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
1528 /* prior to 5.000 stable, this function returned the new OOK-less SvFLAGS
1529 prior to 5.23.4 this function always returned 0
1533 Perl_sv_backoff(SV *const sv)
1536 const char * const s = SvPVX_const(sv);
1538 PERL_ARGS_ASSERT_SV_BACKOFF;
1541 assert(SvTYPE(sv) != SVt_PVHV);
1542 assert(SvTYPE(sv) != SVt_PVAV);
1544 SvOOK_offset(sv, delta);
1546 SvLEN_set(sv, SvLEN(sv) + delta);
1547 SvPV_set(sv, SvPVX(sv) - delta);
1548 SvFLAGS(sv) &= ~SVf_OOK;
1549 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1556 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1557 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1558 Use the C<SvGROW> wrapper instead.
1563 static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1566 Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
1570 PERL_ARGS_ASSERT_SV_GROW;
1574 if (SvTYPE(sv) < SVt_PV) {
1575 sv_upgrade(sv, SVt_PV);
1576 s = SvPVX_mutable(sv);
1578 else if (SvOOK(sv)) { /* pv is offset? */
1580 s = SvPVX_mutable(sv);
1581 if (newlen > SvLEN(sv))
1582 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1586 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
1587 s = SvPVX_mutable(sv);
1590 #ifdef PERL_COPY_ON_WRITE
1591 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
1592 * to store the COW count. So in general, allocate one more byte than
1593 * asked for, to make it likely this byte is always spare: and thus
1594 * make more strings COW-able.
1595 * If the new size is a big power of two, don't bother: we assume the
1596 * caller wanted a nice 2^N sized block and will be annoyed at getting
1598 * Only increment if the allocation isn't MEM_SIZE_MAX,
1599 * otherwise it will wrap to 0.
1601 if (newlen & 0xff && newlen != MEM_SIZE_MAX)
1605 #if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1606 #define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1609 if (newlen > SvLEN(sv)) { /* need more room? */
1610 STRLEN minlen = SvCUR(sv);
1611 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1612 if (newlen < minlen)
1614 #ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1616 /* Don't round up on the first allocation, as odds are pretty good that
1617 * the initial request is accurate as to what is really needed */
1619 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1620 if (rounded > newlen)
1624 if (SvLEN(sv) && s) {
1625 s = (char*)saferealloc(s, newlen);
1628 s = (char*)safemalloc(newlen);
1629 if (SvPVX_const(sv) && SvCUR(sv)) {
1630 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1634 #ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1635 /* Do this here, do it once, do it right, and then we will never get
1636 called back into sv_grow() unless there really is some growing
1638 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1640 SvLEN_set(sv, newlen);
1647 =for apidoc sv_setiv
1649 Copies an integer into the given SV, upgrading first if necessary.
1650 Does not handle 'set' magic. See also C<L</sv_setiv_mg>>.
1656 Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
1658 PERL_ARGS_ASSERT_SV_SETIV;
1660 SV_CHECK_THINKFIRST_COW_DROP(sv);
1661 switch (SvTYPE(sv)) {
1664 sv_upgrade(sv, SVt_IV);
1667 sv_upgrade(sv, SVt_PVIV);
1671 if (!isGV_with_GP(sv))
1678 /* diag_listed_as: Can't coerce %s to %s in %s */
1679 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1684 (void)SvIOK_only(sv); /* validate number */
1690 =for apidoc sv_setiv_mg
1692 Like C<sv_setiv>, but also handles 'set' magic.
1698 Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
1700 PERL_ARGS_ASSERT_SV_SETIV_MG;
1707 =for apidoc sv_setuv
1709 Copies an unsigned integer into the given SV, upgrading first if necessary.
1710 Does not handle 'set' magic. See also C<L</sv_setuv_mg>>.
1716 Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
1718 PERL_ARGS_ASSERT_SV_SETUV;
1720 /* With the if statement to ensure that integers are stored as IVs whenever
1722 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1725 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1727 If you wish to remove the following if statement, so that this routine
1728 (and its callers) always return UVs, please benchmark to see what the
1729 effect is. Modern CPUs may be different. Or may not :-)
1731 if (u <= (UV)IV_MAX) {
1732 sv_setiv(sv, (IV)u);
1741 =for apidoc sv_setuv_mg
1743 Like C<sv_setuv>, but also handles 'set' magic.
1749 Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
1751 PERL_ARGS_ASSERT_SV_SETUV_MG;
1758 =for apidoc sv_setnv
1760 Copies a double into the given SV, upgrading first if necessary.
1761 Does not handle 'set' magic. See also C<L</sv_setnv_mg>>.
1767 Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
1769 PERL_ARGS_ASSERT_SV_SETNV;
1771 SV_CHECK_THINKFIRST_COW_DROP(sv);
1772 switch (SvTYPE(sv)) {
1775 sv_upgrade(sv, SVt_NV);
1779 sv_upgrade(sv, SVt_PVNV);
1783 if (!isGV_with_GP(sv))
1790 /* diag_listed_as: Can't coerce %s to %s in %s */
1791 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1797 (void)SvNOK_only(sv); /* validate number */
1802 =for apidoc sv_setnv_mg
1804 Like C<sv_setnv>, but also handles 'set' magic.
1810 Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
1812 PERL_ARGS_ASSERT_SV_SETNV_MG;
1818 /* Return a cleaned-up, printable version of sv, for non-numeric, or
1819 * not incrementable warning display.
1820 * Originally part of S_not_a_number().
1821 * The return value may be != tmpbuf.
1825 S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1828 PERL_ARGS_ASSERT_SV_DISPLAY;
1831 SV *dsv = newSVpvs_flags("", SVs_TEMP);
1832 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
1835 const char * const limit = tmpbuf + tmpbuf_size - 8;
1836 /* each *s can expand to 4 chars + "...\0",
1837 i.e. need room for 8 chars */
1839 const char *s = SvPVX_const(sv);
1840 const char * const end = s + SvCUR(sv);
1841 for ( ; s < end && d < limit; s++ ) {
1843 if (! isASCII(ch) && !isPRINT_LC(ch)) {
1847 /* Map to ASCII "equivalent" of Latin1 */
1848 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
1854 else if (ch == '\r') {
1858 else if (ch == '\f') {
1862 else if (ch == '\\') {
1866 else if (ch == '\0') {
1870 else if (isPRINT_LC(ch))
1889 /* Print an "isn't numeric" warning, using a cleaned-up,
1890 * printable version of the offending string
1894 S_not_a_number(pTHX_ SV *const sv)
1899 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1901 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1904 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1905 /* diag_listed_as: Argument "%s" isn't numeric%s */
1906 "Argument \"%s\" isn't numeric in %s", pv,
1909 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1910 /* diag_listed_as: Argument "%s" isn't numeric%s */
1911 "Argument \"%s\" isn't numeric", pv);
1915 S_not_incrementable(pTHX_ SV *const sv) {
1919 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1921 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1923 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1924 "Argument \"%s\" treated as 0 in increment (++)", pv);
1928 =for apidoc looks_like_number
1930 Test if the content of an SV looks like a number (or is a number).
1931 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1932 non-numeric warning), even if your C<atof()> doesn't grok them. Get-magic is
1939 Perl_looks_like_number(pTHX_ SV *const sv)
1945 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1947 if (SvPOK(sv) || SvPOKp(sv)) {
1948 sbegin = SvPV_nomg_const(sv, len);
1951 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1952 numtype = grok_number(sbegin, len, NULL);
1953 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
1957 S_glob_2number(pTHX_ GV * const gv)
1959 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1961 /* We know that all GVs stringify to something that is not-a-number,
1962 so no need to test that. */
1963 if (ckWARN(WARN_NUMERIC))
1965 SV *const buffer = sv_newmortal();
1966 gv_efullname3(buffer, gv, "*");
1967 not_a_number(buffer);
1969 /* We just want something true to return, so that S_sv_2iuv_common
1970 can tail call us and return true. */
1974 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1975 until proven guilty, assume that things are not that bad... */
1980 As 64 bit platforms often have an NV that doesn't preserve all bits of
1981 an IV (an assumption perl has been based on to date) it becomes necessary
1982 to remove the assumption that the NV always carries enough precision to
1983 recreate the IV whenever needed, and that the NV is the canonical form.
1984 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1985 precision as a side effect of conversion (which would lead to insanity
1986 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1987 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1988 where precision was lost, and IV/UV/NV slots that have a valid conversion
1989 which has lost no precision
1990 2) to ensure that if a numeric conversion to one form is requested that
1991 would lose precision, the precise conversion (or differently
1992 imprecise conversion) is also performed and cached, to prevent
1993 requests for different numeric formats on the same SV causing
1994 lossy conversion chains. (lossless conversion chains are perfectly
1999 SvIOKp is true if the IV slot contains a valid value
2000 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
2001 SvNOKp is true if the NV slot contains a valid value
2002 SvNOK is true only if the NV value is accurate
2005 while converting from PV to NV, check to see if converting that NV to an
2006 IV(or UV) would lose accuracy over a direct conversion from PV to
2007 IV(or UV). If it would, cache both conversions, return NV, but mark
2008 SV as IOK NOKp (ie not NOK).
2010 While converting from PV to IV, check to see if converting that IV to an
2011 NV would lose accuracy over a direct conversion from PV to NV. If it
2012 would, cache both conversions, flag similarly.
