
1463 length -= difference;
1464 }
1465 assert (length >= 0);
1466
1467 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1468 char);
bd81e77b
NC
1469 }
1470
1471#ifndef NV_ZERO_IS_ALLBITS_ZERO
f2524eef 1472 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
e5ce394c
NC
1473 * correct 0.0 for us. Otherwise, if the old body didn't have an
1474 * NV slot, but the new one does, then we need to initialise the
1475 * freshly created NV slot with whatever the correct bit pattern is
1476 * for 0.0 */
e22a937e
NC
1477 if (old_type_details->zero_nv && !new_type_details->zero_nv
1478 && !isGV_with_GP(sv))
bd81e77b 1479 SvNV_set(sv, 0);
82048762 1480#endif
5e2fc214 1481
2439e033 1482 if (UNLIKELY(new_type == SVt_PVIO)) {
85dca89a 1483 IO * const io = MUTABLE_IO(sv);
d963bf01 1484 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
85dca89a
NC
1485
1486 SvOBJECT_on(io);
1487 /* Clear the stashcache because a new IO could overrule a package
1488 name */
103f5a36 1489 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
85dca89a
NC
1490 hv_clear(PL_stashcache);
1491
85dca89a 1492 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
f2524eef 1493 IoPAGE_LEN(sv) = 60;
85dca89a 1494 }
2439e033 1495 if (UNLIKELY(new_type == SVt_REGEXP))
8d919b0a
FC
1496 sv->sv_u.svu_rx = (regexp *)new_body;
1497 else if (old_type < SVt_PV) {
4df7f6af
NC
1498 /* referant will be NULL unless the old type was SVt_IV emulating
1499 SVt_RV */
1500 sv->sv_u.svu_rv = referant;
1501 }
bd81e77b
NC
1502 break;
1503 default:
afd78fd5
JH
1504 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1505 (unsigned long)new_type);
bd81e77b 1506 }
73171d91 1507
5b306eef
DD
1508 /* if this is zero, this is a body-less SVt_NULL, SVt_IV/SVt_RV,
1509 and sometimes SVt_NV */
1510 if (old_type_details->body_size) {
bd81e77b 1511#ifdef PURIFY
beeec492 1512 safefree(old_body);
bd81e77b 1513#else
bc786448
GG
1514 /* Note that there is an assumption that all bodies of types that
1515 can be upgraded came from arenas. Only the more complex non-
1516 upgradable types are allowed to be directly malloc()ed. */
1517 assert(old_type_details->arena);
bd81e77b
NC
1518 del_body((void*)((char*)old_body + old_type_details->offset),
1519 &PL_body_roots[old_type]);
1520#endif
1521 }
1522}
73171d91 1523
bd81e77b
NC
1524/*
1525=for apidoc sv_backoff
73171d91 1526
fde67290 1527Remove any string offset. You should normally use the C<SvOOK_off> macro
bd81e77b 1528wrapper instead.
73171d91 1529
bd81e77b 1530=cut
73171d91
NC
1531*/
1532
bd81e77b 1533int
ddeaf645 1534Perl_sv_backoff(SV *const sv)
bd81e77b 1535{
69240efd 1536 STRLEN delta;
7a4bba22 1537 const char * const s = SvPVX_const(sv);
7918f24d
NC
1538
1539 PERL_ARGS_ASSERT_SV_BACKOFF;
7918f24d 1540
bd81e77b
NC
1541 assert(SvOOK(sv));
1542 assert(SvTYPE(sv) != SVt_PVHV);
1543 assert(SvTYPE(sv) != SVt_PVAV);
7a4bba22 1544
69240efd
NC
1545 SvOOK_offset(sv, delta);
1546
7a4bba22
NC
1547 SvLEN_set(sv, SvLEN(sv) + delta);
1548 SvPV_set(sv, SvPVX(sv) - delta);
1549 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
bd81e77b
NC
1550 SvFLAGS(sv) &= ~SVf_OOK;
1551 return 0;
1552}
73171d91 1553
bd81e77b
NC
1554/*
1555=for apidoc sv_grow
73171d91 1556
bd81e77b
NC
1557Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1558upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1559Use the C<SvGROW> wrapper instead.
93e68bfb 1560
bd81e77b
NC
1561=cut
1562*/
93e68bfb 1563
e0060e30
FC
1564static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1565
bd81e77b 1566char *
5aaab254 1567Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
bd81e77b 1568{
eb578fdb 1569 char *s;
93e68bfb 1570
7918f24d
NC
1571 PERL_ARGS_ASSERT_SV_GROW;
1572
bd81e77b
NC
1573 if (SvROK(sv))
1574 sv_unref(sv);
1575 if (SvTYPE(sv) < SVt_PV) {
1576 sv_upgrade(sv, SVt_PV);
1577 s = SvPVX_mutable(sv);
1578 }
1579 else if (SvOOK(sv)) { /* pv is offset? */
1580 sv_backoff(sv);
1581 s = SvPVX_mutable(sv);
1582 if (newlen > SvLEN(sv))
1583 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
bd81e77b
NC
1584 }
1585 else
db2c6cb3 1586 {
e0060e30 1587 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
bd81e77b 1588 s = SvPVX_mutable(sv);
db2c6cb3 1589 }
aeb18a1e 1590
cbcb2a16
DM
1591#ifdef PERL_NEW_COPY_ON_WRITE
1592 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
3c239bea 1593 * to store the COW count. So in general, allocate one more byte than
cbcb2a16
DM
1594 * asked for, to make it likely this byte is always spare: and thus
1595 * make more strings COW-able.
1596 * If the new size is a big power of two, don't bother: we assume the
1597 * caller wanted a nice 2^N sized block and will be annoyed at getting
fa8f4f85
TC
1598 * 2^N+1.
1599 * Only increment if the allocation isn't MEM_SIZE_MAX,
1600 * otherwise it will wrap to 0.
1601 */
1602 if (newlen & 0xff && newlen != MEM_SIZE_MAX)
cbcb2a16
DM
1603 newlen++;
1604#endif
1605
ce861ea7
YO
1606#if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1607#define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1608#endif
1609
bd81e77b 1610 if (newlen > SvLEN(sv)) { /* need more room? */
f1200559 1611 STRLEN minlen = SvCUR(sv);
3c239bea 1612 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
f1200559
WH
1613 if (newlen < minlen)
1614 newlen = minlen;
ce861ea7 1615#ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
7c641603
KW
1616
1617 /* Don't round up on the first allocation, as odds are pretty good that
1618 * the initial request is accurate as to what is really needed */
ce861ea7 1619 if (SvLEN(sv)) {
9efda33a
TC
1620 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1621 if (rounded > newlen)
1622 newlen = rounded;
ce861ea7 1623 }
bd81e77b 1624#endif
98653f18 1625 if (SvLEN(sv) && s) {
10edeb5d 1626 s = (char*)saferealloc(s, newlen);
bd81e77b
NC
1627 }
1628 else {
10edeb5d 1629 s = (char*)safemalloc(newlen);
bd81e77b
NC
1630 if (SvPVX_const(sv) && SvCUR(sv)) {
1631 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1632 }
1633 }
1634 SvPV_set(sv, s);
ce861ea7 1635#ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
98653f18
NC
1636 /* Do this here, do it once, do it right, and then we will never get
1637 called back into sv_grow() unless there really is some growing
1638 needed. */
ca7c1a29 1639 SvLEN_set(sv, Perl_safesysmalloc_size(s));
98653f18 1640#else
bd81e77b 1641 SvLEN_set(sv, newlen);
98653f18 1642#endif
bd81e77b
NC
1643 }
1644 return s;
1645}
aeb18a1e 1646
bd81e77b
NC
1647/*
1648=for apidoc sv_setiv
932e9ff9 1649
bd81e77b
NC
1650Copies an integer into the given SV, upgrading first if necessary.
1651Does not handle 'set' magic. See also C<sv_setiv_mg>.
463ee0b2 1652
bd81e77b
NC
1653=cut
1654*/
463ee0b2 1655
bd81e77b 1656void
5aaab254 1657Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
bd81e77b 1658{
7918f24d
NC
1659 PERL_ARGS_ASSERT_SV_SETIV;
1660
bd81e77b
NC
1661 SV_CHECK_THINKFIRST_COW_DROP(sv);
1662 switch (SvTYPE(sv)) {
1663 case SVt_NULL:
bd81e77b 1664 case SVt_NV:
3376de98 1665 sv_upgrade(sv, SVt_IV);
bd81e77b 1666 break;
bd81e77b
NC
1667 case SVt_PV:
1668 sv_upgrade(sv, SVt_PVIV);
1669 break;
463ee0b2 1670
bd81e77b 1671 case SVt_PVGV:
6e592b3a
BM
1672 if (!isGV_with_GP(sv))
1673 break;
bd81e77b
NC
1674 case SVt_PVAV:
1675 case SVt_PVHV:
1676 case SVt_PVCV:
1677 case SVt_PVFM:
1678 case SVt_PVIO:
22e74366 1679 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b
NC
1680 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1681 OP_DESC(PL_op));
42d0e0b7 1682 default: NOOP;
bd81e77b
NC
1683 }
1684 (void)SvIOK_only(sv); /* validate number */
1685 SvIV_set(sv, i);
1686 SvTAINT(sv);
1687}
932e9ff9 1688
bd81e77b
NC
1689/*
1690=for apidoc sv_setiv_mg
d33b2eba 1691
bd81e77b 1692Like C<sv_setiv>, but also handles 'set' magic.
1c846c1f 1693
bd81e77b
NC
1694=cut
1695*/
d33b2eba 1696
bd81e77b 1697void
5aaab254 1698Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
bd81e77b 1699{
7918f24d
NC
1700 PERL_ARGS_ASSERT_SV_SETIV_MG;
1701
bd81e77b
NC
1702 sv_setiv(sv,i);
1703 SvSETMAGIC(sv);
1704}
727879eb 1705
bd81e77b
NC
1706/*
1707=for apidoc sv_setuv
d33b2eba 1708
bd81e77b
NC
1709Copies an unsigned integer into the given SV, upgrading first if necessary.
1710Does not handle 'set' magic. See also C<sv_setuv_mg>.
