This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
perlapi: Use C<> instead of I<> for parameter names, etc
[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
645c22ef
DM
128/* ============================================================================
129
130=head1 Allocation and deallocation of SVs.
d2a0f284
JC
131An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
132sv, av, hv...) contains type and reference count information, and for
133many types, a pointer to the body (struct xrv, xpv, xpviv...), which
134contains fields specific to each type. Some types store all they need
135in the head, so don't have a body.
136
486ec47a 137In all but the most memory-paranoid configurations (ex: PURIFY), heads
d2a0f284
JC
138and bodies are allocated out of arenas, which by default are
139approximately 4K chunks of memory parcelled up into N heads or bodies.
93e68bfb
JC
140Sv-bodies are allocated by their sv-type, guaranteeing size
141consistency needed to allocate safely from arrays.
142
d2a0f284
JC
143For SV-heads, the first slot in each arena is reserved, and holds a
144link to the next arena, some flags, and a note of the number of slots.
145Snaked through each arena chain is a linked list of free items; when
146this becomes empty, an extra arena is allocated and divided up into N
147items which are threaded into the free list.
148
149SV-bodies are similar, but they use arena-sets by default, which
150separate the link and info from the arena itself, and reclaim the 1st
151slot in the arena. SV-bodies are further described later.
645c22ef
DM
152
153The following global variables are associated with arenas:
154
7fefc6c1
KW
155 PL_sv_arenaroot pointer to list of SV arenas
156 PL_sv_root pointer to list of free SV structures
645c22ef 157
7fefc6c1
KW
158 PL_body_arenas head of linked-list of body arenas
159 PL_body_roots[] array of pointers to list of free bodies of svtype
160 arrays are indexed by the svtype needed
93e68bfb 161
d2a0f284
JC
162A few special SV heads are not allocated from an arena, but are
163instead directly created in the interpreter structure, eg PL_sv_undef.
93e68bfb
JC
164The size of arenas can be changed from the default by setting
165PERL_ARENA_SIZE appropriately at compile time.
645c22ef
DM
166
167The SV arena serves the secondary purpose of allowing still-live SVs
168to be located and destroyed during final cleanup.
169
170At the lowest level, the macros new_SV() and del_SV() grab and free
171an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
172to return the SV to the free list with error checking.) new_SV() calls
173more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
174SVs in the free list have their SvTYPE field set to all ones.
175
ff276b08 176At the time of very final cleanup, sv_free_arenas() is called from
645c22ef 177perl_destruct() to physically free all the arenas allocated since the
6a93a7e5 178start of the interpreter.
645c22ef 179
645c22ef
DM
180The function visit() scans the SV arenas list, and calls a specified
181function for each SV it finds which is still live - ie which has an SvTYPE
182other than all 1's, and a non-zero SvREFCNT. visit() is used by the
183following functions (specified as [function that calls visit()] / [function
184called by visit() for each SV]):
185
186 sv_report_used() / do_report_used()
f2524eef 187 dump all remaining SVs (debugging aid)
645c22ef 188
e4487e9b 189 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
e76981f9 190 do_clean_named_io_objs(),do_curse()
645c22ef 191 Attempt to free all objects pointed to by RVs,
e76981f9
FC
192 try to do the same for all objects indir-
193 ectly referenced by typeglobs too, and
194 then do a final sweep, cursing any
195 objects that remain. Called once from
645c22ef
DM
196 perl_destruct(), prior to calling sv_clean_all()
197 below.
198
199 sv_clean_all() / do_clean_all()
200 SvREFCNT_dec(sv) each remaining SV, possibly
201 triggering an sv_free(). It also sets the
202 SVf_BREAK flag on the SV to indicate that the
203 refcnt has been artificially lowered, and thus
204 stopping sv_free() from giving spurious warnings
205 about SVs which unexpectedly have a refcnt
206 of zero. called repeatedly from perl_destruct()
207 until there are no SVs left.
208
93e68bfb 209=head2 Arena allocator API Summary
645c22ef
DM
210
211Private API to rest of sv.c
212
213 new_SV(), del_SV(),
214
df0f0429 215 new_XPVNV(), del_XPVGV(),
645c22ef
DM
216 etc
217
218Public API:
219
8cf8f3d1 220 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
645c22ef 221
645c22ef
DM
222=cut
223
3e8320cc 224 * ========================================================================= */
645c22ef 225
4561caa4
CS
226/*
227 * "A time to plant, and a time to uproot what was planted..."
228 */
229
d7a2c63c
MHM
230#ifdef PERL_MEM_LOG
231# define MEM_LOG_NEW_SV(sv, file, line, func) \
232 Perl_mem_log_new_sv(sv, file, line, func)
233# define MEM_LOG_DEL_SV(sv, file, line, func) \
234 Perl_mem_log_del_sv(sv, file, line, func)
235#else
236# define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
237# define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
238#endif
239
fd0854ff 240#ifdef DEBUG_LEAKING_SCALARS
484e6108
FC
241# define FREE_SV_DEBUG_FILE(sv) STMT_START { \
242 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
243 } STMT_END
d7a2c63c
MHM
244# define DEBUG_SV_SERIAL(sv) \
245 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
246 PTR2UV(sv), (long)(sv)->sv_debug_serial))
fd0854ff
DM
247#else
248# define FREE_SV_DEBUG_FILE(sv)
d7a2c63c 249# define DEBUG_SV_SERIAL(sv) NOOP
fd0854ff
DM
250#endif
251
48614a46
NC
252#ifdef PERL_POISON
253# define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
daba3364 254# define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
48614a46
NC
255/* Whilst I'd love to do this, it seems that things like to check on
256 unreferenced scalars
ce5dbf61 257# define POISON_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
48614a46 258*/
ce5dbf61 259# define POISON_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
7e337ee0 260 PoisonNew(&SvREFCNT(sv), 1, U32)
48614a46
NC
261#else
262# define SvARENA_CHAIN(sv) SvANY(sv)
3eef1deb 263# define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
ce5dbf61 264# define POISON_SV_HEAD(sv)
48614a46
NC
265#endif
266
990198f0
DM
267/* Mark an SV head as unused, and add to free list.
268 *
269 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
270 * its refcount artificially decremented during global destruction, so
271 * there may be dangling pointers to it. The last thing we want in that
272 * case is for it to be reused. */
273
053fc874
GS
274#define plant_SV(p) \
275 STMT_START { \
990198f0 276 const U32 old_flags = SvFLAGS(p); \
d7a2c63c
MHM
277 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
278 DEBUG_SV_SERIAL(p); \
fd0854ff 279 FREE_SV_DEBUG_FILE(p); \
ce5dbf61 280 POISON_SV_HEAD(p); \
053fc874 281 SvFLAGS(p) = SVTYPEMASK; \
990198f0 282 if (!(old_flags & SVf_BREAK)) { \
3eef1deb 283 SvARENA_CHAIN_SET(p, PL_sv_root); \
990198f0
DM
284 PL_sv_root = (p); \
285 } \
053fc874
GS
286 --PL_sv_count; \
287 } STMT_END
a0d0e21e 288
053fc874
GS
289#define uproot_SV(p) \
290 STMT_START { \
291 (p) = PL_sv_root; \
daba3364 292 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
053fc874
GS
293 ++PL_sv_count; \
294 } STMT_END
295
645c22ef 296
cac9b346
NC
297/* make some more SVs by adding another arena */
298
cac9b346
NC
299STATIC SV*
300S_more_sv(pTHX)
301{
302 SV* sv;
9a87bd09
NC
303 char *chunk; /* must use New here to match call to */
304 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
305 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
cac9b346
NC
306 uproot_SV(sv);
307 return sv;
308}
309
645c22ef
DM
310/* new_SV(): return a new, empty SV head */
311
eba0f806
DM
312#ifdef DEBUG_LEAKING_SCALARS
313/* provide a real function for a debugger to play with */
314STATIC SV*
d7a2c63c 315S_new_SV(pTHX_ const char *file, int line, const char *func)
eba0f806
DM
316{
317 SV* sv;
318
eba0f806
DM
319 if (PL_sv_root)
320 uproot_SV(sv);
321 else
cac9b346 322 sv = S_more_sv(aTHX);
eba0f806
DM
323 SvANY(sv) = 0;
324 SvREFCNT(sv) = 1;
325 SvFLAGS(sv) = 0;
fd0854ff 326 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
e385c3bf
DM
327 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
328 ? PL_parser->copline
329 : PL_curcop
f24aceb1
DM
330 ? CopLINE(PL_curcop)
331 : 0
e385c3bf 332 );
fd0854ff 333 sv->sv_debug_inpad = 0;
cd676548 334 sv->sv_debug_parent = NULL;
484e6108 335 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
d7a2c63c
MHM
336
337 sv->sv_debug_serial = PL_sv_serial++;
338
339 MEM_LOG_NEW_SV(sv, file, line, func);
340 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
341 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
342
eba0f806
DM
343 return sv;
344}
d7a2c63c 345# define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
eba0f806
DM
346
347#else
348# define new_SV(p) \
053fc874 349 STMT_START { \
053fc874
GS
350 if (PL_sv_root) \
351 uproot_SV(p); \
352 else \
cac9b346 353 (p) = S_more_sv(aTHX); \
053fc874
GS
354 SvANY(p) = 0; \
355 SvREFCNT(p) = 1; \
356 SvFLAGS(p) = 0; \
d7a2c63c 357 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
053fc874 358 } STMT_END
eba0f806 359#endif
463ee0b2 360
645c22ef
DM
361
362/* del_SV(): return an empty SV head to the free list */
363
a0d0e21e 364#ifdef DEBUGGING
4561caa4 365
053fc874
GS
366#define del_SV(p) \
367 STMT_START { \
aea4f609 368 if (DEBUG_D_TEST) \
053fc874
GS
369 del_sv(p); \
370 else \
371 plant_SV(p); \
053fc874 372 } STMT_END
a0d0e21e 373
76e3520e 374STATIC void
cea2e8a9 375S_del_sv(pTHX_ SV *p)
463ee0b2 376{
7918f24d
NC
377 PERL_ARGS_ASSERT_DEL_SV;
378
aea4f609 379 if (DEBUG_D_TEST) {
4633a7c4 380 SV* sva;
a3b680e6 381 bool ok = 0;
daba3364 382 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
53c1dcc0
AL
383 const SV * const sv = sva + 1;
384 const SV * const svend = &sva[SvREFCNT(sva)];
c0ff570e 385 if (p >= sv && p < svend) {
a0d0e21e 386 ok = 1;
c0ff570e
NC
387 break;
388 }
a0d0e21e
LW
389 }
390 if (!ok) {
9b387841
NC
391 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
392 "Attempt to free non-arena SV: 0x%"UVxf
393 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
a0d0e21e
LW
394 return;
395 }
396 }
4561caa4 397 plant_SV(p);
463ee0b2 398}
a0d0e21e 399
4561caa4
CS
400#else /* ! DEBUGGING */
401
402#define del_SV(p) plant_SV(p)
403
404#endif /* DEBUGGING */
463ee0b2 405
dc6369ef
EH
406/*
407 * Bodyless IVs and NVs!
408 *
409 * Since 5.9.2, we can avoid allocating a body for SVt_IV-type SVs.
410 * Since the larger IV-holding variants of SVs store their integer
411 * values in their respective bodies, the family of SvIV() accessor
412 * macros would naively have to branch on the SV type to find the
413 * integer value either in the HEAD or BODY. In order to avoid this
414 * expensive branch, a clever soul has deployed a great hack:
415 * We set up the SvANY pointer such that instead of pointing to a
416 * real body, it points into the memory before the location of the
417 * head. We compute this pointer such that the location of
418 * the integer member of the hypothetical body struct happens to
419 * be the same as the location of the integer member of the bodyless
420 * SV head. This now means that the SvIV() family of accessors can
421 * always read from the (hypothetical or real) body via SvANY.
422 *
423 * Since the 5.21 dev series, we employ the same trick for NVs
424 * if the architecture can support it (NVSIZE <= IVSIZE).
425 */
426
427/* The following two macros compute the necessary offsets for the above
428 * trick and store them in SvANY for SvIV() (and friends) to use. */
429#define SET_SVANY_FOR_BODYLESS_IV(sv) \
430 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv))
431
432#define SET_SVANY_FOR_BODYLESS_NV(sv) \
433 SvANY(sv) = (XPVNV*)((char*)&(sv->sv_u.svu_nv) - STRUCT_OFFSET(XPVNV, xnv_u.xnv_nv))
645c22ef
DM
434
435/*
ccfc67b7
JH
436=head1 SV Manipulation Functions
437
645c22ef
DM
438=for apidoc sv_add_arena
439
440Given a chunk of memory, link it to the head of the list of arenas,
441and split it into a list of free SVs.
442
443=cut
444*/
445
d2bd4e7f
NC
446static void
447S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
463ee0b2 448{
daba3364 449 SV *const sva = MUTABLE_SV(ptr);
eb578fdb
KW
450 SV* sv;
451 SV* svend;
4633a7c4 452
7918f24d
NC
453 PERL_ARGS_ASSERT_SV_ADD_ARENA;
454
4633a7c4 455 /* The first SV in an arena isn't an SV. */
3280af22 456 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
4633a7c4
LW
457 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
458 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
459
3280af22
NIS
460 PL_sv_arenaroot = sva;
461 PL_sv_root = sva + 1;
4633a7c4
LW
462
463 svend = &sva[SvREFCNT(sva) - 1];
464 sv = sva + 1;
463ee0b2 465 while (sv < svend) {
3eef1deb 466 SvARENA_CHAIN_SET(sv, (sv + 1));
03e36789 467#ifdef DEBUGGING
978b032e 468 SvREFCNT(sv) = 0;
03e36789 469#endif
4b69cbe3 470 /* Must always set typemask because it's always checked in on cleanup
03e36789 471 when the arenas are walked looking for objects. */
8990e307 472 SvFLAGS(sv) = SVTYPEMASK;
463ee0b2
LW
473 sv++;
474 }
3eef1deb 475 SvARENA_CHAIN_SET(sv, 0);
03e36789
NC
476#ifdef DEBUGGING
477 SvREFCNT(sv) = 0;
478#endif
4633a7c4
LW
479 SvFLAGS(sv) = SVTYPEMASK;
480}
481
055972dc
DM
482/* visit(): call the named function for each non-free SV in the arenas
483 * whose flags field matches the flags/mask args. */
645c22ef 484
5226ed68 485STATIC I32
de37a194 486S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
8990e307 487{
4633a7c4 488 SV* sva;
5226ed68 489 I32 visited = 0;
8990e307 490
7918f24d
NC
491 PERL_ARGS_ASSERT_VISIT;
492
daba3364 493 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
eb578fdb
KW
494 const SV * const svend = &sva[SvREFCNT(sva)];
495 SV* sv;
4561caa4 496 for (sv = sva + 1; sv < svend; ++sv) {
e4787c0c 497 if (SvTYPE(sv) != (svtype)SVTYPEMASK
055972dc
DM
498 && (sv->sv_flags & mask) == flags
499 && SvREFCNT(sv))
500 {
942481a7 501 (*f)(aTHX_ sv);
5226ed68
JH
502 ++visited;
503 }
8990e307
LW
504 }
505 }
5226ed68 506 return visited;
8990e307
LW
507}
508
758a08c3
JH
509#ifdef DEBUGGING
510
645c22ef
DM
511/* called by sv_report_used() for each live SV */
512
513static void
5fa45a31 514do_report_used(pTHX_ SV *const sv)
645c22ef 515{
e4787c0c 516 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
645c22ef
DM
517 PerlIO_printf(Perl_debug_log, "****\n");
518 sv_dump(sv);
519 }
520}
758a08c3 521#endif
645c22ef
DM
522
523/*
524=for apidoc sv_report_used
525
fde67290 526Dump the contents of all SVs not yet freed (debugging aid).