2014 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
2015 correctly because if IV & NV were set NV *always* overruled.
2016 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
2017 changes - now IV and NV together means that the two are interchangeable:
2018 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
2020 The benefit of this is that operations such as pp_add know that if
2021 SvIOK is true for both left and right operands, then integer addition
2022 can be used instead of floating point (for cases where the result won't
2023 overflow). Before, floating point was always used, which could lead to
2024 loss of precision compared with integer addition.
2026 * making IV and NV equal status should make maths accurate on 64 bit
2028 * may speed up maths somewhat if pp_add and friends start to use
2029 integers when possible instead of fp. (Hopefully the overhead in
2030 looking for SvIOK and checking for overflow will not outweigh the
2031 fp to integer speedup)
2032 * will slow down integer operations (callers of SvIV) on "inaccurate"
2033 values, as the change from SvIOK to SvIOKp will cause a call into
2034 sv_2iv each time rather than a macro access direct to the IV slot
2035 * should speed up number->string conversion on integers as IV is
2036 favoured when IV and NV are equally accurate
2038 ####################################################################
2039 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2040 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2041 On the other hand, SvUOK is true iff UV.
2042 ####################################################################
2044 Your mileage will vary depending your CPU's relative fp to integer
2048 #ifndef NV_PRESERVES_UV
2049 # define IS_NUMBER_UNDERFLOW_IV 1
2050 # define IS_NUMBER_UNDERFLOW_UV 2
2051 # define IS_NUMBER_IV_AND_UV 2
2052 # define IS_NUMBER_OVERFLOW_IV 4
2053 # define IS_NUMBER_OVERFLOW_UV 5
2055 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
2057 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2059 S_sv_2iuv_non_preserve(pTHX_ SV *const sv
2065 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
2066 PERL_UNUSED_CONTEXT;
2068 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));
2069 if (SvNVX(sv) < (NV)IV_MIN) {
2070 (void)SvIOKp_on(sv);
2072 SvIV_set(sv, IV_MIN);
2073 return IS_NUMBER_UNDERFLOW_IV;
2075 if (SvNVX(sv) > (NV)UV_MAX) {
2076 (void)SvIOKp_on(sv);
2079 SvUV_set(sv, UV_MAX);
2080 return IS_NUMBER_OVERFLOW_UV;
2082 (void)SvIOKp_on(sv);
2084 /* Can't use strtol etc to convert this string. (See truth table in
2086 if (SvNVX(sv) <= (UV)IV_MAX) {
2087 SvIV_set(sv, I_V(SvNVX(sv)));
2088 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2089 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2091 /* Integer is imprecise. NOK, IOKp */
2093 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2096 SvUV_set(sv, U_V(SvNVX(sv)));
2097 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2098 if (SvUVX(sv) == UV_MAX) {
2099 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2100 possibly be preserved by NV. Hence, it must be overflow.
2102 return IS_NUMBER_OVERFLOW_UV;
2104 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2106 /* Integer is imprecise. NOK, IOKp */
2108 return IS_NUMBER_OVERFLOW_IV;
2110 #endif /* !NV_PRESERVES_UV*/
2112 /* If numtype is infnan, set the NV of the sv accordingly.
2113 * If numtype is anything else, try setting the NV using Atof(PV). */
2115 # pragma warning(push)
2116 # pragma warning(disable:4756;disable:4056)
2119 S_sv_setnv(pTHX_ SV* sv, int numtype)
2121 bool pok = cBOOL(SvPOK(sv));
2123 if ((numtype & IS_NUMBER_INFINITY)) {
2124 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
2127 else if ((numtype & IS_NUMBER_NAN)) {
2128 SvNV_set(sv, NV_NAN);
2132 SvNV_set(sv, Atof(SvPVX_const(sv)));
2133 /* Purposefully no true nok here, since we don't want to blow
2134 * away the possible IOK/UV of an existing sv. */
2137 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
2139 SvPOK_on(sv); /* PV is okay, though. */
2143 # pragma warning(pop)
2147 S_sv_2iuv_common(pTHX_ SV *const sv)
2149 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2152 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2153 * without also getting a cached IV/UV from it at the same time
2154 * (ie PV->NV conversion should detect loss of accuracy and cache
2155 * IV or UV at same time to avoid this. */
2156 /* IV-over-UV optimisation - choose to cache IV if possible */
2158 if (SvTYPE(sv) == SVt_NV)
2159 sv_upgrade(sv, SVt_PVNV);
2161 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2162 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2163 certainly cast into the IV range at IV_MAX, whereas the correct
2164 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2166 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2167 if (Perl_isnan(SvNVX(sv))) {
2173 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2174 SvIV_set(sv, I_V(SvNVX(sv)));
2175 if (SvNVX(sv) == (NV) SvIVX(sv)
2176 #ifndef NV_PRESERVES_UV
2177 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
2178 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2179 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2180 /* Don't flag it as "accurately an integer" if the number
2181 came from a (by definition imprecise) NV operation, and
2182 we're outside the range of NV integer precision */
2186 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2188 /* scalar has trailing garbage, eg "42a" */
2190 DEBUG_c(PerlIO_printf(Perl_debug_log,
2191 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
2197 /* IV not precise. No need to convert from PV, as NV
2198 conversion would already have cached IV if it detected
2199 that PV->IV would be better than PV->NV->IV
2200 flags already correct - don't set public IOK. */
2201 DEBUG_c(PerlIO_printf(Perl_debug_log,
2202 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
2207 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2208 but the cast (NV)IV_MIN rounds to a the value less (more
2209 negative) than IV_MIN which happens to be equal to SvNVX ??
2210 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2211 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2212 (NV)UVX == NVX are both true, but the values differ. :-(
2213 Hopefully for 2s complement IV_MIN is something like
2214 0x8000000000000000 which will be exact. NWC */
2217 SvUV_set(sv, U_V(SvNVX(sv)));
2219 (SvNVX(sv) == (NV) SvUVX(sv))
2220 #ifndef NV_PRESERVES_UV
2221 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2222 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2223 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2224 /* Don't flag it as "accurately an integer" if the number
2225 came from a (by definition imprecise) NV operation, and
2226 we're outside the range of NV integer precision */
2232 DEBUG_c(PerlIO_printf(Perl_debug_log,
2233 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2239 else if (SvPOKp(sv)) {
2241 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2242 /* We want to avoid a possible problem when we cache an IV/ a UV which
2243 may be later translated to an NV, and the resulting NV is not
2244 the same as the direct translation of the initial string
2245 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2246 be careful to ensure that the value with the .456 is around if the
2247 NV value is requested in the future).
2249 This means that if we cache such an IV/a UV, we need to cache the
2250 NV as well. Moreover, we trade speed for space, and do not
2251 cache the NV if we are sure it's not needed.
2254 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2255 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2256 == IS_NUMBER_IN_UV) {
2257 /* It's definitely an integer, only upgrade to PVIV */
2258 if (SvTYPE(sv) < SVt_PVIV)
2259 sv_upgrade(sv, SVt_PVIV);
2261 } else if (SvTYPE(sv) < SVt_PVNV)
2262 sv_upgrade(sv, SVt_PVNV);
2264 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
2265 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
2267 S_sv_setnv(aTHX_ sv, numtype);
2271 /* If NVs preserve UVs then we only use the UV value if we know that
2272 we aren't going to call atof() below. If NVs don't preserve UVs
2273 then the value returned may have more precision than atof() will
2274 return, even though value isn't perfectly accurate. */
2275 if ((numtype & (IS_NUMBER_IN_UV
2276 #ifdef NV_PRESERVES_UV
2279 )) == IS_NUMBER_IN_UV) {
2280 /* This won't turn off the public IOK flag if it was set above */
2281 (void)SvIOKp_on(sv);
2283 if (!(numtype & IS_NUMBER_NEG)) {
2285 if (value <= (UV)IV_MAX) {
2286 SvIV_set(sv, (IV)value);
2288 /* it didn't overflow, and it was positive. */
2289 SvUV_set(sv, value);
2293 /* 2s complement assumption */
2294 if (value <= (UV)IV_MIN) {
2295 SvIV_set(sv, value == (UV)IV_MIN
2296 ? IV_MIN : -(IV)value);
2298 /* Too negative for an IV. This is a double upgrade, but
2299 I'm assuming it will be rare. */
2300 if (SvTYPE(sv) < SVt_PVNV)
2301 sv_upgrade(sv, SVt_PVNV);
2305 SvNV_set(sv, -(NV)value);
2306 SvIV_set(sv, IV_MIN);
2310 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2311 will be in the previous block to set the IV slot, and the next
2312 block to set the NV slot. So no else here. */
2314 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2315 != IS_NUMBER_IN_UV) {
2316 /* It wasn't an (integer that doesn't overflow the UV). */
2317 S_sv_setnv(aTHX_ sv, numtype);
2319 if (! numtype && ckWARN(WARN_NUMERIC))
2322 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
2323 PTR2UV(sv), SvNVX(sv)));
2325 #ifdef NV_PRESERVES_UV
2326 (void)SvIOKp_on(sv);
2328 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2329 if (Perl_isnan(SvNVX(sv))) {
2335 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2336 SvIV_set(sv, I_V(SvNVX(sv)));
2337 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2340 NOOP; /* Integer is imprecise. NOK, IOKp */
2342 /* UV will not work better than IV */
2344 if (SvNVX(sv) > (NV)UV_MAX) {
2346 /* Integer is inaccurate. NOK, IOKp, is UV */
2347 SvUV_set(sv, UV_MAX);
2349 SvUV_set(sv, U_V(SvNVX(sv)));
2350 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2351 NV preservse UV so can do correct comparison. */
2352 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2355 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2360 #else /* NV_PRESERVES_UV */
2361 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2362 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2363 /* The IV/UV slot will have been set from value returned by
2364 grok_number above. The NV slot has just been set using
2367 assert (SvIOKp(sv));
2369 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2370 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2371 /* Small enough to preserve all bits. */
2372 (void)SvIOKp_on(sv);
2374 SvIV_set(sv, I_V(SvNVX(sv)));
2375 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2377 /* Assumption: first non-preserved integer is < IV_MAX,
2378 this NV is in the preserved range, therefore: */
2379 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2381 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);
2385 0 0 already failed to read UV.