9b94d1dd 1711
bd81e77b
NC
1712=cut
1713*/
d33b2eba 1714
bd81e77b 1715void
5aaab254 1716Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
bd81e77b 1717{
7918f24d
NC
1718 PERL_ARGS_ASSERT_SV_SETUV;
1719
013abb9b
NC
1720 /* With the if statement to ensure that integers are stored as IVs whenever
1721 possible:
bd81e77b 1722 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
d33b2eba 1723
bd81e77b
NC
1724 without
1725 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1c846c1f 1726
013abb9b
NC
1727 If you wish to remove the following if statement, so that this routine
1728 (and its callers) always return UVs, please benchmark to see what the
1729 effect is. Modern CPUs may be different. Or may not :-)
bd81e77b
NC
1730 */
1731 if (u <= (UV)IV_MAX) {
1732 sv_setiv(sv, (IV)u);
1733 return;
1734 }
1735 sv_setiv(sv, 0);
1736 SvIsUV_on(sv);
1737 SvUV_set(sv, u);
1738}
d33b2eba 1739
bd81e77b
NC
1740/*
1741=for apidoc sv_setuv_mg
727879eb 1742
bd81e77b 1743Like C<sv_setuv>, but also handles 'set' magic.
9b94d1dd 1744
bd81e77b
NC
1745=cut
1746*/
5e2fc214 1747
bd81e77b 1748void
5aaab254 1749Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
bd81e77b 1750{
7918f24d
NC
1751 PERL_ARGS_ASSERT_SV_SETUV_MG;
1752
bd81e77b
NC
1753 sv_setuv(sv,u);
1754 SvSETMAGIC(sv);
1755}
5e2fc214 1756
954c1994 1757/*
bd81e77b 1758=for apidoc sv_setnv
954c1994 1759
bd81e77b
NC
1760Copies a double into the given SV, upgrading first if necessary.
1761Does not handle 'set' magic. See also C<sv_setnv_mg>.
954c1994
GS
1762
1763=cut
1764*/
1765
63f97190 1766void
5aaab254 1767Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
79072805 1768{
7918f24d
NC
1769 PERL_ARGS_ASSERT_SV_SETNV;
1770
bd81e77b
NC
1771 SV_CHECK_THINKFIRST_COW_DROP(sv);
1772 switch (SvTYPE(sv)) {
79072805 1773 case SVt_NULL:
79072805 1774 case SVt_IV:
bd81e77b 1775 sv_upgrade(sv, SVt_NV);
79072805
LW
1776 break;
1777 case SVt_PV:
79072805 1778 case SVt_PVIV:
bd81e77b 1779 sv_upgrade(sv, SVt_PVNV);
79072805 1780 break;
bd4b1eb5 1781
bd4b1eb5 1782 case SVt_PVGV:
6e592b3a
BM
1783 if (!isGV_with_GP(sv))
1784 break;
bd81e77b
NC
1785 case SVt_PVAV:
1786 case SVt_PVHV:
79072805 1787 case SVt_PVCV:
bd81e77b
NC
1788 case SVt_PVFM:
1789 case SVt_PVIO:
22e74366 1790 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b 1791 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
94bbb3f4 1792 OP_DESC(PL_op));
42d0e0b7 1793 default: NOOP;
2068cd4d 1794 }
bd81e77b
NC
1795 SvNV_set(sv, num);
1796 (void)SvNOK_only(sv); /* validate number */
1797 SvTAINT(sv);
79072805
LW
1798}
1799
645c22ef 1800/*
bd81e77b 1801=for apidoc sv_setnv_mg
645c22ef 1802
bd81e77b 1803Like C<sv_setnv>, but also handles 'set' magic.
645c22ef
DM
1804
1805=cut
1806*/
1807
bd81e77b 1808void
5aaab254 1809Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
79072805 1810{
7918f24d
NC
1811 PERL_ARGS_ASSERT_SV_SETNV_MG;
1812
bd81e77b
NC
1813 sv_setnv(sv,num);
1814 SvSETMAGIC(sv);
79072805
LW
1815}
1816
3f7602fa
TC
1817/* Return a cleaned-up, printable version of sv, for non-numeric, or
1818 * not incrementable warning display.
1819 * Originally part of S_not_a_number().
1820 * The return value may be != tmpbuf.
bd81e77b 1821 */
954c1994 1822
3f7602fa
TC
1823STATIC const char *
1824S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1825 const char *pv;
94463019 1826
3f7602fa 1827 PERL_ARGS_ASSERT_SV_DISPLAY;
7918f24d 1828
94463019 1829 if (DO_UTF8(sv)) {
3f7602fa 1830 SV *dsv = newSVpvs_flags("", SVs_TEMP);
37b8cdd1 1831 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
94463019
JH
1832 } else {
1833 char *d = tmpbuf;
3f7602fa 1834 const char * const limit = tmpbuf + tmpbuf_size - 8;
94463019
JH
1835 /* each *s can expand to 4 chars + "...\0",
1836 i.e. need room for 8 chars */
ecdeb87c 1837
00b6aa41
AL
1838 const char *s = SvPVX_const(sv);
1839 const char * const end = s + SvCUR(sv);
1840 for ( ; s < end && d < limit; s++ ) {
94463019 1841 int ch = *s & 0xFF;
bd27cf70 1842 if (! isASCII(ch) && !isPRINT_LC(ch)) {
94463019
JH
1843 *d++ = 'M';
1844 *d++ = '-';
bd27cf70
KW
1845
1846 /* Map to ASCII "equivalent" of Latin1 */
1847 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
94463019
JH
1848 }
1849 if (ch == '\n') {
1850 *d++ = '\\';
1851 *d++ = 'n';
1852 }
1853 else if (ch == '\r') {
1854 *d++ = '\\';
1855 *d++ = 'r';
1856 }
1857 else if (ch == '\f') {
1858 *d++ = '\\';
1859 *d++ = 'f';
1860 }
1861 else if (ch == '\\') {
1862 *d++ = '\\';
1863 *d++ = '\\';
1864 }
1865 else if (ch == '\0') {
1866 *d++ = '\\';
1867 *d++ = '0';
1868 }
1869 else if (isPRINT_LC(ch))
1870 *d++ = ch;
1871 else {
1872 *d++ = '^';
1873 *d++ = toCTRL(ch);
1874 }
1875 }
1876 if (s < end) {
1877 *d++ = '.';
1878 *d++ = '.';
1879 *d++ = '.';
1880 }
1881 *d = '\0';
1882 pv = tmpbuf;
a0d0e21e 1883 }
a0d0e21e 1884
3f7602fa
TC
1885 return pv;
1886}
1887
1888/* Print an "isn't numeric" warning, using a cleaned-up,
1889 * printable version of the offending string
1890 */
1891
1892STATIC void
1893S_not_a_number(pTHX_ SV *const sv)
1894{
3f7602fa
TC
1895 char tmpbuf[64];
1896 const char *pv;
1897
1898 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1899
1900 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1901
533c011a 1902 if (PL_op)
9014280d 1903 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1904 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019
JH
1905 "Argument \"%s\" isn't numeric in %s", pv,
1906 OP_DESC(PL_op));
a0d0e21e 1907 else
9014280d 1908 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1909 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019 1910 "Argument \"%s\" isn't numeric", pv);
a0d0e21e
LW
1911}
1912
3f7602fa
TC
1913STATIC void
1914S_not_incrementable(pTHX_ SV *const sv) {
3f7602fa
TC
1915 char tmpbuf[64];
1916 const char *pv;
1917
1918 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1919
1920 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1921
1922 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1923 "Argument \"%s\" treated as 0 in increment (++)", pv);
1924}
1925
c2988b20
NC
1926/*
1927=for apidoc looks_like_number
1928
645c22ef
DM
1929Test if the content of an SV looks like a number (or is a number).
1930C<Inf> and C<Infinity> are treated as numbers (so will not issue a
f52e41ad
FC
1931non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1932ignored.
c2988b20
NC
1933
1934=cut
1935*/
1936
1937I32
aad570aa 1938Perl_looks_like_number(pTHX_ SV *const sv)
c2988b20 1939{
eb578fdb 1940 const char *sbegin;
c2988b20 1941 STRLEN len;
ea2485eb 1942 int numtype;
c2988b20 1943
7918f24d
NC
1944 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1945
f52e41ad
FC
1946 if (SvPOK(sv) || SvPOKp(sv)) {
1947 sbegin = SvPV_nomg_const(sv, len);
c2988b20 1948 }
c2988b20 1949 else
e0ab1c0e 1950 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
ea2485eb
JH
1951 numtype = grok_number(sbegin, len, NULL);
1952 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
c2988b20 1953}
25da4f38 1954
19f6321d
NC
1955STATIC bool
1956S_glob_2number(pTHX_ GV * const gv)
180488f8 1957{
7918f24d
NC
1958 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1959
675c862f
AL
1960 /* We know that all GVs stringify to something that is not-a-number,
1961 so no need to test that. */
1962 if (ckWARN(WARN_NUMERIC))
8e629ff4
FC
1963 {
1964 SV *const buffer = sv_newmortal();
1965 gv_efullname3(buffer, gv, "*");
675c862f 1966 not_a_number(buffer);
8e629ff4 1967 }
675c862f
AL
1968 /* We just want something true to return, so that S_sv_2iuv_common
1969 can tail call us and return true. */
19f6321d 1970 return TRUE;
675c862f
AL
1971}
1972
25da4f38
IZ
1973/* Actually, ISO C leaves conversion of UV to IV undefined, but
1974 until proven guilty, assume that things are not that bad... */
1975
645c22ef
DM
1976/*
1977 NV_PRESERVES_UV:
1978
1979 As 64 bit platforms often have an NV that doesn't preserve all bits of
28e5dec8
JH
1980 an IV (an assumption perl has been based on to date) it becomes necessary
1981 to remove the assumption that the NV always carries enough precision to
1982 recreate the IV whenever needed, and that the NV is the canonical form.
1983 Instead, IV/UV and NV need to be given equal rights. So as to not lose
645c22ef 1984 precision as a side effect of conversion (which would lead to insanity
28e5dec8 1985 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
8a4a3196
KW
1986 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1987 where precision was lost, and IV/UV/NV slots that have a valid conversion
1988 which has lost no precision
645c22ef 1989 2) to ensure that if a numeric conversion to one form is requested that
28e5dec8
JH
1990 would lose precision, the precise conversion (or differently
1991 imprecise conversion) is also performed and cached, to prevent
1992 requests for different numeric formats on the same SV causing
1993 lossy conversion chains. (lossless conversion chains are perfectly
1994 acceptable (still))
1995
1996
1997 flags are used:
1998 SvIOKp is true if the IV slot contains a valid value
1999 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
2000 SvNOKp is true if the NV slot contains a valid value
2001 SvNOK is true only if the NV value is accurate
2002
2003 so
645c22ef 2004 while converting from PV to NV, check to see if converting that NV to an
28e5dec8
JH
2005 IV(or UV) would lose accuracy over a direct conversion from PV to
2006 IV(or UV). If it would, cache both conversions, return NV, but mark
2007 SV as IOK NOKp (ie not NOK).
2008
645c22ef 2009 While converting from PV to IV, check to see if converting that IV to an
28e5dec8
JH
2010 NV would lose accuracy over a direct conversion from PV to NV. If it
2011 would, cache both conversions, flag similarly.