645c22ef
DM
527
528=cut
529*/
530
8990e307 531void
864dbfa3 532Perl_sv_report_used(pTHX)
4561caa4 533{
ff270d3a 534#ifdef DEBUGGING
055972dc 535 visit(do_report_used, 0, 0);
96a5add6
AL
536#else
537 PERL_UNUSED_CONTEXT;
ff270d3a 538#endif
4561caa4
CS
539}
540
645c22ef
DM
541/* called by sv_clean_objs() for each live SV */
542
543static void
de37a194 544do_clean_objs(pTHX_ SV *const ref)
645c22ef 545{
ea724faa
NC
546 assert (SvROK(ref));
547 {
823a54a3
AL
548 SV * const target = SvRV(ref);
549 if (SvOBJECT(target)) {
550 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
551 if (SvWEAKREF(ref)) {
552 sv_del_backref(target, ref);
553 SvWEAKREF_off(ref);
554 SvRV_set(ref, NULL);
555 } else {
556 SvROK_off(ref);
557 SvRV_set(ref, NULL);
fc2b2dca 558 SvREFCNT_dec_NN(target);
823a54a3 559 }
645c22ef
DM
560 }
561 }
645c22ef
DM
562}
563
645c22ef 564
e4487e9b
DM
565/* clear any slots in a GV which hold objects - except IO;
566 * called by sv_clean_objs() for each live GV */
567
645c22ef 568static void
f30de749 569do_clean_named_objs(pTHX_ SV *const sv)
645c22ef 570{
57ef47cc 571 SV *obj;
ea724faa 572 assert(SvTYPE(sv) == SVt_PVGV);
d011219a 573 assert(isGV_with_GP(sv));
57ef47cc
DM
574 if (!GvGP(sv))
575 return;
576
577 /* freeing GP entries may indirectly free the current GV;
578 * hold onto it while we mess with the GP slots */
579 SvREFCNT_inc(sv);
580
581 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
582 DEBUG_D((PerlIO_printf(Perl_debug_log,
583 "Cleaning named glob SV object:\n "), sv_dump(obj)));
584 GvSV(sv) = NULL;
fc2b2dca 585 SvREFCNT_dec_NN(obj);
57ef47cc
DM
586 }
587 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
588 DEBUG_D((PerlIO_printf(Perl_debug_log,
589 "Cleaning named glob AV object:\n "), sv_dump(obj)));
590 GvAV(sv) = NULL;
fc2b2dca 591 SvREFCNT_dec_NN(obj);
57ef47cc
DM
592 }
593 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
594 DEBUG_D((PerlIO_printf(Perl_debug_log,
595 "Cleaning named glob HV object:\n "), sv_dump(obj)));
596 GvHV(sv) = NULL;
fc2b2dca 597 SvREFCNT_dec_NN(obj);
57ef47cc
DM
598 }
599 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
600 DEBUG_D((PerlIO_printf(Perl_debug_log,
601 "Cleaning named glob CV object:\n "), sv_dump(obj)));
c43ae56f 602 GvCV_set(sv, NULL);
fc2b2dca 603 SvREFCNT_dec_NN(obj);
57ef47cc 604 }
fc2b2dca 605 SvREFCNT_dec_NN(sv); /* undo the inc above */
e4487e9b
DM
606}
607
68b590d9 608/* clear any IO slots in a GV which hold objects (except stderr, defout);
e4487e9b
DM
609 * called by sv_clean_objs() for each live GV */
610
611static void
612do_clean_named_io_objs(pTHX_ SV *const sv)
613{
e4487e9b
DM
614 SV *obj;
615 assert(SvTYPE(sv) == SVt_PVGV);
616 assert(isGV_with_GP(sv));
68b590d9 617 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
e4487e9b
DM
618 return;
619
620 SvREFCNT_inc(sv);
57ef47cc
DM
621 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
622 DEBUG_D((PerlIO_printf(Perl_debug_log,
623 "Cleaning named glob IO object:\n "), sv_dump(obj)));
624 GvIOp(sv) = NULL;
fc2b2dca 625 SvREFCNT_dec_NN(obj);
645c22ef 626 }
fc2b2dca 627 SvREFCNT_dec_NN(sv); /* undo the inc above */
645c22ef 628}
645c22ef 629
4155e4fe
FC
630/* Void wrapper to pass to visit() */
631static void
632do_curse(pTHX_ SV * const sv) {
c2910e6c
FC
633 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
634 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
4155e4fe
FC
635 return;
636 (void)curse(sv, 0);
637}
638
645c22ef
DM
639/*
640=for apidoc sv_clean_objs
641
fde67290 642Attempt to destroy all objects not yet freed.
645c22ef
DM
643
644=cut
645*/
646
4561caa4 647void
864dbfa3 648Perl_sv_clean_objs(pTHX)
4561caa4 649{
68b590d9 650 GV *olddef, *olderr;
3280af22 651 PL_in_clean_objs = TRUE;
055972dc 652 visit(do_clean_objs, SVf_ROK, SVf_ROK);
e4487e9b
DM
653 /* Some barnacles may yet remain, clinging to typeglobs.
654 * Run the non-IO destructors first: they may want to output
655 * error messages, close files etc */
d011219a 656 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
e4487e9b 657 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
4155e4fe
FC
658 /* And if there are some very tenacious barnacles clinging to arrays,
659 closures, or what have you.... */
660 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
68b590d9
DM
661 olddef = PL_defoutgv;
662 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
663 if (olddef && isGV_with_GP(olddef))
664 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
665 olderr = PL_stderrgv;
666 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
667 if (olderr && isGV_with_GP(olderr))
668 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
669 SvREFCNT_dec(olddef);
3280af22 670 PL_in_clean_objs = FALSE;
4561caa4
CS
671}
672
645c22ef
DM
673/* called by sv_clean_all() for each live SV */
674
675static void
de37a194 676do_clean_all(pTHX_ SV *const sv)
645c22ef 677{
daba3364 678 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
cddfcddc 679 /* don't clean pid table and strtab */
d17ea597 680 return;
cddfcddc 681 }
645c22ef
DM
682 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
683 SvFLAGS(sv) |= SVf_BREAK;
fc2b2dca 684 SvREFCNT_dec_NN(sv);
645c22ef
DM
685}
686
687/*
688=for apidoc sv_clean_all
689
690Decrement the refcnt of each remaining SV, possibly triggering a
fde67290 691cleanup. This function may have to be called multiple times to free
ff276b08 692SVs which are in complex self-referential hierarchies.
645c22ef
DM
693
694=cut
695*/
696
5226ed68 697I32
864dbfa3 698Perl_sv_clean_all(pTHX)
8990e307 699{
5226ed68 700 I32 cleaned;
3280af22 701 PL_in_clean_all = TRUE;
055972dc 702 cleaned = visit(do_clean_all, 0,0);
5226ed68 703 return cleaned;
8990e307 704}
463ee0b2 705
5e258f8c
JC
706/*
707 ARENASETS: a meta-arena implementation which separates arena-info
708 into struct arena_set, which contains an array of struct
709 arena_descs, each holding info for a single arena. By separating
710 the meta-info from the arena, we recover the 1st slot, formerly
711 borrowed for list management. The arena_set is about the size of an
39244528 712 arena, avoiding the needless malloc overhead of a naive linked-list.
5e258f8c
JC
713
714 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
715 memory in the last arena-set (1/2 on average). In trade, we get
716 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
d2a0f284 717 smaller types). The recovery of the wasted space allows use of
e15dad31
JC
718 small arenas for large, rare body types, by changing array* fields
719 in body_details_by_type[] below.
5e258f8c 720*/
5e258f8c 721struct arena_desc {
398c677b
NC
722 char *arena; /* the raw storage, allocated aligned */
723 size_t size; /* its size ~4k typ */
e5973ed5 724 svtype utype; /* bodytype stored in arena */
5e258f8c
JC
725};
726
e6148039
NC
727struct arena_set;
728
729/* Get the maximum number of elements in set[] such that struct arena_set
e15dad31 730 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
e6148039
NC
731 therefore likely to be 1 aligned memory page. */
732
733#define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
734 - 2 * sizeof(int)) / sizeof (struct arena_desc))
5e258f8c
JC
735
736struct arena_set {
737 struct arena_set* next;
0a848332
NC
738 unsigned int set_size; /* ie ARENAS_PER_SET */
739 unsigned int curr; /* index of next available arena-desc */
5e258f8c
JC
740 struct arena_desc set[ARENAS_PER_SET];
741};
742
645c22ef
DM
743/*
744=for apidoc sv_free_arenas
745
fde67290 746Deallocate the memory used by all arenas. Note that all the individual SV
645c22ef
DM
747heads and bodies within the arenas must already have been freed.
748
749=cut
7fefc6c1 750
645c22ef 751*/
4633a7c4 752void
864dbfa3 753Perl_sv_free_arenas(pTHX)
4633a7c4
LW
754{
755 SV* sva;
756 SV* svanext;
0a848332 757 unsigned int i;
4633a7c4
LW
758
759 /* Free arenas here, but be careful about fake ones. (We assume
760 contiguity of the fake ones with the corresponding real ones.) */
761
3280af22 762 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
daba3364 763 svanext = MUTABLE_SV(SvANY(sva));
4633a7c4 764 while (svanext && SvFAKE(svanext))
daba3364 765 svanext = MUTABLE_SV(SvANY(svanext));
4633a7c4
LW
766
767 if (!SvFAKE(sva))
1df70142 768 Safefree(sva);
4633a7c4 769 }
93e68bfb 770
5e258f8c 771 {
0a848332
NC
772 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
773
774 while (aroot) {
775 struct arena_set *current = aroot;
776 i = aroot->curr;
777 while (i--) {
5e258f8c
JC
778 assert(aroot->set[i].arena);
779 Safefree(aroot->set[i].arena);
780 }
0a848332
NC
781 aroot = aroot->next;
782 Safefree(current);
5e258f8c
JC
783 }
784 }
dc8220bf 785 PL_body_arenas = 0;
fdda85ca 786
0a848332
NC
787 i = PERL_ARENA_ROOTS_SIZE;
788 while (i--)
93e68bfb 789 PL_body_roots[i] = 0;
93e68bfb 790
3280af22
NIS
791 PL_sv_arenaroot = 0;
792 PL_sv_root = 0;
4633a7c4
LW
793}
794
bd81e77b
NC
795/*
796 Here are mid-level routines that manage the allocation of bodies out
797 of the various arenas. There are 5 kinds of arenas:
29489e7c 798
bd81e77b
NC
799 1. SV-head arenas, which are discussed and handled above
800 2. regular body arenas
801 3. arenas for reduced-size bodies
802 4. Hash-Entry arenas
29489e7c 803
bd81e77b
NC
804 Arena types 2 & 3 are chained by body-type off an array of
805 arena-root pointers, which is indexed by svtype. Some of the
806 larger/less used body types are malloced singly, since a large
807 unused block of them is wasteful. Also, several svtypes dont have
808 bodies; the data fits into the sv-head itself. The arena-root
809 pointer thus has a few unused root-pointers (which may be hijacked
810 later for arena types 4,5)
29489e7c 811
bd81e77b
NC
812 3 differs from 2 as an optimization; some body types have several
813 unused fields in the front of the structure (which are kept in-place
814 for consistency). These bodies can be allocated in smaller chunks,
815 because the leading fields arent accessed. Pointers to such bodies
816 are decremented to point at the unused 'ghost' memory, knowing that
817 the pointers are used with offsets to the real memory.
29489e7c 818
d2a0f284
JC
819
820=head1 SV-Body Allocation
821
f554dfc5
MH
822=cut
823
d2a0f284
JC
824Allocation of SV-bodies is similar to SV-heads, differing as follows;
825the allocation mechanism is used for many body types, so is somewhat
826more complicated, it uses arena-sets, and has no need for still-live
827SV detection.
828
829At the outermost level, (new|del)_X*V macros return bodies of the
830appropriate type. These macros call either (new|del)_body_type or
831(new|del)_body_allocated macro pairs, depending on specifics of the
832type. Most body types use the former pair, the latter pair is used to
833allocate body types with "ghost fields".
834
835"ghost fields" are fields that are unused in certain types, and
69ba284b 836consequently don't need to actually exist. They are declared because
d2a0f284
JC
837they're part of a "base type", which allows use of functions as
838methods. The simplest examples are AVs and HVs, 2 aggregate types
839which don't use the fields which support SCALAR semantics.
840
69ba284b 841For these types, the arenas are carved up into appropriately sized
d2a0f284
JC
842chunks, we thus avoid wasted memory for those unaccessed members.
843When bodies are allocated, we adjust the pointer back in memory by the
69ba284b 844size of the part not allocated, so it's as if we allocated the full
d2a0f284
JC
845structure. (But things will all go boom if you write to the part that
846is "not there", because you'll be overwriting the last members of the
847preceding structure in memory.)
848
69ba284b 849We calculate the correction using the STRUCT_OFFSET macro on the first
a05ea1cf 850member present. If the allocated structure is smaller (no initial NV
69ba284b
NC
851actually allocated) then the net effect is to subtract the size of the NV
852from the pointer, to return a new pointer as if an initial NV were actually
a05ea1cf 853allocated. (We were using structures named *_allocated for this, but
69ba284b
NC
854this turned out to be a subtle bug, because a structure without an NV
855could have a lower alignment constraint, but the compiler is allowed to
856optimised accesses based on the alignment constraint of the actual pointer
857to the full structure, for example, using a single 64 bit load instruction
858because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
d2a0f284 859
a05ea1cf 860This is the same trick as was used for NV and IV bodies. Ironically it
d2a0f284 861doesn't need to be used for NV bodies any more, because NV is now at
5b306eef
DD
862the start of the structure. IV bodies, and also in some builds NV bodies,
863don't need it either, because they are no longer allocated.
d2a0f284
JC
864
865In turn, the new_body_* allocators call S_new_body(), which invokes
866new_body_inline macro, which takes a lock, and takes a body off the
1e30fcd5 867linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
d2a0f284
JC
868necessary to refresh an empty list. Then the lock is released, and
869the body is returned.
870
99816f8d 871Perl_more_bodies allocates a new arena, and carves it up into an array of N
d2a0f284
JC
872bodies, which it strings into a linked list. It looks up arena-size
873and body-size from the body_details table described below, thus
874supporting the multiple body-types.
875
876If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
877the (new|del)_X*V macros are mapped directly to malloc/free.
878
d2a0f284
JC
879For each sv-type, struct body_details bodies_by_type[] carries
880parameters which control these aspects of SV handling:
881
882Arena_size determines whether arenas are used for this body type, and if
883so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
884zero, forcing individual mallocs and frees.
885
886Body_size determines how big a body is, and therefore how many fit into
887each arena. Offset carries the body-pointer adjustment needed for
69ba284b 888"ghost fields", and is used in *_allocated macros.
d2a0f284
JC
889
890But its main purpose is to parameterize info needed in
891Perl_sv_upgrade(). The info here dramatically simplifies the function
69ba284b 892vs the implementation in 5.8.8, making it table-driven. All fields
d2a0f284
JC
893are used for this, except for arena_size.
894
895For the sv-types that have no bodies, arenas are not used, so those
896PL_body_roots[sv_type] are unused, and can be overloaded. In
897something of a special case, SVt_NULL is borrowed for HE arenas;
c6f8b1d0 898PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
d2a0f284 899bodies_by_type[SVt_NULL] slot is not used, as the table is not
c6f8b1d0 900available in hv.c.
d2a0f284 901
29489e7c
DM
902*/
903
bd81e77b 904struct body_details {
0fb58b32 905 U8 body_size; /* Size to allocate */
10666ae3 906 U8 copy; /* Size of structure to copy (may be shorter) */
5b306eef
DD
907 U8 offset; /* Size of unalloced ghost fields to first alloced field*/
908 PERL_BITFIELD8 type : 4; /* We have space for a sanity check. */
909 PERL_BITFIELD8 cant_upgrade : 1;/* Cannot upgrade this type */
910 PERL_BITFIELD8 zero_nv : 1; /* zero the NV when upgrading from this */
911 PERL_BITFIELD8 arena : 1; /* Allocated from an arena */
912 U32 arena_size; /* Size of arena to allocate */
bd81e77b 913};
29489e7c 914
bd81e77b
NC
915#define HADNV FALSE
916#define NONV TRUE
29489e7c 917
d2a0f284 918
bd81e77b
NC
919#ifdef PURIFY
920/* With -DPURFIY we allocate everything directly, and don't use arenas.
921 This seems a rather elegant way to simplify some of the code below. */
922#define HASARENA FALSE
923#else
924#define HASARENA TRUE
925#endif
926#define NOARENA FALSE
29489e7c 927
d2a0f284
JC
928/* Size the arenas to exactly fit a given number of bodies. A count
929 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
930 simplifying the default. If count > 0, the arena is sized to fit
931 only that many bodies, allowing arenas to be used for large, rare
932 bodies (XPVFM, XPVIO) without undue waste. The arena size is
933 limited by PERL_ARENA_SIZE, so we can safely oversize the
934 declarations.