2386 0 1 already failed to read UV.
2387 1 0 you won't get here in this case. IV/UV
2388 slot set, public IOK, Atof() unneeded.
2389 1 1 already read UV.
2390 so there's no point in sv_2iuv_non_preserve() attempting
2391 to use atol, strtol, strtoul etc. */
2393 sv_2iuv_non_preserve (sv, numtype);
2395 sv_2iuv_non_preserve (sv);
2399 #endif /* NV_PRESERVES_UV */
2400 /* It might be more code efficient to go through the entire logic above
2401 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2402 gets complex and potentially buggy, so more programmer efficient
2403 to do it this way, by turning off the public flags: */
2405 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2409 if (isGV_with_GP(sv))
2410 return glob_2number(MUTABLE_GV(sv));
2412 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2414 if (SvTYPE(sv) < SVt_IV)
2415 /* Typically the caller expects that sv_any is not NULL now. */
2416 sv_upgrade(sv, SVt_IV);
2417 /* Return 0 from the caller. */
2424 =for apidoc sv_2iv_flags
2426 Return the integer value of an SV, doing any necessary string
2427 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2428 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2434 Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
2436 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
2438 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2439 && SvTYPE(sv) != SVt_PVFM);
2441 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2447 if (flags & SV_SKIP_OVERLOAD)
2449 tmpstr = AMG_CALLunary(sv, numer_amg);
2450 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2451 return SvIV(tmpstr);
2454 return PTR2IV(SvRV(sv));
2457 if (SvVALID(sv) || isREGEXP(sv)) {
2458 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2459 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2460 In practice they are extremely unlikely to actually get anywhere
2461 accessible by user Perl code - the only way that I'm aware of is when
2462 a constant subroutine which is used as the second argument to index.
2464 Regexps have no SvIVX and SvNVX fields.
2466 assert(isREGEXP(sv) || SvPOKp(sv));
2469 const char * const ptr =
2470 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2472 = grok_number(ptr, SvCUR(sv), &value);
2474 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2475 == IS_NUMBER_IN_UV) {
2476 /* It's definitely an integer */
2477 if (numtype & IS_NUMBER_NEG) {
2478 if (value < (UV)IV_MIN)
2481 if (value < (UV)IV_MAX)
2486 /* Quite wrong but no good choices. */
2487 if ((numtype & IS_NUMBER_INFINITY)) {
2488 return (numtype & IS_NUMBER_NEG) ? IV_MIN : IV_MAX;
2489 } else if ((numtype & IS_NUMBER_NAN)) {
2490 return 0; /* So wrong. */
2494 if (ckWARN(WARN_NUMERIC))
2497 return I_V(Atof(ptr));
2501 if (SvTHINKFIRST(sv)) {
2502 if (SvREADONLY(sv) && !SvOK(sv)) {
2503 if (ckWARN(WARN_UNINITIALIZED))
2510 if (S_sv_2iuv_common(aTHX_ sv))
2514 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2515 PTR2UV(sv),SvIVX(sv)));
2516 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2520 =for apidoc sv_2uv_flags
2522 Return the unsigned integer value of an SV, doing any necessary string
2523 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2524 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2530 Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
2532 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
2534 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2540 if (flags & SV_SKIP_OVERLOAD)
2542 tmpstr = AMG_CALLunary(sv, numer_amg);
2543 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2544 return SvUV(tmpstr);
2547 return PTR2UV(SvRV(sv));
2550 if (SvVALID(sv) || isREGEXP(sv)) {
2551 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2552 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2553 Regexps have no SvIVX and SvNVX fields. */
2554 assert(isREGEXP(sv) || SvPOKp(sv));
2557 const char * const ptr =
2558 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2560 = grok_number(ptr, SvCUR(sv), &value);
2562 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2563 == IS_NUMBER_IN_UV) {
2564 /* It's definitely an integer */
2565 if (!(numtype & IS_NUMBER_NEG))
2569 /* Quite wrong but no good choices. */
2570 if ((numtype & IS_NUMBER_INFINITY)) {
2571 return UV_MAX; /* So wrong. */
2572 } else if ((numtype & IS_NUMBER_NAN)) {
2573 return 0; /* So wrong. */
2577 if (ckWARN(WARN_NUMERIC))
2580 return U_V(Atof(ptr));
2584 if (SvTHINKFIRST(sv)) {
2585 if (SvREADONLY(sv) && !SvOK(sv)) {
2586 if (ckWARN(WARN_UNINITIALIZED))
2593 if (S_sv_2iuv_common(aTHX_ sv))
2597 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2598 PTR2UV(sv),SvUVX(sv)));
2599 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2603 =for apidoc sv_2nv_flags
2605 Return the num value of an SV, doing any necessary string or integer
2606 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2607 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2613 Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
2615 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2617 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2618 && SvTYPE(sv) != SVt_PVFM);
2619 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2620 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2621 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2622 Regexps have no SvIVX and SvNVX fields. */
2624 if (flags & SV_GMAGIC)
2628 if (SvPOKp(sv) && !SvIOKp(sv)) {
2629 ptr = SvPVX_const(sv);
2631 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2632 !grok_number(ptr, SvCUR(sv), NULL))
2638 return (NV)SvUVX(sv);
2640 return (NV)SvIVX(sv);
2646 ptr = RX_WRAPPED((REGEXP *)sv);
2649 assert(SvTYPE(sv) >= SVt_PVMG);
2650 /* This falls through to the report_uninit near the end of the
2652 } else if (SvTHINKFIRST(sv)) {
2657 if (flags & SV_SKIP_OVERLOAD)
2659 tmpstr = AMG_CALLunary(sv, numer_amg);
2660 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2661 return SvNV(tmpstr);
2664 return PTR2NV(SvRV(sv));
2666 if (SvREADONLY(sv) && !SvOK(sv)) {
2667 if (ckWARN(WARN_UNINITIALIZED))
2672 if (SvTYPE(sv) < SVt_NV) {
2673 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2674 sv_upgrade(sv, SVt_NV);
2676 STORE_NUMERIC_LOCAL_SET_STANDARD();
2677 PerlIO_printf(Perl_debug_log,
2678 "0x%"UVxf" num(%" NVgf ")\n",
2679 PTR2UV(sv), SvNVX(sv));
2680 RESTORE_NUMERIC_LOCAL();
2683 else if (SvTYPE(sv) < SVt_PVNV)
2684 sv_upgrade(sv, SVt_PVNV);
2689 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2690 #ifdef NV_PRESERVES_UV
2696 /* Only set the public NV OK flag if this NV preserves the IV */
2697 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2699 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2700 : (SvIVX(sv) == I_V(SvNVX(sv))))
2706 else if (SvPOKp(sv)) {
2708 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2709 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2711 #ifdef NV_PRESERVES_UV
2712 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2713 == IS_NUMBER_IN_UV) {
2714 /* It's definitely an integer */
2715 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2717 S_sv_setnv(aTHX_ sv, numtype);
2724 SvNV_set(sv, Atof(SvPVX_const(sv)));
2725 /* Only set the public NV OK flag if this NV preserves the value in
2726 the PV at least as well as an IV/UV would.
2727 Not sure how to do this 100% reliably. */
2728 /* if that shift count is out of range then Configure's test is
2729 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2731 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2732 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2733 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2734 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2735 /* Can't use strtol etc to convert this string, so don't try.
2736 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2739 /* value has been set. It may not be precise. */
2740 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
2741 /* 2s complement assumption for (UV)IV_MIN */
2742 SvNOK_on(sv); /* Integer is too negative. */
2747 if (numtype & IS_NUMBER_NEG) {
2748 /* -IV_MIN is undefined, but we should never reach
2749 * this point with both IS_NUMBER_NEG and value ==
2751 assert(value != (UV)IV_MIN);
2752 SvIV_set(sv, -(IV)value);
2753 } else if (value <= (UV)IV_MAX) {
2754 SvIV_set(sv, (IV)value);
2756 SvUV_set(sv, value);
2760 if (numtype & IS_NUMBER_NOT_INT) {
2761 /* I believe that even if the original PV had decimals,
2762 they are lost beyond the limit of the FP precision.
