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 #ifdef PERL_OLD_COPY_ON_WRITE
129 #define SV_COW_NEXT_SV(sv) INT2PTR(SV *,SvUVX(sv))
130 #define SV_COW_NEXT_SV_SET(current,next) SvUV_set(current, PTR2UV(next))
133 /* ============================================================================
135 =head1 Allocation and deallocation of SVs.
136 An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
137 sv, av, hv...) contains type and reference count information, and for
138 many types, a pointer to the body (struct xrv, xpv, xpviv...), which
139 contains fields specific to each type. Some types store all they need
140 in the head, so don't have a body.
142 In all but the most memory-paranoid configurations (ex: PURIFY), heads
143 and bodies are allocated out of arenas, which by default are
144 approximately 4K chunks of memory parcelled up into N heads or bodies.
145 Sv-bodies are allocated by their sv-type, guaranteeing size
146 consistency needed to allocate safely from arrays.
148 For SV-heads, the first slot in each arena is reserved, and holds a
149 link to the next arena, some flags, and a note of the number of slots.
150 Snaked through each arena chain is a linked list of free items; when
151 this becomes empty, an extra arena is allocated and divided up into N
152 items which are threaded into the free list.
154 SV-bodies are similar, but they use arena-sets by default, which
155 separate the link and info from the arena itself, and reclaim the 1st
156 slot in the arena. SV-bodies are further described later.
158 The following global variables are associated with arenas:
160 PL_sv_arenaroot pointer to list of SV arenas
161 PL_sv_root pointer to list of free SV structures
163 PL_body_arenas head of linked-list of body arenas
164 PL_body_roots[] array of pointers to list of free bodies of svtype
165 arrays are indexed by the svtype needed
167 A few special SV heads are not allocated from an arena, but are
168 instead directly created in the interpreter structure, eg PL_sv_undef.
169 The size of arenas can be changed from the default by setting
170 PERL_ARENA_SIZE appropriately at compile time.
172 The SV arena serves the secondary purpose of allowing still-live SVs
173 to be located and destroyed during final cleanup.
175 At the lowest level, the macros new_SV() and del_SV() grab and free
176 an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
177 to return the SV to the free list with error checking.) new_SV() calls
178 more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
179 SVs in the free list have their SvTYPE field set to all ones.
181 At the time of very final cleanup, sv_free_arenas() is called from
182 perl_destruct() to physically free all the arenas allocated since the
183 start of the interpreter.
185 The function visit() scans the SV arenas list, and calls a specified
186 function for each SV it finds which is still live - ie which has an SvTYPE
187 other than all 1's, and a non-zero SvREFCNT. visit() is used by the
188 following functions (specified as [function that calls visit()] / [function
189 called by visit() for each SV]):
191 sv_report_used() / do_report_used()
192 dump all remaining SVs (debugging aid)
194 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
195 do_clean_named_io_objs(),do_curse()
196 Attempt to free all objects pointed to by RVs,
197 try to do the same for all objects indir-
198 ectly referenced by typeglobs too, and
199 then do a final sweep, cursing any
200 objects that remain. Called once from
201 perl_destruct(), prior to calling sv_clean_all()
204 sv_clean_all() / do_clean_all()
205 SvREFCNT_dec(sv) each remaining SV, possibly
206 triggering an sv_free(). It also sets the
207 SVf_BREAK flag on the SV to indicate that the
208 refcnt has been artificially lowered, and thus
209 stopping sv_free() from giving spurious warnings
210 about SVs which unexpectedly have a refcnt
211 of zero. called repeatedly from perl_destruct()
212 until there are no SVs left.
214 =head2 Arena allocator API Summary
216 Private API to rest of sv.c
220 new_XPVNV(), del_XPVGV(),
225 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
229 * ========================================================================= */
232 * "A time to plant, and a time to uproot what was planted..."
236 # define MEM_LOG_NEW_SV(sv, file, line, func) \
237 Perl_mem_log_new_sv(sv, file, line, func)
238 # define MEM_LOG_DEL_SV(sv, file, line, func) \
239 Perl_mem_log_del_sv(sv, file, line, func)
241 # define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
242 # define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
245 #ifdef DEBUG_LEAKING_SCALARS
246 # define FREE_SV_DEBUG_FILE(sv) STMT_START { \
247 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
249 # define DEBUG_SV_SERIAL(sv) \
250 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
251 PTR2UV(sv), (long)(sv)->sv_debug_serial))
253 # define FREE_SV_DEBUG_FILE(sv)
254 # define DEBUG_SV_SERIAL(sv) NOOP
258 # define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
259 # define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
260 /* Whilst I'd love to do this, it seems that things like to check on
262 # define POISON_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
264 # define POISON_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
265 PoisonNew(&SvREFCNT(sv), 1, U32)
267 # define SvARENA_CHAIN(sv) SvANY(sv)
268 # define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
269 # define POISON_SV_HEAD(sv)
272 /* Mark an SV head as unused, and add to free list.
