3 * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
4 * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 by Larry Wall
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.
13 * 'I wonder what the Entish is for "yes" and "no",' he thought.
16 * [p.480 of _The Lord of the Rings_, III/iv: "Treebeard"]
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
39 /* Missing proto on LynxOS */
40 char *gconvert(double, int, int, char *);
44 # define SNPRINTF_G(nv, buffer, size, ndig) \
45 quadmath_snprintf(buffer, size, "%.*Qg", (int)ndig, (NV)(nv))
47 # define SNPRINTF_G(nv, buffer, size, ndig) \
48 PERL_UNUSED_RESULT(Gconvert((NV)(nv), (int)ndig, 0, buffer))
51 #ifndef SV_COW_THRESHOLD
52 # define SV_COW_THRESHOLD 0 /* COW iff len > K */
54 #ifndef SV_COWBUF_THRESHOLD
55 # define SV_COWBUF_THRESHOLD 1250 /* COW iff len > K */
57 #ifndef SV_COW_MAX_WASTE_THRESHOLD
58 # define SV_COW_MAX_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
60 #ifndef SV_COWBUF_WASTE_THRESHOLD
61 # define SV_COWBUF_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
63 #ifndef SV_COW_MAX_WASTE_FACTOR_THRESHOLD
64 # define SV_COW_MAX_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
66 #ifndef SV_COWBUF_WASTE_FACTOR_THRESHOLD
67 # define SV_COWBUF_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
69 /* Work around compiler warnings about unsigned >= THRESHOLD when thres-
72 # define GE_COW_THRESHOLD(cur) ((cur) >= SV_COW_THRESHOLD)
74 # define GE_COW_THRESHOLD(cur) 1
76 #if SV_COWBUF_THRESHOLD
77 # define GE_COWBUF_THRESHOLD(cur) ((cur) >= SV_COWBUF_THRESHOLD)
79 # define GE_COWBUF_THRESHOLD(cur) 1
81 #if SV_COW_MAX_WASTE_THRESHOLD
82 # define GE_COW_MAX_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COW_MAX_WASTE_THRESHOLD)
84 # define GE_COW_MAX_WASTE_THRESHOLD(cur,len) 1
86 #if SV_COWBUF_WASTE_THRESHOLD
87 # define GE_COWBUF_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COWBUF_WASTE_THRESHOLD)
89 # define GE_COWBUF_WASTE_THRESHOLD(cur,len) 1
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))
94 # define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) 1
96 #if SV_COWBUF_WASTE_FACTOR_THRESHOLD
97 # define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COWBUF_WASTE_FACTOR_THRESHOLD * (cur))
99 # define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) 1
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)) \
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)) \
113 #ifdef PERL_UTF8_CACHE_ASSERT
114 /* if adding more checks watch out for the following tests:
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
119 # define ASSERT_UTF8_CACHE(cache) \
120 STMT_START { if (cache) { assert((cache)[0] <= (cache)[1]); \
121 assert((cache)[2] <= (cache)[3]); \
122 assert((cache)[3] <= (cache)[1]);} \
125 # define ASSERT_UTF8_CACHE(cache) NOOP
128 /* ============================================================================
130 =head1 Allocation and deallocation of SVs.
131 An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
132 sv, av, hv...) contains type and reference count information, and for
133 many types, a pointer to the body (struct xrv, xpv, xpviv...), which
134 contains fields specific to each type. Some types store all they need
135 in the head, so don't have a body.
137 In all but the most memory-paranoid configurations (ex: PURIFY), heads
138 and bodies are allocated out of arenas, which by default are
139 approximately 4K chunks of memory parcelled up into N heads or bodies.
140 Sv-bodies are allocated by their sv-type, guaranteeing size
141 consistency needed to allocate safely from arrays.
143 For SV-heads, the first slot in each arena is reserved, and holds a
144 link to the next arena, some flags, and a note of the number of slots.
145 Snaked through each arena chain is a linked list of free items; when
146 this becomes empty, an extra arena is allocated and divided up into N
147 items which are threaded into the free list.
149 SV-bodies are similar, but they use arena-sets by default, which
150 separate the link and info from the arena itself, and reclaim the 1st
151 slot in the arena. SV-bodies are further described later.
153 The following global variables are associated with arenas:
155 PL_sv_arenaroot pointer to list of SV arenas
156 PL_sv_root pointer to list of free SV structures
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
162 A few special SV heads are not allocated from an arena, but are
163 instead directly created in the interpreter structure, eg PL_sv_undef.
164 The size of arenas can be changed from the default by setting
165 PERL_ARENA_SIZE appropriately at compile time.
167 The SV arena serves the secondary purpose of allowing still-live SVs
168 to be located and destroyed during final cleanup.
170 At the lowest level, the macros new_SV() and del_SV() grab and free
171 an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
172 to return the SV to the free list with error checking.) new_SV() calls
173 more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
174 SVs in the free list have their SvTYPE field set to all ones.
176 At the time of very final cleanup, sv_free_arenas() is called from
177 perl_destruct() to physically free all the arenas allocated since the
178 start of the interpreter.
180 The function visit() scans the SV arenas list, and calls a specified
181 function for each SV it finds which is still live - ie which has an SvTYPE
182 other than all 1's, and a non-zero SvREFCNT. visit() is used by the
183 following functions (specified as [function that calls visit()] / [function
184 called by visit() for each SV]):
186 sv_report_used() / do_report_used()
187 dump all remaining SVs (debugging aid)
189 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
190 do_clean_named_io_objs(),do_curse()
191 Attempt to free all objects pointed to by RVs,
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
196 perl_destruct(), prior to calling sv_clean_all()
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.
209 =head2 Arena allocator API Summary
211 Private API to rest of sv.c
215 new_XPVNV(), del_XPVGV(),
220 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
224 * ========================================================================= */
227 * "A time to plant, and a time to uproot what was planted..."
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)
236 # define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
237 # define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
240 #ifdef DEBUG_LEAKING_SCALARS
241 # define FREE_SV_DEBUG_FILE(sv) STMT_START { \
242 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
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))
248 # define FREE_SV_DEBUG_FILE(sv)
249 # define DEBUG_SV_SERIAL(sv) NOOP
253 # define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
254 # define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
255 /* Whilst I'd love to do this, it seems that things like to check on
257 # define POISON_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
259 # define POISON_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
260 PoisonNew(&SvREFCNT(sv), 1, U32)
262 # define SvARENA_CHAIN(sv) SvANY(sv)
263 # define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
264 # define POISON_SV_HEAD(sv)
267 /* Mark an SV head as unused, and add to free list.
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. */
274 #define plant_SV(p) \
276 const U32 old_flags = SvFLAGS(p); \
277 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
278 DEBUG_SV_SERIAL(p); \
279 FREE_SV_DEBUG_FILE(p); \
281 SvFLAGS(p) = SVTYPEMASK; \
282 if (!(old_flags & SVf_BREAK)) { \
283 SvARENA_CHAIN_SET(p, PL_sv_root); \
289 #define uproot_SV(p) \
292 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
297 /* make some more SVs by adding another arena */
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);
310 /* new_SV(): return a new, empty SV head */
312 #ifdef DEBUG_LEAKING_SCALARS
313 /* provide a real function for a debugger to play with */
315 S_new_SV(pTHX_ const char *file, int line, const char *func)
322 sv = S_more_sv(aTHX);
326 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
327 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
333 sv->sv_debug_inpad = 0;
334 sv->sv_debug_parent = NULL;
335 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
337 sv->sv_debug_serial = PL_sv_serial++;
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));
345 # define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
353 (p) = S_more_sv(aTHX); \
357 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
362 /* del_SV(): return an empty SV head to the free list */
375 S_del_sv(pTHX_ SV *p)
377 PERL_ARGS_ASSERT_DEL_SV;
382 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
383 const SV * const sv = sva + 1;
384 const SV * const svend = &sva[SvREFCNT(sva)];
385 if (p >= sv && p < svend) {
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);
400 #else /* ! DEBUGGING */
402 #define del_SV(p) plant_SV(p)
404 #endif /* DEBUGGING */
407 * Bodyless IVs and NVs!
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.
423 * Since the 5.21 dev series, we employ the same trick for NVs
424 * if the architecture can support it (NVSIZE <= IVSIZE).
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))
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))
436 =head1 SV Manipulation Functions
438 =for apidoc sv_add_arena
440 Given a chunk of memory, link it to the head of the list of arenas,
441 and split it into a list of free SVs.
447 S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
449 SV *const sva = MUTABLE_SV(ptr);
453 PERL_ARGS_ASSERT_SV_ADD_ARENA;
455 /* The first SV in an arena isn't an SV. */
456 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
457 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
458 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
460 PL_sv_arenaroot = sva;
461 PL_sv_root = sva + 1;
463 svend = &sva[SvREFCNT(sva) - 1];
466 SvARENA_CHAIN_SET(sv, (sv + 1));
470 /* Must always set typemask because it's always checked in on cleanup
471 when the arenas are walked looking for objects. */
472 SvFLAGS(sv) = SVTYPEMASK;
475 SvARENA_CHAIN_SET(sv, 0);
479 SvFLAGS(sv) = SVTYPEMASK;
482 /* visit(): call the named function for each non-free SV in the arenas
483 * whose flags field matches the flags/mask args. */
486 S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
491 PERL_ARGS_ASSERT_VISIT;
493 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
494 const SV * const svend = &sva[SvREFCNT(sva)];
496 for (sv = sva + 1; sv < svend; ++sv) {
497 if (SvTYPE(sv) != (svtype)SVTYPEMASK
498 && (sv->sv_flags & mask) == flags
511 /* called by sv_report_used() for each live SV */
514 do_report_used(pTHX_ SV *const sv)
516 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
517 PerlIO_printf(Perl_debug_log, "****\n");
524 =for apidoc sv_report_used
526 Dump the contents of all SVs not yet freed (debugging aid).
532 Perl_sv_report_used(pTHX)
535 visit(do_report_used, 0, 0);
541 /* called by sv_clean_objs() for each live SV */
544 do_clean_objs(pTHX_ SV *const ref)
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);
558 SvREFCNT_dec_NN(target);
565 /* clear any slots in a GV which hold objects - except IO;
566 * called by sv_clean_objs() for each live GV */
569 do_clean_named_objs(pTHX_ SV *const sv)
572 assert(SvTYPE(sv) == SVt_PVGV);
573 assert(isGV_with_GP(sv));
577 /* freeing GP entries may indirectly free the current GV;
578 * hold onto it while we mess with the GP slots */
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)));
585 SvREFCNT_dec_NN(obj);
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)));
591 SvREFCNT_dec_NN(obj);
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)));
597 SvREFCNT_dec_NN(obj);
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)));
603 SvREFCNT_dec_NN(obj);
605 SvREFCNT_dec_NN(sv); /* undo the inc above */
608 /* clear any IO slots in a GV which hold objects (except stderr, defout);
609 * called by sv_clean_objs() for each live GV */
612 do_clean_named_io_objs(pTHX_ SV *const sv)
615 assert(SvTYPE(sv) == SVt_PVGV);
616 assert(isGV_with_GP(sv));
617 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
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)));
625 SvREFCNT_dec_NN(obj);
627 SvREFCNT_dec_NN(sv); /* undo the inc above */
630 /* Void wrapper to pass to visit() */
632 do_curse(pTHX_ SV * const sv) {
633 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
634 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
640 =for apidoc sv_clean_objs
642 Attempt to destroy all objects not yet freed.
648 Perl_sv_clean_objs(pTHX)
651 PL_in_clean_objs = TRUE;
652 visit(do_clean_objs, SVf_ROK, SVf_ROK);
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 */
656 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
657 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
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);
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);
670 PL_in_clean_objs = FALSE;
673 /* called by sv_clean_all() for each live SV */
676 do_clean_all(pTHX_ SV *const sv)
678 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
679 /* don't clean pid table and strtab */
682 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
683 SvFLAGS(sv) |= SVf_BREAK;
688 =for apidoc sv_clean_all
690 Decrement the refcnt of each remaining SV, possibly triggering a
691 cleanup. This function may have to be called multiple times to free
692 SVs which are in complex self-referential hierarchies.
698 Perl_sv_clean_all(pTHX)
701 PL_in_clean_all = TRUE;
702 cleaned = visit(do_clean_all, 0,0);
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
712 arena, avoiding the needless malloc overhead of a naive linked-list.