2012
2013 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
2014 correctly because if IV & NV were set NV *always* overruled.
645c22ef
DM
2015 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
2016 changes - now IV and NV together means that the two are interchangeable:
28e5dec8 2017 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
d460ef45 2018
645c22ef
DM
2019 The benefit of this is that operations such as pp_add know that if
2020 SvIOK is true for both left and right operands, then integer addition
2021 can be used instead of floating point (for cases where the result won't
2022 overflow). Before, floating point was always used, which could lead to
28e5dec8
JH
2023 loss of precision compared with integer addition.
2024
2025 * making IV and NV equal status should make maths accurate on 64 bit
2026 platforms
2027 * may speed up maths somewhat if pp_add and friends start to use
645c22ef 2028 integers when possible instead of fp. (Hopefully the overhead in
28e5dec8
JH
2029 looking for SvIOK and checking for overflow will not outweigh the
2030 fp to integer speedup)
2031 * will slow down integer operations (callers of SvIV) on "inaccurate"
2032 values, as the change from SvIOK to SvIOKp will cause a call into
2033 sv_2iv each time rather than a macro access direct to the IV slot
2034 * should speed up number->string conversion on integers as IV is
645c22ef 2035 favoured when IV and NV are equally accurate
28e5dec8
JH
2036
2037 ####################################################################
645c22ef
DM
2038 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2039 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2040 On the other hand, SvUOK is true iff UV.
28e5dec8
JH
2041 ####################################################################
2042
645c22ef 2043 Your mileage will vary depending your CPU's relative fp to integer
28e5dec8
JH
2044 performance ratio.
2045*/
2046
2047#ifndef NV_PRESERVES_UV
645c22ef
DM
2048# define IS_NUMBER_UNDERFLOW_IV 1
2049# define IS_NUMBER_UNDERFLOW_UV 2
2050# define IS_NUMBER_IV_AND_UV 2
2051# define IS_NUMBER_OVERFLOW_IV 4
2052# define IS_NUMBER_OVERFLOW_UV 5
2053
2054/* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
28e5dec8
JH
2055
2056/* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2057STATIC int
5aaab254 2058S_sv_2iuv_non_preserve(pTHX_ SV *const sv
47031da6
NC
2059# ifdef DEBUGGING
2060 , I32 numtype
2061# endif
2062 )
28e5dec8 2063{
7918f24d 2064 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
23491f1d 2065 PERL_UNUSED_CONTEXT;
7918f24d 2066
3f7c398e 2067 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
2068 if (SvNVX(sv) < (NV)IV_MIN) {
2069 (void)SvIOKp_on(sv);
2070 (void)SvNOK_on(sv);
45977657 2071 SvIV_set(sv, IV_MIN);
28e5dec8
JH
2072 return IS_NUMBER_UNDERFLOW_IV;
2073 }
2074 if (SvNVX(sv) > (NV)UV_MAX) {
2075 (void)SvIOKp_on(sv);
2076 (void)SvNOK_on(sv);
2077 SvIsUV_on(sv);
607fa7f2 2078 SvUV_set(sv, UV_MAX);
28e5dec8
JH
2079 return IS_NUMBER_OVERFLOW_UV;
2080 }
c2988b20
NC
2081 (void)SvIOKp_on(sv);
2082 (void)SvNOK_on(sv);
2083 /* Can't use strtol etc to convert this string. (See truth table in
2084 sv_2iv */
2085 if (SvNVX(sv) <= (UV)IV_MAX) {
45977657 2086 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2087 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2088 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2089 } else {
2090 /* Integer is imprecise. NOK, IOKp */
2091 }
2092 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2093 }
2094 SvIsUV_on(sv);
607fa7f2 2095 SvUV_set(sv, U_V(SvNVX(sv)));
659c4b96 2096 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2097 if (SvUVX(sv) == UV_MAX) {
2098 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2099 possibly be preserved by NV. Hence, it must be overflow.
2100 NOK, IOKp */
2101 return IS_NUMBER_OVERFLOW_UV;
2102 }
2103 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2104 } else {
2105 /* Integer is imprecise. NOK, IOKp */
28e5dec8 2106 }
c2988b20 2107 return IS_NUMBER_OVERFLOW_IV;
28e5dec8 2108}
645c22ef
DM
2109#endif /* !NV_PRESERVES_UV*/
2110
a13f4dff 2111/* If numtype is infnan, set the NV of the sv accordingly.
5564cd7f 2112 * If numtype is anything else, try setting the NV using Atof(PV). */
3c81f0b3
DD
2113#ifdef USING_MSVC6
2114# pragma warning(push)
2115# pragma warning(disable:4756;disable:4056)
2116#endif
a13f4dff 2117static void
3823048b 2118S_sv_setnv(pTHX_ SV* sv, int numtype)
a13f4dff 2119{
07925c5e 2120 bool pok = cBOOL(SvPOK(sv));
5564cd7f 2121 bool nok = FALSE;
a13f4dff
JH
2122 if ((numtype & IS_NUMBER_INFINITY)) {
2123 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
5564cd7f 2124 nok = TRUE;
a13f4dff
JH
2125 }
2126 else if ((numtype & IS_NUMBER_NAN)) {
3823048b 2127 SvNV_set(sv, NV_NAN);
d48bd569 2128 nok = TRUE;
a13f4dff 2129 }
d48bd569 2130 else if (pok) {
a13f4dff 2131 SvNV_set(sv, Atof(SvPVX_const(sv)));
d48bd569
JH
2132 /* Purposefully no true nok here, since we don't want to blow
2133 * away the possible IOK/UV of an existing sv. */
2134 }
5564cd7f 2135 if (nok) {
d48bd569 2136 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
5564cd7f
JH
2137 if (pok)
2138 SvPOK_on(sv); /* PV is okay, though. */
2139 }
a13f4dff 2140}
3c81f0b3
DD
2141#ifdef USING_MSVC6
2142# pragma warning(pop)
2143#endif
a13f4dff 2144
af359546 2145STATIC bool
7918f24d
NC
2146S_sv_2iuv_common(pTHX_ SV *const sv)
2147{
7918f24d
NC
2148 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2149
af359546 2150 if (SvNOKp(sv)) {
28e5dec8
JH
2151 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2152 * without also getting a cached IV/UV from it at the same time
2153 * (ie PV->NV conversion should detect loss of accuracy and cache
af359546
NC
2154 * IV or UV at same time to avoid this. */
2155 /* IV-over-UV optimisation - choose to cache IV if possible */
25da4f38
IZ
2156
2157 if (SvTYPE(sv) == SVt_NV)
2158 sv_upgrade(sv, SVt_PVNV);
2159
28e5dec8
JH
2160 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2161 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2162 certainly cast into the IV range at IV_MAX, whereas the correct
2163 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2164 cases go to UV */
e91de695
JH
2165#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2166 if (Perl_isnan(SvNVX(sv))) {
2167 SvUV_set(sv, 0);
2168 SvIsUV_on(sv);
2169 return FALSE;
2170 }
2171#endif
28e5dec8 2172 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
45977657 2173 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2174 if (SvNVX(sv) == (NV) SvIVX(sv)
28e5dec8 2175#ifndef NV_PRESERVES_UV
53e2bfb7 2176 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
28e5dec8
JH
2177 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2178 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2179 /* Don't flag it as "accurately an integer" if the number
2180 came from a (by definition imprecise) NV operation, and
2181 we're outside the range of NV integer precision */
2182#endif
2183 ) {
a43d94f2
NC
2184 if (SvNOK(sv))
2185 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2186 else {
2187 /* scalar has trailing garbage, eg "42a" */
2188 }
28e5dec8 2189 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2190 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
28e5dec8
JH
2191 PTR2UV(sv),
2192 SvNVX(sv),
2193 SvIVX(sv)));
2194
2195 } else {
2196 /* IV not precise. No need to convert from PV, as NV
2197 conversion would already have cached IV if it detected
2198 that PV->IV would be better than PV->NV->IV
2199 flags already correct - don't set public IOK. */
2200 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2201 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
28e5dec8
JH
2202 PTR2UV(sv),
2203 SvNVX(sv),
2204 SvIVX(sv)));
2205 }
2206 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2207 but the cast (NV)IV_MIN rounds to a the value less (more
2208 negative) than IV_MIN which happens to be equal to SvNVX ??
2209 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2210 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2211 (NV)UVX == NVX are both true, but the values differ. :-(
2212 Hopefully for 2s complement IV_MIN is something like
2213 0x8000000000000000 which will be exact. NWC */
d460ef45 2214 }
25da4f38 2215 else {
607fa7f2 2216 SvUV_set(sv, U_V(SvNVX(sv)));
28e5dec8 2217 if (
659c4b96 2218 (SvNVX(sv) == (NV) SvUVX(sv))
28e5dec8
JH
2219#ifndef NV_PRESERVES_UV
2220 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2221 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2222 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2223 /* Don't flag it as "accurately an integer" if the number
2224 came from a (by definition imprecise) NV operation, and
2225 we're outside the range of NV integer precision */
2226#endif
a43d94f2 2227 && SvNOK(sv)
28e5dec8
JH
2228 )
2229 SvIOK_on(sv);
25da4f38 2230 SvIsUV_on(sv);
1c846c1f 2231 DEBUG_c(PerlIO_printf(Perl_debug_log,
57def98f 2232 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
56431972 2233 PTR2UV(sv),
57def98f
JH
2234 SvUVX(sv),
2235 SvUVX(sv)));
25da4f38 2236 }
748a9306 2237 }
cd84013a 2238 else if (SvPOKp(sv)) {
c2988b20 2239 UV value;
3823048b 2240 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
af359546 2241 /* We want to avoid a possible problem when we cache an IV/ a UV which
25da4f38 2242 may be later translated to an NV, and the resulting NV is not
c2988b20
NC
2243 the same as the direct translation of the initial string
2244 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2245 be careful to ensure that the value with the .456 is around if the
2246 NV value is requested in the future).
1c846c1f 2247
af359546 2248 This means that if we cache such an IV/a UV, we need to cache the
25da4f38 2249 NV as well. Moreover, we trade speed for space, and do not
28e5dec8 2250 cache the NV if we are sure it's not needed.