274 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
275 * its refcount artificially decremented during global destruction, so
276 * there may be dangling pointers to it. The last thing we want in that
277 * case is for it to be reused. */
279 #define plant_SV(p) \
281 const U32 old_flags = SvFLAGS(p); \
282 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
283 DEBUG_SV_SERIAL(p); \
284 FREE_SV_DEBUG_FILE(p); \
286 SvFLAGS(p) = SVTYPEMASK; \
287 if (!(old_flags & SVf_BREAK)) { \
288 SvARENA_CHAIN_SET(p, PL_sv_root); \
294 #define uproot_SV(p) \
297 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
302 /* make some more SVs by adding another arena */
308 char *chunk; /* must use New here to match call to */
309 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
310 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
315 /* new_SV(): return a new, empty SV head */
317 #ifdef DEBUG_LEAKING_SCALARS
318 /* provide a real function for a debugger to play with */
320 S_new_SV(pTHX_ const char *file, int line, const char *func)
327 sv = S_more_sv(aTHX);
331 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
332 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
338 sv->sv_debug_inpad = 0;
339 sv->sv_debug_parent = NULL;
340 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
342 sv->sv_debug_serial = PL_sv_serial++;
344 MEM_LOG_NEW_SV(sv, file, line, func);
345 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
346 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
350 # define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
358 (p) = S_more_sv(aTHX); \
362 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
367 /* del_SV(): return an empty SV head to the free list */
380 S_del_sv(pTHX_ SV *p)
382 PERL_ARGS_ASSERT_DEL_SV;
387 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
388 const SV * const sv = sva + 1;
389 const SV * const svend = &sva[SvREFCNT(sva)];
390 if (p >= sv && p < svend) {
396 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
397 "Attempt to free non-arena SV: 0x%"UVxf
398 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
405 #else /* ! DEBUGGING */
407 #define del_SV(p) plant_SV(p)
409 #endif /* DEBUGGING */
412 * Bodyless IVs and NVs!
414 * Since 5.9.2, we can avoid allocating a body for SVt_IV-type SVs.
415 * Since the larger IV-holding variants of SVs store their integer
416 * values in their respective bodies, the family of SvIV() accessor
417 * macros would naively have to branch on the SV type to find the
418 * integer value either in the HEAD or BODY. In order to avoid this
419 * expensive branch, a clever soul has deployed a great hack:
420 * We set up the SvANY pointer such that instead of pointing to a
421 * real body, it points into the memory before the location of the
422 * head. We compute this pointer such that the location of
423 * the integer member of the hypothetical body struct happens to
424 * be the same as the location of the integer member of the bodyless
425 * SV head. This now means that the SvIV() family of accessors can
426 * always read from the (hypothetical or real) body via SvANY.
428 * Since the 5.21 dev series, we employ the same trick for NVs
429 * if the architecture can support it (NVSIZE <= IVSIZE).
432 /* The following two macros compute the necessary offsets for the above
433 * trick and store them in SvANY for SvIV() (and friends) to use. */
434 #define SET_SVANY_FOR_BODYLESS_IV(sv) \
435 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv))
437 #define SET_SVANY_FOR_BODYLESS_NV(sv) \
438 SvANY(sv) = (XPVNV*)((char*)&(sv->sv_u.svu_nv) - STRUCT_OFFSET(XPVNV, xnv_u.xnv_nv))
441 =head1 SV Manipulation Functions
443 =for apidoc sv_add_arena
445 Given a chunk of memory, link it to the head of the list of arenas,
446 and split it into a list of free SVs.