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
717 smaller types). The recovery of the wasted space allows use of
718 small arenas for large, rare body types, by changing array* fields
719 in body_details_by_type[] below.
722 char *arena; /* the raw storage, allocated aligned */
723 size_t size; /* its size ~4k typ */
724 svtype utype; /* bodytype stored in arena */
729 /* Get the maximum number of elements in set[] such that struct arena_set
730 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
731 therefore likely to be 1 aligned memory page. */
733 #define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
734 - 2 * sizeof(int)) / sizeof (struct arena_desc))
737 struct arena_set* next;
738 unsigned int set_size; /* ie ARENAS_PER_SET */
739 unsigned int curr; /* index of next available arena-desc */
740 struct arena_desc set[ARENAS_PER_SET];
744 =for apidoc sv_free_arenas
746 Deallocate the memory used by all arenas. Note that all the individual SV
747 heads and bodies within the arenas must already have been freed.
753 Perl_sv_free_arenas(pTHX)
759 /* Free arenas here, but be careful about fake ones. (We assume
760 contiguity of the fake ones with the corresponding real ones.) */
762 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
763 svanext = MUTABLE_SV(SvANY(sva));
764 while (svanext && SvFAKE(svanext))
765 svanext = MUTABLE_SV(SvANY(svanext));
772 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
775 struct arena_set *current = aroot;
778 assert(aroot->set[i].arena);
779 Safefree(aroot->set[i].arena);
787 i = PERL_ARENA_ROOTS_SIZE;
789 PL_body_roots[i] = 0;
796 Here are mid-level routines that manage the allocation of bodies out
797 of the various arenas. There are 5 kinds of arenas:
799 1. SV-head arenas, which are discussed and handled above
800 2. regular body arenas
801 3. arenas for reduced-size bodies
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)
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.
820 =head1 SV-Body Allocation
824 Allocation of SV-bodies is similar to SV-heads, differing as follows;
825 the allocation mechanism is used for many body types, so is somewhat
826 more complicated, it uses arena-sets, and has no need for still-live
829 At the outermost level, (new|del)_X*V macros return bodies of the
830 appropriate type. These macros call either (new|del)_body_type or
831 (new|del)_body_allocated macro pairs, depending on specifics of the
832 type. Most body types use the former pair, the latter pair is used to
833 allocate body types with "ghost fields".
835 "ghost fields" are fields that are unused in certain types, and
836 consequently don't need to actually exist. They are declared because
837 they're part of a "base type", which allows use of functions as
838 methods. The simplest examples are AVs and HVs, 2 aggregate types
839 which don't use the fields which support SCALAR semantics.
841 For these types, the arenas are carved up into appropriately sized
842 chunks, we thus avoid wasted memory for those unaccessed members.
843 When bodies are allocated, we adjust the pointer back in memory by the
844 size of the part not allocated, so it's as if we allocated the full
845 structure. (But things will all go boom if you write to the part that
846 is "not there", because you'll be overwriting the last members of the
847 preceding structure in memory.)
849 We calculate the correction using the STRUCT_OFFSET macro on the first
850 member present. If the allocated structure is smaller (no initial NV
851 actually allocated) then the net effect is to subtract the size of the NV
852 from the pointer, to return a new pointer as if an initial NV were actually
853 allocated. (We were using structures named *_allocated for this, but
854 this turned out to be a subtle bug, because a structure without an NV
855 could have a lower alignment constraint, but the compiler is allowed to
856 optimised accesses based on the alignment constraint of the actual pointer
857 to the full structure, for example, using a single 64 bit load instruction
858 because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
860 This is the same trick as was used for NV and IV bodies. Ironically it
861 doesn't need to be used for NV bodies any more, because NV is now at
862 the start of the structure. IV bodies, and also in some builds NV bodies,
863 don't need it either, because they are no longer allocated.
865 In turn, the new_body_* allocators call S_new_body(), which invokes
866 new_body_inline macro, which takes a lock, and takes a body off the
867 linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
868 necessary to refresh an empty list. Then the lock is released, and
869 the body is returned.
871 Perl_more_bodies allocates a new arena, and carves it up into an array of N
872 bodies, which it strings into a linked list. It looks up arena-size
873 and body-size from the body_details table described below, thus
874 supporting the multiple body-types.
876 If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
877 the (new|del)_X*V macros are mapped directly to malloc/free.
879 For each sv-type, struct body_details bodies_by_type[] carries
880 parameters which control these aspects of SV handling:
882 Arena_size determines whether arenas are used for this body type, and if
883 so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
884 zero, forcing individual mallocs and frees.
886 Body_size determines how big a body is, and therefore how many fit into
887 each arena. Offset carries the body-pointer adjustment needed for
888 "ghost fields", and is used in *_allocated macros.
890 But its main purpose is to parameterize info needed in
891 Perl_sv_upgrade(). The info here dramatically simplifies the function
892 vs the implementation in 5.8.8, making it table-driven. All fields
893 are used for this, except for arena_size.
895 For the sv-types that have no bodies, arenas are not used, so those
896 PL_body_roots[sv_type] are unused, and can be overloaded. In
897 something of a special case, SVt_NULL is borrowed for HE arenas;
898 PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
899 bodies_by_type[SVt_NULL] slot is not used, as the table is not
904 struct body_details {
905 U8 body_size; /* Size to allocate */
906 U8 copy; /* Size of structure to copy (may be shorter) */
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 */
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
924 #define HASARENA TRUE
926 #define NOARENA FALSE
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
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) \
944 ? FIT_ARENAn (count, body_size) \
945 : FIT_ARENA0 (body_size))
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. */
951 #define copy_length(type, last_member) \
952 STRUCT_OFFSET(type, last_member) \
953 + sizeof (((type*)SvANY((const SV *)0))->last_member)
955 static const struct body_details bodies_by_type[] = {
956 /* HEs use this offset for their arena. */
957 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
959 /* IVs are in the head, so the allocation size is 0. */
961 sizeof(IV), /* This is used to copy out the IV body. */
962 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
963 NOARENA /* IVS don't need an arena */, 0
968 STRUCT_OFFSET(XPVNV, xnv_u),
969 SVt_NV, FALSE, HADNV, NOARENA, 0 },
971 { sizeof(NV), sizeof(NV),
972 STRUCT_OFFSET(XPVNV, xnv_u),
973 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
976 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
977 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
978 + STRUCT_OFFSET(XPV, xpv_cur),
979 SVt_PV, FALSE, NONV, HASARENA,
980 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
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)) },
988 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
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)) },
994 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
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)) },
1000 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
1001 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
1006 SVt_REGEXP, TRUE, NONV, HASARENA,
1007 FIT_ARENA(0, sizeof(regexp))
1010 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
1011 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
1013 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
1014 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
1017 copy_length(XPVAV, xav_alloc),
1019 SVt_PVAV, TRUE, NONV, HASARENA,
1020 FIT_ARENA(0, sizeof(XPVAV)) },
1023 copy_length(XPVHV, xhv_max),
1025 SVt_PVHV, TRUE, NONV, HASARENA,
1026 FIT_ARENA(0, sizeof(XPVHV)) },
1031 SVt_PVCV, TRUE, NONV, HASARENA,
1032 FIT_ARENA(0, sizeof(XPVCV)) },
1037 SVt_PVFM, TRUE, NONV, NOARENA,
1038 FIT_ARENA(20, sizeof(XPVFM)) },
1043 SVt_PVIO, TRUE, NONV, HASARENA,
1044 FIT_ARENA(24, sizeof(XPVIO)) },
1047 #define new_body_allocated(sv_type) \
1048 (void *)((char *)S_new_body(aTHX_ sv_type) \
1049 - bodies_by_type[sv_type].offset)
1051 /* return a thing to the free list */
1053 #define del_body(thing, root) \
1055 void ** const thing_copy = (void **)thing; \
1056 *thing_copy = *root; \
1057 *root = (void*)thing_copy; \
1061 #if !(NVSIZE <= IVSIZE)
1062 # define new_XNV() safemalloc(sizeof(XPVNV))
1064 #define new_XPVNV() safemalloc(sizeof(XPVNV))
1065 #define new_XPVMG() safemalloc(sizeof(XPVMG))
1067 #define del_XPVGV(p) safefree(p)
1071 #if !(NVSIZE <= IVSIZE)
1072 # define new_XNV() new_body_allocated(SVt_NV)
1074 #define new_XPVNV() new_body_allocated(SVt_PVNV)
1075 #define new_XPVMG() new_body_allocated(SVt_PVMG)
1077 #define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1078 &PL_body_roots[SVt_PVGV])
1082 /* no arena for you! */
1084 #define new_NOARENA(details) \
1085 safemalloc((details)->body_size + (details)->offset)
1086 #define new_NOARENAZ(details) \
1087 safecalloc((details)->body_size + (details)->offset, 1)
1090 Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1091 const size_t arena_size)
1093 void ** const root = &PL_body_roots[sv_type];
1094 struct arena_desc *adesc;
1095 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1099 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
1100 #if defined(DEBUGGING) && defined(PERL_GLOBAL_STRUCT)
1103 #if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
1104 static bool done_sanity_check;
1106 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1107 * variables like done_sanity_check. */
1108 if (!done_sanity_check) {
1109 unsigned int i = SVt_LAST;
1111 done_sanity_check = TRUE;
1114 assert (bodies_by_type[i].type == i);
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;
1127 PL_body_arenas = (void *) newroot;
1128 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
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);
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));
1142 start = (char *) adesc->arena;
1144 /* Get the address of the byte after the end of the last body we can fit.