935 */
95db5f15
MB
936#define FIT_ARENA0(body_size) \
937 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
938#define FIT_ARENAn(count,body_size) \
939 ( count * body_size <= PERL_ARENA_SIZE) \
940 ? count * body_size \
941 : FIT_ARENA0 (body_size)
942#define FIT_ARENA(count,body_size) \
cd1dc8e2 943 (U32)(count \
95db5f15 944 ? FIT_ARENAn (count, body_size) \
cd1dc8e2 945 : FIT_ARENA0 (body_size))
d2a0f284 946
bd81e77b
NC
947/* Calculate the length to copy. Specifically work out the length less any
948 final padding the compiler needed to add. See the comment in sv_upgrade
949 for why copying the padding proved to be a bug. */
29489e7c 950
bd81e77b
NC
951#define copy_length(type, last_member) \
952 STRUCT_OFFSET(type, last_member) \
daba3364 953 + sizeof (((type*)SvANY((const SV *)0))->last_member)
29489e7c 954
bd81e77b 955static const struct body_details bodies_by_type[] = {
829cd18a
NC
956 /* HEs use this offset for their arena. */
957 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
d2a0f284 958
db93c0c4
NC
959 /* IVs are in the head, so the allocation size is 0. */
960 { 0,
d2a0f284 961 sizeof(IV), /* This is used to copy out the IV body. */
10666ae3 962 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
db93c0c4 963 NOARENA /* IVS don't need an arena */, 0
d2a0f284
JC
964 },
965
5b306eef
DD
966#if NVSIZE <= IVSIZE
967 { 0, sizeof(NV),
968 STRUCT_OFFSET(XPVNV, xnv_u),
969 SVt_NV, FALSE, HADNV, NOARENA, 0 },
970#else
6e128786
NC
971 { sizeof(NV), sizeof(NV),
972 STRUCT_OFFSET(XPVNV, xnv_u),
973 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
5b306eef 974#endif
d2a0f284 975
bc337e5c 976 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
977 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
978 + STRUCT_OFFSET(XPV, xpv_cur),
69ba284b 979 SVt_PV, FALSE, NONV, HASARENA,
889d28b2 980 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 981
d361b004
KW
982 { sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur),
983 copy_length(XINVLIST, is_offset) - STRUCT_OFFSET(XPV, xpv_cur),
984 + STRUCT_OFFSET(XPV, xpv_cur),
985 SVt_INVLIST, TRUE, NONV, HASARENA,
986 FIT_ARENA(0, sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur)) },
e94d9b54 987
bc337e5c 988 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
989 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
990 + STRUCT_OFFSET(XPV, xpv_cur),
991 SVt_PVIV, FALSE, NONV, HASARENA,
992 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 993
bc337e5c 994 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
889d28b2
NC
995 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
996 + STRUCT_OFFSET(XPV, xpv_cur),
997 SVt_PVNV, FALSE, HADNV, HASARENA,
998 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
d2a0f284 999
6e128786 1000 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
d2a0f284 1001 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
4df7f6af 1002
601dfd0a
NC
1003 { sizeof(regexp),
1004 sizeof(regexp),
1005 0,
ecff11eb 1006 SVt_REGEXP, TRUE, NONV, HASARENA,
eaeb1e7f 1007 FIT_ARENA(0, sizeof(regexp))
5c35adbb 1008 },
4df7f6af 1009
10666ae3 1010 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
d2a0f284
JC
1011 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
1012
10666ae3 1013 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
d2a0f284
JC
1014 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
1015
601dfd0a 1016 { sizeof(XPVAV),
4f7003f5 1017 copy_length(XPVAV, xav_alloc),
601dfd0a 1018 0,
69ba284b 1019 SVt_PVAV, TRUE, NONV, HASARENA,
601dfd0a 1020 FIT_ARENA(0, sizeof(XPVAV)) },
d2a0f284 1021
601dfd0a 1022 { sizeof(XPVHV),
359164a0 1023 copy_length(XPVHV, xhv_max),
601dfd0a 1024 0,
69ba284b 1025 SVt_PVHV, TRUE, NONV, HASARENA,
601dfd0a 1026 FIT_ARENA(0, sizeof(XPVHV)) },
d2a0f284 1027
601dfd0a
NC
1028 { sizeof(XPVCV),
1029 sizeof(XPVCV),
1030 0,
69ba284b 1031 SVt_PVCV, TRUE, NONV, HASARENA,
601dfd0a 1032 FIT_ARENA(0, sizeof(XPVCV)) },
69ba284b 1033
601dfd0a
NC
1034 { sizeof(XPVFM),
1035 sizeof(XPVFM),
1036 0,
69ba284b 1037 SVt_PVFM, TRUE, NONV, NOARENA,
601dfd0a 1038 FIT_ARENA(20, sizeof(XPVFM)) },
d2a0f284 1039
601dfd0a
NC
1040 { sizeof(XPVIO),
1041 sizeof(XPVIO),
1042 0,
b6f60916 1043 SVt_PVIO, TRUE, NONV, HASARENA,
601dfd0a 1044 FIT_ARENA(24, sizeof(XPVIO)) },
bd81e77b 1045};
29489e7c 1046
bd81e77b 1047#define new_body_allocated(sv_type) \
d2a0f284 1048 (void *)((char *)S_new_body(aTHX_ sv_type) \
bd81e77b 1049 - bodies_by_type[sv_type].offset)
29489e7c 1050
26359cfa
NC
1051/* return a thing to the free list */
1052
1053#define del_body(thing, root) \
1054 STMT_START { \
1055 void ** const thing_copy = (void **)thing; \
1056 *thing_copy = *root; \
1057 *root = (void*)thing_copy; \
1058 } STMT_END
29489e7c 1059
bd81e77b 1060#ifdef PURIFY
5b306eef
DD
1061#if !(NVSIZE <= IVSIZE)
1062# define new_XNV() safemalloc(sizeof(XPVNV))
1063#endif
beeec492
NC
1064#define new_XPVNV() safemalloc(sizeof(XPVNV))
1065#define new_XPVMG() safemalloc(sizeof(XPVMG))
29489e7c 1066
beeec492 1067#define del_XPVGV(p) safefree(p)
29489e7c 1068
bd81e77b 1069#else /* !PURIFY */
29489e7c 1070
5b306eef
DD
1071#if !(NVSIZE <= IVSIZE)
1072# define new_XNV() new_body_allocated(SVt_NV)
1073#endif
65ac1738 1074#define new_XPVNV() new_body_allocated(SVt_PVNV)
65ac1738 1075#define new_XPVMG() new_body_allocated(SVt_PVMG)
645c22ef 1076
26359cfa
NC
1077#define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1078 &PL_body_roots[SVt_PVGV])
1d7c1841 1079
bd81e77b 1080#endif /* PURIFY */
93e68bfb 1081
bd81e77b 1082/* no arena for you! */
93e68bfb 1083
bd81e77b 1084#define new_NOARENA(details) \
beeec492 1085 safemalloc((details)->body_size + (details)->offset)
bd81e77b 1086#define new_NOARENAZ(details) \
beeec492 1087 safecalloc((details)->body_size + (details)->offset, 1)
d2a0f284 1088
1e30fcd5
NC
1089void *
1090Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1091 const size_t arena_size)
d2a0f284 1092{
d2a0f284 1093 void ** const root = &PL_body_roots[sv_type];
99816f8d
NC
1094 struct arena_desc *adesc;
1095 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1096 unsigned int curr;
d2a0f284
JC
1097 char *start;
1098 const char *end;
02982131 1099 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
20b7effb
JH
1100#if defined(DEBUGGING) && defined(PERL_GLOBAL_STRUCT)
1101 dVAR;
1102#endif
0b2d3faa 1103#if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
23e9d66c
NC
1104 static bool done_sanity_check;
1105
0b2d3faa
JH
1106 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1107 * variables like done_sanity_check. */
10666ae3 1108 if (!done_sanity_check) {
ea471437 1109 unsigned int i = SVt_LAST;
10666ae3
NC
1110
1111 done_sanity_check = TRUE;
1112
1113 while (i--)
1114 assert (bodies_by_type[i].type == i);
1115 }
1116#endif
1117
02982131 1118 assert(arena_size);
23e9d66c 1119
99816f8d
NC
1120 /* may need new arena-set to hold new arena */
1121 if (!aroot || aroot->curr >= aroot->set_size) {
1122 struct arena_set *newroot;
1123 Newxz(newroot, 1, struct arena_set);
1124 newroot->set_size = ARENAS_PER_SET;
1125 newroot->next = aroot;
1126 aroot = newroot;
1127 PL_body_arenas = (void *) newroot;
1128 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1129 }
1130
1131 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1132 curr = aroot->curr++;
1133 adesc = &(aroot->set[curr]);
1134 assert(!adesc->arena);
1135
1136 Newx(adesc->arena, good_arena_size, char);
1137 adesc->size = good_arena_size;
1138 adesc->utype = sv_type;
1139 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %"UVuf"\n",
1140 curr, (void*)adesc->arena, (UV)good_arena_size));
1141
1142 start = (char *) adesc->arena;
d2a0f284 1143
29657bb6
NC
1144 /* Get the address of the byte after the end of the last body we can fit.
1145 Remember, this is integer division: */
02982131 1146 end = start + good_arena_size / body_size * body_size;
d2a0f284 1147
486ec47a 1148 /* computed count doesn't reflect the 1st slot reservation */
d8fca402
NC
1149#if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1150 DEBUG_m(PerlIO_printf(Perl_debug_log,
1151 "arena %p end %p arena-size %d (from %d) type %d "
1152 "size %d ct %d\n",
02982131
NC
1153 (void*)start, (void*)end, (int)good_arena_size,
1154 (int)arena_size, sv_type, (int)body_size,
1155 (int)good_arena_size / (int)body_size));
d8fca402 1156#else
d2a0f284
JC
1157 DEBUG_m(PerlIO_printf(Perl_debug_log,
1158 "arena %p end %p arena-size %d type %d size %d ct %d\n",
6c9570dc 1159 (void*)start, (void*)end,
02982131
NC
1160 (int)arena_size, sv_type, (int)body_size,
1161 (int)good_arena_size / (int)body_size));
d8fca402 1162#endif
d2a0f284
JC
1163 *root = (void *)start;
1164
29657bb6
NC
1165 while (1) {
1166 /* Where the next body would start: */
d2a0f284 1167 char * const next = start + body_size;
29657bb6
NC
1168
1169 if (next >= end) {
1170 /* This is the last body: */
1171 assert(next == end);
1172
1173 *(void **)start = 0;
1174 return *root;
1175 }
1176
d2a0f284
JC
1177 *(void**) start = (void *)next;
1178 start = next;
1179 }
d2a0f284
JC
1180}
1181
1182/* grab a new thing from the free list, allocating more if necessary.
1183 The inline version is used for speed in hot routines, and the
1184 function using it serves the rest (unless PURIFY).
1185*/
1186#define new_body_inline(xpv, sv_type) \
1187 STMT_START { \
1188 void ** const r3wt = &PL_body_roots[sv_type]; \
11b79775 1189 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1e30fcd5 1190 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
02982131
NC
1191 bodies_by_type[sv_type].body_size,\
1192 bodies_by_type[sv_type].arena_size)); \
d2a0f284 1193 *(r3wt) = *(void**)(xpv); \
d2a0f284
JC
1194 } STMT_END
1195
1196#ifndef PURIFY
1197
1198STATIC void *
de37a194 1199S_new_body(pTHX_ const svtype sv_type)
d2a0f284 1200{
d2a0f284
JC
1201 void *xpv;
1202 new_body_inline(xpv, sv_type);
1203 return xpv;
1204}
1205
1206#endif
93e68bfb 1207
238b27b3
NC
1208static const struct body_details fake_rv =
1209 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1210
bd81e77b
NC
1211/*
1212=for apidoc sv_upgrade
93e68bfb 1213
bd81e77b
NC
1214Upgrade an SV to a more complex form. Generally adds a new body type to the
1215SV, then copies across as much information as possible from the old body.
9521ca61
FC
1216It croaks if the SV is already in a more complex form than requested. You
1217generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1218before calling C<sv_upgrade>, and hence does not croak. See also
1219C<svtype>.
93e68bfb 1220
bd81e77b 1221=cut
93e68bfb 1222*/
93e68bfb 1223
bd81e77b 1224void
5aaab254 1225Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
cac9b346 1226{
bd81e77b
NC
1227 void* old_body;
1228 void* new_body;
42d0e0b7 1229 const svtype old_type = SvTYPE(sv);
d2a0f284 1230 const struct body_details *new_type_details;
238b27b3 1231 const struct body_details *old_type_details
bd81e77b 1232 = bodies_by_type + old_type;
4df7f6af 1233 SV *referant = NULL;
cac9b346 1234
7918f24d
NC
1235 PERL_ARGS_ASSERT_SV_UPGRADE;
1236
1776cbe8
NC
1237 if (old_type == new_type)
1238 return;
1239
1240 /* This clause was purposefully added ahead of the early return above to
1241 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1242 inference by Nick I-S that it would fix other troublesome cases. See
1243 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1244
1245 Given that shared hash key scalars are no longer PVIV, but PV, there is
1246 no longer need to unshare so as to free up the IVX slot for its proper
1247 purpose. So it's safe to move the early return earlier. */
1248
093085a8 1249 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
bd81e77b
NC
1250 sv_force_normal_flags(sv, 0);
1251 }
cac9b346 1252
bd81e77b 1253 old_body = SvANY(sv);
de042e1d 1254
bd81e77b
NC
1255 /* Copying structures onto other structures that have been neatly zeroed
1256 has a subtle gotcha. Consider XPVMG
cac9b346 1257
bd81e77b
NC
1258 +------+------+------+------+------+-------+-------+
1259 | NV | CUR | LEN | IV | MAGIC | STASH |
1260 +------+------+------+------+------+-------+-------+
1261 0 4 8 12 16 20 24 28
645c22ef 1262
bd81e77b
NC
1263 where NVs are aligned to 8 bytes, so that sizeof that structure is
1264 actually 32 bytes long, with 4 bytes of padding at the end:
08742458 1265
bd81e77b
NC
1266 +------+------+------+------+------+-------+-------+------+
1267 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1268 +------+------+------+------+------+-------+-------+------+
1269 0 4 8 12 16 20 24 28 32
08742458 1270
bd81e77b 1271 so what happens if you allocate memory for this structure:
30f9da9e 1272
bd81e77b
NC
1273 +------+------+------+------+------+-------+-------+------+------+...
1274 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1275 +------+------+------+------+------+-------+-------+------+------+...
1276 0 4 8 12 16 20 24 28 32 36
bfc44f79 1277
bd81e77b
NC
1278 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1279 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1280 started out as zero once, but it's quite possible that it isn't. So now,
1281 rather than a nicely zeroed GP, you have it pointing somewhere random.
1282 Bugs ensue.
bfc44f79 1283
bd81e77b
NC
1284 (In fact, GP ends up pointing at a previous GP structure, because the
1285 principle cause of the padding in XPVMG getting garbage is a copy of
6c9e42f7
NC
1286 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1287 this happens to be moot because XPVGV has been re-ordered, with GP
1288 no longer after STASH)
30f9da9e 1289
bd81e77b
NC
1290 So we are careful and work out the size of used parts of all the
1291 structures. */
bfc44f79 1292
bd81e77b
NC
1293 switch (old_type) {
1294 case SVt_NULL:
1295 break;
1296 case SVt_IV:
4df7f6af
NC
1297 if (SvROK(sv)) {
1298 referant = SvRV(sv);
238b27b3
NC
1299 old_type_details = &fake_rv;
1300 if (new_type == SVt_NV)
1301 new_type = SVt_PVNV;
4df7f6af
NC
1302 } else {
1303 if (new_type < SVt_PVIV) {
1304 new_type = (new_type == SVt_NV)
1305 ? SVt_PVNV : SVt_PVIV;
1306 }
bd81e77b
NC
1307 }
1308 break;
1309 case SVt_NV:
1310 if (new_type < SVt_PVNV) {
1311 new_type = SVt_PVNV;
bd81e77b
NC
1312 }
1313 break;
bd81e77b
NC
1314 case SVt_PV:
1315 assert(new_type > SVt_PV);
6d59e610
LM
1316 STATIC_ASSERT_STMT(SVt_IV < SVt_PV);
1317 STATIC_ASSERT_STMT(SVt_NV < SVt_PV);
bd81e77b
NC
1318 break;
1319 case SVt_PVIV:
1320 break;
1321 case SVt_PVNV:
1322 break;
1323 case SVt_PVMG:
1324 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1325 there's no way that it can be safely upgraded, because perl.c
1326 expects to Safefree(SvANY(PL_mess_sv)) */
1327 assert(sv != PL_mess_sv);
bd81e77b
NC
1328 break;
1329 default:
2439e033 1330 if (UNLIKELY(old_type_details->cant_upgrade))
c81225bc
NC
1331 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1332 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
bd81e77b 1333 }
3376de98 1334
2439e033 1335 if (UNLIKELY(old_type > new_type))
3376de98
NC
1336 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1337 (int)old_type, (int)new_type);
1338
2fa1109b 1339 new_type_details = bodies_by_type + new_type;
645c22ef 1340
bd81e77b
NC
1341 SvFLAGS(sv) &= ~SVTYPEMASK;
1342 SvFLAGS(sv) |= new_type;
932e9ff9 1343
ab4416c0
NC
1344 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1345 the return statements above will have triggered. */
1346 assert (new_type != SVt_NULL);
bd81e77b 1347 switch (new_type) {
bd81e77b
NC
1348 case SVt_IV:
1349 assert(old_type == SVt_NULL);
dc6369ef 1350 SET_SVANY_FOR_BODYLESS_IV(sv);
bd81e77b
NC
1351 SvIV_set(sv, 0);
1352 return;
1353 case SVt_NV:
1354 assert(old_type == SVt_NULL);
5b306eef 1355#if NVSIZE <= IVSIZE
dc6369ef 1356 SET_SVANY_FOR_BODYLESS_NV(sv);
5b306eef 1357#else
bd81e77b 1358 SvANY(sv) = new_XNV();
5b306eef 1359#endif
bd81e77b
NC
1360 SvNV_set(sv, 0);
1361 return;
bd81e77b 1362 case SVt_PVHV:
bd81e77b 1363 case SVt_PVAV:
d2a0f284 1364 assert(new_type_details->body_size);
c1ae03ae
NC
1365
1366#ifndef PURIFY
1367 assert(new_type_details->arena);
d2a0f284 1368 assert(new_type_details->arena_size);
c1ae03ae 1369 /* This points to the start of the allocated area. */
d2a0f284
JC
1370 new_body_inline(new_body, new_type);
1371 Zero(new_body, new_type_details->body_size, char);
c1ae03ae
NC
1372 new_body = ((char *)new_body) - new_type_details->offset;
1373#else
1374 /* We always allocated the full length item with PURIFY. To do this
1375 we fake things so that arena is false for all 16 types.. */
1376 new_body = new_NOARENAZ(new_type_details);
1377#endif
1378 SvANY(sv) = new_body;
1379 if (new_type == SVt_PVAV) {
1380 AvMAX(sv) = -1;
1381 AvFILLp(sv) = -1;
1382 AvREAL_only(sv);
64484faa 1383 if (old_type_details->body_size) {
ac572bf4
NC
1384 AvALLOC(sv) = 0;
1385 } else {
1386 /* It will have been zeroed when the new body was allocated.
1387 Lets not write to it, in case it confuses a write-back
1388 cache. */
1389 }
78ac7dd9
NC
1390 } else {
1391 assert(!SvOK(sv));
1392 SvOK_off(sv);
1393#ifndef NODEFAULT_SHAREKEYS
1394 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1395#endif
586fc6a3
SM
1396 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1397 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
c1ae03ae 1398 }
aeb18a1e 1399
bd81e77b
NC
1400 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1401 The target created by newSVrv also is, and it can have magic.