2763 However, neither is canonical, so both only get p
2764 flags. NWC, 2000/11/25 */
2765 /* Both already have p flags, so do nothing */
2767 const NV nv = SvNVX(sv);
2768 /* XXX should this spot have NAN_COMPARE_BROKEN, too? */
2769 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2770 if (SvIVX(sv) == I_V(nv)) {
2773 /* It had no "." so it must be integer. */
2777 /* between IV_MAX and NV(UV_MAX).
2778 Could be slightly > UV_MAX */
2780 if (numtype & IS_NUMBER_NOT_INT) {
2781 /* UV and NV both imprecise. */
2783 const UV nv_as_uv = U_V(nv);
2785 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2794 /* It might be more code efficient to go through the entire logic above
2795 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2796 gets complex and potentially buggy, so more programmer efficient
2797 to do it this way, by turning off the public flags: */
2799 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2800 #endif /* NV_PRESERVES_UV */
2803 if (isGV_with_GP(sv)) {
2804 glob_2number(MUTABLE_GV(sv));
2808 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2810 assert (SvTYPE(sv) >= SVt_NV);
2811 /* Typically the caller expects that sv_any is not NULL now. */
2812 /* XXX Ilya implies that this is a bug in callers that assume this
2813 and ideally should be fixed. */
2817 STORE_NUMERIC_LOCAL_SET_STANDARD();
2818 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" NVgf ")\n",
2819 PTR2UV(sv), SvNVX(sv));
2820 RESTORE_NUMERIC_LOCAL();
2828 Return an SV with the numeric value of the source SV, doing any necessary
2829 reference or overload conversion. The caller is expected to have handled
2836 Perl_sv_2num(pTHX_ SV *const sv)
2838 PERL_ARGS_ASSERT_SV_2NUM;
2843 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2844 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2845 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2846 return sv_2num(tmpsv);
2848 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2851 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2852 * UV as a string towards the end of buf, and return pointers to start and
2855 * We assume that buf is at least TYPE_CHARS(UV) long.
2859 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2861 char *ptr = buf + TYPE_CHARS(UV);
2862 char * const ebuf = ptr;
2865 PERL_ARGS_ASSERT_UIV_2BUF;
2873 uv = (iv == IV_MIN) ? (UV)iv : (UV)(-iv);
2877 *--ptr = '0' + (char)(uv % 10);
2885 /* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
2886 * infinity or a not-a-number, writes the appropriate strings to the
2887 * buffer, including a zero byte. On success returns the written length,
2888 * excluding the zero byte, on failure (not an infinity, not a nan)
2889 * returns zero, assert-fails on maxlen being too short.
2891 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2892 * shared string constants we point to, instead of generating a new
2893 * string for each instance. */
2895 S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
2897 assert(maxlen >= 4);
2898 if (Perl_isinf(nv)) {
2900 if (maxlen < 5) /* "-Inf\0" */
2910 else if (Perl_isnan(nv)) {
2914 /* XXX optionally output the payload mantissa bits as
2915 * "(unsigned)" (to match the nan("...") C99 function,
2916 * or maybe as "(0xhhh...)" would make more sense...
2917 * provide a format string so that the user can decide?
2918 * NOTE: would affect the maxlen and assert() logic.*/
2923 assert((s == buffer + 3) || (s == buffer + 4));
2925 return s - buffer - 1; /* -1: excluding the zero byte */
2929 =for apidoc sv_2pv_flags
2931 Returns a pointer to the string value of an SV, and sets C<*lp> to its length.
2932 If flags has the C<SV_GMAGIC> bit set, does an C<mg_get()> first. Coerces C<sv> to a
2933 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2934 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2940 Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2944 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2946 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2947 && SvTYPE(sv) != SVt_PVFM);
2948 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2953 if (flags & SV_SKIP_OVERLOAD)
2955 tmpstr = AMG_CALLunary(sv, string_amg);
2956 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2957 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2959 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2963 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2964 if (flags & SV_CONST_RETURN) {
2965 pv = (char *) SvPVX_const(tmpstr);
2967 pv = (flags & SV_MUTABLE_RETURN)
2968 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2971 *lp = SvCUR(tmpstr);
2973 pv = sv_2pv_flags(tmpstr, lp, flags);
2986 SV *const referent = SvRV(sv);
2990 retval = buffer = savepvn("NULLREF", len);
2991 } else if (SvTYPE(referent) == SVt_REGEXP &&
2992 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
2993 amagic_is_enabled(string_amg))) {
2994 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2998 /* If the regex is UTF-8 we want the containing scalar to
2999 have an UTF-8 flag too */
3006 *lp = RX_WRAPLEN(re);
3008 return RX_WRAPPED(re);
3010 const char *const typestr = sv_reftype(referent, 0);
3011 const STRLEN typelen = strlen(typestr);
3012 UV addr = PTR2UV(referent);
3013 const char *stashname = NULL;
3014 STRLEN stashnamelen = 0; /* hush, gcc */
3015 const char *buffer_end;
3017 if (SvOBJECT(referent)) {
3018 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3021 stashname = HEK_KEY(name);
3022 stashnamelen = HEK_LEN(name);
3024 if (HEK_UTF8(name)) {
3030 stashname = "__ANON__";
3033 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3034 + 2 * sizeof(UV) + 2 /* )\0 */;
3036 len = typelen + 3 /* (0x */
3037 + 2 * sizeof(UV) + 2 /* )\0 */;
3040 Newx(buffer, len, char);
3041 buffer_end = retval = buffer + len;
3043 /* Working backwards */
3047 *--retval = PL_hexdigit[addr & 15];
3048 } while (addr >>= 4);
3054 memcpy(retval, typestr, typelen);
3058 retval -= stashnamelen;
3059 memcpy(retval, stashname, stashnamelen);
3061 /* retval may not necessarily have reached the start of the
3063 assert (retval >= buffer);
3065 len = buffer_end - retval - 1; /* -1 for that \0 */
3077 if (flags & SV_MUTABLE_RETURN)
3078 return SvPVX_mutable(sv);
3079 if (flags & SV_CONST_RETURN)
3080 return (char *)SvPVX_const(sv);
3085 /* I'm assuming that if both IV and NV are equally valid then
3086 converting the IV is going to be more efficient */
3087 const U32 isUIOK = SvIsUV(sv);
3088 char buf[TYPE_CHARS(UV)];
3092 if (SvTYPE(sv) < SVt_PVIV)
3093 sv_upgrade(sv, SVt_PVIV);
3094 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
3096 /* inlined from sv_setpvn */
3097 s = SvGROW_mutable(sv, len + 1);
3098 Move(ptr, s, len, char);
3103 else if (SvNOK(sv)) {
3104 if (SvTYPE(sv) < SVt_PVNV)
3105 sv_upgrade(sv, SVt_PVNV);
3106 if (SvNVX(sv) == 0.0
3107 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3108 && !Perl_isnan(SvNVX(sv))
3111 s = SvGROW_mutable(sv, 2);
3116 STRLEN size = 5; /* "-Inf\0" */
3118 s = SvGROW_mutable(sv, size);
3119 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
3125 /* some Xenix systems wipe out errno here */
3134 5 + /* exponent digits */
3138 s = SvGROW_mutable(sv, size);
3139 #ifndef USE_LOCALE_NUMERIC
3140 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3146 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3147 STORE_LC_NUMERIC_SET_TO_NEEDED();
3151 PL_numeric_radix_sv &&
3152 SvUTF8(PL_numeric_radix_sv);
3153 if (local_radix && SvLEN(PL_numeric_radix_sv) > 1) {
3154 size += SvLEN(PL_numeric_radix_sv) - 1;
3155 s = SvGROW_mutable(sv, size);
3158 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3160 /* If the radix character is UTF-8, and actually is in the
3161 * output, turn on the UTF-8 flag for the scalar */
3163 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3167 RESTORE_LC_NUMERIC();
3170 /* We don't call SvPOK_on(), because it may come to
3171 * pass that the locale changes so that the
3172 * stringification we just did is no longer correct. We
3173 * will have to re-stringify every time it is needed */
3180 else if (isGV_with_GP(sv)) {
3181 GV *const gv = MUTABLE_GV(sv);
3182 SV *const buffer = sv_newmortal();
3184 gv_efullname3(buffer, gv, "*");
3186 assert(SvPOK(buffer));
3190 *lp = SvCUR(buffer);
3191 return SvPVX(buffer);
3193 else if (isREGEXP(sv)) {
3194 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3195 return RX_WRAPPED((REGEXP *)sv);
3200 if (flags & SV_UNDEF_RETURNS_NULL)
3202 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
3204 /* Typically the caller expects that sv_any is not NULL now. */
3205 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
3206 sv_upgrade(sv, SVt_PV);
3211 const STRLEN len = s - SvPVX_const(sv);
3216 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3217 PTR2UV(sv),SvPVX_const(sv)));
3218 if (flags & SV_CONST_RETURN)
3219 return (char *)SvPVX_const(sv);
3220 if (flags & SV_MUTABLE_RETURN)
3221 return SvPVX_mutable(sv);
3226 =for apidoc sv_copypv
3228 Copies a stringified representation of the source SV into the
3229 destination SV. Automatically performs any necessary C<mg_get> and
3230 coercion of numeric values into strings. Guaranteed to preserve
3231 C<UTF8> flag even from overloaded objects. Similar in nature to
3232 C<sv_2pv[_flags]> but operates directly on an SV instead of just the
3233 string. Mostly uses C<sv_2pv_flags> to do its work, except when that
3234 would lose the UTF-8'ness of the PV.