1145 Remember, this is integer division: */
1146 end = start + good_arena_size / body_size * body_size;
1148 /* computed count doesn't reflect the 1st slot reservation */
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 "
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));
1157 DEBUG_m(PerlIO_printf(Perl_debug_log,
1158 "arena %p end %p arena-size %d type %d size %d ct %d\n",
1159 (void*)start, (void*)end,
1160 (int)arena_size, sv_type, (int)body_size,
1161 (int)good_arena_size / (int)body_size));
1163 *root = (void *)start;
1166 /* Where the next body would start: */
1167 char * const next = start + body_size;
1170 /* This is the last body: */
1171 assert(next == end);
1173 *(void **)start = 0;
1177 *(void**) start = (void *)next;
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).
1186 #define new_body_inline(xpv, sv_type) \
1188 void ** const r3wt = &PL_body_roots[sv_type]; \
1189 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1190 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
1191 bodies_by_type[sv_type].body_size,\
1192 bodies_by_type[sv_type].arena_size)); \
1193 *(r3wt) = *(void**)(xpv); \
1199 S_new_body(pTHX_ const svtype sv_type)
1202 new_body_inline(xpv, sv_type);
1208 static const struct body_details fake_rv =
1209 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1212 =for apidoc sv_upgrade
1214 Upgrade an SV to a more complex form. Generally adds a new body type to the
1215 SV, then copies across as much information as possible from the old body.
1216 It croaks if the SV is already in a more complex form than requested. You
1217 generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1218 before calling C<sv_upgrade>, and hence does not croak. See also
1225 Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
1229 const svtype old_type = SvTYPE(sv);
1230 const struct body_details *new_type_details;
1231 const struct body_details *old_type_details
1232 = bodies_by_type + old_type;
1233 SV *referant = NULL;
1235 PERL_ARGS_ASSERT_SV_UPGRADE;
1237 if (old_type == new_type)
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)
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. */
1249 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
1250 sv_force_normal_flags(sv, 0);
1253 old_body = SvANY(sv);
1255 /* Copying structures onto other structures that have been neatly zeroed
1256 has a subtle gotcha. Consider XPVMG
1258 +------+------+------+------+------+-------+-------+
1259 | NV | CUR | LEN | IV | MAGIC | STASH |
1260 +------+------+------+------+------+-------+-------+
1261 0 4 8 12 16 20 24 28
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:
1266 +------+------+------+------+------+-------+-------+------+
1267 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1268 +------+------+------+------+------+-------+-------+------+
1269 0 4 8 12 16 20 24 28 32
1271 so what happens if you allocate memory for this structure:
1273 +------+------+------+------+------+-------+-------+------+------+...
1274 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1275 +------+------+------+------+------+-------+-------+------+------+...
1276 0 4 8 12 16 20 24 28 32 36
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.
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
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)
1290 So we are careful and work out the size of used parts of all the
1298 referant = SvRV(sv);
1299 old_type_details = &fake_rv;
1300 if (new_type == SVt_NV)
1301 new_type = SVt_PVNV;
1303 if (new_type < SVt_PVIV) {
1304 new_type = (new_type == SVt_NV)
1305 ? SVt_PVNV : SVt_PVIV;
1310 if (new_type < SVt_PVNV) {
1311 new_type = SVt_PVNV;
1315 assert(new_type > SVt_PV);
1316 STATIC_ASSERT_STMT(SVt_IV < SVt_PV);
1317 STATIC_ASSERT_STMT(SVt_NV < SVt_PV);
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);
1330 if (UNLIKELY(old_type_details->cant_upgrade))
1331 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1332 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1335 if (UNLIKELY(old_type > new_type))
1336 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1337 (int)old_type, (int)new_type);
1339 new_type_details = bodies_by_type + new_type;
1341 SvFLAGS(sv) &= ~SVTYPEMASK;
1342 SvFLAGS(sv) |= new_type;
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);
1349 assert(old_type == SVt_NULL);
1350 SET_SVANY_FOR_BODYLESS_IV(sv);
1354 assert(old_type == SVt_NULL);
1355 #if NVSIZE <= IVSIZE
1356 SET_SVANY_FOR_BODYLESS_NV(sv);
1358 SvANY(sv) = new_XNV();
1364 assert(new_type_details->body_size);
1367 assert(new_type_details->arena);
1368 assert(new_type_details->arena_size);
1369 /* This points to the start of the allocated area. */
1370 new_body_inline(new_body, new_type);
1371 Zero(new_body, new_type_details->body_size, char);
1372 new_body = ((char *)new_body) - new_type_details->offset;
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);
1378 SvANY(sv) = new_body;
1379 if (new_type == SVt_PVAV) {
1383 if (old_type_details->body_size) {
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
1393 #ifndef NODEFAULT_SHAREKEYS
1394 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1396 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1397 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
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.
1404 if (old_type == SVt_IV) {
1406 } else if (old_type >= SVt_PV) {
1407 assert(SvPVX_const(sv) == 0);
1410 if (old_type >= SVt_PVMG) {
1411 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1412 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1414 sv->sv_u.svu_array = NULL; /* or svu_hash */
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));
1435 assert(new_type_details->body_size);
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. */
1440 new_body_inline(new_body, new_type);
1441 Zero(new_body, new_type_details->body_size, char);
1442 new_body = ((char *)new_body) - new_type_details->offset;
1444 new_body = new_NOARENAZ(new_type_details);
1446 SvANY(sv) = new_body;
1448 if (old_type_details->copy) {
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;
1454 if (new_type_details->offset > old_type_details->offset) {
1455 const int difference
1456 = new_type_details->offset - old_type_details->offset;
1457 offset += difference;
1458 length -= difference;
1460 assert (length >= 0);
1462 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1466 #ifndef NV_ZERO_IS_ALLBITS_ZERO
1467 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
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
1472 if (old_type_details->zero_nv && !new_type_details->zero_nv
1473 && !isGV_with_GP(sv))
1477 if (UNLIKELY(new_type == SVt_PVIO)) {
1478 IO * const io = MUTABLE_IO(sv);
1479 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
1482 /* Clear the stashcache because a new IO could overrule a package
1484 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
1485 hv_clear(PL_stashcache);
1487 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1488 IoPAGE_LEN(sv) = 60;
1490 if (UNLIKELY(new_type == SVt_REGEXP))
1491 sv->sv_u.svu_rx = (regexp *)new_body;
1492 else if (old_type < SVt_PV) {
1493 /* referant will be NULL unless the old type was SVt_IV emulating
1495 sv->sv_u.svu_rv = referant;
1499 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1500 (unsigned long)new_type);
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) {
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);
1513 del_body((void*)((char*)old_body + old_type_details->offset),
1514 &PL_body_roots[old_type]);
1520 =for apidoc sv_backoff
1522 Remove any string offset. You should normally use the C<SvOOK_off> macro
1529 Perl_sv_backoff(SV *const sv)
1532 const char * const s = SvPVX_const(sv);
1534 PERL_ARGS_ASSERT_SV_BACKOFF;
1537 assert(SvTYPE(sv) != SVt_PVHV);
1538 assert(SvTYPE(sv) != SVt_PVAV);
1540 SvOOK_offset(sv, delta);
1542 SvLEN_set(sv, SvLEN(sv) + delta);
1543 SvPV_set(sv, SvPVX(sv) - delta);
1544 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1545 SvFLAGS(sv) &= ~SVf_OOK;
1552 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1553 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1554 Use the C<SvGROW> wrapper instead.
1559 static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1562 Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
1566 PERL_ARGS_ASSERT_SV_GROW;
1570 if (SvTYPE(sv) < SVt_PV) {
1571 sv_upgrade(sv, SVt_PV);
1572 s = SvPVX_mutable(sv);
1574 else if (SvOOK(sv)) { /* pv is offset? */
1576 s = SvPVX_mutable(sv);
1577 if (newlen > SvLEN(sv))
1578 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1582 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
1583 s = SvPVX_mutable(sv);
1586 #ifdef PERL_COPY_ON_WRITE
1587 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
1588 * to store the COW count. So in general, allocate one more byte than
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
1594 * Only increment if the allocation isn't MEM_SIZE_MAX,
1595 * otherwise it will wrap to 0.
1597 if (newlen & 0xff && newlen != MEM_SIZE_MAX)
1601 #if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1602 #define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1605 if (newlen > SvLEN(sv)) { /* need more room? */
1606 STRLEN minlen = SvCUR(sv);
1607 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1608 if (newlen < minlen)
1610 #ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
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 */
1615 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1616 if (rounded > newlen)
1620 if (SvLEN(sv) && s) {
1621 s = (char*)saferealloc(s, newlen);
1624 s = (char*)safemalloc(newlen);
1625 if (SvPVX_const(sv) && SvCUR(sv)) {
1626 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1630 #ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
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
1634 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1636 SvLEN_set(sv, newlen);
1643 =for apidoc sv_setiv
1645 Copies an integer into the given SV, upgrading first if necessary.
1646 Does not handle 'set' magic. See also C<sv_setiv_mg>.
1652 Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
1654 PERL_ARGS_ASSERT_SV_SETIV;
1656 SV_CHECK_THINKFIRST_COW_DROP(sv);
1657 switch (SvTYPE(sv)) {
1660 sv_upgrade(sv, SVt_IV);
1663 sv_upgrade(sv, SVt_PVIV);
1667 if (!isGV_with_GP(sv))
1674 /* diag_listed_as: Can't coerce %s to %s in %s */
1675 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1680 (void)SvIOK_only(sv); /* validate number */
1686 =for apidoc sv_setiv_mg
1688 Like C<sv_setiv>, but also handles 'set' magic.
1694 Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
1696 PERL_ARGS_ASSERT_SV_SETIV_MG;
1703 =for apidoc sv_setuv
1705 Copies an unsigned integer into the given SV, upgrading first if necessary.
1706 Does not handle 'set' magic. See also C<sv_setuv_mg>.
1712 Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
1714 PERL_ARGS_ASSERT_SV_SETUV;
1716 /* With the if statement to ensure that integers are stored as IVs whenever
1718 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1721 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
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 :-)
1727 if (u <= (UV)IV_MAX) {
1728 sv_setiv(sv, (IV)u);
1737 =for apidoc sv_setuv_mg
1739 Like C<sv_setuv>, but also handles 'set' magic.
1745 Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
1747 PERL_ARGS_ASSERT_SV_SETUV_MG;
1754 =for apidoc sv_setnv
1756 Copies a double into the given SV, upgrading first if necessary.
1757 Does not handle 'set' magic. See also C<sv_setnv_mg>.
1763 Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
1765 PERL_ARGS_ASSERT_SV_SETNV;
1767 SV_CHECK_THINKFIRST_COW_DROP(sv);
1768 switch (SvTYPE(sv)) {
1771 sv_upgrade(sv, SVt_NV);
1775 sv_upgrade(sv, SVt_PVNV);
1779 if (!isGV_with_GP(sv))
1786 /* diag_listed_as: Can't coerce %s to %s in %s */
1787 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1793 (void)SvNOK_only(sv); /* validate number */
1798 =for apidoc sv_setnv_mg
1800 Like C<sv_setnv>, but also handles 'set' magic.
1806 Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
1808 PERL_ARGS_ASSERT_SV_SETNV_MG;
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.