452 S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
454 SV *const sva = MUTABLE_SV(ptr);
458 PERL_ARGS_ASSERT_SV_ADD_ARENA;
460 /* The first SV in an arena isn't an SV. */
461 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
462 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
463 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
465 PL_sv_arenaroot = sva;
466 PL_sv_root = sva + 1;
468 svend = &sva[SvREFCNT(sva) - 1];
471 SvARENA_CHAIN_SET(sv, (sv + 1));
475 /* Must always set typemask because it's always checked in on cleanup
476 when the arenas are walked looking for objects. */
477 SvFLAGS(sv) = SVTYPEMASK;
480 SvARENA_CHAIN_SET(sv, 0);
484 SvFLAGS(sv) = SVTYPEMASK;
487 /* visit(): call the named function for each non-free SV in the arenas
488 * whose flags field matches the flags/mask args. */
491 S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
496 PERL_ARGS_ASSERT_VISIT;
498 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
499 const SV * const svend = &sva[SvREFCNT(sva)];
501 for (sv = sva + 1; sv < svend; ++sv) {
502 if (SvTYPE(sv) != (svtype)SVTYPEMASK
503 && (sv->sv_flags & mask) == flags
516 /* called by sv_report_used() for each live SV */
519 do_report_used(pTHX_ SV *const sv)
521 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
522 PerlIO_printf(Perl_debug_log, "****\n");
529 =for apidoc sv_report_used
531 Dump the contents of all SVs not yet freed (debugging aid).
537 Perl_sv_report_used(pTHX)
540 visit(do_report_used, 0, 0);
546 /* called by sv_clean_objs() for each live SV */
549 do_clean_objs(pTHX_ SV *const ref)
553 SV * const target = SvRV(ref);
554 if (SvOBJECT(target)) {
555 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
556 if (SvWEAKREF(ref)) {
557 sv_del_backref(target, ref);
563 SvREFCNT_dec_NN(target);
570 /* clear any slots in a GV which hold objects - except IO;
571 * called by sv_clean_objs() for each live GV */
574 do_clean_named_objs(pTHX_ SV *const sv)
577 assert(SvTYPE(sv) == SVt_PVGV);
578 assert(isGV_with_GP(sv));
582 /* freeing GP entries may indirectly free the current GV;
583 * hold onto it while we mess with the GP slots */
586 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
587 DEBUG_D((PerlIO_printf(Perl_debug_log,
588 "Cleaning named glob SV object:\n "), sv_dump(obj)));
590 SvREFCNT_dec_NN(obj);
592 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
593 DEBUG_D((PerlIO_printf(Perl_debug_log,
594 "Cleaning named glob AV object:\n "), sv_dump(obj)));
596 SvREFCNT_dec_NN(obj);
598 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
599 DEBUG_D((PerlIO_printf(Perl_debug_log,
600 "Cleaning named glob HV object:\n "), sv_dump(obj)));
602 SvREFCNT_dec_NN(obj);
604 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
605 DEBUG_D((PerlIO_printf(Perl_debug_log,
606 "Cleaning named glob CV object:\n "), sv_dump(obj)));
608 SvREFCNT_dec_NN(obj);
610 SvREFCNT_dec_NN(sv); /* undo the inc above */
613 /* clear any IO slots in a GV which hold objects (except stderr, defout);
614 * called by sv_clean_objs() for each live GV */
617 do_clean_named_io_objs(pTHX_ SV *const sv)
620 assert(SvTYPE(sv) == SVt_PVGV);
621 assert(isGV_with_GP(sv));
622 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
626 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
627 DEBUG_D((PerlIO_printf(Perl_debug_log,
628 "Cleaning named glob IO object:\n "), sv_dump(obj)));
630 SvREFCNT_dec_NN(obj);
632 SvREFCNT_dec_NN(sv); /* undo the inc above */
635 /* Void wrapper to pass to visit() */
637 do_curse(pTHX_ SV * const sv) {
638 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
639 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
645 =for apidoc sv_clean_objs
647 Attempt to destroy all objects not yet freed.
653 Perl_sv_clean_objs(pTHX)
656 PL_in_clean_objs = TRUE;
657 visit(do_clean_objs, SVf_ROK, SVf_ROK);
658 /* Some barnacles may yet remain, clinging to typeglobs.
659 * Run the non-IO destructors first: they may want to output
660 * error messages, close files etc */
661 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
662 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
663 /* And if there are some very tenacious barnacles clinging to arrays,
664 closures, or what have you.... */
665 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
666 olddef = PL_defoutgv;
667 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
668 if (olddef && isGV_with_GP(olddef))
669 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
670 olderr = PL_stderrgv;
671 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
672 if (olderr && isGV_with_GP(olderr))
673 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
674 SvREFCNT_dec(olddef);
675 PL_in_clean_objs = FALSE;
678 /* called by sv_clean_all() for each live SV */
681 do_clean_all(pTHX_ SV *const sv)