25da4f38 2251 */
16b7a9a4 2252
c2988b20
NC
2253 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2254 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2255 == IS_NUMBER_IN_UV) {
5e045b90 2256 /* It's definitely an integer, only upgrade to PVIV */
28e5dec8
JH
2257 if (SvTYPE(sv) < SVt_PVIV)
2258 sv_upgrade(sv, SVt_PVIV);
f7bbb42a 2259 (void)SvIOK_on(sv);
c2988b20
NC
2260 } else if (SvTYPE(sv) < SVt_PVNV)
2261 sv_upgrade(sv, SVt_PVNV);
28e5dec8 2262
a13f4dff 2263 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
75a57a38 2264 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
6b322424 2265 not_a_number(sv);
3823048b 2266 S_sv_setnv(aTHX_ sv, numtype);
a13f4dff
JH
2267 return FALSE;
2268 }
2269
f2524eef 2270 /* If NVs preserve UVs then we only use the UV value if we know that
c2988b20
NC
2271 we aren't going to call atof() below. If NVs don't preserve UVs
2272 then the value returned may have more precision than atof() will
2273 return, even though value isn't perfectly accurate. */
2274 if ((numtype & (IS_NUMBER_IN_UV
2275#ifdef NV_PRESERVES_UV
2276 | IS_NUMBER_NOT_INT
2277#endif
2278 )) == IS_NUMBER_IN_UV) {
2279 /* This won't turn off the public IOK flag if it was set above */
2280 (void)SvIOKp_on(sv);
2281
2282 if (!(numtype & IS_NUMBER_NEG)) {
2283 /* positive */;
2284 if (value <= (UV)IV_MAX) {
45977657 2285 SvIV_set(sv, (IV)value);
c2988b20 2286 } else {
af359546 2287 /* it didn't overflow, and it was positive. */
607fa7f2 2288 SvUV_set(sv, value);
c2988b20
NC
2289 SvIsUV_on(sv);
2290 }
2291 } else {
2292 /* 2s complement assumption */
2293 if (value <= (UV)IV_MIN) {
53e2bfb7
DM
2294 SvIV_set(sv, value == (UV)IV_MIN
2295 ? IV_MIN : -(IV)value);
c2988b20
NC
2296 } else {
2297 /* Too negative for an IV. This is a double upgrade, but
d1be9408 2298 I'm assuming it will be rare. */
c2988b20
NC
2299 if (SvTYPE(sv) < SVt_PVNV)
2300 sv_upgrade(sv, SVt_PVNV);
2301 SvNOK_on(sv);
2302 SvIOK_off(sv);
2303 SvIOKp_on(sv);
9d6ce603 2304 SvNV_set(sv, -(NV)value);
45977657 2305 SvIV_set(sv, IV_MIN);
c2988b20
NC
2306 }
2307 }
2308 }
2309 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2310 will be in the previous block to set the IV slot, and the next
2311 block to set the NV slot. So no else here. */
2312
2313 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2314 != IS_NUMBER_IN_UV) {
2315 /* It wasn't an (integer that doesn't overflow the UV). */
3823048b 2316 S_sv_setnv(aTHX_ sv, numtype);
28e5dec8 2317
c2988b20
NC
2318 if (! numtype && ckWARN(WARN_NUMERIC))
2319 not_a_number(sv);
28e5dec8 2320
88cb8500 2321 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
c2988b20 2322 PTR2UV(sv), SvNVX(sv)));
28e5dec8 2323
28e5dec8 2324#ifdef NV_PRESERVES_UV
af359546
NC
2325 (void)SvIOKp_on(sv);
2326 (void)SvNOK_on(sv);
e91de695
JH
2327#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2328 if (Perl_isnan(SvNVX(sv))) {
2329 SvUV_set(sv, 0);
2330 SvIsUV_on(sv);
2331 return FALSE;
2332 }
2333#endif
af359546
NC
2334 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2335 SvIV_set(sv, I_V(SvNVX(sv)));
2336 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2337 SvIOK_on(sv);
2338 } else {
6f207bd3 2339 NOOP; /* Integer is imprecise. NOK, IOKp */
af359546
NC
2340 }
2341 /* UV will not work better than IV */
2342 } else {
2343 if (SvNVX(sv) > (NV)UV_MAX) {
2344 SvIsUV_on(sv);
2345 /* Integer is inaccurate. NOK, IOKp, is UV */
2346 SvUV_set(sv, UV_MAX);
af359546
NC
2347 } else {
2348 SvUV_set(sv, U_V(SvNVX(sv)));
2349 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2350 NV preservse UV so can do correct comparison. */
2351 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2352 SvIOK_on(sv);
af359546 2353 } else {
6f207bd3 2354 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
af359546
NC
2355 }
2356 }
4b0c9573 2357 SvIsUV_on(sv);
af359546 2358 }
28e5dec8 2359#else /* NV_PRESERVES_UV */
c2988b20
NC
2360 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2361 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
af359546 2362 /* The IV/UV slot will have been set from value returned by
c2988b20
NC
2363 grok_number above. The NV slot has just been set using
2364 Atof. */
560b0c46 2365 SvNOK_on(sv);
c2988b20
NC
2366 assert (SvIOKp(sv));
2367 } else {
2368 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2369 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2370 /* Small enough to preserve all bits. */
2371 (void)SvIOKp_on(sv);
2372 SvNOK_on(sv);
45977657 2373 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2374 if ((NV)(SvIVX(sv)) == SvNVX(sv))
c2988b20
NC
2375 SvIOK_on(sv);
2376 /* Assumption: first non-preserved integer is < IV_MAX,
2377 this NV is in the preserved range, therefore: */
2378 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2379 < (UV)IV_MAX)) {
32fdb065 2380 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
2381 }
2382 } else {
2383 /* IN_UV NOT_INT
2384 0 0 already failed to read UV.
2385 0 1 already failed to read UV.
2386 1 0 you won't get here in this case. IV/UV
2387 slot set, public IOK, Atof() unneeded.
2388 1 1 already read UV.
2389 so there's no point in sv_2iuv_non_preserve() attempting
2390 to use atol, strtol, strtoul etc. */
47031da6 2391# ifdef DEBUGGING
40a17c4c 2392 sv_2iuv_non_preserve (sv, numtype);
47031da6
NC
2393# else
2394 sv_2iuv_non_preserve (sv);
2395# endif
c2988b20
NC
2396 }
2397 }
28e5dec8 2398#endif /* NV_PRESERVES_UV */
a43d94f2
NC
2399 /* It might be more code efficient to go through the entire logic above
2400 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2401 gets complex and potentially buggy, so more programmer efficient
2402 to do it this way, by turning off the public flags: */
2403 if (!numtype)
2404 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
25da4f38 2405 }
af359546
NC
2406 }
2407 else {
675c862f 2408 if (isGV_with_GP(sv))
159b6efe 2409 return glob_2number(MUTABLE_GV(sv));
180488f8 2410
4f62cd62 2411 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
af359546 2412 report_uninit(sv);
25da4f38
IZ
2413 if (SvTYPE(sv) < SVt_IV)
2414 /* Typically the caller expects that sv_any is not NULL now. */
2415 sv_upgrade(sv, SVt_IV);
af359546
NC
2416 /* Return 0 from the caller. */
2417 return TRUE;
2418 }
2419 return FALSE;
2420}
2421
2422/*
2423=for apidoc sv_2iv_flags
2424
2425Return the integer value of an SV, doing any necessary string
2426conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2427Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2428
2429=cut
2430*/
2431
2432IV
5aaab254 2433Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
af359546 2434{
1061065f 2435 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
4bac9ae4 2436
217f6fa3
FC
2437 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2438 && SvTYPE(sv) != SVt_PVFM);
2439
4bac9ae4
CS
2440 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2441 mg_get(sv);
2442
2443 if (SvROK(sv)) {
2444 if (SvAMAGIC(sv)) {
2445 SV * tmpstr;
2446 if (flags & SV_SKIP_OVERLOAD)
2447 return 0;
2448 tmpstr = AMG_CALLunary(sv, numer_amg);
2449 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2450 return SvIV(tmpstr);
2451 }
2452 }
2453 return PTR2IV(SvRV(sv));
2454 }
2455
8d919b0a 2456 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d
NC
2457 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2458 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2459 In practice they are extremely unlikely to actually get anywhere
2460 accessible by user Perl code - the only way that I'm aware of is when
2461 a constant subroutine which is used as the second argument to index.