1821 S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1824 PERL_ARGS_ASSERT_SV_DISPLAY;
1827 SV *dsv = newSVpvs_flags("", SVs_TEMP);
1828 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
1831 const char * const limit = tmpbuf + tmpbuf_size - 8;
1832 /* each *s can expand to 4 chars + "...\0",
1833 i.e. need room for 8 chars */
1835 const char *s = SvPVX_const(sv);
1836 const char * const end = s + SvCUR(sv);
1837 for ( ; s < end && d < limit; s++ ) {
1839 if (! isASCII(ch) && !isPRINT_LC(ch)) {
1843 /* Map to ASCII "equivalent" of Latin1 */
1844 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
1850 else if (ch == '\r') {
1854 else if (ch == '\f') {
1858 else if (ch == '\\') {
1862 else if (ch == '\0') {
1866 else if (isPRINT_LC(ch))
1885 /* Print an "isn't numeric" warning, using a cleaned-up,
1886 * printable version of the offending string
1890 S_not_a_number(pTHX_ SV *const sv)
1895 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1897 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1900 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1901 /* diag_listed_as: Argument "%s" isn't numeric%s */
1902 "Argument \"%s\" isn't numeric in %s", pv,
1905 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1906 /* diag_listed_as: Argument "%s" isn't numeric%s */
1907 "Argument \"%s\" isn't numeric", pv);
1911 S_not_incrementable(pTHX_ SV *const sv) {
1915 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1917 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1919 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1920 "Argument \"%s\" treated as 0 in increment (++)", pv);
1924 =for apidoc looks_like_number
1926 Test if the content of an SV looks like a number (or is a number).
1927 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1928 non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1935 Perl_looks_like_number(pTHX_ SV *const sv)
1941 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1943 if (SvPOK(sv) || SvPOKp(sv)) {
1944 sbegin = SvPV_nomg_const(sv, len);
1947 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1948 numtype = grok_number(sbegin, len, NULL);
1949 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
1953 S_glob_2number(pTHX_ GV * const gv)
1955 PERL_ARGS_ASSERT_GLOB_2NUMBER;
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))
1961 SV *const buffer = sv_newmortal();
1962 gv_efullname3(buffer, gv, "*");
1963 not_a_number(buffer);
1965 /* We just want something true to return, so that S_sv_2iuv_common
1966 can tail call us and return true. */
1970 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1971 until proven guilty, assume that things are not that bad... */
1976 As 64 bit platforms often have an NV that doesn't preserve all bits of
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
1981 precision as a side effect of conversion (which would lead to insanity
1982 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
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
1986 2) to ensure that if a numeric conversion to one form is requested that
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
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
2001 while converting from PV to NV, check to see if converting that NV to an
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).
2006 While converting from PV to IV, check to see if converting that IV to an
2007 NV would lose accuracy over a direct conversion from PV to NV. If it
2008 would, cache both conversions, flag similarly.
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.
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:
2014 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
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
2020 loss of precision compared with integer addition.
2022 * making IV and NV equal status should make maths accurate on 64 bit
2024 * may speed up maths somewhat if pp_add and friends start to use
2025 integers when possible instead of fp. (Hopefully the overhead in
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
2032 favoured when IV and NV are equally accurate
2034 ####################################################################
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.
2038 ####################################################################
2040 Your mileage will vary depending your CPU's relative fp to integer
2044 #ifndef NV_PRESERVES_UV
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
2051 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
2053 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2055 S_sv_2iuv_non_preserve(pTHX_ SV *const sv
2061 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
2062 PERL_UNUSED_CONTEXT;
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));
2065 if (SvNVX(sv) < (NV)IV_MIN) {
2066 (void)SvIOKp_on(sv);
2068 SvIV_set(sv, IV_MIN);
2069 return IS_NUMBER_UNDERFLOW_IV;
2071 if (SvNVX(sv) > (NV)UV_MAX) {
2072 (void)SvIOKp_on(sv);
2075 SvUV_set(sv, UV_MAX);
2076 return IS_NUMBER_OVERFLOW_UV;
2078 (void)SvIOKp_on(sv);
2080 /* Can't use strtol etc to convert this string. (See truth table in
2082 if (SvNVX(sv) <= (UV)IV_MAX) {
2083 SvIV_set(sv, I_V(SvNVX(sv)));
2084 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2085 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2087 /* Integer is imprecise. NOK, IOKp */
2089 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2092 SvUV_set(sv, U_V(SvNVX(sv)));
2093 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
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.
2098 return IS_NUMBER_OVERFLOW_UV;
2100 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2102 /* Integer is imprecise. NOK, IOKp */
2104 return IS_NUMBER_OVERFLOW_IV;
2106 #endif /* !NV_PRESERVES_UV*/
2108 /* If numtype is infnan, set the NV of the sv accordingly.
2109 * If numtype is anything else, try setting the NV using Atof(PV). */
2111 # pragma warning(push)
2112 # pragma warning(disable:4756;disable:4056)
2115 S_sv_setnv(pTHX_ SV* sv, int numtype)
2117 bool pok = cBOOL(SvPOK(sv));
2119 if ((numtype & IS_NUMBER_INFINITY)) {
2120 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
2123 else if ((numtype & IS_NUMBER_NAN)) {
2124 SvNV_set(sv, NV_NAN);
2128 SvNV_set(sv, Atof(SvPVX_const(sv)));
2129 /* Purposefully no true nok here, since we don't want to blow
2130 * away the possible IOK/UV of an existing sv. */
2133 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
2135 SvPOK_on(sv); /* PV is okay, though. */
2139 # pragma warning(pop)
2143 S_sv_2iuv_common(pTHX_ SV *const sv)
2145 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
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
2151 * IV or UV at same time to avoid this. */
2152 /* IV-over-UV optimisation - choose to cache IV if possible */
2154 if (SvTYPE(sv) == SVt_NV)
2155 sv_upgrade(sv, SVt_PVNV);
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
2162 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2163 if (Perl_isnan(SvNVX(sv))) {
2169 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2170 SvIV_set(sv, I_V(SvNVX(sv)));
2171 if (SvNVX(sv) == (NV) SvIVX(sv)
2172 #ifndef NV_PRESERVES_UV
2173 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
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 */
2182 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2184 /* scalar has trailing garbage, eg "42a" */
2186 DEBUG_c(PerlIO_printf(Perl_debug_log,
2187 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
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,
2198 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
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 */
2213 SvUV_set(sv, U_V(SvNVX(sv)));
2215 (SvNVX(sv) == (NV) SvUVX(sv))
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 */
2228 DEBUG_c(PerlIO_printf(Perl_debug_log,
2229 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2235 else if (SvPOKp(sv)) {
2237 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2238 /* We want to avoid a possible problem when we cache an IV/ a UV which
2239 may be later translated to an NV, and the resulting NV is not
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).
2245 This means that if we cache such an IV/a UV, we need to cache the
2246 NV as well. Moreover, we trade speed for space, and do not
2247 cache the NV if we are sure it's not needed.
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) {
2253 /* It's definitely an integer, only upgrade to PVIV */
2254 if (SvTYPE(sv) < SVt_PVIV)
2255 sv_upgrade(sv, SVt_PVIV);
2257 } else if (SvTYPE(sv) < SVt_PVNV)
2258 sv_upgrade(sv, SVt_PVNV);
2260 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
2261 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
2263 S_sv_setnv(aTHX_ sv, numtype);
2267 /* If NVs preserve UVs then we only use the UV value if we know that
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
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);
2279 if (!(numtype & IS_NUMBER_NEG)) {
2281 if (value <= (UV)IV_MAX) {
2282 SvIV_set(sv, (IV)value);
2284 /* it didn't overflow, and it was positive. */
2285 SvUV_set(sv, value);
2289 /* 2s complement assumption */
2290 if (value <= (UV)IV_MIN) {
2291 SvIV_set(sv, value == (UV)IV_MIN
2292 ? IV_MIN : -(IV)value);
2294 /* Too negative for an IV. This is a double upgrade, but
2295 I'm assuming it will be rare. */
2296 if (SvTYPE(sv) < SVt_PVNV)
2297 sv_upgrade(sv, SVt_PVNV);
2301 SvNV_set(sv, -(NV)value);
2302 SvIV_set(sv, IV_MIN);
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. */
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). */
2313 S_sv_setnv(aTHX_ sv, numtype);
2315 if (! numtype && ckWARN(WARN_NUMERIC))
2318 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
2319 PTR2UV(sv), SvNVX(sv)));
2321 #ifdef NV_PRESERVES_UV
2322 (void)SvIOKp_on(sv);
2324 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2325 if (Perl_isnan(SvNVX(sv))) {
2331 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2332 SvIV_set(sv, I_V(SvNVX(sv)));
2333 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2336 NOOP; /* Integer is imprecise. NOK, IOKp */
2338 /* UV will not work better than IV */
2340 if (SvNVX(sv) > (NV)UV_MAX) {
2342 /* Integer is inaccurate. NOK, IOKp, is UV */
2343 SvUV_set(sv, UV_MAX);
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)) {
2351 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2356 #else /* NV_PRESERVES_UV */
2357 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2358 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2359 /* The IV/UV slot will have been set from value returned by
2360 grok_number above. The NV slot has just been set using
2363 assert (SvIOKp(sv));
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);
2370 SvIV_set(sv, I_V(SvNVX(sv)));
2371 if ((NV)(SvIVX(sv)) == SvNVX(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))
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);
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. */
2389 sv_2iuv_non_preserve (sv, numtype);
2391 sv_2iuv_non_preserve (sv);
2395 #endif /* NV_PRESERVES_UV */
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: */
2401 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2405 if (isGV_with_GP(sv))
2406 return glob_2number(MUTABLE_GV(sv));
2408 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2410 if (SvTYPE(sv) < SVt_IV)
2411 /* Typically the caller expects that sv_any is not NULL now. */
2412 sv_upgrade(sv, SVt_IV);
2413 /* Return 0 from the caller. */
2420 =for apidoc sv_2iv_flags
2422 Return the integer value of an SV, doing any necessary string
2423 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2424 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2430 Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
2432 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
2434 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2435 && SvTYPE(sv) != SVt_PVFM);
2437 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2443 if (flags & SV_SKIP_OVERLOAD)
2445 tmpstr = AMG_CALLunary(sv, numer_amg);
2446 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2447 return SvIV(tmpstr);
2450 return PTR2IV(SvRV(sv));
2453 if (SvVALID(sv) || isREGEXP(sv)) {
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.