1402 However, it never has SvPVX set.
1403 */
4df7f6af
NC
1404 if (old_type == SVt_IV) {
1405 assert(!SvROK(sv));
1406 } else if (old_type >= SVt_PV) {
bd81e77b
NC
1407 assert(SvPVX_const(sv) == 0);
1408 }
aeb18a1e 1409
bd81e77b 1410 if (old_type >= SVt_PVMG) {
e736a858 1411 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
bd81e77b 1412 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
797c7171
NC
1413 } else {
1414 sv->sv_u.svu_array = NULL; /* or svu_hash */
bd81e77b
NC
1415 }
1416 break;
93e68bfb 1417
bd81e77b
NC
1418 case SVt_PVIV:
1419 /* XXX Is this still needed? Was it ever needed? Surely as there is
1420 no route from NV to PVIV, NOK can never be true */
1421 assert(!SvNOKp(sv));
1422 assert(!SvNOK(sv));
2b5060ae 1423 /* FALLTHROUGH */
bd81e77b
NC
1424 case SVt_PVIO:
1425 case SVt_PVFM:
bd81e77b
NC
1426 case SVt_PVGV:
1427 case SVt_PVCV:
1428 case SVt_PVLV:
d361b004 1429 case SVt_INVLIST:
12c45b25 1430 case SVt_REGEXP:
bd81e77b
NC
1431 case SVt_PVMG:
1432 case SVt_PVNV:
1433 case SVt_PV:
93e68bfb 1434
d2a0f284 1435 assert(new_type_details->body_size);
bd81e77b
NC
1436 /* We always allocated the full length item with PURIFY. To do this
1437 we fake things so that arena is false for all 16 types.. */
1438 if(new_type_details->arena) {
1439 /* This points to the start of the allocated area. */
d2a0f284
JC
1440 new_body_inline(new_body, new_type);
1441 Zero(new_body, new_type_details->body_size, char);
bd81e77b
NC
1442 new_body = ((char *)new_body) - new_type_details->offset;
1443 } else {
1444 new_body = new_NOARENAZ(new_type_details);
1445 }
1446 SvANY(sv) = new_body;
5e2fc214 1447
bd81e77b 1448 if (old_type_details->copy) {
f9ba3d20
NC
1449 /* There is now the potential for an upgrade from something without
1450 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1451 int offset = old_type_details->offset;
1452 int length = old_type_details->copy;
1453
1454 if (new_type_details->offset > old_type_details->offset) {
d4c19fe8 1455 const int difference
f9ba3d20
NC
1456 = new_type_details->offset - old_type_details->offset;
1457 offset += difference;
1458 length -= difference;
1459 }
1460 assert (length >= 0);
1461
1462 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1463 char);
bd81e77b
NC
1464 }
1465
1466#ifndef NV_ZERO_IS_ALLBITS_ZERO
f2524eef 1467 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
e5ce394c
NC
1468 * correct 0.0 for us. Otherwise, if the old body didn't have an
1469 * NV slot, but the new one does, then we need to initialise the
1470 * freshly created NV slot with whatever the correct bit pattern is
1471 * for 0.0 */
e22a937e
NC
1472 if (old_type_details->zero_nv && !new_type_details->zero_nv
1473 && !isGV_with_GP(sv))
bd81e77b 1474 SvNV_set(sv, 0);
82048762 1475#endif
5e2fc214 1476
2439e033 1477 if (UNLIKELY(new_type == SVt_PVIO)) {
85dca89a 1478 IO * const io = MUTABLE_IO(sv);
d963bf01 1479 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
85dca89a
NC
1480
1481 SvOBJECT_on(io);
1482 /* Clear the stashcache because a new IO could overrule a package
1483 name */
103f5a36 1484 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
85dca89a
NC
1485 hv_clear(PL_stashcache);
1486
85dca89a 1487 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
f2524eef 1488 IoPAGE_LEN(sv) = 60;
85dca89a 1489 }
2439e033 1490 if (UNLIKELY(new_type == SVt_REGEXP))
8d919b0a
FC
1491 sv->sv_u.svu_rx = (regexp *)new_body;
1492 else if (old_type < SVt_PV) {
4df7f6af
NC
1493 /* referant will be NULL unless the old type was SVt_IV emulating
1494 SVt_RV */
1495 sv->sv_u.svu_rv = referant;
1496 }
bd81e77b
NC
1497 break;
1498 default:
afd78fd5
JH
1499 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1500 (unsigned long)new_type);
bd81e77b 1501 }
73171d91 1502
5b306eef
DD
1503 /* if this is zero, this is a body-less SVt_NULL, SVt_IV/SVt_RV,
1504 and sometimes SVt_NV */
1505 if (old_type_details->body_size) {
bd81e77b 1506#ifdef PURIFY
beeec492 1507 safefree(old_body);
bd81e77b 1508#else
bc786448
GG
1509 /* Note that there is an assumption that all bodies of types that
1510 can be upgraded came from arenas. Only the more complex non-
1511 upgradable types are allowed to be directly malloc()ed. */
1512 assert(old_type_details->arena);
bd81e77b
NC
1513 del_body((void*)((char*)old_body + old_type_details->offset),
1514 &PL_body_roots[old_type]);
1515#endif
1516 }
1517}
73171d91 1518
bd81e77b
NC
1519/*
1520=for apidoc sv_backoff
73171d91 1521
fde67290 1522Remove any string offset. You should normally use the C<SvOOK_off> macro
bd81e77b 1523wrapper instead.
73171d91 1524
bd81e77b 1525=cut
73171d91
NC
1526*/
1527
bd81e77b 1528int
ddeaf645 1529Perl_sv_backoff(SV *const sv)
bd81e77b 1530{
69240efd 1531 STRLEN delta;
7a4bba22 1532 const char * const s = SvPVX_const(sv);
7918f24d
NC
1533
1534 PERL_ARGS_ASSERT_SV_BACKOFF;
7918f24d 1535
bd81e77b
NC
1536 assert(SvOOK(sv));
1537 assert(SvTYPE(sv) != SVt_PVHV);
1538 assert(SvTYPE(sv) != SVt_PVAV);
7a4bba22 1539
69240efd
NC
1540 SvOOK_offset(sv, delta);
1541
7a4bba22
NC
1542 SvLEN_set(sv, SvLEN(sv) + delta);
1543 SvPV_set(sv, SvPVX(sv) - delta);
1544 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
bd81e77b
NC
1545 SvFLAGS(sv) &= ~SVf_OOK;
1546 return 0;
1547}
73171d91 1548
bd81e77b
NC
1549/*
1550=for apidoc sv_grow
73171d91 1551
bd81e77b
NC
1552Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1553upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1554Use the C<SvGROW> wrapper instead.
93e68bfb 1555
bd81e77b
NC
1556=cut
1557*/
93e68bfb 1558
e0060e30
FC
1559static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1560
bd81e77b 1561char *
5aaab254 1562Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
bd81e77b 1563{
eb578fdb 1564 char *s;
93e68bfb 1565
7918f24d
NC
1566 PERL_ARGS_ASSERT_SV_GROW;
1567
bd81e77b
NC
1568 if (SvROK(sv))
1569 sv_unref(sv);
1570 if (SvTYPE(sv) < SVt_PV) {
1571 sv_upgrade(sv, SVt_PV);
1572 s = SvPVX_mutable(sv);
1573 }
1574 else if (SvOOK(sv)) { /* pv is offset? */
1575 sv_backoff(sv);
1576 s = SvPVX_mutable(sv);
1577 if (newlen > SvLEN(sv))
1578 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
bd81e77b
NC
1579 }
1580 else
db2c6cb3 1581 {
e0060e30 1582 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
bd81e77b 1583 s = SvPVX_mutable(sv);
db2c6cb3 1584 }
aeb18a1e 1585
93c10d60 1586#ifdef PERL_COPY_ON_WRITE
cbcb2a16 1587 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
3c239bea 1588 * to store the COW count. So in general, allocate one more byte than
cbcb2a16
DM
1589 * asked for, to make it likely this byte is always spare: and thus
1590 * make more strings COW-able.
1591 * If the new size is a big power of two, don't bother: we assume the
1592 * caller wanted a nice 2^N sized block and will be annoyed at getting
fa8f4f85
TC
1593 * 2^N+1.
1594 * Only increment if the allocation isn't MEM_SIZE_MAX,
1595 * otherwise it will wrap to 0.
1596 */
1597 if (newlen & 0xff && newlen != MEM_SIZE_MAX)
cbcb2a16
DM
1598 newlen++;
1599#endif
1600
ce861ea7
YO
1601#if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1602#define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1603#endif
1604
bd81e77b 1605 if (newlen > SvLEN(sv)) { /* need more room? */
f1200559 1606 STRLEN minlen = SvCUR(sv);
3c239bea 1607 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
f1200559
WH
1608 if (newlen < minlen)
1609 newlen = minlen;
ce861ea7 1610#ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
7c641603
KW
1611
1612 /* Don't round up on the first allocation, as odds are pretty good that
1613 * the initial request is accurate as to what is really needed */
ce861ea7 1614 if (SvLEN(sv)) {
9efda33a
TC
1615 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1616 if (rounded > newlen)
1617 newlen = rounded;
ce861ea7 1618 }
bd81e77b 1619#endif
98653f18 1620 if (SvLEN(sv) && s) {
10edeb5d 1621 s = (char*)saferealloc(s, newlen);
bd81e77b
NC
1622 }
1623 else {
10edeb5d 1624 s = (char*)safemalloc(newlen);
bd81e77b
NC
1625 if (SvPVX_const(sv) && SvCUR(sv)) {
1626 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1627 }
1628 }
1629 SvPV_set(sv, s);
ce861ea7 1630#ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
98653f18
NC
1631 /* Do this here, do it once, do it right, and then we will never get
1632 called back into sv_grow() unless there really is some growing
1633 needed. */
ca7c1a29 1634 SvLEN_set(sv, Perl_safesysmalloc_size(s));
98653f18 1635#else
bd81e77b 1636 SvLEN_set(sv, newlen);
98653f18 1637#endif
bd81e77b
NC
1638 }
1639 return s;
1640}
aeb18a1e 1641
bd81e77b
NC
1642/*
1643=for apidoc sv_setiv
932e9ff9 1644
bd81e77b
NC
1645Copies an integer into the given SV, upgrading first if necessary.
1646Does not handle 'set' magic. See also C<sv_setiv_mg>.
463ee0b2 1647
bd81e77b
NC
1648=cut
1649*/
463ee0b2 1650
bd81e77b 1651void
5aaab254 1652Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
bd81e77b 1653{
7918f24d
NC
1654 PERL_ARGS_ASSERT_SV_SETIV;
1655
bd81e77b
NC
1656 SV_CHECK_THINKFIRST_COW_DROP(sv);
1657 switch (SvTYPE(sv)) {
1658 case SVt_NULL:
bd81e77b 1659 case SVt_NV:
3376de98 1660 sv_upgrade(sv, SVt_IV);
bd81e77b 1661 break;
bd81e77b
NC
1662 case SVt_PV:
1663 sv_upgrade(sv, SVt_PVIV);
1664 break;
463ee0b2 1665
bd81e77b 1666 case SVt_PVGV:
6e592b3a
BM
1667 if (!isGV_with_GP(sv))
1668 break;
bd81e77b
NC
1669 case SVt_PVAV:
1670 case SVt_PVHV:
1671 case SVt_PVCV:
1672 case SVt_PVFM:
1673 case SVt_PVIO:
22e74366 1674 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b
NC
1675 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1676 OP_DESC(PL_op));
0103ca14 1677 break;
42d0e0b7 1678 default: NOOP;
bd81e77b
NC
1679 }
1680 (void)SvIOK_only(sv); /* validate number */
1681 SvIV_set(sv, i);
1682 SvTAINT(sv);
1683}
932e9ff9 1684
bd81e77b
NC
1685/*
1686=for apidoc sv_setiv_mg
d33b2eba 1687
bd81e77b 1688Like C<sv_setiv>, but also handles 'set' magic.
1c846c1f 1689
bd81e77b
NC
1690=cut
1691*/
d33b2eba 1692
bd81e77b 1693void
5aaab254 1694Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
bd81e77b 1695{
7918f24d
NC
1696 PERL_ARGS_ASSERT_SV_SETIV_MG;
1697
bd81e77b
NC
1698 sv_setiv(sv,i);
1699 SvSETMAGIC(sv);
1700}
727879eb 1701
bd81e77b
NC
1702/*
1703=for apidoc sv_setuv
d33b2eba 1704
bd81e77b
NC
1705Copies an unsigned integer into the given SV, upgrading first if necessary.
1706Does not handle 'set' magic. See also C<sv_setuv_mg>.
9b94d1dd 1707
bd81e77b
NC
1708=cut
1709*/
d33b2eba 1710
bd81e77b 1711void
5aaab254 1712Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
bd81e77b 1713{
7918f24d
NC
1714 PERL_ARGS_ASSERT_SV_SETUV;
1715
013abb9b
NC
1716 /* With the if statement to ensure that integers are stored as IVs whenever
1717 possible:
bd81e77b 1718 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
d33b2eba 1719
bd81e77b
NC
1720 without
1721 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1c846c1f 1722
013abb9b
NC
1723 If you wish to remove the following if statement, so that this routine
1724 (and its callers) always return UVs, please benchmark to see what the
1725 effect is. Modern CPUs may be different. Or may not :-)
bd81e77b
NC
1726 */
1727 if (u <= (UV)IV_MAX) {
1728 sv_setiv(sv, (IV)u);
1729 return;
1730 }
1731 sv_setiv(sv, 0);
1732 SvIsUV_on(sv);
1733 SvUV_set(sv, u);
1734}
d33b2eba 1735
bd81e77b
NC
1736/*
1737=for apidoc sv_setuv_mg
727879eb 1738
bd81e77b 1739Like C<sv_setuv>, but also handles 'set' magic.
9b94d1dd 1740
bd81e77b
NC
1741=cut
1742*/
5e2fc214 1743
bd81e77b 1744void
5aaab254 1745Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
bd81e77b 1746{
7918f24d
NC
1747 PERL_ARGS_ASSERT_SV_SETUV_MG;
1748
bd81e77b
NC
1749 sv_setuv(sv,u);
1750 SvSETMAGIC(sv);
1751}
5e2fc214 1752
954c1994 1753/*
bd81e77b 1754=for apidoc sv_setnv
954c1994 1755
bd81e77b
NC
1756Copies a double into the given SV, upgrading first if necessary.
1757Does not handle 'set' magic. See also C<sv_setnv_mg>.
954c1994
GS
1758
1759=cut
1760*/
1761
63f97190 1762void
5aaab254 1763Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
79072805 1764{
7918f24d
NC
1765 PERL_ARGS_ASSERT_SV_SETNV;
1766
bd81e77b
NC
1767 SV_CHECK_THINKFIRST_COW_DROP(sv);
1768 switch (SvTYPE(sv)) {
79072805 1769 case SVt_NULL:
79072805 1770 case SVt_IV:
bd81e77b 1771 sv_upgrade(sv, SVt_NV);
79072805
LW
1772 break;
1773 case SVt_PV:
79072805 1774 case SVt_PVIV:
bd81e77b 1775 sv_upgrade(sv, SVt_PVNV);
79072805 1776 break;
bd4b1eb5 1777
bd4b1eb5 1778 case SVt_PVGV:
6e592b3a
BM
1779 if (!isGV_with_GP(sv))
1780 break;
bd81e77b
NC
1781 case SVt_PVAV:
1782 case SVt_PVHV:
79072805 1783 case SVt_PVCV:
bd81e77b
NC
1784 case SVt_PVFM:
1785 case SVt_PVIO:
22e74366 1786 /* diag_listed_as: Can't coerce %s to %s in %s */
bd81e77b 1787 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
94bbb3f4 1788 OP_DESC(PL_op));
0103ca14 1789 break;
42d0e0b7 1790 default: NOOP;
2068cd4d 1791 }
bd81e77b
NC
1792 SvNV_set(sv, num);
1793 (void)SvNOK_only(sv); /* validate number */
1794 SvTAINT(sv);
79072805
LW
1795}
1796
645c22ef 1797/*
bd81e77b 1798=for apidoc sv_setnv_mg
645c22ef 1799
bd81e77b 1800Like C<sv_setnv>, but also handles 'set' magic.
645c22ef
DM
1801
1802=cut
1803*/
1804
bd81e77b 1805void
5aaab254 1806Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
79072805 1807{
7918f24d
NC
1808 PERL_ARGS_ASSERT_SV_SETNV_MG;
1809
bd81e77b
NC
1810 sv_setnv(sv,num);
1811 SvSETMAGIC(sv);
79072805
LW
1812}
1813
3f7602fa
TC
1814/* Return a cleaned-up, printable version of sv, for non-numeric, or
1815 * not incrementable warning display.
1816 * Originally part of S_not_a_number().