cd84013a
FC
2462
2463 Regexps have no SvIVX and SvNVX fields.
2b2b6d6d 2464 */
8d919b0a 2465 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2466 {
71c558c3 2467 UV value;
8d919b0a
FC
2468 const char * const ptr =
2469 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
e91de695
JH
2470 const int numtype
2471 = grok_number(ptr, SvCUR(sv), &value);
71c558c3
NC
2472
2473 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2474 == IS_NUMBER_IN_UV) {
2475 /* It's definitely an integer */
2476 if (numtype & IS_NUMBER_NEG) {
2477 if (value < (UV)IV_MIN)
2478 return -(IV)value;
2479 } else {
2480 if (value < (UV)IV_MAX)
2481 return (IV)value;
2482 }
2483 }
058b8ae2 2484
e91de695
JH
2485 /* Quite wrong but no good choices. */
2486 if ((numtype & IS_NUMBER_INFINITY)) {
2487 return (numtype & IS_NUMBER_NEG) ? IV_MIN : IV_MAX;
2488 } else if ((numtype & IS_NUMBER_NAN)) {
2489 return 0; /* So wrong. */
2490 }
2491
71c558c3
NC
2492 if (!numtype) {
2493 if (ckWARN(WARN_NUMERIC))
2494 not_a_number(sv);
2495 }
8d919b0a 2496 return I_V(Atof(ptr));
e20b6c3b 2497 }
4bac9ae4
CS
2498 }
2499
2500 if (SvTHINKFIRST(sv)) {
73381561 2501#ifdef PERL_OLD_COPY_ON_WRITE
af359546
NC
2502 if (SvIsCOW(sv)) {
2503 sv_force_normal_flags(sv, 0);
2504 }
73381561 2505#endif
af359546
NC
2506 if (SvREADONLY(sv) && !SvOK(sv)) {
2507 if (ckWARN(WARN_UNINITIALIZED))
2508 report_uninit(sv);
2509 return 0;
2510 }
2511 }
4bac9ae4 2512
af359546
NC
2513 if (!SvIOKp(sv)) {
2514 if (S_sv_2iuv_common(aTHX_ sv))
2515 return 0;
79072805 2516 }
4bac9ae4 2517
1d7c1841
GS
2518 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2519 PTR2UV(sv),SvIVX(sv)));
25da4f38 2520 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
79072805
LW
2521}
2522
645c22ef 2523/*
891f9566 2524=for apidoc sv_2uv_flags
645c22ef
DM
2525
2526Return the unsigned integer value of an SV, doing any necessary string
891f9566
YST
2527conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2528Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
645c22ef
DM
2529
2530=cut
2531*/
2532
ff68c719 2533UV
5aaab254 2534Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
ff68c719 2535{
1061065f 2536 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
4bac9ae4
CS
2537
2538 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2539 mg_get(sv);
2540
2541 if (SvROK(sv)) {
2542 if (SvAMAGIC(sv)) {
2543 SV *tmpstr;
2544 if (flags & SV_SKIP_OVERLOAD)
2545 return 0;
2546 tmpstr = AMG_CALLunary(sv, numer_amg);
2547 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2548 return SvUV(tmpstr);
2549 }
2550 }
2551 return PTR2UV(SvRV(sv));
2552 }
2553
8d919b0a 2554 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2555 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2556 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2557 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2558 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2559 {
71c558c3 2560 UV value;
8d919b0a
FC
2561 const char * const ptr =
2562 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
e91de695
JH
2563 const int numtype
2564 = grok_number(ptr, SvCUR(sv), &value);
71c558c3
NC
2565
2566 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2567 == IS_NUMBER_IN_UV) {
2568 /* It's definitely an integer */
2569 if (!(numtype & IS_NUMBER_NEG))
2570 return value;
2571 }
058b8ae2 2572
e91de695
JH
2573 /* Quite wrong but no good choices. */
2574 if ((numtype & IS_NUMBER_INFINITY)) {
2575 return UV_MAX; /* So wrong. */
2576 } else if ((numtype & IS_NUMBER_NAN)) {
2577 return 0; /* So wrong. */
2578 }
2579
71c558c3
NC
2580 if (!numtype) {
2581 if (ckWARN(WARN_NUMERIC))
2582 not_a_number(sv);
2583 }
8d919b0a 2584 return U_V(Atof(ptr));
e20b6c3b 2585 }
4bac9ae4
CS
2586 }
2587
2588 if (SvTHINKFIRST(sv)) {
73381561 2589#ifdef PERL_OLD_COPY_ON_WRITE
765f542d
NC
2590 if (SvIsCOW(sv)) {
2591 sv_force_normal_flags(sv, 0);
8a818333 2592 }
73381561 2593#endif
0336b60e 2594 if (SvREADONLY(sv) && !SvOK(sv)) {
0336b60e 2595 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2596 report_uninit(sv);
ff68c719
PP
2597 return 0;
2598 }
2599 }
4bac9ae4 2600
af359546
NC
2601 if (!SvIOKp(sv)) {
2602 if (S_sv_2iuv_common(aTHX_ sv))
2603 return 0;
ff68c719 2604 }
25da4f38 2605
1d7c1841
GS
2606 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2607 PTR2UV(sv),SvUVX(sv)));
25da4f38 2608 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
ff68c719
PP
2609}
2610
645c22ef 2611/*
196007d1 2612=for apidoc sv_2nv_flags
645c22ef
DM
2613
2614Return the num value of an SV, doing any necessary string or integer
fde67290 2615conversion. If flags includes SV_GMAGIC, does an mg_get() first.
39d5de13 2616Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
645c22ef
DM
2617
2618=cut
2619*/
2620
65202027 2621NV
5aaab254 2622Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
79072805 2623{
1061065f
DD
2624 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2625
217f6fa3
FC
2626 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2627 && SvTYPE(sv) != SVt_PVFM);
8d919b0a 2628 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2629 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2630 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2631 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2632 const char *ptr;
39d5de13
DM
2633 if (flags & SV_GMAGIC)
2634 mg_get(sv);
463ee0b2
LW
2635 if (SvNOKp(sv))
2636 return SvNVX(sv);
cd84013a 2637 if (SvPOKp(sv) && !SvIOKp(sv)) {
8d919b0a
FC
2638 ptr = SvPVX_const(sv);
2639 grokpv:
041457d9 2640 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
8d919b0a 2641 !grok_number(ptr, SvCUR(sv), NULL))
a0d0e21e 2642 not_a_number(sv);
8d919b0a 2643 return Atof(ptr);
a0d0e21e 2644 }
25da4f38 2645 if (SvIOKp(sv)) {
1c846c1f 2646 if (SvIsUV(sv))
65202027 2647 return (NV)SvUVX(sv);
25da4f38 2648 else
65202027 2649 return (NV)SvIVX(sv);
47a72cb8
NC
2650 }
2651 if (SvROK(sv)) {
2652 goto return_rok;
2653 }
8d919b0a
FC
2654 if (isREGEXP(sv)) {
2655 ptr = RX_WRAPPED((REGEXP *)sv);
2656 goto grokpv;
2657 }
47a72cb8
NC
2658 assert(SvTYPE(sv) >= SVt_PVMG);
2659 /* This falls through to the report_uninit near the end of the
2660 function. */
2661 } else if (SvTHINKFIRST(sv)) {
a0d0e21e 2662 if (SvROK(sv)) {
47a72cb8 2663 return_rok:
deb46114 2664 if (SvAMAGIC(sv)) {
aee036bb
DM
2665 SV *tmpstr;
2666 if (flags & SV_SKIP_OVERLOAD)
2667 return 0;
31d632c3 2668 tmpstr = AMG_CALLunary(sv, numer_amg);
deb46114
NC
2669 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2670 return SvNV(tmpstr);
2671 }
2672 }
2673 return PTR2NV(SvRV(sv));
a0d0e21e 2674 }
73381561 2675#ifdef PERL_OLD_COPY_ON_WRITE
765f542d
NC
2676 if (SvIsCOW(sv)) {
2677 sv_force_normal_flags(sv, 0);
8a818333 2678 }
73381561 2679#endif
0336b60e 2680 if (SvREADONLY(sv) && !SvOK(sv)) {
599cee73 2681 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2682 report_uninit(sv);
ed6116ce
LW
2683 return 0.0;
2684 }
79072805
LW
2685 }
2686 if (SvTYPE(sv) < SVt_NV) {
7e25a7e9
NC
2687 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2688 sv_upgrade(sv, SVt_NV);
097ee67d 2689 DEBUG_c({
f93f4e46 2690 STORE_NUMERIC_LOCAL_SET_STANDARD();
1d7c1841 2691 PerlIO_printf(Perl_debug_log,
88cb8500 2692 "0x%"UVxf" num(%" NVgf ")\n",
1d7c1841 2693 PTR2UV(sv), SvNVX(sv));
097ee67d
JH
2694 RESTORE_NUMERIC_LOCAL();
2695 });
79072805
LW
2696 }
2697 else if (SvTYPE(sv) < SVt_PVNV)
2698 sv_upgrade(sv, SVt_PVNV);
59d8ce62
NC
2699 if (SvNOKp(sv)) {
2700 return SvNVX(sv);
61604483 2701 }
59d8ce62 2702 if (SvIOKp(sv)) {
9d6ce603 2703 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
28e5dec8 2704#ifdef NV_PRESERVES_UV
a43d94f2
NC
2705 if (SvIOK(sv))
2706 SvNOK_on(sv);
2707 else
2708 SvNOKp_on(sv);
28e5dec8
JH
2709#else
2710 /* Only set the public NV OK flag if this NV preserves the IV */
2711 /* Check it's not 0xFFFFFFFFFFFFFFFF */
a43d94f2
NC
2712 if (SvIOK(sv) &&
2713 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
28e5dec8
JH
2714 : (SvIVX(sv) == I_V(SvNVX(sv))))
2715 SvNOK_on(sv);
2716 else
2717 SvNOKp_on(sv);
2718#endif
93a17b20 2719 }
cd84013a 2720 else if (SvPOKp(sv)) {
c2988b20 2721 UV value;
3823048b 2722 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
041457d9 2723 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
a0d0e21e 2724 not_a_number(sv);
28e5dec8 2725#ifdef NV_PRESERVES_UV
c2988b20
NC
2726 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2727 == IS_NUMBER_IN_UV) {
5e045b90 2728 /* It's definitely an integer */
9d6ce603 2729 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
66d83377 2730 } else {
3823048b 2731 S_sv_setnv(aTHX_ sv, numtype);
66d83377 2732 }
a43d94f2
NC
2733 if (numtype)
2734 SvNOK_on(sv);
2735 else
2736 SvNOKp_on(sv);
28e5dec8 2737#else
e91de695
JH
2738 SvNV_set(sv, Atof(SvPVX_const(sv)));
2739 /* Only set the public NV OK flag if this NV preserves the value in
2740 the PV at least as well as an IV/UV would.
2741 Not sure how to do this 100% reliably. */
2742 /* if that shift count is out of range then Configure's test is
2743 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2744 UV_BITS */
2745 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2746 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2747 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2748 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2749 /* Can't use strtol etc to convert this string, so don't try.
2750 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
c2988b20
NC
2751 SvNOK_on(sv);
2752 } else {
e91de695 2753 /* value has been set. It may not be precise. */
53e2bfb7 2754 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
e91de695
JH
2755 /* 2s complement assumption for (UV)IV_MIN */
2756 SvNOK_on(sv); /* Integer is too negative. */
c2988b20 2757 } else {
e91de695
JH
2758 SvNOKp_on(sv);
2759 SvIOKp_on(sv);
6fa402ec 2760
e91de695 2761 if (numtype & IS_NUMBER_NEG) {
02b08bbc
DM
2762 /* -IV_MIN is undefined, but we should never reach
2763 * this point with both IS_NUMBER_NEG and value ==
2764 * (UV)IV_MIN */
2765 assert(value != (UV)IV_MIN);
e91de695
JH
2766 SvIV_set(sv, -(IV)value);
2767 } else if (value <= (UV)IV_MAX) {
2768 SvIV_set(sv, (IV)value);
2769 } else {
2770 SvUV_set(sv, value);
2771 SvIsUV_on(sv);
2772 }
c2988b20 2773
e91de695
JH
2774 if (numtype & IS_NUMBER_NOT_INT) {
2775 /* I believe that even if the original PV had decimals,
2776 they are lost beyond the limit of the FP precision.
2777 However, neither is canonical, so both only get p
2778 flags. NWC, 2000/11/25 */
2779 /* Both already have p flags, so do nothing */
2780 } else {
2781 const NV nv = SvNVX(sv);
2782 /* XXX should this spot have NAN_COMPARE_BROKEN, too? */
2783 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2784 if (SvIVX(sv) == I_V(nv)) {
2785 SvNOK_on(sv);
2786 } else {
2787 /* It had no "." so it must be integer. */
2788 }
2789 SvIOK_on(sv);
0f83c5a4 2790 } else {
e91de695
JH
2791 /* between IV_MAX and NV(UV_MAX).