683 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
684 /* don't clean pid table and strtab */
687 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
688 SvFLAGS(sv) |= SVf_BREAK;
693 =for apidoc sv_clean_all
695 Decrement the refcnt of each remaining SV, possibly triggering a
696 cleanup. This function may have to be called multiple times to free
697 SVs which are in complex self-referential hierarchies.
703 Perl_sv_clean_all(pTHX)
706 PL_in_clean_all = TRUE;
707 cleaned = visit(do_clean_all, 0,0);
712 ARENASETS: a meta-arena implementation which separates arena-info
713 into struct arena_set, which contains an array of struct
714 arena_descs, each holding info for a single arena. By separating
715 the meta-info from the arena, we recover the 1st slot, formerly
716 borrowed for list management. The arena_set is about the size of an
717 arena, avoiding the needless malloc overhead of a naive linked-list.
719 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
720 memory in the last arena-set (1/2 on average). In trade, we get
721 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
722 smaller types). The recovery of the wasted space allows use of
723 small arenas for large, rare body types, by changing array* fields
724 in body_details_by_type[] below.
727 char *arena; /* the raw storage, allocated aligned */
728 size_t size; /* its size ~4k typ */
729 svtype utype; /* bodytype stored in arena */
734 /* Get the maximum number of elements in set[] such that struct arena_set
735 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
736 therefore likely to be 1 aligned memory page. */
738 #define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
739 - 2 * sizeof(int)) / sizeof (struct arena_desc))
742 struct arena_set* next;
743 unsigned int set_size; /* ie ARENAS_PER_SET */
744 unsigned int curr; /* index of next available arena-desc */
745 struct arena_desc set[ARENAS_PER_SET];
749 =for apidoc sv_free_arenas
751 Deallocate the memory used by all arenas. Note that all the individual SV
752 heads and bodies within the arenas must already have been freed.
758 Perl_sv_free_arenas(pTHX)
764 /* Free arenas here, but be careful about fake ones. (We assume
765 contiguity of the fake ones with the corresponding real ones.) */
767 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
768 svanext = MUTABLE_SV(SvANY(sva));
769 while (svanext && SvFAKE(svanext))
770 svanext = MUTABLE_SV(SvANY(svanext));
777 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
780 struct arena_set *current = aroot;
783 assert(aroot->set[i].arena);
784 Safefree(aroot->set[i].arena);
792 i = PERL_ARENA_ROOTS_SIZE;
794 PL_body_roots[i] = 0;
801 Here are mid-level routines that manage the allocation of bodies out
802 of the various arenas. There are 5 kinds of arenas:
804 1. SV-head arenas, which are discussed and handled above
805 2. regular body arenas
806 3. arenas for reduced-size bodies
809 Arena types 2 & 3 are chained by body-type off an array of
810 arena-root pointers, which is indexed by svtype. Some of the
811 larger/less used body types are malloced singly, since a large
812 unused block of them is wasteful. Also, several svtypes dont have
813 bodies; the data fits into the sv-head itself. The arena-root
814 pointer thus has a few unused root-pointers (which may be hijacked
815 later for arena types 4,5)
817 3 differs from 2 as an optimization; some body types have several
818 unused fields in the front of the structure (which are kept in-place
819 for consistency). These bodies can be allocated in smaller chunks,
820 because the leading fields arent accessed. Pointers to such bodies
821 are decremented to point at the unused 'ghost' memory, knowing that
822 the pointers are used with offsets to the real memory.
825 =head1 SV-Body Allocation
829 Allocation of SV-bodies is similar to SV-heads, differing as follows;
830 the allocation mechanism is used for many body types, so is somewhat
831 more complicated, it uses arena-sets, and has no need for still-live
834 At the outermost level, (new|del)_X*V macros return bodies of the
835 appropriate type. These macros call either (new|del)_body_type or
836 (new|del)_body_allocated macro pairs, depending on specifics of the
837 type. Most body types use the former pair, the latter pair is used to
838 allocate body types with "ghost fields".
840 "ghost fields" are fields that are unused in certain types, and
841 consequently don't need to actually exist. They are declared because
842 they're part of a "base type", which allows use of functions as
843 methods. The simplest examples are AVs and HVs, 2 aggregate types
844 which don't use the fields which support SCALAR semantics.
846 For these types, the arenas are carved up into appropriately sized
847 chunks, we thus avoid wasted memory for those unaccessed members.
848 When bodies are allocated, we adjust the pointer back in memory by the
849 size of the part not allocated, so it's as if we allocated the full
850 structure. (But things will all go boom if you write to the part that
851 is "not there", because you'll be overwriting the last members of the
852 preceding structure in memory.)