1817 * The return value may be != tmpbuf.
bd81e77b 1818 */
954c1994 1819
3f7602fa
TC
1820STATIC const char *
1821S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1822 const char *pv;
94463019 1823
3f7602fa 1824 PERL_ARGS_ASSERT_SV_DISPLAY;
7918f24d 1825
94463019 1826 if (DO_UTF8(sv)) {
3f7602fa 1827 SV *dsv = newSVpvs_flags("", SVs_TEMP);
37b8cdd1 1828 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
94463019
JH
1829 } else {
1830 char *d = tmpbuf;
3f7602fa 1831 const char * const limit = tmpbuf + tmpbuf_size - 8;
94463019
JH
1832 /* each *s can expand to 4 chars + "...\0",
1833 i.e. need room for 8 chars */
ecdeb87c 1834
00b6aa41
AL
1835 const char *s = SvPVX_const(sv);
1836 const char * const end = s + SvCUR(sv);
1837 for ( ; s < end && d < limit; s++ ) {
94463019 1838 int ch = *s & 0xFF;
bd27cf70 1839 if (! isASCII(ch) && !isPRINT_LC(ch)) {
94463019
JH
1840 *d++ = 'M';
1841 *d++ = '-';
bd27cf70
KW
1842
1843 /* Map to ASCII "equivalent" of Latin1 */
1844 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
94463019
JH
1845 }
1846 if (ch == '\n') {
1847 *d++ = '\\';
1848 *d++ = 'n';
1849 }
1850 else if (ch == '\r') {
1851 *d++ = '\\';
1852 *d++ = 'r';
1853 }
1854 else if (ch == '\f') {
1855 *d++ = '\\';
1856 *d++ = 'f';
1857 }
1858 else if (ch == '\\') {
1859 *d++ = '\\';
1860 *d++ = '\\';
1861 }
1862 else if (ch == '\0') {
1863 *d++ = '\\';
1864 *d++ = '0';
1865 }
1866 else if (isPRINT_LC(ch))
1867 *d++ = ch;
1868 else {
1869 *d++ = '^';
1870 *d++ = toCTRL(ch);
1871 }
1872 }
1873 if (s < end) {
1874 *d++ = '.';
1875 *d++ = '.';
1876 *d++ = '.';
1877 }
1878 *d = '\0';
1879 pv = tmpbuf;
a0d0e21e 1880 }
a0d0e21e 1881
3f7602fa
TC
1882 return pv;
1883}
1884
1885/* Print an "isn't numeric" warning, using a cleaned-up,
1886 * printable version of the offending string
1887 */
1888
1889STATIC void
1890S_not_a_number(pTHX_ SV *const sv)
1891{
3f7602fa
TC
1892 char tmpbuf[64];
1893 const char *pv;
1894
1895 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1896
1897 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1898
533c011a 1899 if (PL_op)
9014280d 1900 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1901 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019
JH
1902 "Argument \"%s\" isn't numeric in %s", pv,
1903 OP_DESC(PL_op));
a0d0e21e 1904 else
9014280d 1905 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
734856a2 1906 /* diag_listed_as: Argument "%s" isn't numeric%s */
94463019 1907 "Argument \"%s\" isn't numeric", pv);
a0d0e21e
LW
1908}
1909
3f7602fa
TC
1910STATIC void
1911S_not_incrementable(pTHX_ SV *const sv) {
3f7602fa
TC
1912 char tmpbuf[64];
1913 const char *pv;
1914
1915 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1916
1917 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1918
1919 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1920 "Argument \"%s\" treated as 0 in increment (++)", pv);
1921}
1922
c2988b20
NC
1923/*
1924=for apidoc looks_like_number
1925
645c22ef
DM
1926Test if the content of an SV looks like a number (or is a number).
1927C<Inf> and C<Infinity> are treated as numbers (so will not issue a
f52e41ad
FC
1928non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1929ignored.
c2988b20
NC
1930
1931=cut
1932*/
1933
1934I32
aad570aa 1935Perl_looks_like_number(pTHX_ SV *const sv)
c2988b20 1936{
eb578fdb 1937 const char *sbegin;
c2988b20 1938 STRLEN len;
ea2485eb 1939 int numtype;
c2988b20 1940
7918f24d
NC
1941 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1942
f52e41ad
FC
1943 if (SvPOK(sv) || SvPOKp(sv)) {
1944 sbegin = SvPV_nomg_const(sv, len);
c2988b20 1945 }
c2988b20 1946 else
e0ab1c0e 1947 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
ea2485eb
JH
1948 numtype = grok_number(sbegin, len, NULL);
1949 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
c2988b20 1950}
25da4f38 1951
19f6321d
NC
1952STATIC bool
1953S_glob_2number(pTHX_ GV * const gv)
180488f8 1954{
7918f24d
NC
1955 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1956
675c862f
AL
1957 /* We know that all GVs stringify to something that is not-a-number,
1958 so no need to test that. */
1959 if (ckWARN(WARN_NUMERIC))
8e629ff4
FC
1960 {
1961 SV *const buffer = sv_newmortal();
1962 gv_efullname3(buffer, gv, "*");
675c862f 1963 not_a_number(buffer);
8e629ff4 1964 }
675c862f
AL
1965 /* We just want something true to return, so that S_sv_2iuv_common
1966 can tail call us and return true. */
19f6321d 1967 return TRUE;
675c862f
AL
1968}
1969
25da4f38
IZ
1970/* Actually, ISO C leaves conversion of UV to IV undefined, but
1971 until proven guilty, assume that things are not that bad... */
1972
645c22ef
DM
1973/*
1974 NV_PRESERVES_UV:
1975
1976 As 64 bit platforms often have an NV that doesn't preserve all bits of
28e5dec8
JH
1977 an IV (an assumption perl has been based on to date) it becomes necessary
1978 to remove the assumption that the NV always carries enough precision to
1979 recreate the IV whenever needed, and that the NV is the canonical form.
1980 Instead, IV/UV and NV need to be given equal rights. So as to not lose
645c22ef 1981 precision as a side effect of conversion (which would lead to insanity
28e5dec8 1982 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
8a4a3196
KW
1983 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1984 where precision was lost, and IV/UV/NV slots that have a valid conversion
1985 which has lost no precision
645c22ef 1986 2) to ensure that if a numeric conversion to one form is requested that
28e5dec8
JH
1987 would lose precision, the precise conversion (or differently
1988 imprecise conversion) is also performed and cached, to prevent
1989 requests for different numeric formats on the same SV causing
1990 lossy conversion chains. (lossless conversion chains are perfectly
1991 acceptable (still))
1992
1993
1994 flags are used:
1995 SvIOKp is true if the IV slot contains a valid value
1996 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1997 SvNOKp is true if the NV slot contains a valid value
1998 SvNOK is true only if the NV value is accurate
1999
2000 so
645c22ef 2001 while converting from PV to NV, check to see if converting that NV to an
28e5dec8
JH
2002 IV(or UV) would lose accuracy over a direct conversion from PV to
2003 IV(or UV). If it would, cache both conversions, return NV, but mark
2004 SV as IOK NOKp (ie not NOK).
2005
645c22ef 2006 While converting from PV to IV, check to see if converting that IV to an
28e5dec8
JH
2007 NV would lose accuracy over a direct conversion from PV to NV. If it
2008 would, cache both conversions, flag similarly.
2009
2010 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
2011 correctly because if IV & NV were set NV *always* overruled.
645c22ef
DM
2012 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
2013 changes - now IV and NV together means that the two are interchangeable:
28e5dec8 2014 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
d460ef45 2015
645c22ef
DM
2016 The benefit of this is that operations such as pp_add know that if
2017 SvIOK is true for both left and right operands, then integer addition
2018 can be used instead of floating point (for cases where the result won't
2019 overflow). Before, floating point was always used, which could lead to
28e5dec8
JH
2020 loss of precision compared with integer addition.
2021
2022 * making IV and NV equal status should make maths accurate on 64 bit
2023 platforms
2024 * may speed up maths somewhat if pp_add and friends start to use
645c22ef 2025 integers when possible instead of fp. (Hopefully the overhead in
28e5dec8
JH
2026 looking for SvIOK and checking for overflow will not outweigh the
2027 fp to integer speedup)
2028 * will slow down integer operations (callers of SvIV) on "inaccurate"
2029 values, as the change from SvIOK to SvIOKp will cause a call into
2030 sv_2iv each time rather than a macro access direct to the IV slot
2031 * should speed up number->string conversion on integers as IV is
645c22ef 2032 favoured when IV and NV are equally accurate
28e5dec8
JH
2033
2034 ####################################################################
645c22ef
DM
2035 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2036 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2037 On the other hand, SvUOK is true iff UV.
28e5dec8
JH
2038 ####################################################################
2039
645c22ef 2040 Your mileage will vary depending your CPU's relative fp to integer
28e5dec8
JH
2041 performance ratio.
2042*/
2043
2044#ifndef NV_PRESERVES_UV
645c22ef
DM
2045# define IS_NUMBER_UNDERFLOW_IV 1
2046# define IS_NUMBER_UNDERFLOW_UV 2
2047# define IS_NUMBER_IV_AND_UV 2
2048# define IS_NUMBER_OVERFLOW_IV 4
2049# define IS_NUMBER_OVERFLOW_UV 5
2050
2051/* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
28e5dec8
JH
2052
2053/* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2054STATIC int
5aaab254 2055S_sv_2iuv_non_preserve(pTHX_ SV *const sv
47031da6
NC
2056# ifdef DEBUGGING
2057 , I32 numtype
2058# endif
2059 )
28e5dec8 2060{
7918f24d 2061 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
23491f1d 2062 PERL_UNUSED_CONTEXT;
7918f24d 2063
3f7c398e 2064 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
2065 if (SvNVX(sv) < (NV)IV_MIN) {
2066 (void)SvIOKp_on(sv);
2067 (void)SvNOK_on(sv);
45977657 2068 SvIV_set(sv, IV_MIN);
28e5dec8
JH
2069 return IS_NUMBER_UNDERFLOW_IV;
2070 }
2071 if (SvNVX(sv) > (NV)UV_MAX) {
2072 (void)SvIOKp_on(sv);
2073 (void)SvNOK_on(sv);
2074 SvIsUV_on(sv);
607fa7f2 2075 SvUV_set(sv, UV_MAX);
28e5dec8
JH
2076 return IS_NUMBER_OVERFLOW_UV;
2077 }
c2988b20
NC
2078 (void)SvIOKp_on(sv);
2079 (void)SvNOK_on(sv);
2080 /* Can't use strtol etc to convert this string. (See truth table in
2081 sv_2iv */
2082 if (SvNVX(sv) <= (UV)IV_MAX) {
45977657 2083 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2084 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2085 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2086 } else {
2087 /* Integer is imprecise. NOK, IOKp */
2088 }
2089 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2090 }
2091 SvIsUV_on(sv);
607fa7f2 2092 SvUV_set(sv, U_V(SvNVX(sv)));
659c4b96 2093 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
c2988b20
NC
2094 if (SvUVX(sv) == UV_MAX) {
2095 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2096 possibly be preserved by NV. Hence, it must be overflow.
2097 NOK, IOKp */
2098 return IS_NUMBER_OVERFLOW_UV;
2099 }
2100 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2101 } else {
2102 /* Integer is imprecise. NOK, IOKp */
28e5dec8 2103 }
c2988b20 2104 return IS_NUMBER_OVERFLOW_IV;
28e5dec8 2105}
645c22ef
DM
2106#endif /* !NV_PRESERVES_UV*/
2107
a13f4dff 2108/* If numtype is infnan, set the NV of the sv accordingly.
5564cd7f 2109 * If numtype is anything else, try setting the NV using Atof(PV). */
3c81f0b3
DD
2110#ifdef USING_MSVC6
2111# pragma warning(push)
2112# pragma warning(disable:4756;disable:4056)
2113#endif
a13f4dff 2114static void
3823048b 2115S_sv_setnv(pTHX_ SV* sv, int numtype)
a13f4dff 2116{
07925c5e 2117 bool pok = cBOOL(SvPOK(sv));
5564cd7f 2118 bool nok = FALSE;
a13f4dff
JH
2119 if ((numtype & IS_NUMBER_INFINITY)) {
2120 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
5564cd7f 2121 nok = TRUE;
a13f4dff
JH
2122 }
2123 else if ((numtype & IS_NUMBER_NAN)) {
3823048b 2124 SvNV_set(sv, NV_NAN);
d48bd569 2125 nok = TRUE;
a13f4dff 2126 }
d48bd569 2127 else if (pok) {
a13f4dff 2128 SvNV_set(sv, Atof(SvPVX_const(sv)));
d48bd569
JH
2129 /* Purposefully no true nok here, since we don't want to blow
2130 * away the possible IOK/UV of an existing sv. */
2131 }
5564cd7f 2132 if (nok) {
d48bd569 2133 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
5564cd7f
JH
2134 if (pok)
2135 SvPOK_on(sv); /* PV is okay, though. */
2136 }
a13f4dff 2137}
3c81f0b3
DD
2138#ifdef USING_MSVC6
2139# pragma warning(pop)
2140#endif
a13f4dff 2141
af359546 2142STATIC bool
7918f24d
NC
2143S_sv_2iuv_common(pTHX_ SV *const sv)
2144{
7918f24d
NC
2145 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2146
af359546 2147 if (SvNOKp(sv)) {
28e5dec8
JH
2148 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2149 * without also getting a cached IV/UV from it at the same time
2150 * (ie PV->NV conversion should detect loss of accuracy and cache
af359546
NC
2151 * IV or UV at same time to avoid this. */
2152 /* IV-over-UV optimisation - choose to cache IV if possible */
25da4f38
IZ
2153
2154 if (SvTYPE(sv) == SVt_NV)
2155 sv_upgrade(sv, SVt_PVNV);
2156
28e5dec8
JH
2157 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2158 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2159 certainly cast into the IV range at IV_MAX, whereas the correct
2160 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2161 cases go to UV */
e91de695
JH
2162#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2163 if (Perl_isnan(SvNVX(sv))) {
2164 SvUV_set(sv, 0);
2165 SvIsUV_on(sv);
2166 return FALSE;
2167 }
2168#endif
28e5dec8 2169 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
45977657 2170 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2171 if (SvNVX(sv) == (NV) SvIVX(sv)
28e5dec8 2172#ifndef NV_PRESERVES_UV
53e2bfb7 2173 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
28e5dec8
JH
2174 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2175 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2176 /* Don't flag it as "accurately an integer" if the number
2177 came from a (by definition imprecise) NV operation, and
2178 we're outside the range of NV integer precision */
2179#endif
2180 ) {
a43d94f2
NC
2181 if (SvNOK(sv))
2182 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2183 else {
2184 /* scalar has trailing garbage, eg "42a" */
2185 }
28e5dec8 2186 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2187 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
28e5dec8
JH
2188 PTR2UV(sv),
2189 SvNVX(sv),
2190 SvIVX(sv)));
2191
2192 } else {
2193 /* IV not precise. No need to convert from PV, as NV
2194 conversion would already have cached IV if it detected
2195 that PV->IV would be better than PV->NV->IV
2196 flags already correct - don't set public IOK. */
2197 DEBUG_c(PerlIO_printf(Perl_debug_log,
7234c960 2198 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
28e5dec8
JH
2199 PTR2UV(sv),
2200 SvNVX(sv),
2201 SvIVX(sv)));
2202 }
2203 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2204 but the cast (NV)IV_MIN rounds to a the value less (more
2205 negative) than IV_MIN which happens to be equal to SvNVX ??
2206 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2207 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2208 (NV)UVX == NVX are both true, but the values differ. :-(
2209 Hopefully for 2s complement IV_MIN is something like
2210 0x8000000000000000 which will be exact. NWC */
d460ef45 2211 }
25da4f38 2212 else {
607fa7f2 2213 SvUV_set(sv, U_V(SvNVX(sv)));
28e5dec8 2214 if (
659c4b96 2215 (SvNVX(sv) == (NV) SvUVX(sv))
28e5dec8
JH
2216#ifndef NV_PRESERVES_UV
2217 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2218 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2219 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2220 /* Don't flag it as "accurately an integer" if the number
2221 came from a (by definition imprecise) NV operation, and
2222 we're outside the range of NV integer precision */
2223#endif
a43d94f2 2224 && SvNOK(sv)
28e5dec8
JH
2225 )
2226 SvIOK_on(sv);
25da4f38 2227 SvIsUV_on(sv);
1c846c1f 2228 DEBUG_c(PerlIO_printf(Perl_debug_log,
57def98f 2229 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
56431972 2230 PTR2UV(sv),
57def98f
JH
2231 SvUVX(sv),
2232 SvUVX(sv)));
25da4f38 2233 }
748a9306 2234 }
cd84013a 2235 else if (SvPOKp(sv)) {
c2988b20 2236 UV value;
3823048b 2237 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
af359546 2238 /* We want to avoid a possible problem when we cache an IV/ a UV which
25da4f38 2239 may be later translated to an NV, and the resulting NV is not
c2988b20
NC
2240 the same as the direct translation of the initial string
2241 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2242 be careful to ensure that the value with the .456 is around if the
2243 NV value is requested in the future).
1c846c1f 2244
af359546 2245 This means that if we cache such an IV/a UV, we need to cache the
25da4f38 2246 NV as well. Moreover, we trade speed for space, and do not
28e5dec8 2247 cache the NV if we are sure it's not needed.