2460 Regexps have no SvIVX and SvNVX fields.
2462 assert(isREGEXP(sv) || SvPOKp(sv));
2465 const char * const ptr =
2466 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2468 = grok_number(ptr, SvCUR(sv), &value);
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)
2477 if (value < (UV)IV_MAX)
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. */
2490 if (ckWARN(WARN_NUMERIC))
2493 return I_V(Atof(ptr));
2497 if (SvTHINKFIRST(sv)) {
2498 if (SvREADONLY(sv) && !SvOK(sv)) {
2499 if (ckWARN(WARN_UNINITIALIZED))
2506 if (S_sv_2iuv_common(aTHX_ sv))
2510 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2511 PTR2UV(sv),SvIVX(sv)));
2512 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2516 =for apidoc sv_2uv_flags
2518 Return the unsigned integer value of an SV, doing any necessary string
2519 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2520 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2526 Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
2528 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
2530 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2536 if (flags & SV_SKIP_OVERLOAD)
2538 tmpstr = AMG_CALLunary(sv, numer_amg);
2539 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2540 return SvUV(tmpstr);
2543 return PTR2UV(SvRV(sv));
2546 if (SvVALID(sv) || isREGEXP(sv)) {
2547 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2548 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2549 Regexps have no SvIVX and SvNVX fields. */
2550 assert(isREGEXP(sv) || SvPOKp(sv));
2553 const char * const ptr =
2554 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2556 = grok_number(ptr, SvCUR(sv), &value);
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))
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. */
2573 if (ckWARN(WARN_NUMERIC))
2576 return U_V(Atof(ptr));
2580 if (SvTHINKFIRST(sv)) {
2581 if (SvREADONLY(sv) && !SvOK(sv)) {
2582 if (ckWARN(WARN_UNINITIALIZED))
2589 if (S_sv_2iuv_common(aTHX_ sv))
2593 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2594 PTR2UV(sv),SvUVX(sv)));
2595 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2599 =for apidoc sv_2nv_flags
2601 Return the num value of an SV, doing any necessary string or integer
2602 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2603 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2609 Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
2611 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2613 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2614 && SvTYPE(sv) != SVt_PVFM);
2615 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2616 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2617 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2618 Regexps have no SvIVX and SvNVX fields. */
2620 if (flags & SV_GMAGIC)
2624 if (SvPOKp(sv) && !SvIOKp(sv)) {
2625 ptr = SvPVX_const(sv);
2627 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2628 !grok_number(ptr, SvCUR(sv), NULL))
2634 return (NV)SvUVX(sv);
2636 return (NV)SvIVX(sv);
2642 ptr = RX_WRAPPED((REGEXP *)sv);
2645 assert(SvTYPE(sv) >= SVt_PVMG);
2646 /* This falls through to the report_uninit near the end of the
2648 } else if (SvTHINKFIRST(sv)) {
2653 if (flags & SV_SKIP_OVERLOAD)
2655 tmpstr = AMG_CALLunary(sv, numer_amg);
2656 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2657 return SvNV(tmpstr);
2660 return PTR2NV(SvRV(sv));
2662 if (SvREADONLY(sv) && !SvOK(sv)) {
2663 if (ckWARN(WARN_UNINITIALIZED))
2668 if (SvTYPE(sv) < SVt_NV) {
2669 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2670 sv_upgrade(sv, SVt_NV);
2672 STORE_NUMERIC_LOCAL_SET_STANDARD();
2673 PerlIO_printf(Perl_debug_log,
2674 "0x%"UVxf" num(%" NVgf ")\n",
2675 PTR2UV(sv), SvNVX(sv));
2676 RESTORE_NUMERIC_LOCAL();
2679 else if (SvTYPE(sv) < SVt_PVNV)
2680 sv_upgrade(sv, SVt_PVNV);
2685 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2686 #ifdef NV_PRESERVES_UV
2692 /* Only set the public NV OK flag if this NV preserves the IV */
2693 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2695 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2696 : (SvIVX(sv) == I_V(SvNVX(sv))))
2702 else if (SvPOKp(sv)) {
2704 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2705 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2707 #ifdef NV_PRESERVES_UV
2708 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2709 == IS_NUMBER_IN_UV) {
2710 /* It's definitely an integer */
2711 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2713 S_sv_setnv(aTHX_ sv, numtype);
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 ==
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. */
2735 /* value has been set. It may not be precise. */
2736 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
2737 /* 2s complement assumption for (UV)IV_MIN */
2738 SvNOK_on(sv); /* Integer is too negative. */
2743 if (numtype & IS_NUMBER_NEG) {
2744 /* -IV_MIN is undefined, but we should never reach
2745 * this point with both IS_NUMBER_NEG and value ==
2747 assert(value != (UV)IV_MIN);
2748 SvIV_set(sv, -(IV)value);
2749 } else if (value <= (UV)IV_MAX) {
2750 SvIV_set(sv, (IV)value);
2752 SvUV_set(sv, value);
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 */
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)) {
2769 /* It had no "." so it must be integer. */
2773 /* between IV_MAX and NV(UV_MAX).
2774 Could be slightly > UV_MAX */
2776 if (numtype & IS_NUMBER_NOT_INT) {
2777 /* UV and NV both imprecise. */
2779 const UV nv_as_uv = U_V(nv);
2781 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
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: */
2795 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2796 #endif /* NV_PRESERVES_UV */
2799 if (isGV_with_GP(sv)) {
2800 glob_2number(MUTABLE_GV(sv));
2804 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
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. */
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();
2824 Return an SV with the numeric value of the source SV, doing any necessary
2825 reference or overload conversion. The caller is expected to have handled
2832 Perl_sv_2num(pTHX_ SV *const sv)
2834 PERL_ARGS_ASSERT_SV_2NUM;
2839 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2840 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2841 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2842 return sv_2num(tmpsv);
2844 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
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
2851 * We assume that buf is at least TYPE_CHARS(UV) long.
2855 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2857 char *ptr = buf + TYPE_CHARS(UV);
2858 char * const ebuf = ptr;
2861 PERL_ARGS_ASSERT_UIV_2BUF;
2869 uv = (iv == IV_MIN) ? (UV)iv : (UV)(-iv);
2873 *--ptr = '0' + (char)(uv % 10);
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,
2884 * excluding the zero byte, on failure (not an infinity, not a nan)
2885 * returns zero, assert-fails on maxlen being too short.
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. */
2891 S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
2893 assert(maxlen >= 4);
2894 if (Perl_isinf(nv)) {
2896 if (maxlen < 5) /* "-Inf\0" */
2906 else if (Perl_isnan(nv)) {
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.*/
2919 assert((s == buffer + 3) || (s == buffer + 4));
2921 return s - buffer - 1; /* -1: excluding the zero byte */
2925 =for apidoc sv_2pv_flags
2927 Returns a pointer to the string value of an SV, and sets *lp to its length.
2928 If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2929 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2930 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2936 Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2940 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2942 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2943 && SvTYPE(sv) != SVt_PVFM);
2944 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2949 if (flags & SV_SKIP_OVERLOAD)
2951 tmpstr = AMG_CALLunary(sv, string_amg);
2952 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2953 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2955 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2959 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2960 if (flags & SV_CONST_RETURN) {
2961 pv = (char *) SvPVX_const(tmpstr);
2963 pv = (flags & SV_MUTABLE_RETURN)
2964 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2967 *lp = SvCUR(tmpstr);
2969 pv = sv_2pv_flags(tmpstr, lp, flags);
2982 SV *const referent = SvRV(sv);
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);
2994 /* If the regex is UTF-8 we want the containing scalar to
2995 have an UTF-8 flag too */
3002 *lp = RX_WRAPLEN(re);
3004 return RX_WRAPPED(re);
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;
3013 if (SvOBJECT(referent)) {
3014 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3017 stashname = HEK_KEY(name);
3018 stashnamelen = HEK_LEN(name);
3020 if (HEK_UTF8(name)) {
3026 stashname = "__ANON__";
3029 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3030 + 2 * sizeof(UV) + 2 /* )\0 */;
3032 len = typelen + 3 /* (0x */
3033 + 2 * sizeof(UV) + 2 /* )\0 */;
3036 Newx(buffer, len, char);
3037 buffer_end = retval = buffer + len;
3039 /* Working backwards */
3043 *--retval = PL_hexdigit[addr & 15];
3044 } while (addr >>= 4);
3050 memcpy(retval, typestr, typelen);
3054 retval -= stashnamelen;
3055 memcpy(retval, stashname, stashnamelen);
3057 /* retval may not necessarily have reached the start of the
3059 assert (retval >= buffer);
3061 len = buffer_end - retval - 1; /* -1 for that \0 */
3073 if (flags & SV_MUTABLE_RETURN)
3074 return SvPVX_mutable(sv);
3075 if (flags & SV_CONST_RETURN)
3076 return (char *)SvPVX_const(sv);
3081 /* I'm assuming that if both IV and NV are equally valid then
3082 converting the IV is going to be more efficient */
3083 const U32 isUIOK = SvIsUV(sv);
3084 char buf[TYPE_CHARS(UV)];
3088 if (SvTYPE(sv) < SVt_PVIV)
3089 sv_upgrade(sv, SVt_PVIV);
3090 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
3092 /* inlined from sv_setpvn */
3093 s = SvGROW_mutable(sv, len + 1);
3094 Move(ptr, s, len, char);
3099 else if (SvNOK(sv)) {
3100 if (SvTYPE(sv) < SVt_PVNV)
3101 sv_upgrade(sv, SVt_PVNV);
3102 if (SvNVX(sv) == 0.0
3103 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3104 && !Perl_isnan(SvNVX(sv))
3107 s = SvGROW_mutable(sv, 2);
3112 STRLEN size = 5; /* "-Inf\0" */
3114 s = SvGROW_mutable(sv, size);
3115 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
3121 /* some Xenix systems wipe out errno here */
3130 5 + /* exponent digits */
3134 s = SvGROW_mutable(sv, size);
3135 #ifndef USE_LOCALE_NUMERIC
3136 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3142 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3143 STORE_LC_NUMERIC_SET_TO_NEEDED();
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);
3154 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3156 /* If the radix character is UTF-8, and actually is in the
3157 * output, turn on the UTF-8 flag for the scalar */
3159 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3163 RESTORE_LC_NUMERIC();
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 */
3176 else if (isGV_with_GP(sv)) {
3177 GV *const gv = MUTABLE_GV(sv);
3178 SV *const buffer = sv_newmortal();
3180 gv_efullname3(buffer, gv, "*");
3182 assert(SvPOK(buffer));
3186 *lp = SvCUR(buffer);
3187 return SvPVX(buffer);
3189 else if (isREGEXP(sv)) {
3190 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3191 return RX_WRAPPED((REGEXP *)sv);
3196 if (flags & SV_UNDEF_RETURNS_NULL)
3198 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
3200 /* Typically the caller expects that sv_any is not NULL now. */
3201 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
3202 sv_upgrade(sv, SVt_PV);
3207 const STRLEN len = s - SvPVX_const(sv);
3212 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3213 PTR2UV(sv),SvPVX_const(sv)));
3214 if (flags & SV_CONST_RETURN)
3215 return (char *)SvPVX_const(sv);
3216 if (flags & SV_MUTABLE_RETURN)
3217 return SvPVX_mutable(sv);
3222 =for apidoc sv_copypv
3224 Copies a stringified representation of the source SV into the
3225 destination SV. Automatically performs any necessary mg_get and
3226 coercion of numeric values into strings. Guaranteed to preserve
3227 UTF8 flag even from overloaded objects. Similar in nature to
3228 sv_2pv[_flags] but operates directly on an SV instead of just the
3229 string. Mostly uses sv_2pv_flags to do its work, except when that
3230 would lose the UTF-8'ness of the PV.