854 We calculate the correction using the STRUCT_OFFSET macro on the first
855 member present. If the allocated structure is smaller (no initial NV
856 actually allocated) then the net effect is to subtract the size of the NV
857 from the pointer, to return a new pointer as if an initial NV were actually
858 allocated. (We were using structures named *_allocated for this, but
859 this turned out to be a subtle bug, because a structure without an NV
860 could have a lower alignment constraint, but the compiler is allowed to
861 optimised accesses based on the alignment constraint of the actual pointer
862 to the full structure, for example, using a single 64 bit load instruction
863 because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
865 This is the same trick as was used for NV and IV bodies. Ironically it
866 doesn't need to be used for NV bodies any more, because NV is now at
867 the start of the structure. IV bodies, and also in some builds NV bodies,
868 don't need it either, because they are no longer allocated.
870 In turn, the new_body_* allocators call S_new_body(), which invokes
871 new_body_inline macro, which takes a lock, and takes a body off the
872 linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
873 necessary to refresh an empty list. Then the lock is released, and
874 the body is returned.
876 Perl_more_bodies allocates a new arena, and carves it up into an array of N
877 bodies, which it strings into a linked list. It looks up arena-size
878 and body-size from the body_details table described below, thus
879 supporting the multiple body-types.
881 If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
882 the (new|del)_X*V macros are mapped directly to malloc/free.
884 For each sv-type, struct body_details bodies_by_type[] carries
885 parameters which control these aspects of SV handling:
887 Arena_size determines whether arenas are used for this body type, and if
888 so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
889 zero, forcing individual mallocs and frees.
891 Body_size determines how big a body is, and therefore how many fit into
892 each arena. Offset carries the body-pointer adjustment needed for
893 "ghost fields", and is used in *_allocated macros.
895 But its main purpose is to parameterize info needed in
896 Perl_sv_upgrade(). The info here dramatically simplifies the function
897 vs the implementation in 5.8.8, making it table-driven. All fields
898 are used for this, except for arena_size.
900 For the sv-types that have no bodies, arenas are not used, so those
901 PL_body_roots[sv_type] are unused, and can be overloaded. In
902 something of a special case, SVt_NULL is borrowed for HE arenas;
903 PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
904 bodies_by_type[SVt_NULL] slot is not used, as the table is not
909 struct body_details {
910 U8 body_size; /* Size to allocate */
911 U8 copy; /* Size of structure to copy (may be shorter) */
912 U8 offset; /* Size of unalloced ghost fields to first alloced field*/
913 PERL_BITFIELD8 type : 4; /* We have space for a sanity check. */
914 PERL_BITFIELD8 cant_upgrade : 1;/* Cannot upgrade this type */
915 PERL_BITFIELD8 zero_nv : 1; /* zero the NV when upgrading from this */
916 PERL_BITFIELD8 arena : 1; /* Allocated from an arena */
917 U32 arena_size; /* Size of arena to allocate */
925 /* With -DPURFIY we allocate everything directly, and don't use arenas.
926 This seems a rather elegant way to simplify some of the code below. */
927 #define HASARENA FALSE
929 #define HASARENA TRUE
931 #define NOARENA FALSE
933 /* Size the arenas to exactly fit a given number of bodies. A count
934 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
935 simplifying the default. If count > 0, the arena is sized to fit
936 only that many bodies, allowing arenas to be used for large, rare
937 bodies (XPVFM, XPVIO) without undue waste. The arena size is
938 limited by PERL_ARENA_SIZE, so we can safely oversize the
941 #define FIT_ARENA0(body_size) \
942 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
943 #define FIT_ARENAn(count,body_size) \
944 ( count * body_size <= PERL_ARENA_SIZE) \
945 ? count * body_size \
946 : FIT_ARENA0 (body_size)
947 #define FIT_ARENA(count,body_size) \
949 ? FIT_ARENAn (count, body_size) \
950 : FIT_ARENA0 (body_size))
952 /* Calculate the length to copy. Specifically work out the length less any
953 final padding the compiler needed to add. See the comment in sv_upgrade
954 for why copying the padding proved to be a bug. */
956 #define copy_length(type, last_member) \
957 STRUCT_OFFSET(type, last_member) \
958 + sizeof (((type*)SvANY((const SV *)0))->last_member)
960 static const struct body_details bodies_by_type[] = {
961 /* HEs use this offset for their arena. */
962 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