25da4f38 2248 */
16b7a9a4 2249
c2988b20
NC
2250 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2251 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2252 == IS_NUMBER_IN_UV) {
5e045b90 2253 /* It's definitely an integer, only upgrade to PVIV */
28e5dec8
JH
2254 if (SvTYPE(sv) < SVt_PVIV)
2255 sv_upgrade(sv, SVt_PVIV);
f7bbb42a 2256 (void)SvIOK_on(sv);
c2988b20
NC
2257 } else if (SvTYPE(sv) < SVt_PVNV)
2258 sv_upgrade(sv, SVt_PVNV);
28e5dec8 2259
a13f4dff 2260 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
75a57a38 2261 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
6b322424 2262 not_a_number(sv);
3823048b 2263 S_sv_setnv(aTHX_ sv, numtype);
a13f4dff
JH
2264 return FALSE;
2265 }
2266
f2524eef 2267 /* If NVs preserve UVs then we only use the UV value if we know that
c2988b20
NC
2268 we aren't going to call atof() below. If NVs don't preserve UVs
2269 then the value returned may have more precision than atof() will
2270 return, even though value isn't perfectly accurate. */
2271 if ((numtype & (IS_NUMBER_IN_UV
2272#ifdef NV_PRESERVES_UV
2273 | IS_NUMBER_NOT_INT
2274#endif
2275 )) == IS_NUMBER_IN_UV) {
2276 /* This won't turn off the public IOK flag if it was set above */
2277 (void)SvIOKp_on(sv);
2278
2279 if (!(numtype & IS_NUMBER_NEG)) {
2280 /* positive */;
2281 if (value <= (UV)IV_MAX) {
45977657 2282 SvIV_set(sv, (IV)value);
c2988b20 2283 } else {
af359546 2284 /* it didn't overflow, and it was positive. */
607fa7f2 2285 SvUV_set(sv, value);
c2988b20
NC
2286 SvIsUV_on(sv);
2287 }
2288 } else {
2289 /* 2s complement assumption */
2290 if (value <= (UV)IV_MIN) {
53e2bfb7
DM
2291 SvIV_set(sv, value == (UV)IV_MIN
2292 ? IV_MIN : -(IV)value);
c2988b20
NC
2293 } else {
2294 /* Too negative for an IV. This is a double upgrade, but
d1be9408 2295 I'm assuming it will be rare. */
c2988b20
NC
2296 if (SvTYPE(sv) < SVt_PVNV)
2297 sv_upgrade(sv, SVt_PVNV);
2298 SvNOK_on(sv);
2299 SvIOK_off(sv);
2300 SvIOKp_on(sv);
9d6ce603 2301 SvNV_set(sv, -(NV)value);
45977657 2302 SvIV_set(sv, IV_MIN);
c2988b20
NC
2303 }
2304 }
2305 }
2306 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2307 will be in the previous block to set the IV slot, and the next
2308 block to set the NV slot. So no else here. */
2309
2310 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2311 != IS_NUMBER_IN_UV) {
2312 /* It wasn't an (integer that doesn't overflow the UV). */
3823048b 2313 S_sv_setnv(aTHX_ sv, numtype);
28e5dec8 2314
c2988b20
NC
2315 if (! numtype && ckWARN(WARN_NUMERIC))
2316 not_a_number(sv);
28e5dec8 2317
88cb8500 2318 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
c2988b20 2319 PTR2UV(sv), SvNVX(sv)));
28e5dec8 2320
28e5dec8 2321#ifdef NV_PRESERVES_UV
af359546
NC
2322 (void)SvIOKp_on(sv);
2323 (void)SvNOK_on(sv);
e91de695
JH
2324#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2325 if (Perl_isnan(SvNVX(sv))) {
2326 SvUV_set(sv, 0);
2327 SvIsUV_on(sv);
2328 return FALSE;
2329 }
2330#endif
af359546
NC
2331 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2332 SvIV_set(sv, I_V(SvNVX(sv)));
2333 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2334 SvIOK_on(sv);
2335 } else {
6f207bd3 2336 NOOP; /* Integer is imprecise. NOK, IOKp */
af359546
NC
2337 }
2338 /* UV will not work better than IV */
2339 } else {
2340 if (SvNVX(sv) > (NV)UV_MAX) {
2341 SvIsUV_on(sv);
2342 /* Integer is inaccurate. NOK, IOKp, is UV */
2343 SvUV_set(sv, UV_MAX);
af359546
NC
2344 } else {
2345 SvUV_set(sv, U_V(SvNVX(sv)));
2346 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2347 NV preservse UV so can do correct comparison. */
2348 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2349 SvIOK_on(sv);
af359546 2350 } else {
6f207bd3 2351 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
af359546
NC
2352 }
2353 }
4b0c9573 2354 SvIsUV_on(sv);
af359546 2355 }
28e5dec8 2356#else /* NV_PRESERVES_UV */
c2988b20
NC
2357 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2358 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
af359546 2359 /* The IV/UV slot will have been set from value returned by
c2988b20
NC
2360 grok_number above. The NV slot has just been set using
2361 Atof. */
560b0c46 2362 SvNOK_on(sv);
c2988b20
NC
2363 assert (SvIOKp(sv));
2364 } else {
2365 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2366 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2367 /* Small enough to preserve all bits. */
2368 (void)SvIOKp_on(sv);
2369 SvNOK_on(sv);
45977657 2370 SvIV_set(sv, I_V(SvNVX(sv)));
659c4b96 2371 if ((NV)(SvIVX(sv)) == SvNVX(sv))
c2988b20
NC
2372 SvIOK_on(sv);
2373 /* Assumption: first non-preserved integer is < IV_MAX,
2374 this NV is in the preserved range, therefore: */
2375 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2376 < (UV)IV_MAX)) {
32fdb065 2377 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
2378 }
2379 } else {
2380 /* IN_UV NOT_INT
2381 0 0 already failed to read UV.
2382 0 1 already failed to read UV.
2383 1 0 you won't get here in this case. IV/UV
2384 slot set, public IOK, Atof() unneeded.
2385 1 1 already read UV.
2386 so there's no point in sv_2iuv_non_preserve() attempting
2387 to use atol, strtol, strtoul etc. */
47031da6 2388# ifdef DEBUGGING
40a17c4c 2389 sv_2iuv_non_preserve (sv, numtype);
47031da6
NC
2390# else
2391 sv_2iuv_non_preserve (sv);
2392# endif
c2988b20
NC
2393 }
2394 }
28e5dec8 2395#endif /* NV_PRESERVES_UV */
a43d94f2
NC
2396 /* It might be more code efficient to go through the entire logic above
2397 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2398 gets complex and potentially buggy, so more programmer efficient
2399 to do it this way, by turning off the public flags: */
2400 if (!numtype)
2401 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
25da4f38 2402 }
af359546
NC
2403 }
2404 else {
675c862f 2405 if (isGV_with_GP(sv))
159b6efe 2406 return glob_2number(MUTABLE_GV(sv));
180488f8 2407
4f62cd62 2408 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
af359546 2409 report_uninit(sv);
25da4f38
IZ
2410 if (SvTYPE(sv) < SVt_IV)
2411 /* Typically the caller expects that sv_any is not NULL now. */
2412 sv_upgrade(sv, SVt_IV);
af359546
NC
2413 /* Return 0 from the caller. */
2414 return TRUE;
2415 }
2416 return FALSE;
2417}
2418
2419/*
2420=for apidoc sv_2iv_flags
2421
2422Return the integer value of an SV, doing any necessary string
2423conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2424Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2425
2426=cut
2427*/
2428
2429IV
5aaab254 2430Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
af359546 2431{
1061065f 2432 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
4bac9ae4 2433
217f6fa3
FC
2434 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2435 && SvTYPE(sv) != SVt_PVFM);
2436
4bac9ae4
CS
2437 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2438 mg_get(sv);
2439
2440 if (SvROK(sv)) {
2441 if (SvAMAGIC(sv)) {
2442 SV * tmpstr;
2443 if (flags & SV_SKIP_OVERLOAD)
2444 return 0;
2445 tmpstr = AMG_CALLunary(sv, numer_amg);
2446 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2447 return SvIV(tmpstr);
2448 }
2449 }
2450 return PTR2IV(SvRV(sv));
2451 }
2452
8d919b0a 2453 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d
NC
2454 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2455 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2456 In practice they are extremely unlikely to actually get anywhere
2457 accessible by user Perl code - the only way that I'm aware of is when
2458 a constant subroutine which is used as the second argument to index.
cd84013a
FC
2459
2460 Regexps have no SvIVX and SvNVX fields.
2b2b6d6d 2461 */
8d919b0a 2462 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2463 {
71c558c3 2464 UV value;
8d919b0a
FC
2465 const char * const ptr =
2466 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
e91de695
JH
2467 const int numtype
2468 = grok_number(ptr, SvCUR(sv), &value);
71c558c3
NC
2469
2470 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2471 == IS_NUMBER_IN_UV) {
2472 /* It's definitely an integer */
2473 if (numtype & IS_NUMBER_NEG) {
2474 if (value < (UV)IV_MIN)
2475 return -(IV)value;
2476 } else {
2477 if (value < (UV)IV_MAX)
2478 return (IV)value;
2479 }
2480 }
058b8ae2 2481
e91de695
JH
2482 /* Quite wrong but no good choices. */
2483 if ((numtype & IS_NUMBER_INFINITY)) {
2484 return (numtype & IS_NUMBER_NEG) ? IV_MIN : IV_MAX;
2485 } else if ((numtype & IS_NUMBER_NAN)) {
2486 return 0; /* So wrong. */
2487 }
2488
71c558c3
NC
2489 if (!numtype) {
2490 if (ckWARN(WARN_NUMERIC))
2491 not_a_number(sv);
2492 }
8d919b0a 2493 return I_V(Atof(ptr));
e20b6c3b 2494 }
4bac9ae4
CS
2495 }
2496
2497 if (SvTHINKFIRST(sv)) {
af359546
NC
2498 if (SvREADONLY(sv) && !SvOK(sv)) {
2499 if (ckWARN(WARN_UNINITIALIZED))
2500 report_uninit(sv);
2501 return 0;
2502 }
2503 }
4bac9ae4 2504
af359546
NC
2505 if (!SvIOKp(sv)) {
2506 if (S_sv_2iuv_common(aTHX_ sv))
2507 return 0;
79072805 2508 }
4bac9ae4 2509
1d7c1841
GS
2510 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2511 PTR2UV(sv),SvIVX(sv)));
25da4f38 2512 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
79072805
LW
2513}
2514
645c22ef 2515/*
891f9566 2516=for apidoc sv_2uv_flags
645c22ef
DM
2517
2518Return the unsigned integer value of an SV, doing any necessary string
891f9566
YST
2519conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2520Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
645c22ef
DM
2521
2522=cut
2523*/
2524
ff68c719 2525UV
5aaab254 2526Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
ff68c719 2527{
1061065f 2528 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
4bac9ae4
CS
2529
2530 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2531 mg_get(sv);
2532
2533 if (SvROK(sv)) {
2534 if (SvAMAGIC(sv)) {
2535 SV *tmpstr;
2536 if (flags & SV_SKIP_OVERLOAD)
2537 return 0;
2538 tmpstr = AMG_CALLunary(sv, numer_amg);
2539 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2540 return SvUV(tmpstr);
2541 }
2542 }
2543 return PTR2UV(SvRV(sv));
2544 }
2545
8d919b0a 2546 if (SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2547 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2548 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2549 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2550 assert(isREGEXP(sv) || SvPOKp(sv));
e20b6c3b 2551 {
71c558c3 2552 UV value;
8d919b0a
FC
2553 const char * const ptr =
2554 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
e91de695
JH
2555 const int numtype
2556 = grok_number(ptr, SvCUR(sv), &value);
71c558c3
NC
2557
2558 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2559 == IS_NUMBER_IN_UV) {
2560 /* It's definitely an integer */
2561 if (!(numtype & IS_NUMBER_NEG))
2562 return value;
2563 }
058b8ae2 2564
e91de695
JH
2565 /* Quite wrong but no good choices. */
2566 if ((numtype & IS_NUMBER_INFINITY)) {
2567 return UV_MAX; /* So wrong. */
2568 } else if ((numtype & IS_NUMBER_NAN)) {
2569 return 0; /* So wrong. */
2570 }
2571
71c558c3
NC
2572 if (!numtype) {
2573 if (ckWARN(WARN_NUMERIC))
2574 not_a_number(sv);
2575 }
8d919b0a 2576 return U_V(Atof(ptr));
e20b6c3b 2577 }
4bac9ae4
CS
2578 }
2579
2580 if (SvTHINKFIRST(sv)) {
0336b60e 2581 if (SvREADONLY(sv) && !SvOK(sv)) {
0336b60e 2582 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2583 report_uninit(sv);
ff68c719
PP
2584 return 0;
2585 }
2586 }
4bac9ae4 2587
af359546
NC
2588 if (!SvIOKp(sv)) {
2589 if (S_sv_2iuv_common(aTHX_ sv))
2590 return 0;
ff68c719 2591 }
25da4f38 2592
1d7c1841
GS
2593 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2594 PTR2UV(sv),SvUVX(sv)));
25da4f38 2595 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
ff68c719
PP
2596}
2597
645c22ef 2598/*
196007d1 2599=for apidoc sv_2nv_flags
645c22ef
DM
2600
2601Return the num value of an SV, doing any necessary string or integer
fde67290 2602conversion. If flags includes SV_GMAGIC, does an mg_get() first.
39d5de13 2603Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
645c22ef
DM
2604
2605=cut
2606*/
2607
65202027 2608NV
5aaab254 2609Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
79072805 2610{
1061065f
DD
2611 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2612
217f6fa3
FC
2613 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2614 && SvTYPE(sv) != SVt_PVFM);
8d919b0a 2615 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2b2b6d6d 2616 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
cd84013a
FC
2617 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2618 Regexps have no SvIVX and SvNVX fields. */
8d919b0a 2619 const char *ptr;
39d5de13
DM
2620 if (flags & SV_GMAGIC)
2621 mg_get(sv);
463ee0b2
LW
2622 if (SvNOKp(sv))
2623 return SvNVX(sv);
cd84013a 2624 if (SvPOKp(sv) && !SvIOKp(sv)) {
8d919b0a
FC
2625 ptr = SvPVX_const(sv);
2626 grokpv:
041457d9 2627 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
8d919b0a 2628 !grok_number(ptr, SvCUR(sv), NULL))
a0d0e21e 2629 not_a_number(sv);
8d919b0a 2630 return Atof(ptr);
a0d0e21e 2631 }
25da4f38 2632 if (SvIOKp(sv)) {
1c846c1f 2633 if (SvIsUV(sv))
65202027 2634 return (NV)SvUVX(sv);
25da4f38 2635 else
65202027 2636 return (NV)SvIVX(sv);
47a72cb8
NC
2637 }
2638 if (SvROK(sv)) {
2639 goto return_rok;
2640 }
8d919b0a
FC
2641 if (isREGEXP(sv)) {
2642 ptr = RX_WRAPPED((REGEXP *)sv);
2643 goto grokpv;
2644 }
47a72cb8
NC
2645 assert(SvTYPE(sv) >= SVt_PVMG);
2646 /* This falls through to the report_uninit near the end of the
2647 function. */
2648 } else if (SvTHINKFIRST(sv)) {
a0d0e21e 2649 if (SvROK(sv)) {
47a72cb8 2650 return_rok:
deb46114 2651 if (SvAMAGIC(sv)) {
aee036bb
DM
2652 SV *tmpstr;
2653 if (flags & SV_SKIP_OVERLOAD)
2654 return 0;
31d632c3 2655 tmpstr = AMG_CALLunary(sv, numer_amg);
deb46114
NC
2656 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2657 return SvNV(tmpstr);
2658 }
2659 }
2660 return PTR2NV(SvRV(sv));
a0d0e21e 2661 }
0336b60e 2662 if (SvREADONLY(sv) && !SvOK(sv)) {
599cee73 2663 if (ckWARN(WARN_UNINITIALIZED))
29489e7c 2664 report_uninit(sv);
ed6116ce
LW
2665 return 0.0;
2666 }
79072805
LW
2667 }
2668 if (SvTYPE(sv) < SVt_NV) {
7e25a7e9
NC
2669 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2670 sv_upgrade(sv, SVt_NV);
097ee67d 2671 DEBUG_c({
f93f4e46 2672 STORE_NUMERIC_LOCAL_SET_STANDARD();
1d7c1841 2673 PerlIO_printf(Perl_debug_log,
88cb8500 2674 "0x%"UVxf" num(%" NVgf ")\n",
1d7c1841 2675 PTR2UV(sv), SvNVX(sv));
097ee67d
JH
2676 RESTORE_NUMERIC_LOCAL();
2677 });
79072805
LW
2678 }
2679 else if (SvTYPE(sv) < SVt_PVNV)
2680 sv_upgrade(sv, SVt_PVNV);
59d8ce62
NC
2681 if (SvNOKp(sv)) {
2682 return SvNVX(sv);
61604483 2683 }
59d8ce62 2684 if (SvIOKp(sv)) {
9d6ce603 2685 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
28e5dec8 2686#ifdef NV_PRESERVES_UV
a43d94f2
NC
2687 if (SvIOK(sv))
2688 SvNOK_on(sv);
2689 else
2690 SvNOKp_on(sv);
28e5dec8
JH
2691#else
2692 /* Only set the public NV OK flag if this NV preserves the IV */
2693 /* Check it's not 0xFFFFFFFFFFFFFFFF */
a43d94f2
NC
2694 if (SvIOK(sv) &&
2695 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
28e5dec8
JH
2696 : (SvIVX(sv) == I_V(SvNVX(sv))))
2697 SvNOK_on(sv);
2698 else
2699 SvNOKp_on(sv);
2700#endif
93a17b20 2701 }
cd84013a 2702 else if (SvPOKp(sv)) {
c2988b20 2703 UV value;
3823048b 2704 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
041457d9 2705 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
a0d0e21e 2706 not_a_number(sv);
28e5dec8 2707#ifdef NV_PRESERVES_UV
c2988b20
NC
2708 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2709 == IS_NUMBER_IN_UV) {
5e045b90 2710 /* It's definitely an integer */
9d6ce603 2711 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
66d83377 2712 } else {
3823048b 2713 S_sv_setnv(aTHX_ sv, numtype);
66d83377 2714 }
a43d94f2
NC
2715 if (numtype)
2716 SvNOK_on(sv);
2717 else
2718 SvNOKp_on(sv);
28e5dec8 2719#else
e91de695
JH
2720 SvNV_set(sv, Atof(SvPVX_const(sv)));
2721 /* Only set the public NV OK flag if this NV preserves the value in
2722 the PV at least as well as an IV/UV would.