2792 Could be slightly > UV_MAX */
2793
2794 if (numtype & IS_NUMBER_NOT_INT) {
2795 /* UV and NV both imprecise. */
0f83c5a4 2796 } else {
e91de695
JH
2797 const UV nv_as_uv = U_V(nv);
2798
2799 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2800 SvNOK_on(sv);
c2988b20 2801 }
e91de695 2802 SvIOK_on(sv);
c2988b20
NC
2803 }
2804 }
2805 }
2806 }
0f83c5a4 2807 }
e91de695
JH
2808 /* It might be more code efficient to go through the entire logic above
2809 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2810 gets complex and potentially buggy, so more programmer efficient
2811 to do it this way, by turning off the public flags: */
2812 if (!numtype)
2813 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
28e5dec8 2814#endif /* NV_PRESERVES_UV */
93a17b20 2815 }
79072805 2816 else {
e91de695
JH
2817 if (isGV_with_GP(sv)) {
2818 glob_2number(MUTABLE_GV(sv));
2819 return 0.0;
2820 }
180488f8 2821
e91de695
JH
2822 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2823 report_uninit(sv);
2824 assert (SvTYPE(sv) >= SVt_NV);
2825 /* Typically the caller expects that sv_any is not NULL now. */
2826 /* XXX Ilya implies that this is a bug in callers that assume this
2827 and ideally should be fixed. */
2828 return 0.0;
79072805 2829 }
097ee67d 2830 DEBUG_c({
e91de695
JH
2831 STORE_NUMERIC_LOCAL_SET_STANDARD();
2832 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" NVgf ")\n",
2833 PTR2UV(sv), SvNVX(sv));
2834 RESTORE_NUMERIC_LOCAL();
2835 });
463ee0b2 2836 return SvNVX(sv);
79072805
LW
2837}
2838
800401ee
JH
2839/*
2840=for apidoc sv_2num
2841
2842Return an SV with the numeric value of the source SV, doing any necessary
d024d1a7
FC
2843reference or overload conversion. The caller is expected to have handled
2844get-magic already.
800401ee
JH
2845
2846=cut
2847*/
2848
2849SV *
5aaab254 2850Perl_sv_2num(pTHX_ SV *const sv)
800401ee 2851{
7918f24d
NC
2852 PERL_ARGS_ASSERT_SV_2NUM;
2853
b9ee0594
RGS
2854 if (!SvROK(sv))
2855 return sv;
800401ee 2856 if (SvAMAGIC(sv)) {
31d632c3 2857 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
a02ec77a 2858 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
800401ee
JH
2859 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2860 return sv_2num(tmpsv);
2861 }
2862 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2863}
2864
645c22ef
DM
2865/* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2866 * UV as a string towards the end of buf, and return pointers to start and
2867 * end of it.
2868 *
2869 * We assume that buf is at least TYPE_CHARS(UV) long.
2870 */
2871
864dbfa3 2872static char *
5de3775c 2873S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
25da4f38 2874{
25da4f38 2875 char *ptr = buf + TYPE_CHARS(UV);
823a54a3 2876 char * const ebuf = ptr;
25da4f38 2877 int sign;
25da4f38 2878
7918f24d
NC
2879 PERL_ARGS_ASSERT_UIV_2BUF;
2880
25da4f38
IZ
2881 if (is_uv)
2882 sign = 0;
2883 else if (iv >= 0) {
2884 uv = iv;
2885 sign = 0;
2886 } else {
53e2bfb7 2887 uv = (iv == IV_MIN) ? (UV)iv : (UV)(-iv);
25da4f38
IZ
2888 sign = 1;
2889 }
2890 do {
eb160463 2891 *--ptr = '0' + (char)(uv % 10);
25da4f38
IZ
2892 } while (uv /= 10);
2893 if (sign)
2894 *--ptr = '-';
2895 *peob = ebuf;
2896 return ptr;
2897}
2898
bfaa02d5
JH
2899/* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
2900 * infinity or a not-a-number, writes the appropriate strings to the
2901 * buffer, including a zero byte. On success returns the written length,
2902 * excluding the zero byte, on failure (not an infinity, not a nan, or the
3823048b
JH
2903 * maxlen too small) returns zero.
2904 *
2905 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2906 * shared string constants we point to, instead of generating a new
2907 * string for each instance. */
bfaa02d5 2908STATIC size_t
3823048b 2909S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
bfaa02d5
JH
2910 assert(maxlen >= 4);
2911 if (maxlen < 4) /* "Inf\0", "NaN\0" */
2912 return 0;
2913 else {
2914 char* s = buffer;
2915 if (Perl_isinf(nv)) {
2916 if (nv < 0) {
2917 if (maxlen < 5) /* "-Inf\0" */
2918 return 0;
2919 *s++ = '-';
2920 } else if (plus) {
2921 *s++ = '+';
2922 }
2923 *s++ = 'I';
2924 *s++ = 'n';
2925 *s++ = 'f';
2926 } else if (Perl_isnan(nv)) {
2927 *s++ = 'N';
2928 *s++ = 'a';
2929 *s++ = 'N';
3823048b
JH
2930 /* XXX optionally output the payload mantissa bits as
2931 * "(unsigned)" (to match the nan("...") C99 function,
2932 * or maybe as "(0xhhh...)" would make more sense...
2933 * provide a format string so that the user can decide?
2934 * NOTE: would affect the maxlen and assert() logic.*/
6e915616 2935 }
3823048b 2936
bfaa02d5
JH
2937 else
2938 return 0;
3823048b 2939 assert((s == buffer + 3) || (s == buffer + 4));
bfaa02d5
JH
2940 *s++ = 0;
2941 return s - buffer - 1; /* -1: excluding the zero byte */
2942 }
2943}
2944
2945/*
2946=for apidoc sv_2pv_flags
2947
2948Returns a pointer to the string value of an SV, and sets *lp to its length.
2949If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2950string if necessary. Normally invoked via the C<SvPV_flags> macro.
2951C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2952
2953=cut
2954*/
2955
2956char *
2957Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2958{
2959 char *s;
2960
2961 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2962
2963 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2964 && SvTYPE(sv) != SVt_PVFM);
2965 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2966 mg_get(sv);
2967 if (SvROK(sv)) {
2968 if (SvAMAGIC(sv)) {
2969 SV *tmpstr;
2970 if (flags & SV_SKIP_OVERLOAD)
2971 return NULL;
2972 tmpstr = AMG_CALLunary(sv, string_amg);
2973 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2974 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2975 /* Unwrap this: */
2976 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2977 */
2978
2979 char *pv;
2980 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2981 if (flags & SV_CONST_RETURN) {
2982 pv = (char *) SvPVX_const(tmpstr);
2983 } else {
2984 pv = (flags & SV_MUTABLE_RETURN)
2985 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2986 }
2987 if (lp)
2988 *lp = SvCUR(tmpstr);
2989 } else {
2990 pv = sv_2pv_flags(tmpstr, lp, flags);
2991 }
2992 if (SvUTF8(tmpstr))
2993 SvUTF8_on(sv);
2994 else
2995 SvUTF8_off(sv);
2996 return pv;
2997 }
2998 }
2999 {
3000 STRLEN len;
3001 char *retval;
3002 char *buffer;
3003 SV *const referent = SvRV(sv);
3004
3005 if (!referent) {
3006 len = 7;
3007 retval = buffer = savepvn("NULLREF", len);
3008 } else if (SvTYPE(referent) == SVt_REGEXP &&
3009 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
3010 amagic_is_enabled(string_amg))) {
3011 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
3012
3013 assert(re);
3014
3015 /* If the regex is UTF-8 we want the containing scalar to
3016 have an UTF-8 flag too */
3017 if (RX_UTF8(re))
3018 SvUTF8_on(sv);
3019 else
3020 SvUTF8_off(sv);
3021
3022 if (lp)
3023 *lp = RX_WRAPLEN(re);
3024
3025 return RX_WRAPPED(re);
3026 } else {
3027 const char *const typestr = sv_reftype(referent, 0);
3028 const STRLEN typelen = strlen(typestr);
3029 UV addr = PTR2UV(referent);
3030 const char *stashname = NULL;
3031 STRLEN stashnamelen = 0; /* hush, gcc */
3032 const char *buffer_end;
3033
3034 if (SvOBJECT(referent)) {
3035 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3036
3037 if (name) {
3038 stashname = HEK_KEY(name);
3039 stashnamelen = HEK_LEN(name);
3040
3041 if (HEK_UTF8(name)) {
3042 SvUTF8_on(sv);
3043 } else {
3044 SvUTF8_off(sv);
3045 }
3046 } else {
3047 stashname = "__ANON__";
3048 stashnamelen = 8;
3049 }
3050 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3051 + 2 * sizeof(UV) + 2 /* )\0 */;
3052 } else {
3053 len = typelen + 3 /* (0x */
3054 + 2 * sizeof(UV) + 2 /* )\0 */;
3055 }
fafee734 3056
4bac9ae4
CS
3057 Newx(buffer, len, char);
3058 buffer_end = retval = buffer + len;
3059
3060 /* Working backwards */
3061 *--retval = '\0';
3062 *--retval = ')';
3063 do {
3064 *--retval = PL_hexdigit[addr & 15];
3065 } while (addr >>= 4);
3066 *--retval = 'x';
3067 *--retval = '0';
3068 *--retval = '(';
3069
3070 retval -= typelen;
3071 memcpy(retval, typestr, typelen);
3072
3073 if (stashname) {
3074 *--retval = '=';
3075 retval -= stashnamelen;
3076 memcpy(retval, stashname, stashnamelen);
c080367d 3077 }
4bac9ae4
CS
3078 /* retval may not necessarily have reached the start of the
3079 buffer here. */
3080 assert (retval >= buffer);
3081
3082 len = buffer_end - retval - 1; /* -1 for that \0 */
463ee0b2 3083 }
cdb061a3 3084 if (lp)
4bac9ae4
CS
3085 *lp = len;
3086 SAVEFREEPV(buffer);
3087 return retval;
79072805 3088 }
79072805 3089 }
4bac9ae4
CS
3090
3091 if (SvPOKp(sv)) {
3092 if (lp)
3093 *lp = SvCUR(sv);
3094 if (flags & SV_MUTABLE_RETURN)
3095 return SvPVX_mutable(sv);
3096 if (flags & SV_CONST_RETURN)
3097 return (char *)SvPVX_const(sv);
3098 return SvPVX(sv);
3099 }
3100
3101 if (SvIOK(sv)) {
28e5dec8
JH
3102 /* I'm assuming that if both IV and NV are equally valid then
3103 converting the IV is going to be more efficient */
e1ec3a88 3104 const U32 isUIOK = SvIsUV(sv);
28e5dec8
JH
3105 char buf[TYPE_CHARS(UV)];