3236 =for apidoc sv_copypv_nomg
3238 Like C<sv_copypv>, but doesn't invoke get magic first.
3240 =for apidoc sv_copypv_flags
3242 Implementation of C<sv_copypv> and C<sv_copypv_nomg>. Calls get magic iff flags
3243 has the C<SV_GMAGIC> bit set.
3249 Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
3254 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
3256 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
3257 sv_setpvn(dsv,s,len);
3265 =for apidoc sv_2pvbyte
3267 Return a pointer to the byte-encoded representation of the SV, and set C<*lp>
3268 to its length. May cause the SV to be downgraded from UTF-8 as a
3271 Usually accessed via the C<SvPVbyte> macro.
3277 Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
3279 PERL_ARGS_ASSERT_SV_2PVBYTE;
3282 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3283 || isGV_with_GP(sv) || SvROK(sv)) {
3284 SV *sv2 = sv_newmortal();
3285 sv_copypv_nomg(sv2,sv);
3288 sv_utf8_downgrade(sv,0);
3289 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3293 =for apidoc sv_2pvutf8
3295 Return a pointer to the UTF-8-encoded representation of the SV, and set C<*lp>
3296 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3298 Usually accessed via the C<SvPVutf8> macro.
3304 Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
3306 PERL_ARGS_ASSERT_SV_2PVUTF8;
3308 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3309 || isGV_with_GP(sv) || SvROK(sv))
3310 sv = sv_mortalcopy(sv);
3313 sv_utf8_upgrade_nomg(sv);
3314 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3319 =for apidoc sv_2bool
3321 This macro is only used by C<sv_true()> or its macro equivalent, and only if
3322 the latter's argument is neither C<SvPOK>, C<SvIOK> nor C<SvNOK>.
3323 It calls C<sv_2bool_flags> with the C<SV_GMAGIC> flag.
3325 =for apidoc sv_2bool_flags
3327 This function is only used by C<sv_true()> and friends, and only if
3328 the latter's argument is neither C<SvPOK>, C<SvIOK> nor C<SvNOK>. If the flags
3329 contain C<SV_GMAGIC>, then it does an C<mg_get()> first.
3336 Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
3338 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3341 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3347 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3348 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3351 if(SvGMAGICAL(sv)) {
3353 goto restart; /* call sv_2bool */
3355 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3356 else if(!SvOK(sv)) {
3359 else if(SvPOK(sv)) {
3360 svb = SvPVXtrue(sv);
3362 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3363 svb = (SvIOK(sv) && SvIVX(sv) != 0)
3364 || (SvNOK(sv) && SvNVX(sv) != 0.0);
3368 goto restart; /* call sv_2bool_nomg */
3373 return SvRV(sv) != 0;
3377 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
3378 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
3382 =for apidoc sv_utf8_upgrade
3384 Converts the PV of an SV to its UTF-8-encoded form.
3385 Forces the SV to string form if it is not already.
3386 Will C<mg_get> on C<sv> if appropriate.
3387 Always sets the C<SvUTF8> flag to avoid future validity checks even
3388 if the whole string is the same in UTF-8 as not.
3389 Returns the number of bytes in the converted string
3391 This is not a general purpose byte encoding to Unicode interface:
3392 use the Encode extension for that.
3394 =for apidoc sv_utf8_upgrade_nomg
3396 Like C<sv_utf8_upgrade>, but doesn't do magic on C<sv>.
3398 =for apidoc sv_utf8_upgrade_flags
3400 Converts the PV of an SV to its UTF-8-encoded form.
3401 Forces the SV to string form if it is not already.
3402 Always sets the SvUTF8 flag to avoid future validity checks even
3403 if all the bytes are invariant in UTF-8.
3404 If C<flags> has C<SV_GMAGIC> bit set,
3405 will C<mg_get> on C<sv> if appropriate, else not.
3407 If C<flags> has C<SV_FORCE_UTF8_UPGRADE> set, this function assumes that the PV
3408 will expand when converted to UTF-8, and skips the extra work of checking for
3409 that. Typically this flag is used by a routine that has already parsed the
3410 string and found such characters, and passes this information on so that the
3411 work doesn't have to be repeated.
3413 Returns the number of bytes in the converted string.
3415 This is not a general purpose byte encoding to Unicode interface:
3416 use the Encode extension for that.
3418 =for apidoc sv_utf8_upgrade_flags_grow
3420 Like C<sv_utf8_upgrade_flags>, but has an additional parameter C<extra>, which is
3421 the number of unused bytes the string of C<sv> is guaranteed to have free after
3422 it upon return. This allows the caller to reserve extra space that it intends
3423 to fill, to avoid extra grows.
3425 C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3426 are implemented in terms of this function.
3428 Returns the number of bytes in the converted string (not including the spares).
3432 (One might think that the calling routine could pass in the position of the
3433 first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3434 have to be found again. But that is not the case, because typically when the
3435 caller is likely to use this flag, it won't be calling this routine unless it
3436 finds something that won't fit into a byte. Otherwise it tries to not upgrade
3437 and just use bytes. But some things that do fit into a byte are variants in
3438 utf8, and the caller may not have been keeping track of these.)
3440 If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3441 C<NUL> isn't guaranteed due to having other routines do the work in some input
3442 cases, or if the input is already flagged as being in utf8.
3444 The speed of this could perhaps be improved for many cases if someone wanted to
3445 write a fast function that counts the number of variant characters in a string,
3446 especially if it could return the position of the first one.
3451 Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
3453 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3455 if (sv == &PL_sv_undef)
3457 if (!SvPOK_nog(sv)) {
3459 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3460 (void) sv_2pv_flags(sv,&len, flags);
3462 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3466 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3471 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3476 S_sv_uncow(aTHX_ sv, 0);
3479 if (IN_ENCODING && !(flags & SV_UTF8_NO_ENCODING)) {
3480 sv_recode_to_utf8(sv, _get_encoding());
3481 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3485 if (SvCUR(sv) == 0) {
3486 if (extra) SvGROW(sv, extra);
3487 } else { /* Assume Latin-1/EBCDIC */
3488 /* This function could be much more efficient if we
3489 * had a FLAG in SVs to signal if there are any variant
3490 * chars in the PV. Given that there isn't such a flag
3491 * make the loop as fast as possible (although there are certainly ways
3492 * to speed this up, eg. through vectorization) */
3493 U8 * s = (U8 *) SvPVX_const(sv);
3494 U8 * e = (U8 *) SvEND(sv);
3496 STRLEN two_byte_count = 0;
3498 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3500 /* See if really will need to convert to utf8. We mustn't rely on our
3501 * incoming SV being well formed and having a trailing '\0', as certain
3502 * code in pp_formline can send us partially built SVs. */
3506 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
3508 t--; /* t already incremented; re-point to first variant */
3513 /* utf8 conversion not needed because all are invariants. Mark as
3514 * UTF-8 even if no variant - saves scanning loop */
3516 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3521 /* Here, the string should be converted to utf8, either because of an
3522 * input flag (two_byte_count = 0), or because a character that
3523 * requires 2 bytes was found (two_byte_count = 1). t points either to
3524 * the beginning of the string (if we didn't examine anything), or to
3525 * the first variant. In either case, everything from s to t - 1 will
3526 * occupy only 1 byte each on output.
3528 * There are two main ways to convert. One is to create a new string
3529 * and go through the input starting from the beginning, appending each
3530 * converted value onto the new string as we go along. It's probably
3531 * best to allocate enough space in the string for the worst possible
3532 * case rather than possibly running out of space and having to
3533 * reallocate and then copy what we've done so far. Since everything
3534 * from s to t - 1 is invariant, the destination can be initialized
3535 * with these using a fast memory copy
3537 * The other way is to figure out exactly how big the string should be
3538 * by parsing the entire input. Then you don't have to make it big
3539 * enough to handle the worst possible case, and more importantly, if
3540 * the string you already have is large enough, you don't have to
3541 * allocate a new string, you can copy the last character in the input
3542 * string to the final position(s) that will be occupied by the
3543 * converted string and go backwards, stopping at t, since everything
3544 * before that is invariant.
3546 * There are advantages and disadvantages to each method.
3548 * In the first method, we can allocate a new string, do the memory
3549 * copy from the s to t - 1, and then proceed through the rest of the
3550 * string byte-by-byte.
3552 * In the second method, we proceed through the rest of the input
3553 * string just calculating how big the converted string will be. Then
3554 * there are two cases:
3555 * 1) if the string has enough extra space to handle the converted
3556 * value. We go backwards through the string, converting until we
3557 * get to the position we are at now, and then stop. If this
3558 * position is far enough along in the string, this method is
3559 * faster than the other method. If the memory copy were the same
3560 * speed as the byte-by-byte loop, that position would be about
3561 * half-way, as at the half-way mark, parsing to the end and back
3562 * is one complete string's parse, the same amount as starting
3563 * over and going all the way through. Actually, it would be
3564 * somewhat less than half-way, as it's faster to just count bytes
3565 * than to also copy, and we don't have the overhead of allocating
3566 * a new string, changing the scalar to use it, and freeing the
3567 * existing one. But if the memory copy is fast, the break-even
3568 * point is somewhere after half way. The counting loop could be
3569 * sped up by vectorization, etc, to move the break-even point
3570 * further towards the beginning.