3232 =for apidoc sv_copypv_nomg
3234 Like sv_copypv, but doesn't invoke get magic first.
3236 =for apidoc sv_copypv_flags
3238 Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3245 Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
3250 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
3252 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
3253 sv_setpvn(dsv,s,len);
3261 =for apidoc sv_2pvbyte
3263 Return a pointer to the byte-encoded representation of the SV, and set *lp
3264 to its length. May cause the SV to be downgraded from UTF-8 as a
3267 Usually accessed via the C<SvPVbyte> macro.
3273 Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
3275 PERL_ARGS_ASSERT_SV_2PVBYTE;
3278 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3279 || isGV_with_GP(sv) || SvROK(sv)) {
3280 SV *sv2 = sv_newmortal();
3281 sv_copypv_nomg(sv2,sv);
3284 sv_utf8_downgrade(sv,0);
3285 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3289 =for apidoc sv_2pvutf8
3291 Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3292 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3294 Usually accessed via the C<SvPVutf8> macro.
3300 Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
3302 PERL_ARGS_ASSERT_SV_2PVUTF8;
3304 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3305 || isGV_with_GP(sv) || SvROK(sv))
3306 sv = sv_mortalcopy(sv);
3309 sv_utf8_upgrade_nomg(sv);
3310 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3315 =for apidoc sv_2bool
3317 This macro is only used by sv_true() or its macro equivalent, and only if
3318 the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3319 It calls sv_2bool_flags with the SV_GMAGIC flag.
3321 =for apidoc sv_2bool_flags
3323 This function is only used by sv_true() and friends, and only if
3324 the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
3325 contain SV_GMAGIC, then it does an mg_get() first.
3332 Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
3334 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3337 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3343 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3344 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3347 if(SvGMAGICAL(sv)) {
3349 goto restart; /* call sv_2bool */
3351 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3352 else if(!SvOK(sv)) {
3355 else if(SvPOK(sv)) {
3356 svb = SvPVXtrue(sv);
3358 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3359 svb = (SvIOK(sv) && SvIVX(sv) != 0)
3360 || (SvNOK(sv) && SvNVX(sv) != 0.0);
3364 goto restart; /* call sv_2bool_nomg */
3369 return SvRV(sv) != 0;
3373 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
3374 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
3378 =for apidoc sv_utf8_upgrade
3380 Converts the PV of an SV to its UTF-8-encoded form.
3381 Forces the SV to string form if it is not already.
3382 Will C<mg_get> on C<sv> if appropriate.
3383 Always sets the SvUTF8 flag to avoid future validity checks even
3384 if the whole string is the same in UTF-8 as not.
3385 Returns the number of bytes in the converted string
3387 This is not a general purpose byte encoding to Unicode interface:
3388 use the Encode extension for that.
3390 =for apidoc sv_utf8_upgrade_nomg
3392 Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
3394 =for apidoc sv_utf8_upgrade_flags
3396 Converts the PV of an SV to its UTF-8-encoded form.
3397 Forces the SV to string form if it is not already.
3398 Always sets the SvUTF8 flag to avoid future validity checks even
3399 if all the bytes are invariant in UTF-8.
3400 If C<flags> has C<SV_GMAGIC> bit set,
3401 will C<mg_get> on C<sv> if appropriate, else not.
3403 If C<flags> has SV_FORCE_UTF8_UPGRADE set, this function assumes that the PV
3404 will expand when converted to UTF-8, and skips the extra work of checking for
3405 that. Typically this flag is used by a routine that has already parsed the
3406 string and found such characters, and passes this information on so that the
3407 work doesn't have to be repeated.
3409 Returns the number of bytes in the converted string.
3411 This is not a general purpose byte encoding to Unicode interface:
3412 use the Encode extension for that.
3414 =for apidoc sv_utf8_upgrade_flags_grow
3416 Like sv_utf8_upgrade_flags, but has an additional parameter C<extra>, which is
3417 the number of unused bytes the string of 'sv' is guaranteed to have free after
3418 it upon return. This allows the caller to reserve extra space that it intends
3419 to fill, to avoid extra grows.
3421 C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3422 are implemented in terms of this function.
3424 Returns the number of bytes in the converted string (not including the spares).
3428 (One might think that the calling routine could pass in the position of the
3429 first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3430 have to be found again. But that is not the case, because typically when the
3431 caller is likely to use this flag, it won't be calling this routine unless it
3432 finds something that won't fit into a byte. Otherwise it tries to not upgrade
3433 and just use bytes. But some things that do fit into a byte are variants in
3434 utf8, and the caller may not have been keeping track of these.)
3436 If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3437 C<NUL> isn't guaranteed due to having other routines do the work in some input
3438 cases, or if the input is already flagged as being in utf8.
3440 The speed of this could perhaps be improved for many cases if someone wanted to
3441 write a fast function that counts the number of variant characters in a string,
3442 especially if it could return the position of the first one.
3447 Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
3449 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3451 if (sv == &PL_sv_undef)
3453 if (!SvPOK_nog(sv)) {
3455 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3456 (void) sv_2pv_flags(sv,&len, flags);
3458 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3462 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3467 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3472 S_sv_uncow(aTHX_ sv, 0);
3475 if (IN_ENCODING && !(flags & SV_UTF8_NO_ENCODING)) {
3476 sv_recode_to_utf8(sv, _get_encoding());
3477 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3481 if (SvCUR(sv) == 0) {
3482 if (extra) SvGROW(sv, extra);
3483 } else { /* Assume Latin-1/EBCDIC */
3484 /* This function could be much more efficient if we
3485 * had a FLAG in SVs to signal if there are any variant
3486 * chars in the PV. Given that there isn't such a flag
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);
3492 STRLEN two_byte_count = 0;
3494 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
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. */
3502 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
3504 t--; /* t already incremented; re-point to first variant */
3509 /* utf8 conversion not needed because all are invariants. Mark as
3510 * UTF-8 even if no variant - saves scanning loop */
3512 if (extra) SvGROW(sv, SvCUR(sv) + extra);
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.
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
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.
3542 * There are advantages and disadvantages to each method.
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.
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 */
3587 STRLEN invariant_head = t - s;
3588 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3589 if (SvLEN(sv) < size) {
3591 /* Here, have decided to allocate a new string */
3596 Newx(dst, size, U8);
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 */
3602 if (invariant_head == 0) {
3605 Copy(s, dst, invariant_head, char);
3606 d = dst + invariant_head;
3610 append_utf8_from_native_byte(*t, &d);
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);
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
3623 * don't have to worry about growing. If two_byte_count is 0,
3624 * then t points to the first byte of the string which hasn't
3625 * been examined yet. Otherwise two_byte_count is 1, and t
3626 * points to the first byte in the string that will expand to
3627 * two. Depending on this, start examining at t or 1 after t.
3630 U8 *d = t + two_byte_count;
3633 /* Count up the remaining bytes that expand to two */
3636 const U8 chr = *d++;
3637 if (! NATIVE_BYTE_IS_INVARIANT(chr)) two_byte_count++;
3640 /* The string will expand by just the number of bytes that
3641 * occupy two positions. But we are one afterwards because of
3642 * the increment just above. This is the place to put the
3643 * trailing NUL, and to set the length before we decrement */
3645 d += two_byte_count;
3646 SvCUR_set(sv, d - s);
3650 /* Having decremented d, it points to the position to put the
3651 * very last byte of the expanded string. Go backwards through
3652 * the string, copying and expanding as we go, stopping when we
3653 * get to the part that is invariant the rest of the way down */
3657 if (NATIVE_BYTE_IS_INVARIANT(*e)) {
3660 *d-- = UTF8_EIGHT_BIT_LO(*e);
3661 *d-- = UTF8_EIGHT_BIT_HI(*e);
3667 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3668 /* Update pos. We do it at the end rather than during
3669 * the upgrade, to avoid slowing down the common case
3670 * (upgrade without pos).
3671 * pos can be stored as either bytes or characters. Since
3672 * this was previously a byte string we can just turn off
3673 * the bytes flag. */
3674 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3676 mg->mg_flags &= ~MGf_BYTES;
3678 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3679 magic_setutf8(sv,mg); /* clear UTF8 cache */
3684 /* Mark as UTF-8 even if no variant - saves scanning loop */
3690 =for apidoc sv_utf8_downgrade
3692 Attempts to convert the PV of an SV from characters to bytes.
3693 If the PV contains a character that cannot fit
3694 in a byte, this conversion will fail;
3695 in this case, either returns false or, if C<fail_ok> is not
3698 This is not a general purpose Unicode to byte encoding interface:
3699 use the Encode extension for that.
3705 Perl_sv_utf8_downgrade(pTHX_ SV *const sv, const bool fail_ok)
3707 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3709 if (SvPOKp(sv) && SvUTF8(sv)) {
3713 int mg_flags = SV_GMAGIC;
3716 S_sv_uncow(aTHX_ sv, 0);
3718 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3720 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3721 if (mg && mg->mg_len > 0 && mg->mg_flags & MGf_BYTES) {
3722 mg->mg_len = sv_pos_b2u_flags(sv, mg->mg_len,
3723 SV_GMAGIC|SV_CONST_RETURN);
3724 mg_flags = 0; /* sv_pos_b2u does get magic */
3726 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3727 magic_setutf8(sv,mg); /* clear UTF8 cache */
3730 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3732 if (!utf8_to_bytes(s, &len)) {
3737 Perl_croak(aTHX_ "Wide character in %s",
3740 Perl_croak(aTHX_ "Wide character");
3751 =for apidoc sv_utf8_encode
3753 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3754 flag off so that it looks like octets again.
3760 Perl_sv_utf8_encode(pTHX_ SV *const sv)
3762 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3764 if (SvREADONLY(sv)) {
3765 sv_force_normal_flags(sv, 0);
3767 (void) sv_utf8_upgrade(sv);
3772 =for apidoc sv_utf8_decode
3774 If the PV of the SV is an octet sequence in UTF-8
3775 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3776 so that it looks like a character. If the PV contains only single-byte
3777 characters, the C<SvUTF8> flag stays off.