3106 char *ebuf, *ptr;
97a130b8 3107 STRLEN len;
28e5dec8
JH
3108
3109 if (SvTYPE(sv) < SVt_PVIV)
3110 sv_upgrade(sv, SVt_PVIV);
4ea1d550 3111 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
97a130b8 3112 len = ebuf - ptr;
5902b6a9 3113 /* inlined from sv_setpvn */
97a130b8
NC
3114 s = SvGROW_mutable(sv, len + 1);
3115 Move(ptr, s, len, char);
3116 s += len;
28e5dec8 3117 *s = '\0';
b127e37e 3118 SvPOK_on(sv);
28e5dec8 3119 }
4bac9ae4 3120 else if (SvNOK(sv)) {
79072805
LW
3121 if (SvTYPE(sv) < SVt_PVNV)
3122 sv_upgrade(sv, SVt_PVNV);
128eeacb
DD
3123 if (SvNVX(sv) == 0.0
3124#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3125 && !Perl_isnan(SvNVX(sv))
3126#endif
3127 ) {
29912d93
NC
3128 s = SvGROW_mutable(sv, 2);
3129 *s++ = '0';
3130 *s = '\0';
3131 } else {
5e85836e 3132 STRLEN len;
fb8cdbc5 3133 STRLEN size = 5; /* "-Inf\0" */
0c7e610f 3134
fb8cdbc5 3135 s = SvGROW_mutable(sv, size);
3823048b 3136 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
fb8cdbc5 3137 if (len > 0) {
0c7e610f 3138 s += len;
fb8cdbc5
JH
3139 SvPOK_on(sv);
3140 }
0c7e610f 3141 else {
0c7e610f 3142 /* some Xenix systems wipe out errno here */
fb8cdbc5
JH
3143 dSAVE_ERRNO;
3144
3840bff0
JH
3145 size =
3146 1 + /* sign */
3147 1 + /* "." */
3148 NV_DIG +
3149 1 + /* "e" */
3150 1 + /* sign */
3151 5 + /* exponent digits */
3152 1 + /* \0 */
3153 2; /* paranoia */
b127e37e 3154
fb8cdbc5 3155 s = SvGROW_mutable(sv, size);
b127e37e 3156#ifndef USE_LOCALE_NUMERIC
a4eca1d4
JH
3157 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3158
0c7e610f
JH
3159 SvPOK_on(sv);
3160#else
28acfe03 3161 {
3840bff0 3162 bool local_radix;
67d796ae
KW
3163 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3164 STORE_LC_NUMERIC_SET_TO_NEEDED();
3840bff0
JH
3165
3166 local_radix =
3167 PL_numeric_local &&
3168 PL_numeric_radix_sv &&
3169 SvUTF8(PL_numeric_radix_sv);
3170 if (local_radix && SvLEN(PL_numeric_radix_sv) > 1) {
3171 size += SvLEN(PL_numeric_radix_sv) - 1;
3172 s = SvGROW_mutable(sv, size);
3173 }
3174
a4eca1d4 3175 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
0c7e610f
JH
3176
3177 /* If the radix character is UTF-8, and actually is in the
3178 * output, turn on the UTF-8 flag for the scalar */
3840bff0
JH
3179 if (local_radix &&
3180 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3181 SvUTF8_on(sv);
3182 }
3183
0c7e610f 3184 RESTORE_LC_NUMERIC();
28acfe03 3185 }
68e8f474 3186
0c7e610f
JH
3187 /* We don't call SvPOK_on(), because it may come to
3188 * pass that the locale changes so that the
3189 * stringification we just did is no longer correct. We
3190 * will have to re-stringify every time it is needed */
b127e37e 3191#endif
0c7e610f
JH
3192 RESTORE_ERRNO;
3193 }
3194 while (*s) s++;
bbce6d69 3195 }
79072805 3196 }
4bac9ae4
CS
3197 else if (isGV_with_GP(sv)) {
3198 GV *const gv = MUTABLE_GV(sv);
3199 SV *const buffer = sv_newmortal();
8d1c3e26 3200
4bac9ae4 3201 gv_efullname3(buffer, gv, "*");
180488f8 3202
4bac9ae4
CS
3203 assert(SvPOK(buffer));
3204 if (SvUTF8(buffer))
3205 SvUTF8_on(sv);
3206 if (lp)
3207 *lp = SvCUR(buffer);
3208 return SvPVX(buffer);
3209 }
8d919b0a
FC
3210 else if (isREGEXP(sv)) {
3211 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3212 return RX_WRAPPED((REGEXP *)sv);
3213 }
4bac9ae4 3214 else {
cdb061a3 3215 if (lp)
00b6aa41 3216 *lp = 0;
9f621bb0
NC
3217 if (flags & SV_UNDEF_RETURNS_NULL)
3218 return NULL;
4f62cd62 3219 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
9f621bb0 3220 report_uninit(sv);
4bac9ae4
CS
3221 /* Typically the caller expects that sv_any is not NULL now. */
3222 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
25da4f38 3223 sv_upgrade(sv, SVt_PV);
73d840c0 3224 return (char *)"";
79072805 3225 }
4bac9ae4 3226
cdb061a3 3227 {
823a54a3 3228 const STRLEN len = s - SvPVX_const(sv);
cdb061a3
NC
3229 if (lp)
3230 *lp = len;
3231 SvCUR_set(sv, len);
3232 }
1d7c1841 3233 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3f7c398e 3234 PTR2UV(sv),SvPVX_const(sv)));
4d84ee25
NC
3235 if (flags & SV_CONST_RETURN)
3236 return (char *)SvPVX_const(sv);
10516c54
NC
3237 if (flags & SV_MUTABLE_RETURN)
3238 return SvPVX_mutable(sv);
463ee0b2
LW
3239 return SvPVX(sv);
3240}
3241
645c22ef 3242/*
6050d10e
JP
3243=for apidoc sv_copypv
3244
3245Copies a stringified representation of the source SV into the
3246destination SV. Automatically performs any necessary mg_get and
54f0641b 3247coercion of numeric values into strings. Guaranteed to preserve
2575c402 3248UTF8 flag even from overloaded objects. Similar in nature to
54f0641b
NIS
3249sv_2pv[_flags] but operates directly on an SV instead of just the
3250string. Mostly uses sv_2pv_flags to do its work, except when that
6050d10e
JP
3251would lose the UTF-8'ness of the PV.
3252
4bac9ae4
CS
3253=for apidoc sv_copypv_nomg
3254
3255Like sv_copypv, but doesn't invoke get magic first.
3256
3257=for apidoc sv_copypv_flags
3258
3259Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3260include SV_GMAGIC.
3261
6050d10e
JP
3262=cut
3263*/
3264
3265void
5aaab254 3266Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
4bac9ae4 3267{
446eaa42 3268 STRLEN len;
4bac9ae4 3269 const char *s;
7918f24d 3270
4bac9ae4 3271 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
7918f24d 3272
c77ed9ca 3273 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
cb50f42d 3274 sv_setpvn(dsv,s,len);
446eaa42 3275 if (SvUTF8(ssv))
cb50f42d 3276 SvUTF8_on(dsv);
446eaa42 3277 else
cb50f42d 3278 SvUTF8_off(dsv);
6050d10e
JP
3279}
3280
3281/*
645c22ef
DM
3282=for apidoc sv_2pvbyte
3283
3284Return a pointer to the byte-encoded representation of the SV, and set *lp
1e54db1a 3285to its length. May cause the SV to be downgraded from UTF-8 as a
645c22ef
DM
3286side-effect.
3287
3288Usually accessed via the C<SvPVbyte> macro.
3289
3290=cut
3291*/
3292
7340a771 3293char *
5aaab254 3294Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3295{
7918f24d
NC
3296 PERL_ARGS_ASSERT_SV_2PVBYTE;
3297
48120f8f 3298 SvGETMAGIC(sv);
4499db73
FC
3299 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3300 || isGV_with_GP(sv) || SvROK(sv)) {
a901b181 3301 SV *sv2 = sv_newmortal();
48120f8f 3302 sv_copypv_nomg(sv2,sv);
a901b181
FC
3303 sv = sv2;
3304 }
0875d2fe 3305 sv_utf8_downgrade(sv,0);
71eb6d8c 3306 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771
GS
3307}
3308
645c22ef 3309/*
035cbb0e
RGS
3310=for apidoc sv_2pvutf8
3311
3312Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3313to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3314
3315Usually accessed via the C<SvPVutf8> macro.
3316
3317=cut
3318*/
645c22ef 3319
7340a771 3320char *
5aaab254 3321Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3322{
7918f24d
NC
3323 PERL_ARGS_ASSERT_SV_2PVUTF8;
3324
4499db73
FC
3325 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3326 || isGV_with_GP(sv) || SvROK(sv))
fe46cbda 3327 sv = sv_mortalcopy(sv);
4bac9ae4
CS
3328 else
3329 SvGETMAGIC(sv);
3330 sv_utf8_upgrade_nomg(sv);
c3ec315f 3331 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771 3332}
1c846c1f 3333
7ee2227d 3334
645c22ef
DM
3335/*
3336=for apidoc sv_2bool
3337
06c841cf
FC
3338This macro is only used by sv_true() or its macro equivalent, and only if
3339the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3340It calls sv_2bool_flags with the SV_GMAGIC flag.
3341
3342=for apidoc sv_2bool_flags
3343
3344This function is only used by sv_true() and friends, and only if
fde67290 3345the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
06c841cf
FC
3346contain SV_GMAGIC, then it does an mg_get() first.
3347
645c22ef
DM
3348
3349=cut
3350*/
3351
463ee0b2 3352bool
9d176cd8 3353Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
463ee0b2 3354{
06c841cf 3355 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
7918f24d 3356
9d176cd8 3357 restart:
06c841cf 3358 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
463ee0b2 3359
a0d0e21e
LW
3360 if (!SvOK(sv))
3361 return 0;
3362 if (SvROK(sv)) {
fabdb6c0 3363 if (SvAMAGIC(sv)) {
31d632c3 3364 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
9d176cd8
DD
3365 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3366 bool svb;
3367 sv = tmpsv;
3368 if(SvGMAGICAL(sv)) {
3369 flags = SV_GMAGIC;
3370 goto restart; /* call sv_2bool */
3371 }
3372 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3373 else if(!SvOK(sv)) {
3374 svb = 0;
3375 }
3376 else if(SvPOK(sv)) {
3377 svb = SvPVXtrue(sv);
3378 }
3379 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3380 svb = (SvIOK(sv) && SvIVX(sv) != 0)
659c4b96 3381 || (SvNOK(sv) && SvNVX(sv) != 0.0);
9d176cd8
DD
3382 }
3383 else {
3384 flags = 0;
3385 goto restart; /* call sv_2bool_nomg */
3386 }
3387 return cBOOL(svb);
3388 }
fabdb6c0
AL
3389 }
3390 return SvRV(sv) != 0;
a0d0e21e 3391 }
85b7d9b3
FC
3392 if (isREGEXP(sv))
3393 return
3394 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
4bac9ae4 3395 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
79072805
LW
3396}
3397
c461cf8f
JH
3398/*
3399=for apidoc sv_utf8_upgrade
3400
78ea37eb 3401Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3402Forces the SV to string form if it is not already.