2723 Not sure how to do this 100% reliably. */
2724 /* if that shift count is out of range then Configure's test is
2725 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2726 UV_BITS */
2727 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2728 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2729 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2730 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2731 /* Can't use strtol etc to convert this string, so don't try.
2732 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
c2988b20
NC
2733 SvNOK_on(sv);
2734 } else {
e91de695 2735 /* value has been set. It may not be precise. */
53e2bfb7 2736 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
e91de695
JH
2737 /* 2s complement assumption for (UV)IV_MIN */
2738 SvNOK_on(sv); /* Integer is too negative. */
c2988b20 2739 } else {
e91de695
JH
2740 SvNOKp_on(sv);
2741 SvIOKp_on(sv);
6fa402ec 2742
e91de695 2743 if (numtype & IS_NUMBER_NEG) {
02b08bbc
DM
2744 /* -IV_MIN is undefined, but we should never reach
2745 * this point with both IS_NUMBER_NEG and value ==
2746 * (UV)IV_MIN */
2747 assert(value != (UV)IV_MIN);
e91de695
JH
2748 SvIV_set(sv, -(IV)value);
2749 } else if (value <= (UV)IV_MAX) {
2750 SvIV_set(sv, (IV)value);
2751 } else {
2752 SvUV_set(sv, value);
2753 SvIsUV_on(sv);
2754 }
c2988b20 2755
e91de695
JH
2756 if (numtype & IS_NUMBER_NOT_INT) {
2757 /* I believe that even if the original PV had decimals,
2758 they are lost beyond the limit of the FP precision.
2759 However, neither is canonical, so both only get p
2760 flags. NWC, 2000/11/25 */
2761 /* Both already have p flags, so do nothing */
2762 } else {
2763 const NV nv = SvNVX(sv);
2764 /* XXX should this spot have NAN_COMPARE_BROKEN, too? */
2765 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2766 if (SvIVX(sv) == I_V(nv)) {
2767 SvNOK_on(sv);
2768 } else {
2769 /* It had no "." so it must be integer. */
2770 }
2771 SvIOK_on(sv);
0f83c5a4 2772 } else {
e91de695
JH
2773 /* between IV_MAX and NV(UV_MAX).
2774 Could be slightly > UV_MAX */
2775
2776 if (numtype & IS_NUMBER_NOT_INT) {
2777 /* UV and NV both imprecise. */
0f83c5a4 2778 } else {
e91de695
JH
2779 const UV nv_as_uv = U_V(nv);
2780
2781 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2782 SvNOK_on(sv);
c2988b20 2783 }
e91de695 2784 SvIOK_on(sv);
c2988b20
NC
2785 }
2786 }
2787 }
2788 }
0f83c5a4 2789 }
e91de695
JH
2790 /* It might be more code efficient to go through the entire logic above
2791 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2792 gets complex and potentially buggy, so more programmer efficient
2793 to do it this way, by turning off the public flags: */
2794 if (!numtype)
2795 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
28e5dec8 2796#endif /* NV_PRESERVES_UV */
93a17b20 2797 }
79072805 2798 else {
e91de695
JH
2799 if (isGV_with_GP(sv)) {
2800 glob_2number(MUTABLE_GV(sv));
2801 return 0.0;
2802 }
180488f8 2803
e91de695
JH
2804 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2805 report_uninit(sv);
2806 assert (SvTYPE(sv) >= SVt_NV);
2807 /* Typically the caller expects that sv_any is not NULL now. */
2808 /* XXX Ilya implies that this is a bug in callers that assume this
2809 and ideally should be fixed. */
2810 return 0.0;
79072805 2811 }
097ee67d 2812 DEBUG_c({
e91de695
JH
2813 STORE_NUMERIC_LOCAL_SET_STANDARD();
2814 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" NVgf ")\n",
2815 PTR2UV(sv), SvNVX(sv));
2816 RESTORE_NUMERIC_LOCAL();
2817 });
463ee0b2 2818 return SvNVX(sv);
79072805
LW
2819}
2820
800401ee
JH
2821/*
2822=for apidoc sv_2num
2823
2824Return an SV with the numeric value of the source SV, doing any necessary
d024d1a7
FC
2825reference or overload conversion. The caller is expected to have handled
2826get-magic already.
800401ee
JH
2827
2828=cut
2829*/
2830
2831SV *
5aaab254 2832Perl_sv_2num(pTHX_ SV *const sv)
800401ee 2833{
7918f24d
NC
2834 PERL_ARGS_ASSERT_SV_2NUM;
2835
b9ee0594
RGS
2836 if (!SvROK(sv))
2837 return sv;
800401ee 2838 if (SvAMAGIC(sv)) {
31d632c3 2839 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
a02ec77a 2840 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
800401ee
JH
2841 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2842 return sv_2num(tmpsv);
2843 }
2844 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2845}
2846
645c22ef
DM
2847/* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2848 * UV as a string towards the end of buf, and return pointers to start and
2849 * end of it.
2850 *
2851 * We assume that buf is at least TYPE_CHARS(UV) long.
2852 */
2853
864dbfa3 2854static char *
5de3775c 2855S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
25da4f38 2856{
25da4f38 2857 char *ptr = buf + TYPE_CHARS(UV);
823a54a3 2858 char * const ebuf = ptr;
25da4f38 2859 int sign;
25da4f38 2860
7918f24d
NC
2861 PERL_ARGS_ASSERT_UIV_2BUF;
2862
25da4f38
IZ
2863 if (is_uv)
2864 sign = 0;
2865 else if (iv >= 0) {
2866 uv = iv;
2867 sign = 0;
2868 } else {
53e2bfb7 2869 uv = (iv == IV_MIN) ? (UV)iv : (UV)(-iv);
25da4f38
IZ
2870 sign = 1;
2871 }
2872 do {
eb160463 2873 *--ptr = '0' + (char)(uv % 10);
25da4f38
IZ
2874 } while (uv /= 10);
2875 if (sign)
2876 *--ptr = '-';
2877 *peob = ebuf;
2878 return ptr;
2879}
2880
bfaa02d5
JH
2881/* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
2882 * infinity or a not-a-number, writes the appropriate strings to the
2883 * buffer, including a zero byte. On success returns the written length,
3bde2d43
JH
2884 * excluding the zero byte, on failure (not an infinity, not a nan)
2885 * returns zero, assert-fails on maxlen being too short.
3823048b
JH
2886 *
2887 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2888 * shared string constants we point to, instead of generating a new
2889 * string for each instance. */
bfaa02d5 2890STATIC size_t
3823048b 2891S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
3bde2d43 2892 char* s = buffer;
bfaa02d5 2893 assert(maxlen >= 4);
3bde2d43
JH
2894 if (Perl_isinf(nv)) {
2895 if (nv < 0) {
2896 if (maxlen < 5) /* "-Inf\0" */
2897 return 0;
2898 *s++ = '-';
2899 } else if (plus) {
2900 *s++ = '+';
6e915616 2901 }
3bde2d43
JH
2902 *s++ = 'I';
2903 *s++ = 'n';
2904 *s++ = 'f';
2905 }
2906 else if (Perl_isnan(nv)) {
2907 *s++ = 'N';
2908 *s++ = 'a';
2909 *s++ = 'N';
2910 /* XXX optionally output the payload mantissa bits as
2911 * "(unsigned)" (to match the nan("...") C99 function,
2912 * or maybe as "(0xhhh...)" would make more sense...
2913 * provide a format string so that the user can decide?
2914 * NOTE: would affect the maxlen and assert() logic.*/
2915 }
2916 else {
2917 return 0;
bfaa02d5 2918 }
3bde2d43
JH
2919 assert((s == buffer + 3) || (s == buffer + 4));
2920 *s++ = 0;
2921 return s - buffer - 1; /* -1: excluding the zero byte */
bfaa02d5
JH
2922}
2923
2924/*
2925=for apidoc sv_2pv_flags
2926
2927Returns a pointer to the string value of an SV, and sets *lp to its length.
2928If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2929string if necessary. Normally invoked via the C<SvPV_flags> macro.
2930C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2931
2932=cut
2933*/
2934
2935char *
2936Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2937{
2938 char *s;
2939
2940 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2941
2942 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2943 && SvTYPE(sv) != SVt_PVFM);
2944 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2945 mg_get(sv);
2946 if (SvROK(sv)) {
2947 if (SvAMAGIC(sv)) {
2948 SV *tmpstr;
2949 if (flags & SV_SKIP_OVERLOAD)
2950 return NULL;
2951 tmpstr = AMG_CALLunary(sv, string_amg);
2952 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2953 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2954 /* Unwrap this: */
2955 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2956 */
2957
2958 char *pv;
2959 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2960 if (flags & SV_CONST_RETURN) {
2961 pv = (char *) SvPVX_const(tmpstr);
2962 } else {
2963 pv = (flags & SV_MUTABLE_RETURN)
2964 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2965 }
2966 if (lp)
2967 *lp = SvCUR(tmpstr);
2968 } else {
2969 pv = sv_2pv_flags(tmpstr, lp, flags);
2970 }
2971 if (SvUTF8(tmpstr))
2972 SvUTF8_on(sv);
2973 else
2974 SvUTF8_off(sv);
2975 return pv;
2976 }
2977 }
2978 {
2979 STRLEN len;
2980 char *retval;
2981 char *buffer;
2982 SV *const referent = SvRV(sv);
2983
2984 if (!referent) {
2985 len = 7;
2986 retval = buffer = savepvn("NULLREF", len);
2987 } else if (SvTYPE(referent) == SVt_REGEXP &&
2988 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
2989 amagic_is_enabled(string_amg))) {
2990 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2991
2992 assert(re);
2993
2994 /* If the regex is UTF-8 we want the containing scalar to
2995 have an UTF-8 flag too */
2996 if (RX_UTF8(re))
2997 SvUTF8_on(sv);
2998 else
2999 SvUTF8_off(sv);
3000
3001 if (lp)
3002 *lp = RX_WRAPLEN(re);
3003
3004 return RX_WRAPPED(re);
3005 } else {
3006 const char *const typestr = sv_reftype(referent, 0);
3007 const STRLEN typelen = strlen(typestr);
3008 UV addr = PTR2UV(referent);
3009 const char *stashname = NULL;
3010 STRLEN stashnamelen = 0; /* hush, gcc */
3011 const char *buffer_end;
3012
3013 if (SvOBJECT(referent)) {
3014 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3015
3016 if (name) {
3017 stashname = HEK_KEY(name);
3018 stashnamelen = HEK_LEN(name);
3019
3020 if (HEK_UTF8(name)) {
3021 SvUTF8_on(sv);
3022 } else {
3023 SvUTF8_off(sv);
3024 }
3025 } else {
3026 stashname = "__ANON__";
3027 stashnamelen = 8;
3028 }
3029 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3030 + 2 * sizeof(UV) + 2 /* )\0 */;
3031 } else {
3032 len = typelen + 3 /* (0x */
3033 + 2 * sizeof(UV) + 2 /* )\0 */;
3034 }
fafee734 3035
4bac9ae4
CS
3036 Newx(buffer, len, char);
3037 buffer_end = retval = buffer + len;
3038
3039 /* Working backwards */
3040 *--retval = '\0';
3041 *--retval = ')';
3042 do {
3043 *--retval = PL_hexdigit[addr & 15];
3044 } while (addr >>= 4);
3045 *--retval = 'x';
3046 *--retval = '0';
3047 *--retval = '(';
3048
3049 retval -= typelen;
3050 memcpy(retval, typestr, typelen);
3051
3052 if (stashname) {
3053 *--retval = '=';
3054 retval -= stashnamelen;
3055 memcpy(retval, stashname, stashnamelen);
c080367d 3056 }
4bac9ae4
CS
3057 /* retval may not necessarily have reached the start of the
3058 buffer here. */
3059 assert (retval >= buffer);
3060
3061 len = buffer_end - retval - 1; /* -1 for that \0 */
463ee0b2 3062 }
cdb061a3 3063 if (lp)
4bac9ae4
CS
3064 *lp = len;
3065 SAVEFREEPV(buffer);
3066 return retval;
79072805 3067 }
79072805 3068 }
4bac9ae4
CS
3069
3070 if (SvPOKp(sv)) {
3071 if (lp)
3072 *lp = SvCUR(sv);
3073 if (flags & SV_MUTABLE_RETURN)
3074 return SvPVX_mutable(sv);
3075 if (flags & SV_CONST_RETURN)
3076 return (char *)SvPVX_const(sv);
3077 return SvPVX(sv);
3078 }
3079
3080 if (SvIOK(sv)) {
28e5dec8
JH
3081 /* I'm assuming that if both IV and NV are equally valid then
3082 converting the IV is going to be more efficient */
e1ec3a88 3083 const U32 isUIOK = SvIsUV(sv);
28e5dec8
JH
3084 char buf[TYPE_CHARS(UV)];
3085 char *ebuf, *ptr;
97a130b8 3086 STRLEN len;
28e5dec8
JH
3087
3088 if (SvTYPE(sv) < SVt_PVIV)
3089 sv_upgrade(sv, SVt_PVIV);
4ea1d550 3090 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
97a130b8 3091 len = ebuf - ptr;
5902b6a9 3092 /* inlined from sv_setpvn */
97a130b8
NC
3093 s = SvGROW_mutable(sv, len + 1);
3094 Move(ptr, s, len, char);
3095 s += len;
28e5dec8 3096 *s = '\0';
b127e37e 3097 SvPOK_on(sv);
28e5dec8 3098 }
4bac9ae4 3099 else if (SvNOK(sv)) {
79072805
LW
3100 if (SvTYPE(sv) < SVt_PVNV)
3101 sv_upgrade(sv, SVt_PVNV);
128eeacb
DD
3102 if (SvNVX(sv) == 0.0
3103#if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3104 && !Perl_isnan(SvNVX(sv))
3105#endif
3106 ) {
29912d93
NC
3107 s = SvGROW_mutable(sv, 2);
3108 *s++ = '0';
3109 *s = '\0';
3110 } else {
5e85836e 3111 STRLEN len;
fb8cdbc5 3112 STRLEN size = 5; /* "-Inf\0" */
0c7e610f 3113
fb8cdbc5 3114 s = SvGROW_mutable(sv, size);
3823048b 3115 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
fb8cdbc5 3116 if (len > 0) {
0c7e610f 3117 s += len;
fb8cdbc5
JH
3118 SvPOK_on(sv);
3119 }
0c7e610f 3120 else {
0c7e610f 3121 /* some Xenix systems wipe out errno here */
fb8cdbc5
JH
3122 dSAVE_ERRNO;
3123
3840bff0
JH
3124 size =
3125 1 + /* sign */
3126 1 + /* "." */
3127 NV_DIG +
3128 1 + /* "e" */
3129 1 + /* sign */
3130 5 + /* exponent digits */
3131 1 + /* \0 */
3132 2; /* paranoia */
b127e37e 3133
fb8cdbc5 3134 s = SvGROW_mutable(sv, size);
b127e37e 3135#ifndef USE_LOCALE_NUMERIC
a4eca1d4
JH
3136 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3137
0c7e610f
JH
3138 SvPOK_on(sv);
3139#else
28acfe03 3140 {
3840bff0 3141 bool local_radix;
67d796ae
KW
3142 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3143 STORE_LC_NUMERIC_SET_TO_NEEDED();
3840bff0
JH
3144
3145 local_radix =
3146 PL_numeric_local &&
3147 PL_numeric_radix_sv &&
3148 SvUTF8(PL_numeric_radix_sv);
3149 if (local_radix && SvLEN(PL_numeric_radix_sv) > 1) {
3150 size += SvLEN(PL_numeric_radix_sv) - 1;
3151 s = SvGROW_mutable(sv, size);
3152 }
3153
a4eca1d4 3154 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
0c7e610f
JH
3155
3156 /* If the radix character is UTF-8, and actually is in the
3157 * output, turn on the UTF-8 flag for the scalar */
3840bff0
JH
3158 if (local_radix &&
3159 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3160 SvUTF8_on(sv);
3161 }
3162
0c7e610f 3163 RESTORE_LC_NUMERIC();
28acfe03 3164 }
68e8f474 3165
0c7e610f
JH
3166 /* We don't call SvPOK_on(), because it may come to
3167 * pass that the locale changes so that the
3168 * stringification we just did is no longer correct. We
3169 * will have to re-stringify every time it is needed */
b127e37e 3170#endif
0c7e610f
JH
3171 RESTORE_ERRNO;
3172 }
3173 while (*s) s++;
bbce6d69 3174 }
79072805 3175 }
4bac9ae4
CS
3176 else if (isGV_with_GP(sv)) {
3177 GV *const gv = MUTABLE_GV(sv);
3178 SV *const buffer = sv_newmortal();
8d1c3e26 3179
4bac9ae4 3180 gv_efullname3(buffer, gv, "*");
180488f8 3181
4bac9ae4
CS
3182 assert(SvPOK(buffer));
3183 if (SvUTF8(buffer))
3184 SvUTF8_on(sv);
3185 if (lp)
3186 *lp = SvCUR(buffer);
3187 return SvPVX(buffer);
3188 }
8d919b0a
FC
3189 else if (isREGEXP(sv)) {
3190 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3191 return RX_WRAPPED((REGEXP *)sv);
3192 }
4bac9ae4 3193 else {
cdb061a3 3194 if (lp)
00b6aa41 3195 *lp = 0;
9f621bb0
NC
3196 if (flags & SV_UNDEF_RETURNS_NULL)
3197 return NULL;
4f62cd62 3198 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
9f621bb0 3199 report_uninit(sv);
4bac9ae4
CS
3200 /* Typically the caller expects that sv_any is not NULL now. */
3201 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
25da4f38 3202 sv_upgrade(sv, SVt_PV);
73d840c0 3203 return (char *)"";
79072805 3204 }
4bac9ae4 3205
cdb061a3 3206 {
823a54a3 3207 const STRLEN len = s - SvPVX_const(sv);
cdb061a3
NC
3208 if (lp)
3209 *lp = len;
3210 SvCUR_set(sv, len);
3211 }
1d7c1841 3212 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3f7c398e 3213 PTR2UV(sv),SvPVX_const(sv)));
4d84ee25
NC
3214 if (flags & SV_CONST_RETURN)
3215 return (char *)SvPVX_const(sv);
10516c54
NC
3216 if (flags & SV_MUTABLE_RETURN)
3217 return SvPVX_mutable(sv);
463ee0b2
LW
3218 return SvPVX(sv);
3219}
3220
645c22ef 3221/*
6050d10e
JP
3222=for apidoc sv_copypv
3223
3224Copies a stringified representation of the source SV into the
3225destination SV. Automatically performs any necessary mg_get and
54f0641b 3226coercion of numeric values into strings. Guaranteed to preserve
2575c402 3227UTF8 flag even from overloaded objects. Similar in nature to
54f0641b
NIS
3228sv_2pv[_flags] but operates directly on an SV instead of just the
3229string. Mostly uses sv_2pv_flags to do its work, except when that
6050d10e
JP
3230would lose the UTF-8'ness of the PV.