3571 * 2) if the string doesn't have enough space to handle the converted
3572 * value. A new string will have to be allocated, and one might
3573 * as well, given that, start from the beginning doing the first
3574 * method. We've spent extra time parsing the string and in
3575 * exchange all we've gotten is that we know precisely how big to
3576 * make the new one. Perl is more optimized for time than space,
3577 * so this case is a loser.
3578 * So what I've decided to do is not use the 2nd method unless it is
3579 * guaranteed that a new string won't have to be allocated, assuming
3580 * the worst case. I also decided not to put any more conditions on it
3581 * than this, for now. It seems likely that, since the worst case is
3582 * twice as big as the unknown portion of the string (plus 1), we won't
3583 * be guaranteed enough space, causing us to go to the first method,
3584 * unless the string is short, or the first variant character is near
3585 * the end of it. In either of these cases, it seems best to use the
3586 * 2nd method. The only circumstance I can think of where this would
3587 * be really slower is if the string had once had much more data in it
3588 * than it does now, but there is still a substantial amount in it */
3591 STRLEN invariant_head = t - s;
3592 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3593 if (SvLEN(sv) < size) {
3595 /* Here, have decided to allocate a new string */
3600 Newx(dst, size, U8);
3602 /* If no known invariants at the beginning of the input string,
3603 * set so starts from there. Otherwise, can use memory copy to
3604 * get up to where we are now, and then start from here */
3606 if (invariant_head == 0) {
3609 Copy(s, dst, invariant_head, char);
3610 d = dst + invariant_head;
3614 append_utf8_from_native_byte(*t, &d);
3618 SvPV_free(sv); /* No longer using pre-existing string */
3619 SvPV_set(sv, (char*)dst);
3620 SvCUR_set(sv, d - dst);
3621 SvLEN_set(sv, size);
3624 /* Here, have decided to get the exact size of the string.
3625 * Currently this happens only when we know that there is
3626 * guaranteed enough space to fit the converted string, so
3627 * don't have to worry about growing. If two_byte_count is 0,
3628 * then t points to the first byte of the string which hasn't
3629 * been examined yet. Otherwise two_byte_count is 1, and t
3630 * points to the first byte in the string that will expand to
3631 * two. Depending on this, start examining at t or 1 after t.
3634 U8 *d = t + two_byte_count;
3637 /* Count up the remaining bytes that expand to two */
3640 const U8 chr = *d++;
3641 if (! NATIVE_BYTE_IS_INVARIANT(chr)) two_byte_count++;
3644 /* The string will expand by just the number of bytes that
3645 * occupy two positions. But we are one afterwards because of
3646 * the increment just above. This is the place to put the
3647 * trailing NUL, and to set the length before we decrement */
3649 d += two_byte_count;
3650 SvCUR_set(sv, d - s);
3654 /* Having decremented d, it points to the position to put the
3655 * very last byte of the expanded string. Go backwards through
3656 * the string, copying and expanding as we go, stopping when we
3657 * get to the part that is invariant the rest of the way down */
3661 if (NATIVE_BYTE_IS_INVARIANT(*e)) {
3664 *d-- = UTF8_EIGHT_BIT_LO(*e);
3665 *d-- = UTF8_EIGHT_BIT_HI(*e);
3671 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3672 /* Update pos. We do it at the end rather than during
3673 * the upgrade, to avoid slowing down the common case
3674 * (upgrade without pos).
3675 * pos can be stored as either bytes or characters. Since
3676 * this was previously a byte string we can just turn off
3677 * the bytes flag. */
3678 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3680 mg->mg_flags &= ~MGf_BYTES;
3682 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3683 magic_setutf8(sv,mg); /* clear UTF8 cache */
3688 /* Mark as UTF-8 even if no variant - saves scanning loop */
3694 =for apidoc sv_utf8_downgrade
3696 Attempts to convert the PV of an SV from characters to bytes.
3697 If the PV contains a character that cannot fit
3698 in a byte, this conversion will fail;
3699 in this case, either returns false or, if C<fail_ok> is not
3702 This is not a general purpose Unicode to byte encoding interface:
3703 use the C<Encode> extension for that.
3709 Perl_sv_utf8_downgrade(pTHX_ SV *const sv, const bool fail_ok)
3711 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3713 if (SvPOKp(sv) && SvUTF8(sv)) {
3717 int mg_flags = SV_GMAGIC;
3720 S_sv_uncow(aTHX_ sv, 0);
3722 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3724 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3725 if (mg && mg->mg_len > 0 && mg->mg_flags & MGf_BYTES) {
3726 mg->mg_len = sv_pos_b2u_flags(sv, mg->mg_len,
3727 SV_GMAGIC|SV_CONST_RETURN);
3728 mg_flags = 0; /* sv_pos_b2u does get magic */
3730 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3731 magic_setutf8(sv,mg); /* clear UTF8 cache */
3734 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3736 if (!utf8_to_bytes(s, &len)) {
3741 Perl_croak(aTHX_ "Wide character in %s",
3744 Perl_croak(aTHX_ "Wide character");
3755 =for apidoc sv_utf8_encode
3757 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3758 flag off so that it looks like octets again.
3764 Perl_sv_utf8_encode(pTHX_ SV *const sv)
3766 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3768 if (SvREADONLY(sv)) {
3769 sv_force_normal_flags(sv, 0);
3771 (void) sv_utf8_upgrade(sv);
3776 =for apidoc sv_utf8_decode
3778 If the PV of the SV is an octet sequence in UTF-8
3779 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3780 so that it looks like a character. If the PV contains only single-byte
3781 characters, the C<SvUTF8> flag stays off.
3782 Scans PV for validity and returns false if the PV is invalid UTF-8.
3788 Perl_sv_utf8_decode(pTHX_ SV *const sv)
3790 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3793 const U8 *start, *c;
3796 /* The octets may have got themselves encoded - get them back as
3799 if (!sv_utf8_downgrade(sv, TRUE))
3802 /* it is actually just a matter of turning the utf8 flag on, but
3803 * we want to make sure everything inside is valid utf8 first.
3805 c = start = (const U8 *) SvPVX_const(sv);
3806 if (!is_utf8_string(c, SvCUR(sv)))
3808 e = (const U8 *) SvEND(sv);
3811 if (!UTF8_IS_INVARIANT(ch)) {
3816 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3817 /* XXX Is this dead code? XS_utf8_decode calls SvSETMAGIC
3818 after this, clearing pos. Does anything on CPAN
3820 /* adjust pos to the start of a UTF8 char sequence */
3821 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3823 I32 pos = mg->mg_len;
3825 for (c = start + pos; c > start; c--) {
3826 if (UTF8_IS_START(*c))
3829 mg->mg_len = c - start;
3832 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3833 magic_setutf8(sv,mg); /* clear UTF8 cache */
3840 =for apidoc sv_setsv
3842 Copies the contents of the source SV C<ssv> into the destination SV
3843 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3844 function if the source SV needs to be reused. Does not handle 'set' magic on
3845 destination SV. Calls 'get' magic on source SV. Loosely speaking, it
3846 performs a copy-by-value, obliterating any previous content of the
3849 You probably want to use one of the assortment of wrappers, such as
3850 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3851 C<SvSetMagicSV_nosteal>.
3853 =for apidoc sv_setsv_flags
3855 Copies the contents of the source SV C<ssv> into the destination SV
3856 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3857 function if the source SV needs to be reused. Does not handle 'set' magic.
3858 Loosely speaking, it performs a copy-by-value, obliterating any previous
3859 content of the destination.
3860 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3861 C<ssv> if appropriate, else not. If the C<flags>
3862 parameter has the C<SV_NOSTEAL> bit set then the
3863 buffers of temps will not be stolen. C<sv_setsv>
3864 and C<sv_setsv_nomg> are implemented in terms of this function.
3866 You probably want to use one of the assortment of wrappers, such as
3867 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3868 C<SvSetMagicSV_nosteal>.
3870 This is the primary function for copying scalars, and most other
3871 copy-ish functions and macros use this underneath.