2bbc8d55 3403Will C<mg_get> on C<sv> if appropriate.
4411f3b6 3404Always sets the SvUTF8 flag to avoid future validity checks even
2bbc8d55
SP
3405if the whole string is the same in UTF-8 as not.
3406Returns the number of bytes in the converted string
c461cf8f 3407
0efd0472 3408This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3409use the Encode extension for that.
3410
fe749c9a
KW
3411=for apidoc sv_utf8_upgrade_nomg
3412
fde67290 3413Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
fe749c9a 3414
8d6d96c1
HS
3415=for apidoc sv_utf8_upgrade_flags
3416
78ea37eb 3417Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3418Forces the SV to string form if it is not already.
8d6d96c1 3419Always sets the SvUTF8 flag to avoid future validity checks even
960b0271
FC
3420if all the bytes are invariant in UTF-8.
3421If C<flags> has C<SV_GMAGIC> bit set,
2bbc8d55 3422will C<mg_get> on C<sv> if appropriate, else not.
2a590426
KW
3423
3424If C<flags> has SV_FORCE_UTF8_UPGRADE set, this function assumes that the PV
3425will expand when converted to UTF-8, and skips the extra work of checking for
3426that. Typically this flag is used by a routine that has already parsed the
3427string and found such characters, and passes this information on so that the
3428work doesn't have to be repeated.
3429
3430Returns the number of bytes in the converted string.
8d6d96c1 3431
0efd0472 3432This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3433use the Encode extension for that.
3434
2a590426 3435=for apidoc sv_utf8_upgrade_flags_grow
b3ab6785 3436
2a590426
KW
3437Like sv_utf8_upgrade_flags, but has an additional parameter C<extra>, which is
3438the number of unused bytes the string of 'sv' is guaranteed to have free after
3439it upon return. This allows the caller to reserve extra space that it intends
3440to fill, to avoid extra grows.
b3ab6785 3441
2a590426
KW
3442C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3443are implemented in terms of this function.
3444
3445Returns the number of bytes in the converted string (not including the spares).
3446
3447=cut
b3ab6785
KW
3448
3449(One might think that the calling routine could pass in the position of the
2a590426
KW
3450first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3451have to be found again. But that is not the case, because typically when the
3452caller is likely to use this flag, it won't be calling this routine unless it
3453finds something that won't fit into a byte. Otherwise it tries to not upgrade
3454and just use bytes. But some things that do fit into a byte are variants in
3455utf8, and the caller may not have been keeping track of these.)
b3ab6785 3456
6602b933
KW
3457If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3458C<NUL> isn't guaranteed due to having other routines do the work in some input
3459cases, or if the input is already flagged as being in utf8.
b3ab6785
KW
3460
3461The speed of this could perhaps be improved for many cases if someone wanted to
3462write a fast function that counts the number of variant characters in a string,
3463especially if it could return the position of the first one.
3464
8d6d96c1
HS
3465*/
3466
3467STRLEN
5aaab254 3468Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
8d6d96c1 3469{
b3ab6785 3470 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
7918f24d 3471
808c356f
RGS
3472 if (sv == &PL_sv_undef)
3473 return 0;
892f9127 3474 if (!SvPOK_nog(sv)) {
e0e62c2a 3475 STRLEN len = 0;
d52b7888
NC
3476 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3477 (void) sv_2pv_flags(sv,&len, flags);
b3ab6785
KW
3478 if (SvUTF8(sv)) {
3479 if (extra) SvGROW(sv, SvCUR(sv) + extra);
d52b7888 3480 return len;
b3ab6785 3481 }
d52b7888 3482 } else {
33fb6f35 3483 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
d52b7888 3484 }
e0e62c2a 3485 }
4411f3b6 3486
f5cee72b 3487 if (SvUTF8(sv)) {
b3ab6785 3488 if (extra) SvGROW(sv, SvCUR(sv) + extra);
5fec3b1d 3489 return SvCUR(sv);
f5cee72b 3490 }
5fec3b1d 3491
765f542d 3492 if (SvIsCOW(sv)) {
c56ed9f6 3493 S_sv_uncow(aTHX_ sv, 0);
db42d148
NIS
3494 }
3495
47e13f24 3496 if (IN_ENCODING && !(flags & SV_UTF8_NO_ENCODING)) {
ad2de1b2 3497 sv_recode_to_utf8(sv, _get_encoding());
b3ab6785
KW
3498 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3499 return SvCUR(sv);
3500 }
3501
4e93345f
KW
3502 if (SvCUR(sv) == 0) {
3503 if (extra) SvGROW(sv, extra);
3504 } else { /* Assume Latin-1/EBCDIC */
c4e7c712 3505 /* This function could be much more efficient if we
2bbc8d55 3506 * had a FLAG in SVs to signal if there are any variant
c4e7c712 3507 * chars in the PV. Given that there isn't such a flag
b3ab6785
KW
3508 * make the loop as fast as possible (although there are certainly ways
3509 * to speed this up, eg. through vectorization) */
3510 U8 * s = (U8 *) SvPVX_const(sv);
3511 U8 * e = (U8 *) SvEND(sv);
3512 U8 *t = s;
3513 STRLEN two_byte_count = 0;
c4e7c712 3514
b3ab6785
KW
3515 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3516
3517 /* See if really will need to convert to utf8. We mustn't rely on our
3518 * incoming SV being well formed and having a trailing '\0', as certain
3519 * code in pp_formline can send us partially built SVs. */
3520
c4e7c712 3521 while (t < e) {
53c1dcc0 3522 const U8 ch = *t++;
6f2d5cbc 3523 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
b3ab6785
KW
3524
3525 t--; /* t already incremented; re-point to first variant */
3526 two_byte_count = 1;
3527 goto must_be_utf8;
c4e7c712 3528 }
b3ab6785
KW
3529
3530 /* utf8 conversion not needed because all are invariants. Mark as
3531 * UTF-8 even if no variant - saves scanning loop */
c4e7c712 3532 SvUTF8_on(sv);
7f0bfbea 3533 if (extra) SvGROW(sv, SvCUR(sv) + extra);
b3ab6785
KW
3534 return SvCUR(sv);
3535
7b52d656 3536 must_be_utf8:
b3ab6785
KW
3537
3538 /* Here, the string should be converted to utf8, either because of an
3539 * input flag (two_byte_count = 0), or because a character that
3540 * requires 2 bytes was found (two_byte_count = 1). t points either to
3541 * the beginning of the string (if we didn't examine anything), or to
3542 * the first variant. In either case, everything from s to t - 1 will
3543 * occupy only 1 byte each on output.
3544 *
3545 * There are two main ways to convert. One is to create a new string
3546 * and go through the input starting from the beginning, appending each
3547 * converted value onto the new string as we go along. It's probably
3548 * best to allocate enough space in the string for the worst possible
3549 * case rather than possibly running out of space and having to
3550 * reallocate and then copy what we've done so far. Since everything
3551 * from s to t - 1 is invariant, the destination can be initialized
3552 * with these using a fast memory copy
3553 *
3554 * The other way is to figure out exactly how big the string should be
3555 * by parsing the entire input. Then you don't have to make it big
3556 * enough to handle the worst possible case, and more importantly, if
3557 * the string you already have is large enough, you don't have to
3558 * allocate a new string, you can copy the last character in the input
3559 * string to the final position(s) that will be occupied by the
3560 * converted string and go backwards, stopping at t, since everything
3561 * before that is invariant.
3562 *
3563 * There are advantages and disadvantages to each method.
3564 *
3565 * In the first method, we can allocate a new string, do the memory
3566 * copy from the s to t - 1, and then proceed through the rest of the
3567 * string byte-by-byte.
3568 *
3569 * In the second method, we proceed through the rest of the input
3570 * string just calculating how big the converted string will be. Then
3571 * there are two cases:
3572 * 1) if the string has enough extra space to handle the converted
3573 * value. We go backwards through the string, converting until we
3574 * get to the position we are at now, and then stop. If this
3575 * position is far enough along in the string, this method is
3576 * faster than the other method. If the memory copy were the same
3577 * speed as the byte-by-byte loop, that position would be about
3578 * half-way, as at the half-way mark, parsing to the end and back
3579 * is one complete string's parse, the same amount as starting
3580 * over and going all the way through. Actually, it would be
3581 * somewhat less than half-way, as it's faster to just count bytes
3582 * than to also copy, and we don't have the overhead of allocating
3583 * a new string, changing the scalar to use it, and freeing the
3584 * existing one. But if the memory copy is fast, the break-even
3585 * point is somewhere after half way. The counting loop could be
3586 * sped up by vectorization, etc, to move the break-even point
3587 * further towards the beginning.
3588 * 2) if the string doesn't have enough space to handle the converted
3589 * value. A new string will have to be allocated, and one might
3590 * as well, given that, start from the beginning doing the first
3591 * method. We've spent extra time parsing the string and in
3592 * exchange all we've gotten is that we know precisely how big to
3593 * make the new one. Perl is more optimized for time than space,
3594 * so this case is a loser.
3595 * So what I've decided to do is not use the 2nd method unless it is
3596 * guaranteed that a new string won't have to be allocated, assuming
3597 * the worst case. I also decided not to put any more conditions on it
3598 * than this, for now. It seems likely that, since the worst case is
3599 * twice as big as the unknown portion of the string (plus 1), we won't
3600 * be guaranteed enough space, causing us to go to the first method,
3601 * unless the string is short, or the first variant character is near
3602 * the end of it. In either of these cases, it seems best to use the
3603 * 2nd method. The only circumstance I can think of where this would
3604 * be really slower is if the string had once had much more data in it
3605 * than it does now, but there is still a substantial amount in it */
3606
3607 {
3608 STRLEN invariant_head = t - s;
3609 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3610 if (SvLEN(sv) < size) {
3611
3612 /* Here, have decided to allocate a new string */
3613
3614 U8 *dst;
3615 U8 *d;
3616
3617 Newx(dst, size, U8);
3618