3778 Scans PV for validity and returns false if the PV is invalid UTF-8.
3784 Perl_sv_utf8_decode(pTHX_ SV *const sv)
3786 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3789 const U8 *start, *c;
3792 /* The octets may have got themselves encoded - get them back as
3795 if (!sv_utf8_downgrade(sv, TRUE))
3798 /* it is actually just a matter of turning the utf8 flag on, but
3799 * we want to make sure everything inside is valid utf8 first.
3801 c = start = (const U8 *) SvPVX_const(sv);
3802 if (!is_utf8_string(c, SvCUR(sv)))
3804 e = (const U8 *) SvEND(sv);
3807 if (!UTF8_IS_INVARIANT(ch)) {
3812 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3813 /* XXX Is this dead code? XS_utf8_decode calls SvSETMAGIC
3814 after this, clearing pos. Does anything on CPAN
3816 /* adjust pos to the start of a UTF8 char sequence */
3817 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3819 I32 pos = mg->mg_len;
3821 for (c = start + pos; c > start; c--) {
3822 if (UTF8_IS_START(*c))
3825 mg->mg_len = c - start;
3828 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3829 magic_setutf8(sv,mg); /* clear UTF8 cache */
3836 =for apidoc sv_setsv
3838 Copies the contents of the source SV C<ssv> into the destination SV
3839 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3840 function if the source SV needs to be reused. Does not handle 'set' magic on
3841 destination SV. Calls 'get' magic on source SV. Loosely speaking, it
3842 performs a copy-by-value, obliterating any previous content of the
3845 You probably want to use one of the assortment of wrappers, such as
3846 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3847 C<SvSetMagicSV_nosteal>.
3849 =for apidoc sv_setsv_flags
3851 Copies the contents of the source SV C<ssv> into the destination SV
3852 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3853 function if the source SV needs to be reused. Does not handle 'set' magic.
3854 Loosely speaking, it performs a copy-by-value, obliterating any previous
3855 content of the destination.
3856 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3857 C<ssv> if appropriate, else not. If the C<flags>
3858 parameter has the C<SV_NOSTEAL> bit set then the
3859 buffers of temps will not be stolen. <sv_setsv>
3860 and C<sv_setsv_nomg> are implemented in terms of this function.
3862 You probably want to use one of the assortment of wrappers, such as
3863 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3864 C<SvSetMagicSV_nosteal>.
3866 This is the primary function for copying scalars, and most other
3867 copy-ish functions and macros use this underneath.
3873 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3875 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3876 HV *old_stash = NULL;
3878 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3880 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3881 const char * const name = GvNAME(sstr);
3882 const STRLEN len = GvNAMELEN(sstr);
3884 if (dtype >= SVt_PV) {
3890 SvUPGRADE(dstr, SVt_PVGV);
3891 (void)SvOK_off(dstr);
3892 isGV_with_GP_on(dstr);
3894 GvSTASH(dstr) = GvSTASH(sstr);
3896 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3897 gv_name_set(MUTABLE_GV(dstr), name, len,
3898 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3899 SvFAKE_on(dstr); /* can coerce to non-glob */
3902 if(GvGP(MUTABLE_GV(sstr))) {
3903 /* If source has method cache entry, clear it */
3905 SvREFCNT_dec(GvCV(sstr));
3906 GvCV_set(sstr, NULL);
3909 /* If source has a real method, then a method is
3912 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3918 /* If dest already had a real method, that's a change as well */
3920 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3921 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3926 /* We don't need to check the name of the destination if it was not a
3927 glob to begin with. */
3928 if(dtype == SVt_PVGV) {
3929 const char * const name = GvNAME((const GV *)dstr);
3932 /* The stash may have been detached from the symbol table, so
3934 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3938 const STRLEN len = GvNAMELEN(dstr);
3939 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3940 || (len == 1 && name[0] == ':')) {
3943 /* Set aside the old stash, so we can reset isa caches on
3945 if((old_stash = GvHV(dstr)))
3946 /* Make sure we do not lose it early. */
3947 SvREFCNT_inc_simple_void_NN(
3948 sv_2mortal((SV *)old_stash)
3953 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
3956 gp_free(MUTABLE_GV(dstr));
3957 GvINTRO_off(dstr); /* one-shot flag */
3958 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3959 if (SvTAINTED(sstr))
3961 if (GvIMPORTED(dstr) != GVf_IMPORTED
3962 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3964 GvIMPORTED_on(dstr);
3967 if(mro_changes == 2) {
3968 if (GvAV((const GV *)sstr)) {
3970 SV * const sref = (SV *)GvAV((const GV *)dstr);
3971 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3972 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3973 AV * const ary = newAV();
3974 av_push(ary, mg->mg_obj); /* takes the refcount */
3975 mg->mg_obj = (SV *)ary;
3977 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3979 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3981 mro_isa_changed_in(GvSTASH(dstr));
3983 else if(mro_changes == 3) {
3984 HV * const stash = GvHV(dstr);
3985 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
3991 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
3992 if (GvIO(dstr) && dtype == SVt_PVGV) {
3993 DEBUG_o(Perl_deb(aTHX_
3994 "glob_assign_glob clearing PL_stashcache\n"));
3995 /* It's a cache. It will rebuild itself quite happily.
3996 It's a lot of effort to work out exactly which key (or keys)
3997 might be invalidated by the creation of the this file handle.
3999 hv_clear(PL_stashcache);
4005 Perl_gv_setref(pTHX_ SV *const dstr, SV *const sstr)
4007 SV * const sref = SvRV(sstr);
4009 const int intro = GvINTRO(dstr);
4012 const U32 stype = SvTYPE(sref);
4014 PERL_ARGS_ASSERT_GV_SETREF;
4017 GvINTRO_off(dstr); /* one-shot flag */
4018 GvLINE(dstr) = CopLINE(PL_curcop);
4019 GvEGV(dstr) = MUTABLE_GV(dstr);
4024 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
4025 import_flag = GVf_IMPORTED_CV;
4028 location = (SV **) &GvHV(dstr);
4029 import_flag = GVf_IMPORTED_HV;
4032 location = (SV **) &GvAV(dstr);
4033 import_flag = GVf_IMPORTED_AV;
4036 location = (SV **) &GvIOp(dstr);
4039 location = (SV **) &GvFORM(dstr);
4042 location = &GvSV(dstr);
4043 import_flag = GVf_IMPORTED_SV;
4046 if (stype == SVt_PVCV) {
4047 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
4048 if (GvCVGEN(dstr)) {
4049 SvREFCNT_dec(GvCV(dstr));
4050 GvCV_set(dstr, NULL);
4051 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4054 /* SAVEt_GVSLOT takes more room on the savestack and has more
4055 overhead in leave_scope than SAVEt_GENERIC_SV. But for CVs
4056 leave_scope needs access to the GV so it can reset method
4057 caches. We must use SAVEt_GVSLOT whenever the type is
4058 SVt_PVCV, even if the stash is anonymous, as the stash may
4059 gain a name somehow before leave_scope. */
4060 if (stype == SVt_PVCV) {
4061 /* There is no save_pushptrptrptr. Creating it for this
4062 one call site would be overkill. So inline the ss add
4066 SS_ADD_PTR(location);
4067 SS_ADD_PTR(SvREFCNT_inc(*location));
4068 SS_ADD_UV(SAVEt_GVSLOT);
4071 else SAVEGENERICSV(*location);
4074 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
4075 CV* const cv = MUTABLE_CV(*location);
4077 if (!GvCVGEN((const GV *)dstr) &&
4078 (CvROOT(cv) || CvXSUB(cv)) &&
4079 /* redundant check that avoids creating the extra SV
4080 most of the time: */
4081 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
4083 SV * const new_const_sv =
4084 CvCONST((const CV *)sref)
4085 ? cv_const_sv((const CV *)sref)
4087 report_redefined_cv(
4088 sv_2mortal(Perl_newSVpvf(aTHX_
4091 HvNAME_HEK(GvSTASH((const GV *)dstr))
4093 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr)))
4096 CvCONST((const CV *)sref) ? &new_const_sv : NULL
4100 cv_ckproto_len_flags(cv, (const GV *)dstr,
4101 SvPOK(sref) ? CvPROTO(sref) : NULL,
4102 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
4103 SvPOK(sref) ? SvUTF8(sref) : 0);
4105 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4106 GvASSUMECV_on(dstr);
4107 if(GvSTASH(dstr)) { /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
4108 if (intro && GvREFCNT(dstr) > 1) {
4109 /* temporary remove extra savestack's ref */
4111 gv_method_changed(dstr);
4114 else gv_method_changed(dstr);
4117 *location = SvREFCNT_inc_simple_NN(sref);
4118 if (import_flag && !(GvFLAGS(dstr) & import_flag)
4119 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
4120 GvFLAGS(dstr) |= import_flag;
4122 if (import_flag == GVf_IMPORTED_SV) {
4124 save_aliased_sv((GV *)dstr);
4126 /* Turn off the flag if sref is not referenced elsewhere,
4127 even by weak refs. (SvRMAGICAL is a pessimistic check for
4129 if (SvREFCNT(sref) <= 2 && !SvRMAGICAL(sref))
4130 GvALIASED_SV_off(dstr);
4132 GvALIASED_SV_on(dstr);
4134 if (stype == SVt_PVHV) {
4135 const char * const name = GvNAME((GV*)dstr);
4136 const STRLEN len = GvNAMELEN(dstr);
4139 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
4140 || (len == 1 && name[0] == ':')
4142 && (!dref || HvENAME_get(dref))
4145 (HV *)sref, (HV *)dref,
4151 stype == SVt_PVAV && sref != dref
4152 && strEQ(GvNAME((GV*)dstr), "ISA")
4153 /* The stash may have been detached from the symbol table, so
4154 check its name before doing anything. */
4155 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
4158 MAGIC * const omg = dref && SvSMAGICAL(dref)
4159 ? mg_find(dref, PERL_MAGIC_isa)
4161 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
4162 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
4163 AV * const ary = newAV();
4164 av_push(ary, mg->mg_obj); /* takes the refcount */
4165 mg->mg_obj = (SV *)ary;
4168 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
4169 SV **svp = AvARRAY((AV *)omg->mg_obj);
4170 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
4174 SvREFCNT_inc_simple_NN(*svp++)
4180 SvREFCNT_inc_simple_NN(omg->mg_obj)
4184 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
4189 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
4191 mg = mg_find(sref, PERL_MAGIC_isa);
4193 /* Since the *ISA assignment could have affected more than
4194 one stash, don't call mro_isa_changed_in directly, but let
4195 magic_clearisa do it for us, as it already has the logic for
4196 dealing with globs vs arrays of globs. */
4198 Perl_magic_clearisa(aTHX_ NULL, mg);
4200 else if (stype == SVt_PVIO) {
4201 DEBUG_o(Perl_deb(aTHX_ "gv_setref clearing PL_stashcache\n"));
4202 /* It's a cache. It will rebuild itself quite happily.