3877 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3879 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3880 HV *old_stash = NULL;
3882 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3884 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3885 const char * const name = GvNAME(sstr);
3886 const STRLEN len = GvNAMELEN(sstr);
3888 if (dtype >= SVt_PV) {
3894 SvUPGRADE(dstr, SVt_PVGV);
3895 (void)SvOK_off(dstr);
3896 isGV_with_GP_on(dstr);
3898 GvSTASH(dstr) = GvSTASH(sstr);
3900 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3901 gv_name_set(MUTABLE_GV(dstr), name, len,
3902 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3903 SvFAKE_on(dstr); /* can coerce to non-glob */
3906 if(GvGP(MUTABLE_GV(sstr))) {
3907 /* If source has method cache entry, clear it */
3909 SvREFCNT_dec(GvCV(sstr));
3910 GvCV_set(sstr, NULL);
3913 /* If source has a real method, then a method is
3916 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3922 /* If dest already had a real method, that's a change as well */
3924 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3925 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3930 /* We don't need to check the name of the destination if it was not a
3931 glob to begin with. */
3932 if(dtype == SVt_PVGV) {
3933 const char * const name = GvNAME((const GV *)dstr);
3936 /* The stash may have been detached from the symbol table, so
3938 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3942 const STRLEN len = GvNAMELEN(dstr);
3943 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3944 || (len == 1 && name[0] == ':')) {
3947 /* Set aside the old stash, so we can reset isa caches on
3949 if((old_stash = GvHV(dstr)))
3950 /* Make sure we do not lose it early. */
3951 SvREFCNT_inc_simple_void_NN(
3952 sv_2mortal((SV *)old_stash)
3957 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
3960 /* freeing dstr's GP might free sstr (e.g. *x = $x),
3961 * so temporarily protect it */
3963 SAVEFREESV(SvREFCNT_inc_simple_NN(sstr));
3964 gp_free(MUTABLE_GV(dstr));
3965 GvINTRO_off(dstr); /* one-shot flag */
3966 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3969 if (SvTAINTED(sstr))
3971 if (GvIMPORTED(dstr) != GVf_IMPORTED
3972 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3974 GvIMPORTED_on(dstr);
3977 if(mro_changes == 2) {
3978 if (GvAV((const GV *)sstr)) {
3980 SV * const sref = (SV *)GvAV((const GV *)dstr);
3981 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3982 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3983 AV * const ary = newAV();
3984 av_push(ary, mg->mg_obj); /* takes the refcount */
3985 mg->mg_obj = (SV *)ary;
3987 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3989 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3991 mro_isa_changed_in(GvSTASH(dstr));
3993 else if(mro_changes == 3) {
3994 HV * const stash = GvHV(dstr);
3995 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
4001 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
4002 if (GvIO(dstr) && dtype == SVt_PVGV) {
4003 DEBUG_o(Perl_deb(aTHX_
4004 "glob_assign_glob clearing PL_stashcache\n"));
4005 /* It's a cache. It will rebuild itself quite happily.
4006 It's a lot of effort to work out exactly which key (or keys)
4007 might be invalidated by the creation of the this file handle.
4009 hv_clear(PL_stashcache);
4015 Perl_gv_setref(pTHX_ SV *const dstr, SV *const sstr)
4017 SV * const sref = SvRV(sstr);
4019 const int intro = GvINTRO(dstr);
4022 const U32 stype = SvTYPE(sref);
4024 PERL_ARGS_ASSERT_GV_SETREF;
4027 GvINTRO_off(dstr); /* one-shot flag */
4028 GvLINE(dstr) = CopLINE(PL_curcop);
4029 GvEGV(dstr) = MUTABLE_GV(dstr);
4034 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
4035 import_flag = GVf_IMPORTED_CV;
4038 location = (SV **) &GvHV(dstr);
4039 import_flag = GVf_IMPORTED_HV;
4042 location = (SV **) &GvAV(dstr);
4043 import_flag = GVf_IMPORTED_AV;
4046 location = (SV **) &GvIOp(dstr);
4049 location = (SV **) &GvFORM(dstr);
4052 location = &GvSV(dstr);
4053 import_flag = GVf_IMPORTED_SV;
4056 if (stype == SVt_PVCV) {
4057 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
4058 if (GvCVGEN(dstr)) {
4059 SvREFCNT_dec(GvCV(dstr));
4060 GvCV_set(dstr, NULL);
4061 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4064 /* SAVEt_GVSLOT takes more room on the savestack and has more
4065 overhead in leave_scope than SAVEt_GENERIC_SV. But for CVs
4066 leave_scope needs access to the GV so it can reset method
4067 caches. We must use SAVEt_GVSLOT whenever the type is
4068 SVt_PVCV, even if the stash is anonymous, as the stash may
4069 gain a name somehow before leave_scope. */
4070 if (stype == SVt_PVCV) {
4071 /* There is no save_pushptrptrptr. Creating it for this
4072 one call site would be overkill. So inline the ss add
4076 SS_ADD_PTR(location);
4077 SS_ADD_PTR(SvREFCNT_inc(*location));
4078 SS_ADD_UV(SAVEt_GVSLOT);
4081 else SAVEGENERICSV(*location);
4084 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
4085 CV* const cv = MUTABLE_CV(*location);
4087 if (!GvCVGEN((const GV *)dstr) &&
4088 (CvROOT(cv) || CvXSUB(cv)) &&
4089 /* redundant check that avoids creating the extra SV
4090 most of the time: */
4091 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
4093 SV * const new_const_sv =
4094 CvCONST((const CV *)sref)
4095 ? cv_const_sv((const CV *)sref)
4097 report_redefined_cv(
4098 sv_2mortal(Perl_newSVpvf(aTHX_
4101 HvNAME_HEK(GvSTASH((const GV *)dstr))
4103 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr)))
4106 CvCONST((const CV *)sref) ? &new_const_sv : NULL
4110 cv_ckproto_len_flags(cv, (const GV *)dstr,
4111 SvPOK(sref) ? CvPROTO(sref) : NULL,
4112 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
4113 SvPOK(sref) ? SvUTF8(sref) : 0);
4115 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4116 GvASSUMECV_on(dstr);
4117 if(GvSTASH(dstr)) { /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
4118 if (intro && GvREFCNT(dstr) > 1) {
4119 /* temporary remove extra savestack's ref */
4121 gv_method_changed(dstr);
4124 else gv_method_changed(dstr);
4127 *location = SvREFCNT_inc_simple_NN(sref);
4128 if (import_flag && !(GvFLAGS(dstr) & import_flag)
4129 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
4130 GvFLAGS(dstr) |= import_flag;
4133 if (stype == SVt_PVHV) {
4134 const char * const name = GvNAME((GV*)dstr);
4135 const STRLEN len = GvNAMELEN(dstr);
4138 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
4139 || (len == 1 && name[0] == ':')
4141 && (!dref || HvENAME_get(dref))
4144 (HV *)sref, (HV *)dref,
4150 stype == SVt_PVAV && sref != dref
4151 && strEQ(GvNAME((GV*)dstr), "ISA")
4152 /* The stash may have been detached from the symbol table, so
4153 check its name before doing anything. */
4154 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
4157 MAGIC * const omg = dref && SvSMAGICAL(dref)
4158 ? mg_find(dref, PERL_MAGIC_isa)
4160 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
4161 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
4162 AV * const ary = newAV();
4163 av_push(ary, mg->mg_obj); /* takes the refcount */
4164 mg->mg_obj = (SV *)ary;
4167 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
4168 SV **svp = AvARRAY((AV *)omg->mg_obj);
4169 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
4173 SvREFCNT_inc_simple_NN(*svp++)
4179 SvREFCNT_inc_simple_NN(omg->mg_obj)
4183 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
4189 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
4191 for (i = 0; i <= AvFILL(sref); ++i) {
4192 SV **elem = av_fetch ((AV*)sref, i, 0);
4195 *elem, sref, PERL_MAGIC_isaelem, NULL, i
4199 mg = mg_find(sref, PERL_MAGIC_isa);
4201 /* Since the *ISA assignment could have affected more than
4202 one stash, don't call mro_isa_changed_in directly, but let
4203 magic_clearisa do it for us, as it already has the logic for
4204 dealing with globs vs arrays of globs. */
4206 Perl_magic_clearisa(aTHX_ NULL, mg);
4208 else if (stype == SVt_PVIO) {
4209 DEBUG_o(Perl_deb(aTHX_ "gv_setref clearing PL_stashcache\n"));
4210 /* It's a cache. It will rebuild itself quite happily.
4211 It's a lot of effort to work out exactly which key (or keys)
4212 might be invalidated by the creation of the this file handle.