3231
4bac9ae4
CS
3232=for apidoc sv_copypv_nomg
3233
3234Like sv_copypv, but doesn't invoke get magic first.
3235
3236=for apidoc sv_copypv_flags
3237
3238Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3239include SV_GMAGIC.
3240
6050d10e
JP
3241=cut
3242*/
3243
3244void
5aaab254 3245Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
4bac9ae4 3246{
446eaa42 3247 STRLEN len;
4bac9ae4 3248 const char *s;
7918f24d 3249
4bac9ae4 3250 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
7918f24d 3251
c77ed9ca 3252 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
cb50f42d 3253 sv_setpvn(dsv,s,len);
446eaa42 3254 if (SvUTF8(ssv))
cb50f42d 3255 SvUTF8_on(dsv);
446eaa42 3256 else
cb50f42d 3257 SvUTF8_off(dsv);
6050d10e
JP
3258}
3259
3260/*
645c22ef
DM
3261=for apidoc sv_2pvbyte
3262
3263Return a pointer to the byte-encoded representation of the SV, and set *lp
1e54db1a 3264to its length. May cause the SV to be downgraded from UTF-8 as a
645c22ef
DM
3265side-effect.
3266
3267Usually accessed via the C<SvPVbyte> macro.
3268
3269=cut
3270*/
3271
7340a771 3272char *
5aaab254 3273Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3274{
7918f24d
NC
3275 PERL_ARGS_ASSERT_SV_2PVBYTE;
3276
48120f8f 3277 SvGETMAGIC(sv);
4499db73
FC
3278 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3279 || isGV_with_GP(sv) || SvROK(sv)) {
a901b181 3280 SV *sv2 = sv_newmortal();
48120f8f 3281 sv_copypv_nomg(sv2,sv);
a901b181
FC
3282 sv = sv2;
3283 }
0875d2fe 3284 sv_utf8_downgrade(sv,0);
71eb6d8c 3285 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771
GS
3286}
3287
645c22ef 3288/*
035cbb0e
RGS
3289=for apidoc sv_2pvutf8
3290
3291Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3292to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3293
3294Usually accessed via the C<SvPVutf8> macro.
3295
3296=cut
3297*/
645c22ef 3298
7340a771 3299char *
5aaab254 3300Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
7340a771 3301{
7918f24d
NC
3302 PERL_ARGS_ASSERT_SV_2PVUTF8;
3303
4499db73
FC
3304 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3305 || isGV_with_GP(sv) || SvROK(sv))
fe46cbda 3306 sv = sv_mortalcopy(sv);
4bac9ae4
CS
3307 else
3308 SvGETMAGIC(sv);
3309 sv_utf8_upgrade_nomg(sv);
c3ec315f 3310 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
7340a771 3311}
1c846c1f 3312
7ee2227d 3313
645c22ef
DM
3314/*
3315=for apidoc sv_2bool
3316
06c841cf
FC
3317This macro is only used by sv_true() or its macro equivalent, and only if
3318the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3319It calls sv_2bool_flags with the SV_GMAGIC flag.
3320
3321=for apidoc sv_2bool_flags
3322
3323This function is only used by sv_true() and friends, and only if
fde67290 3324the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
06c841cf
FC
3325contain SV_GMAGIC, then it does an mg_get() first.
3326
645c22ef
DM
3327
3328=cut
3329*/
3330
463ee0b2 3331bool
9d176cd8 3332Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
463ee0b2 3333{
06c841cf 3334 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
7918f24d 3335
9d176cd8 3336 restart:
06c841cf 3337 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
463ee0b2 3338
a0d0e21e
LW
3339 if (!SvOK(sv))
3340 return 0;
3341 if (SvROK(sv)) {
fabdb6c0 3342 if (SvAMAGIC(sv)) {
31d632c3 3343 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
9d176cd8
DD
3344 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3345 bool svb;
3346 sv = tmpsv;
3347 if(SvGMAGICAL(sv)) {
3348 flags = SV_GMAGIC;
3349 goto restart; /* call sv_2bool */
3350 }
3351 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3352 else if(!SvOK(sv)) {
3353 svb = 0;
3354 }
3355 else if(SvPOK(sv)) {
3356 svb = SvPVXtrue(sv);
3357 }
3358 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3359 svb = (SvIOK(sv) && SvIVX(sv) != 0)
659c4b96 3360 || (SvNOK(sv) && SvNVX(sv) != 0.0);
9d176cd8
DD
3361 }
3362 else {
3363 flags = 0;
3364 goto restart; /* call sv_2bool_nomg */
3365 }
3366 return cBOOL(svb);
3367 }
fabdb6c0
AL
3368 }
3369 return SvRV(sv) != 0;
a0d0e21e 3370 }
85b7d9b3
FC
3371 if (isREGEXP(sv))
3372 return
3373 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
4bac9ae4 3374 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
79072805
LW
3375}
3376
c461cf8f
JH
3377/*
3378=for apidoc sv_utf8_upgrade
3379
78ea37eb 3380Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3381Forces the SV to string form if it is not already.
2bbc8d55 3382Will C<mg_get> on C<sv> if appropriate.
4411f3b6 3383Always sets the SvUTF8 flag to avoid future validity checks even
2bbc8d55
SP
3384if the whole string is the same in UTF-8 as not.
3385Returns the number of bytes in the converted string
c461cf8f 3386
0efd0472 3387This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3388use the Encode extension for that.
3389
fe749c9a
KW
3390=for apidoc sv_utf8_upgrade_nomg
3391
fde67290 3392Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
fe749c9a 3393
8d6d96c1
HS
3394=for apidoc sv_utf8_upgrade_flags
3395
78ea37eb 3396Converts the PV of an SV to its UTF-8-encoded form.
645c22ef 3397Forces the SV to string form if it is not already.
8d6d96c1 3398Always sets the SvUTF8 flag to avoid future validity checks even
960b0271
FC
3399if all the bytes are invariant in UTF-8.
3400If C<flags> has C<SV_GMAGIC> bit set,
2bbc8d55 3401will C<mg_get> on C<sv> if appropriate, else not.
2a590426
KW
3402
3403If C<flags> has SV_FORCE_UTF8_UPGRADE set, this function assumes that the PV
3404will expand when converted to UTF-8, and skips the extra work of checking for
3405that. Typically this flag is used by a routine that has already parsed the
3406string and found such characters, and passes this information on so that the
3407work doesn't have to be repeated.
3408
3409Returns the number of bytes in the converted string.
8d6d96c1 3410
0efd0472 3411This is not a general purpose byte encoding to Unicode interface:
13a6c0e0
JH
3412use the Encode extension for that.
3413
2a590426 3414=for apidoc sv_utf8_upgrade_flags_grow
b3ab6785 3415
2a590426
KW
3416Like sv_utf8_upgrade_flags, but has an additional parameter C<extra>, which is
3417the number of unused bytes the string of 'sv' is guaranteed to have free after
3418it upon return. This allows the caller to reserve extra space that it intends
3419to fill, to avoid extra grows.
b3ab6785 3420
2a590426
KW
3421C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3422are implemented in terms of this function.
3423
3424Returns the number of bytes in the converted string (not including the spares).
3425
3426=cut
b3ab6785
KW
3427
3428(One might think that the calling routine could pass in the position of the
2a590426
KW
3429first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3430have to be found again. But that is not the case, because typically when the
3431caller is likely to use this flag, it won't be calling this routine unless it
3432finds something that won't fit into a byte. Otherwise it tries to not upgrade
3433and just use bytes. But some things that do fit into a byte are variants in
3434utf8, and the caller may not have been keeping track of these.)
b3ab6785 3435
6602b933
KW
3436If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3437C<NUL> isn't guaranteed due to having other routines do the work in some input
3438cases, or if the input is already flagged as being in utf8.
b3ab6785
KW
3439
3440The speed of this could perhaps be improved for many cases if someone wanted to
3441write a fast function that counts the number of variant characters in a string,
3442especially if it could return the position of the first one.
3443
8d6d96c1
HS
3444*/
3445
3446STRLEN
5aaab254 3447Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
8d6d96c1 3448{
b3ab6785 3449 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
7918f24d 3450
808c356f
RGS
3451 if (sv == &PL_sv_undef)
3452 return 0;
892f9127 3453 if (!SvPOK_nog(sv)) {
e0e62c2a 3454 STRLEN len = 0;
d52b7888
NC
3455 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3456 (void) sv_2pv_flags(sv,&len, flags);
b3ab6785
KW
3457 if (SvUTF8(sv)) {
3458 if (extra) SvGROW(sv, SvCUR(sv) + extra);
d52b7888 3459 return len;
b3ab6785 3460 }
d52b7888 3461 } else {
33fb6f35 3462 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
d52b7888 3463 }
e0e62c2a 3464 }
4411f3b6 3465
f5cee72b 3466 if (SvUTF8(sv)) {
b3ab6785 3467 if (extra) SvGROW(sv, SvCUR(sv) + extra);
5fec3b1d 3468 return SvCUR(sv);
f5cee72b 3469 }
5fec3b1d 3470
765f542d 3471 if (SvIsCOW(sv)) {
c56ed9f6 3472 S_sv_uncow(aTHX_ sv, 0);
db42d148
NIS
3473 }
3474
47e13f24 3475 if (IN_ENCODING && !(flags & SV_UTF8_NO_ENCODING)) {
ad2de1b2 3476 sv_recode_to_utf8(sv, _get_encoding());
b3ab6785
KW
3477 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3478 return SvCUR(sv);
3479 }
3480
4e93345f
KW
3481 if (SvCUR(sv) == 0) {
3482 if (extra) SvGROW(sv, extra);
3483 } else { /* Assume Latin-1/EBCDIC */
c4e7c712 3484 /* This function could be much more efficient if we
2bbc8d55 3485 * had a FLAG in SVs to signal if there are any variant
c4e7c712 3486 * chars in the PV. Given that there isn't such a flag
b3ab6785
KW
3487 * make the loop as fast as possible (although there are certainly ways
3488 * to speed this up, eg. through vectorization) */
3489 U8 * s = (U8 *) SvPVX_const(sv);
3490 U8 * e = (U8 *) SvEND(sv);
3491 U8 *t = s;
3492 STRLEN two_byte_count = 0;
c4e7c712 3493
b3ab6785
KW
3494 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3495
3496 /* See if really will need to convert to utf8. We mustn't rely on our
3497 * incoming SV being well formed and having a trailing '\0', as certain
3498 * code in pp_formline can send us partially built SVs. */
3499
c4e7c712 3500 while (t < e) {
53c1dcc0 3501 const U8 ch = *t++;
6f2d5cbc 3502 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
b3ab6785
KW
3503
3504 t--; /* t already incremented; re-point to first variant */
3505 two_byte_count = 1;
3506 goto must_be_utf8;
c4e7c712 3507 }
b3ab6785
KW
3508
3509 /* utf8 conversion not needed because all are invariants. Mark as
3510 * UTF-8 even if no variant - saves scanning loop */
c4e7c712 3511 SvUTF8_on(sv);
7f0bfbea 3512 if (extra) SvGROW(sv, SvCUR(sv) + extra);
b3ab6785
KW
3513 return SvCUR(sv);
3514
7b52d656 3515 must_be_utf8:
b3ab6785
KW
3516
3517 /* Here, the string should be converted to utf8, either because of an
3518 * input flag (two_byte_count = 0), or because a character that
3519 * requires 2 bytes was found (two_byte_count = 1). t points either to
3520 * the beginning of the string (if we didn't examine anything), or to
3521 * the first variant. In either case, everything from s to t - 1 will
3522 * occupy only 1 byte each on output.
3523 *
3524 * There are two main ways to convert. One is to create a new string
3525 * and go through the input starting from the beginning, appending each
3526 * converted value onto the new string as we go along. It's probably
3527 * best to allocate enough space in the string for the worst possible
3528 * case rather than possibly running out of space and having to
3529 * reallocate and then copy what we've done so far. Since everything
3530 * from s to t - 1 is invariant, the destination can be initialized
3531 * with these using a fast memory copy
3532 *
3533 * The other way is to figure out exactly how big the string should be
3534 * by parsing the entire input. Then you don't have to make it big
3535 * enough to handle the worst possible case, and more importantly, if
3536 * the string you already have is large enough, you don't have to
3537 * allocate a new string, you can copy the last character in the input
3538 * string to the final position(s) that will be occupied by the
3539 * converted string and go backwards, stopping at t, since everything
3540 * before that is invariant.
3541 *
3542 * There are advantages and disadvantages to each method.
3543 *
3544 * In the first method, we can allocate a new string, do the memory
3545 * copy from the s to t - 1, and then proceed through the rest of the
3546 * string byte-by-byte.
3547 *
3548 * In the second method, we proceed through the rest of the input
3549 * string just calculating how big the converted string will be. Then
3550 * there are two cases:
3551 * 1) if the string has enough extra space to handle the converted
3552 * value. We go backwards through the string, converting until we
3553 * get to the position we are at now, and then stop. If this
3554 * position is far enough along in the string, this method is
3555 * faster than the other method. If the memory copy were the same
3556 * speed as the byte-by-byte loop, that position would be about
3557 * half-way, as at the half-way mark, parsing to the end and back
3558 * is one complete string's parse, the same amount as starting
3559 * over and going all the way through. Actually, it would be
3560 * somewhat less than half-way, as it's faster to just count bytes
3561 * than to also copy, and we don't have the overhead of allocating
3562 * a new string, changing the scalar to use it, and freeing the
3563 * existing one. But if the memory copy is fast, the break-even
3564 * point is somewhere after half way. The counting loop could be
3565 * sped up by vectorization, etc, to move the break-even point
3566 * further towards the beginning.
3567 * 2) if the string doesn't have enough space to handle the converted
3568 * value. A new string will have to be allocated, and one might
3569 * as well, given that, start from the beginning doing the first
3570 * method. We've spent extra time parsing the string and in
3571 * exchange all we've gotten is that we know precisely how big to
3572 * make the new one. Perl is more optimized for time than space,
3573 * so this case is a loser.
3574 * So what I've decided to do is not use the 2nd method unless it is
3575 * guaranteed that a new string won't have to be allocated, assuming
3576 * the worst case. I also decided not to put any more conditions on it
3577 * than this, for now. It seems likely that, since the worst case is
3578 * twice as big as the unknown portion of the string (plus 1), we won't
3579 * be guaranteed enough space, causing us to go to the first method,
3580 * unless the string is short, or the first variant character is near
3581 * the end of it. In either of these cases, it seems best to use the
3582 * 2nd method. The only circumstance I can think of where this would
3583 * be really slower is if the string had once had much more data in it
3584 * than it does now, but there is still a substantial amount in it */
3585
3586 {
3587 STRLEN invariant_head = t - s;
3588 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3589 if (SvLEN(sv) < size) {
3590
3591 /* Here, have decided to allocate a new string */
3592
3593 U8 *dst;
3594 U8 *d;
3595
3596 Newx(dst, size, U8);
3597
3598 /* If no known invariants at the beginning of the input string,
3599 * set so starts from there. Otherwise, can use memory copy to
3600 * get up to where we are now, and then start from here */
3601
5b26a7b3 3602 if (invariant_head == 0) {
b3ab6785
KW
3603 d = dst;
3604 } else {
3605 Copy(s, dst, invariant_head, char);
3606 d = dst + invariant_head;
3607 }
3608
3609 while (t < e) {
55d09dc8
KW
3610 append_utf8_from_native_byte(*t, &d);
3611 t++;
b3ab6785
KW
3612 }
3613 *d = '\0';
3614 SvPV_free(sv); /* No longer using pre-existing string */
3615 SvPV_set(sv, (char*)dst);
3616 SvCUR_set(sv, d - dst);
3617 SvLEN_set(sv, size);
3618 } else {
3619
3620 /* Here, have decided to get the exact size of the string.
3621 * Currently this happens only when we know that there is
3622 * guaranteed enough space to fit the converted string, so