4203 It's a lot of effort to work out exactly which key (or keys)
4204 might be invalidated by the creation of the this file handle.
4206 hv_clear(PL_stashcache);
4210 if (!intro) SvREFCNT_dec(dref);
4211 if (SvTAINTED(sstr))
4219 #ifdef PERL_DEBUG_READONLY_COW
4220 # include <sys/mman.h>
4222 # ifndef PERL_MEMORY_DEBUG_HEADER_SIZE
4223 # define PERL_MEMORY_DEBUG_HEADER_SIZE 0
4227 Perl_sv_buf_to_ro(pTHX_ SV *sv)
4229 struct perl_memory_debug_header * const header =
4230 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4231 const MEM_SIZE len = header->size;
4232 PERL_ARGS_ASSERT_SV_BUF_TO_RO;
4233 # ifdef PERL_TRACK_MEMPOOL
4234 if (!header->readonly) header->readonly = 1;
4236 if (mprotect(header, len, PROT_READ))
4237 Perl_warn(aTHX_ "mprotect RW for COW string %p %lu failed with %d",
4238 header, len, errno);
4242 S_sv_buf_to_rw(pTHX_ SV *sv)
4244 struct perl_memory_debug_header * const header =
4245 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4246 const MEM_SIZE len = header->size;
4247 PERL_ARGS_ASSERT_SV_BUF_TO_RW;
4248 if (mprotect(header, len, PROT_READ|PROT_WRITE))
4249 Perl_warn(aTHX_ "mprotect for COW string %p %lu failed with %d",
4250 header, len, errno);
4251 # ifdef PERL_TRACK_MEMPOOL
4252 header->readonly = 0;
4257 # define sv_buf_to_ro(sv) NOOP
4258 # define sv_buf_to_rw(sv) NOOP
4262 Perl_sv_setsv_flags(pTHX_ SV *dstr, SV* sstr, const I32 flags)
4268 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
4270 if (UNLIKELY( sstr == dstr ))
4273 if (SvIS_FREED(dstr)) {
4274 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
4275 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
4277 SV_CHECK_THINKFIRST_COW_DROP(dstr);
4278 if (UNLIKELY( !sstr ))
4279 sstr = &PL_sv_undef;
4280 if (SvIS_FREED(sstr)) {
4281 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
4282 (void*)sstr, (void*)dstr);
4284 stype = SvTYPE(sstr);
4285 dtype = SvTYPE(dstr);
4287 /* There's a lot of redundancy below but we're going for speed here */
4292 if (LIKELY( dtype != SVt_PVGV && dtype != SVt_PVLV )) {
4293 (void)SvOK_off(dstr);
4301 /* For performance, we inline promoting to type SVt_IV. */
4302 /* We're starting from SVt_NULL, so provided that define is
4303 * actual 0, we don't have to unset any SV type flags
4304 * to promote to SVt_IV. */
4305 STATIC_ASSERT_STMT(SVt_NULL == 0);
4306 SET_SVANY_FOR_BODYLESS_IV(dstr);
4307 SvFLAGS(dstr) |= SVt_IV;
4311 sv_upgrade(dstr, SVt_PVIV);
4315 goto end_of_first_switch;
4317 (void)SvIOK_only(dstr);
4318 SvIV_set(dstr, SvIVX(sstr));
4321 /* SvTAINTED can only be true if the SV has taint magic, which in
4322 turn means that the SV type is PVMG (or greater). This is the
4323 case statement for SVt_IV, so this cannot be true (whatever gcov
4325 assert(!SvTAINTED(sstr));
4330 if (dtype < SVt_PV && dtype != SVt_IV)
4331 sv_upgrade(dstr, SVt_IV);
4335 if (LIKELY( SvNOK(sstr) )) {
4339 sv_upgrade(dstr, SVt_NV);
4343 sv_upgrade(dstr, SVt_PVNV);
4347 goto end_of_first_switch;
4349 SvNV_set(dstr, SvNVX(sstr));
4350 (void)SvNOK_only(dstr);
4351 /* SvTAINTED can only be true if the SV has taint magic, which in
4352 turn means that the SV type is PVMG (or greater). This is the
4353 case statement for SVt_NV, so this cannot be true (whatever gcov
4355 assert(!SvTAINTED(sstr));
4362 sv_upgrade(dstr, SVt_PV);
4365 if (dtype < SVt_PVIV)
4366 sv_upgrade(dstr, SVt_PVIV);
4369 if (dtype < SVt_PVNV)
4370 sv_upgrade(dstr, SVt_PVNV);
4374 const char * const type = sv_reftype(sstr,0);
4376 /* diag_listed_as: Bizarre copy of %s */
4377 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4379 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4381 NOT_REACHED; /* NOTREACHED */
4385 if (dtype < SVt_REGEXP)
4387 if (dtype >= SVt_PV) {
4393 sv_upgrade(dstr, SVt_REGEXP);
4401 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4403 if (SvTYPE(sstr) != stype)
4404 stype = SvTYPE(sstr);
4406 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4407 glob_assign_glob(dstr, sstr, dtype);
4410 if (stype == SVt_PVLV)
4412 if (isREGEXP(sstr)) goto upgregexp;
4413 SvUPGRADE(dstr, SVt_PVNV);
4416 SvUPGRADE(dstr, (svtype)stype);
4418 end_of_first_switch:
4420 /* dstr may have been upgraded. */
4421 dtype = SvTYPE(dstr);
4422 sflags = SvFLAGS(sstr);
4424 if (UNLIKELY( dtype == SVt_PVCV )) {
4425 /* Assigning to a subroutine sets the prototype. */
4428 const char *const ptr = SvPV_const(sstr, len);
4430 SvGROW(dstr, len + 1);
4431 Copy(ptr, SvPVX(dstr), len + 1, char);
4432 SvCUR_set(dstr, len);
4434 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4435 CvAUTOLOAD_off(dstr);
4440 else if (UNLIKELY(dtype == SVt_PVAV || dtype == SVt_PVHV
4441 || dtype == SVt_PVFM))
4443 const char * const type = sv_reftype(dstr,0);
4445 /* diag_listed_as: Cannot copy to %s */
4446 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4448 Perl_croak(aTHX_ "Cannot copy to %s", type);
4449 } else if (sflags & SVf_ROK) {
4450 if (isGV_with_GP(dstr)
4451 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4454 if (GvIMPORTED(dstr) != GVf_IMPORTED
4455 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4457 GvIMPORTED_on(dstr);
4462 glob_assign_glob(dstr, sstr, dtype);
4466 if (dtype >= SVt_PV) {
4467 if (isGV_with_GP(dstr)) {
4468 gv_setref(dstr, sstr);
4471 if (SvPVX_const(dstr)) {
4477 (void)SvOK_off(dstr);
4478 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4479 SvFLAGS(dstr) |= sflags & SVf_ROK;
4480 assert(!(sflags & SVp_NOK));
4481 assert(!(sflags & SVp_IOK));
4482 assert(!(sflags & SVf_NOK));
4483 assert(!(sflags & SVf_IOK));
4485 else if (isGV_with_GP(dstr)) {
4486 if (!(sflags & SVf_OK)) {
4487 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4488 "Undefined value assigned to typeglob");
4491 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4492 if (dstr != (const SV *)gv) {
4493 const char * const name = GvNAME((const GV *)dstr);
4494 const STRLEN len = GvNAMELEN(dstr);
4495 HV *old_stash = NULL;
4496 bool reset_isa = FALSE;
4497 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4498 || (len == 1 && name[0] == ':')) {
4499 /* Set aside the old stash, so we can reset isa caches
4500 on its subclasses. */
4501 if((old_stash = GvHV(dstr))) {
4502 /* Make sure we do not lose it early. */
4503 SvREFCNT_inc_simple_void_NN(
4504 sv_2mortal((SV *)old_stash)
4511 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
4512 gp_free(MUTABLE_GV(dstr));
4514 GvGP_set(dstr, gp_ref(GvGP(gv)));
4517 HV * const stash = GvHV(dstr);
4519 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4529 else if ((dtype == SVt_REGEXP || dtype == SVt_PVLV)
4530 && (stype == SVt_REGEXP || isREGEXP(sstr))) {
4531 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4533 else if (sflags & SVp_POK) {
4534 const STRLEN cur = SvCUR(sstr);
4535 const STRLEN len = SvLEN(sstr);
4538 * We have three basic ways to copy the string:
4544 * Which we choose is based on various factors. The following
4545 * things are listed in order of speed, fastest to slowest:
4547 * - Copying a short string
4548 * - Copy-on-write bookkeeping
4550 * - Copying a long string
4552 * We swipe the string (steal the string buffer) if the SV on the
4553 * rhs is about to be freed anyway (TEMP and refcnt==1). This is a
4554 * big win on long strings. It should be a win on short strings if
4555 * SvPVX_const(dstr) has to be allocated. If not, it should not
4556 * slow things down, as SvPVX_const(sstr) would have been freed
4559 * We also steal the buffer from a PADTMP (operator target) if it
4560 * is ‘long enough’. For short strings, a swipe does not help
4561 * here, as it causes more malloc calls the next time the target
4562 * is used. Benchmarks show that even if SvPVX_const(dstr) has to
4563 * be allocated it is still not worth swiping PADTMPs for short
4564 * strings, as the savings here are small.
4566 * If swiping is not an option, then we see whether it is
4567 * worth using copy-on-write. If the lhs already has a buf-
4568 * fer big enough and the string is short, we skip it and fall back
4569 * to method 3, since memcpy is faster for short strings than the
4570 * later bookkeeping overhead that copy-on-write entails.
4572 * If the rhs is not a copy-on-write string yet, then we also
4573 * consider whether the buffer is too large relative to the string
4574 * it holds. Some operations such as readline allocate a large
4575 * buffer in the expectation of reusing it. But turning such into
4576 * a COW buffer is counter-productive because it increases memory
4577 * usage by making readline allocate a new large buffer the sec-
4578 * ond time round. So, if the buffer is too large, again, we use
4581 * Finally, if there is no buffer on the left, or the buffer is too
4582 * small, then we use copy-on-write and make both SVs share the
4587 /* Whichever path we take through the next code, we want this true,
4588 and doing it now facilitates the COW check. */
4589 (void)SvPOK_only(dstr);
4593 /* slated for free anyway (and not COW)? */
4594 (sflags & (SVs_TEMP|SVf_IsCOW)) == SVs_TEMP
4595 /* or a swipable TARG */
4597 (SVs_PADTMP|SVf_READONLY|SVf_PROTECT|SVf_IsCOW))
4599 /* whose buffer is worth stealing */
4600 && CHECK_COWBUF_THRESHOLD(cur,len)
4603 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4604 (!(flags & SV_NOSTEAL)) &&
4605 /* and we're allowed to steal temps */