4214 hv_clear(PL_stashcache);
4218 if (!intro) SvREFCNT_dec(dref);
4219 if (SvTAINTED(sstr))
4227 #ifdef PERL_DEBUG_READONLY_COW
4228 # include <sys/mman.h>
4230 # ifndef PERL_MEMORY_DEBUG_HEADER_SIZE
4231 # define PERL_MEMORY_DEBUG_HEADER_SIZE 0
4235 Perl_sv_buf_to_ro(pTHX_ SV *sv)
4237 struct perl_memory_debug_header * const header =
4238 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4239 const MEM_SIZE len = header->size;
4240 PERL_ARGS_ASSERT_SV_BUF_TO_RO;
4241 # ifdef PERL_TRACK_MEMPOOL
4242 if (!header->readonly) header->readonly = 1;
4244 if (mprotect(header, len, PROT_READ))
4245 Perl_warn(aTHX_ "mprotect RW for COW string %p %lu failed with %d",
4246 header, len, errno);
4250 S_sv_buf_to_rw(pTHX_ SV *sv)
4252 struct perl_memory_debug_header * const header =
4253 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4254 const MEM_SIZE len = header->size;
4255 PERL_ARGS_ASSERT_SV_BUF_TO_RW;
4256 if (mprotect(header, len, PROT_READ|PROT_WRITE))
4257 Perl_warn(aTHX_ "mprotect for COW string %p %lu failed with %d",
4258 header, len, errno);
4259 # ifdef PERL_TRACK_MEMPOOL
4260 header->readonly = 0;
4265 # define sv_buf_to_ro(sv) NOOP
4266 # define sv_buf_to_rw(sv) NOOP
4270 Perl_sv_setsv_flags(pTHX_ SV *dstr, SV* sstr, const I32 flags)
4276 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
4278 if (UNLIKELY( sstr == dstr ))
4281 if (SvIS_FREED(dstr)) {
4282 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
4283 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
4285 SV_CHECK_THINKFIRST_COW_DROP(dstr);
4286 if (UNLIKELY( !sstr ))
4287 sstr = &PL_sv_undef;
4288 if (SvIS_FREED(sstr)) {
4289 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
4290 (void*)sstr, (void*)dstr);
4292 stype = SvTYPE(sstr);
4293 dtype = SvTYPE(dstr);
4295 /* There's a lot of redundancy below but we're going for speed here */
4300 if (LIKELY( dtype != SVt_PVGV && dtype != SVt_PVLV )) {
4301 (void)SvOK_off(dstr);
4309 /* For performance, we inline promoting to type SVt_IV. */
4310 /* We're starting from SVt_NULL, so provided that define is
4311 * actual 0, we don't have to unset any SV type flags
4312 * to promote to SVt_IV. */
4313 STATIC_ASSERT_STMT(SVt_NULL == 0);
4314 SET_SVANY_FOR_BODYLESS_IV(dstr);
4315 SvFLAGS(dstr) |= SVt_IV;
4319 sv_upgrade(dstr, SVt_PVIV);
4323 goto end_of_first_switch;
4325 (void)SvIOK_only(dstr);
4326 SvIV_set(dstr, SvIVX(sstr));
4329 /* SvTAINTED can only be true if the SV has taint magic, which in
4330 turn means that the SV type is PVMG (or greater). This is the
4331 case statement for SVt_IV, so this cannot be true (whatever gcov
4333 assert(!SvTAINTED(sstr));
4338 if (dtype < SVt_PV && dtype != SVt_IV)
4339 sv_upgrade(dstr, SVt_IV);
4343 if (LIKELY( SvNOK(sstr) )) {
4347 sv_upgrade(dstr, SVt_NV);
4351 sv_upgrade(dstr, SVt_PVNV);
4355 goto end_of_first_switch;
4357 SvNV_set(dstr, SvNVX(sstr));
4358 (void)SvNOK_only(dstr);
4359 /* SvTAINTED can only be true if the SV has taint magic, which in
4360 turn means that the SV type is PVMG (or greater). This is the
4361 case statement for SVt_NV, so this cannot be true (whatever gcov
4363 assert(!SvTAINTED(sstr));
4370 sv_upgrade(dstr, SVt_PV);
4373 if (dtype < SVt_PVIV)
4374 sv_upgrade(dstr, SVt_PVIV);
4377 if (dtype < SVt_PVNV)
4378 sv_upgrade(dstr, SVt_PVNV);
4382 const char * const type = sv_reftype(sstr,0);
4384 /* diag_listed_as: Bizarre copy of %s */
4385 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4387 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4389 NOT_REACHED; /* NOTREACHED */
4393 if (dtype < SVt_REGEXP)
4395 if (dtype >= SVt_PV) {
4401 sv_upgrade(dstr, SVt_REGEXP);
4409 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4411 if (SvTYPE(sstr) != stype)
4412 stype = SvTYPE(sstr);
4414 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4415 glob_assign_glob(dstr, sstr, dtype);
4418 if (stype == SVt_PVLV)
4420 if (isREGEXP(sstr)) goto upgregexp;
4421 SvUPGRADE(dstr, SVt_PVNV);
4424 SvUPGRADE(dstr, (svtype)stype);
4426 end_of_first_switch:
4428 /* dstr may have been upgraded. */
4429 dtype = SvTYPE(dstr);
4430 sflags = SvFLAGS(sstr);
4432 if (UNLIKELY( dtype == SVt_PVCV )) {
4433 /* Assigning to a subroutine sets the prototype. */
4436 const char *const ptr = SvPV_const(sstr, len);
4438 SvGROW(dstr, len + 1);
4439 Copy(ptr, SvPVX(dstr), len + 1, char);
4440 SvCUR_set(dstr, len);
4442 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4443 CvAUTOLOAD_off(dstr);
4448 else if (UNLIKELY(dtype == SVt_PVAV || dtype == SVt_PVHV
4449 || dtype == SVt_PVFM))
4451 const char * const type = sv_reftype(dstr,0);
4453 /* diag_listed_as: Cannot copy to %s */
4454 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4456 Perl_croak(aTHX_ "Cannot copy to %s", type);
4457 } else if (sflags & SVf_ROK) {
4458 if (isGV_with_GP(dstr)
4459 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4462 if (GvIMPORTED(dstr) != GVf_IMPORTED
4463 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4465 GvIMPORTED_on(dstr);
4470 glob_assign_glob(dstr, sstr, dtype);
4474 if (dtype >= SVt_PV) {
4475 if (isGV_with_GP(dstr)) {
4476 gv_setref(dstr, sstr);
4479 if (SvPVX_const(dstr)) {
4485 (void)SvOK_off(dstr);
4486 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4487 SvFLAGS(dstr) |= sflags & SVf_ROK;
4488 assert(!(sflags & SVp_NOK));
4489 assert(!(sflags & SVp_IOK));
4490 assert(!(sflags & SVf_NOK));
4491 assert(!(sflags & SVf_IOK));
4493 else if (isGV_with_GP(dstr)) {
4494 if (!(sflags & SVf_OK)) {
4495 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4496 "Undefined value assigned to typeglob");
4499 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4500 if (dstr != (const SV *)gv) {
4501 const char * const name = GvNAME((const GV *)dstr);
4502 const STRLEN len = GvNAMELEN(dstr);
4503 HV *old_stash = NULL;
4504 bool reset_isa = FALSE;
4505 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4506 || (len == 1 && name[0] == ':')) {
4507 /* Set aside the old stash, so we can reset isa caches
4508 on its subclasses. */
4509 if((old_stash = GvHV(dstr))) {
4510 /* Make sure we do not lose it early. */
4511 SvREFCNT_inc_simple_void_NN(
4512 sv_2mortal((SV *)old_stash)
4519 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
4520 gp_free(MUTABLE_GV(dstr));
4522 GvGP_set(dstr, gp_ref(GvGP(gv)));
4525 HV * const stash = GvHV(dstr);
4527 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4537 else if ((dtype == SVt_REGEXP || dtype == SVt_PVLV)
4538 && (stype == SVt_REGEXP || isREGEXP(sstr))) {
4539 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4541 else if (sflags & SVp_POK) {
4542 const STRLEN cur = SvCUR(sstr);
4543 const STRLEN len = SvLEN(sstr);
4546 * We have three basic ways to copy the string:
4552 * Which we choose is based on various factors. The following
4553 * things are listed in order of speed, fastest to slowest:
4555 * - Copying a short string
4556 * - Copy-on-write bookkeeping
4558 * - Copying a long string
4560 * We swipe the string (steal the string buffer) if the SV on the
4561 * rhs is about to be freed anyway (TEMP and refcnt==1). This is a
4562 * big win on long strings. It should be a win on short strings if
4563 * SvPVX_const(dstr) has to be allocated. If not, it should not
4564 * slow things down, as SvPVX_const(sstr) would have been freed
4567 * We also steal the buffer from a PADTMP (operator target) if it
4568 * is ‘long enough’. For short strings, a swipe does not help
4569 * here, as it causes more malloc calls the next time the target
4570 * is used. Benchmarks show that even if SvPVX_const(dstr) has to
4571 * be allocated it is still not worth swiping PADTMPs for short
4572 * strings, as the savings here are small.
4574 * If swiping is not an option, then we see whether it is
4575 * worth using copy-on-write. If the lhs already has a buf-
4576 * fer big enough and the string is short, we skip it and fall back
4577 * to method 3, since memcpy is faster for short strings than the
4578 * later bookkeeping overhead that copy-on-write entails.
4580 * If the rhs is not a copy-on-write string yet, then we also
4581 * consider whether the buffer is too large relative to the string
4582 * it holds. Some operations such as readline allocate a large
4583 * buffer in the expectation of reusing it. But turning such into
4584 * a COW buffer is counter-productive because it increases memory
4585 * usage by making readline allocate a new large buffer the sec-
4586 * ond time round. So, if the buffer is too large, again, we use
4589 * Finally, if there is no buffer on the left, or the buffer is too
4590 * small, then we use copy-on-write and make both SVs share the
4595 /* Whichever path we take through the next code, we want this true,
4596 and doing it now facilitates the COW check. */
4597 (void)SvPOK_only(dstr);
4601 /* slated for free anyway (and not COW)? */
4602 (sflags & (SVs_TEMP|SVf_IsCOW)) == SVs_TEMP
4603 /* or a swipable TARG */
4605 (SVs_PADTMP|SVf_READONLY|SVf_PROTECT|SVf_IsCOW))
4607 /* whose buffer is worth stealing */
4608 && CHECK_COWBUF_THRESHOLD(cur,len)
4611 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4612 (!(flags & SV_NOSTEAL)) &&
4613 /* and we're allowed to steal temps */
4614 SvREFCNT(sstr) == 1 && /* and no other references to it? */
4615 len) /* and really is a string */
4616 { /* Passes the swipe test. */
4617 if (SvPVX_const(dstr)) /* we know that dtype >= SVt_PV */