964 /* IVs are in the head, so the allocation size is 0. */
966 sizeof(IV), /* This is used to copy out the IV body. */
967 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
968 NOARENA /* IVS don't need an arena */, 0
973 STRUCT_OFFSET(XPVNV, xnv_u),
974 SVt_NV, FALSE, HADNV, NOARENA, 0 },
976 { sizeof(NV), sizeof(NV),
977 STRUCT_OFFSET(XPVNV, xnv_u),
978 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
981 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
982 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
983 + STRUCT_OFFSET(XPV, xpv_cur),
984 SVt_PV, FALSE, NONV, HASARENA,
985 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
987 { sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur),
988 copy_length(XINVLIST, is_offset) - STRUCT_OFFSET(XPV, xpv_cur),
989 + STRUCT_OFFSET(XPV, xpv_cur),
990 SVt_INVLIST, TRUE, NONV, HASARENA,
991 FIT_ARENA(0, sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur)) },
993 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
994 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
995 + STRUCT_OFFSET(XPV, xpv_cur),
996 SVt_PVIV, FALSE, NONV, HASARENA,
997 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
999 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
1000 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
1001 + STRUCT_OFFSET(XPV, xpv_cur),
1002 SVt_PVNV, FALSE, HADNV, HASARENA,
1003 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
1005 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
1006 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
1011 SVt_REGEXP, TRUE, NONV, HASARENA,
1012 FIT_ARENA(0, sizeof(regexp))
1015 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
1016 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
1018 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
1019 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
1022 copy_length(XPVAV, xav_alloc),
1024 SVt_PVAV, TRUE, NONV, HASARENA,
1025 FIT_ARENA(0, sizeof(XPVAV)) },
1028 copy_length(XPVHV, xhv_max),
1030 SVt_PVHV, TRUE, NONV, HASARENA,
1031 FIT_ARENA(0, sizeof(XPVHV)) },
1036 SVt_PVCV, TRUE, NONV, HASARENA,
1037 FIT_ARENA(0, sizeof(XPVCV)) },
1042 SVt_PVFM, TRUE, NONV, NOARENA,
1043 FIT_ARENA(20, sizeof(XPVFM)) },
1048 SVt_PVIO, TRUE, NONV, HASARENA,
1049 FIT_ARENA(24, sizeof(XPVIO)) },
1052 #define new_body_allocated(sv_type) \
1053 (void *)((char *)S_new_body(aTHX_ sv_type) \
1054 - bodies_by_type[sv_type].offset)
1056 /* return a thing to the free list */
1058 #define del_body(thing, root) \
1060 void ** const thing_copy = (void **)thing; \
1061 *thing_copy = *root; \
1062 *root = (void*)thing_copy; \
1066 #if !(NVSIZE <= IVSIZE)
1067 # define new_XNV() safemalloc(sizeof(XPVNV))
1069 #define new_XPVNV() safemalloc(sizeof(XPVNV))
1070 #define new_XPVMG() safemalloc(sizeof(XPVMG))
1072 #define del_XPVGV(p) safefree(p)
1076 #if !(NVSIZE <= IVSIZE)
1077 # define new_XNV() new_body_allocated(SVt_NV)
1079 #define new_XPVNV() new_body_allocated(SVt_PVNV)
1080 #define new_XPVMG() new_body_allocated(SVt_PVMG)
1082 #define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1083 &PL_body_roots[SVt_PVGV])
1087 /* no arena for you! */
1089 #define new_NOARENA(details) \
1090 safemalloc((details)->body_size + (details)->offset)
1091 #define new_NOARENAZ(details) \
1092 safecalloc((details)->body_size + (details)->offset, 1)
1095 Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1096 const size_t arena_size)
1098 void ** const root = &PL_body_roots[sv_type];
1099 struct arena_desc *adesc;
1100 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1104 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
1105 #if defined(DEBUGGING) && defined(PERL_GLOBAL_STRUCT)
1108 #if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
1109 static bool done_sanity_check;
1111 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1112 * variables like done_sanity_check. */
1113 if (!done_sanity_check) {
1114 unsigned int i = SVt_LAST;
1116 done_sanity_check = TRUE;
1119 assert (bodies_by_type[i].type == i);
1125 /* may need new arena-set to hold new arena */
1126 if (!aroot || aroot->curr >= aroot->set_size) {
1127 struct arena_set *newroot;
1128 Newxz(newroot, 1, struct arena_set);
1129 newroot->set_size = ARENAS_PER_SET;
1130 newroot->next = aroot;
1132 PL_body_arenas = (void *) newroot;
1133 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1136 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1137 curr = aroot->curr++;
1138 adesc = &(aroot->set[curr]);
1139 assert(!adesc->arena);
1141 Newx(adesc->arena, good_arena_size, char);
1142 adesc->size = good_arena_size;
1143 adesc->utype = sv_type;
1144 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %"UVuf"\n",
1145 curr, (void*)adesc->arena, (UV)good_arena_size));
1147 start = (char *) adesc->arena;
1149 /* Get the address of the byte after the end of the last body we can fit.
1150 Remember, this is integer division: */
1151 end = start + good_arena_size / body_size * body_size;
1153 /* computed count doesn't reflect the 1st slot reservation */
1154 #if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1155 DEBUG_m(PerlIO_printf(Perl_debug_log,
1156 "arena %p end %p arena-size %d (from %d) type %d "
1158 (void*)start, (void*)end, (int)good_arena_size,
1159 (int)arena_size, sv_type, (int)body_size,
1160 (int)good_arena_size / (int)body_size));
1162 DEBUG_m(PerlIO_printf(Perl_debug_log,
1163 "arena %p end %p arena-size %d type %d size %d ct %d\n",
1164 (void*)start, (void*)end,
1165 (int)arena_size, sv_type, (int)body_size,
1166 (int)good_arena_size / (int)body_size));
1168 *root = (void *)start;
1171 /* Where the next body would start: */
1172 char * const next = start + body_size;
1175 /* This is the last body: */
1176 assert(next == end);
1178 *(void **)start = 0;
1182 *(void**) start = (void *)next;
1187 /* grab a new thing from the free list, allocating more if necessary.
1188 The inline version is used for speed in hot routines, and the
1189 function using it serves the rest (unless PURIFY).
1191 #define new_body_inline(xpv, sv_type) \
1193 void ** const r3wt = &PL_body_roots[sv_type]; \
1194 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1195 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
1196 bodies_by_type[sv_type].body_size,\
1197 bodies_by_type[sv_type].arena_size)); \
1198 *(r3wt) = *(void**)(xpv); \
1204 S_new_body(pTHX_ const svtype sv_type)
1207 new_body_inline(xpv, sv_type);
1213 static const struct body_details fake_rv =
1214 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1217 =for apidoc sv_upgrade
1219 Upgrade an SV to a more complex form. Generally adds a new body type to the
1220 SV, then copies across as much information as possible from the old body.
1221 It croaks if the SV is already in a more complex form than requested. You
1222 generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1223 before calling C<sv_upgrade>, and hence does not croak. See also
1230 Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
1234 const svtype old_type = SvTYPE(sv);
1235 const struct body_details *new_type_details;
1236 const struct body_details *old_type_details
1237 = bodies_by_type + old_type;
1238 SV *referant = NULL;
1240 PERL_ARGS_ASSERT_SV_UPGRADE;
1242 if (old_type == new_type)
1245 /* This clause was purposefully added ahead of the early return above to
1246 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1247 inference by Nick I-S that it would fix other troublesome cases. See
1248 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1250 Given that shared hash key scalars are no longer PVIV, but PV, there is
1251 no longer need to unshare so as to free up the IVX slot for its proper
1252 purpose. So it's safe to move the early return earlier. */
1254 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
1255 sv_force_normal_flags(sv, 0);
1258 old_body = SvANY(sv);
1260 /* Copying structures onto other structures that have been neatly zeroed
1261 has a subtle gotcha. Consider XPVMG
1263 +------+------+------+------+------+-------+-------+
1264 | NV | CUR | LEN | IV | MAGIC | STASH |
1265 +------+------+------+------+------+-------+-------+
1266 0 4 8 12 16 20 24 28
1268 where NVs are aligned to 8 bytes, so that sizeof that structure is
1269 actually 32 bytes long, with 4 bytes of padding at the end:
1271 +------+------+------+------+------+-------+-------+------+
1272 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1273 +------+------+------+------+------+-------+-------+------+
1274 0 4 8 12 16 20 24 28 32
1276 so what happens if you allocate memory for this structure:
1278 +------+------+------+------+------+-------+-------+------+------+...
1279 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1280 +------+------+------+------+------+-------+-------+------+------+...
1281 0 4 8 12 16 20 24 28 32 36
1283 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1284 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1285 started out as zero once, but it's quite possible that it isn't. So now,
1286 rather than a nicely zeroed GP, you have it pointing somewhere random.
1289 (In fact, GP ends up pointing at a previous GP structure, because the
1290 principle cause of the padding in XPVMG getting garbage is a copy of
1291 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1292 this happens to be moot because XPVGV has been re-ordered, with GP
1293 no longer after STASH)
1295 So we are careful and work out the size of used parts of all the
1303 referant = SvRV(sv);
1304 old_type_details = &fake_rv;
1305 if (new_type == SVt_NV)
1306 new_type = SVt_PVNV;
1308 if (new_type < SVt_PVIV) {
1309 new_type = (new_type == SVt_NV)
1310 ? SVt_PVNV : SVt_PVIV;
1315 if (new_type < SVt_PVNV) {
1316 new_type = SVt_PVNV;
1320 assert(new_type > SVt_PV);
1321 STATIC_ASSERT_STMT(SVt_IV < SVt_PV);
1322 STATIC_ASSERT_STMT(SVt_NV < SVt_PV);
1329 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1330 there's no way that it can be safely upgraded, because perl.c
1331 expects to Safefree(SvANY(PL_mess_sv)) */
1332 assert(sv != PL_mess_sv);
1335 if (UNLIKELY(old_type_details->cant_upgrade))
1336 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1337 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1340 if (UNLIKELY(old_type > new_type))
1341 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1342 (int)old_type, (int)new_type);
1344 new_type_details = bodies_by_type + new_type;
1346 SvFLAGS(sv) &= ~SVTYPEMASK;
1347 SvFLAGS(sv) |= new_type;
1349 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1350 the return statements above will have triggered. */
1351 assert (new_type != SVt_NULL);
1354 assert(old_type == SVt_NULL);
1355 SET_SVANY_FOR_BODYLESS_IV(sv);
1359 assert(old_type == SVt_NULL);
1360 #if NVSIZE <= IVSIZE
1361 SET_SVANY_FOR_BODYLESS_NV(sv);
1363 SvANY(sv) = new_XNV();
1369 assert(new_type_details->body_size);
1372 assert(new_type_details->arena);
1373 assert(new_type_details->arena_size);
1374 /* This points to the start of the allocated area. */
1375 new_body_inline(new_body, new_type);
1376 Zero(new_body, new_type_details->body_size, char);
1377 new_body = ((char *)new_body) - new_type_details->offset;
1379 /* We always allocated the full length item with PURIFY. To do this
1380 we fake things so that arena is false for all 16 types.. */
1381 new_body = new_NOARENAZ(new_type_details);
1383 SvANY(sv) = new_body;
1384 if (new_type == SVt_PVAV) {
1388 if (old_type_details->body_size) {
1391 /* It will have been zeroed when the new body was allocated.
1392 Lets not write to it, in case it confuses a write-back
1398 #ifndef NODEFAULT_SHAREKEYS
1399 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1401 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1402 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
1405 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1406 The target created by newSVrv also is, and it can have magic.
1407 However, it never has SvPVX set.
1409 if (old_type == SVt_IV) {
1411 } else if (old_type >= SVt_PV) {
1412 assert(SvPVX_const(sv) == 0);
1415 if (old_type >= SVt_PVMG) {
1416 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1417 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1419 sv->sv_u.svu_array = NULL; /* or svu_hash */
1424 /* XXX Is this still needed? Was it ever needed? Surely as there is
1425 no route from NV to PVIV, NOK can never be true */
1426 assert(!SvNOKp(sv));
1440 assert(new_type_details->body_size);
1441 /* We always allocated the full length item with PURIFY. To do this
1442 we fake things so that arena is false for all 16 types.. */
1443 if(new_type_details->arena) {
1444 /* This points to the start of the allocated area. */
1445 new_body_inline(new_body, new_type);
1446 Zero(new_body, new_type_details->body_size, char);
1447 new_body = ((char *)new_body) - new_type_details->offset;
1449 new_body = new_NOARENAZ(new_type_details);
1451 SvANY(sv) = new_body;
1453 if (old_type_details->copy) {
1454 /* There is now the potential for an upgrade from something without
1455 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1456 int offset = old_type_details->offset;
1457 int length = old_type_details->copy;
1459 if (new_type_details->offset > old_type_details->offset) {
1460 const int difference
1461 = new_type_details->offset - old_type_details->offset;
1462 offset += difference;
1463 length -= difference;
1465 assert (length >= 0);
1467 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1471 #ifndef NV_ZERO_IS_ALLBITS_ZERO
1472 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
1473 * correct 0.0 for us. Otherwise, if the old body didn't have an
1474 * NV slot, but the new one does, then we need to initialise the
1475 * freshly created NV slot with whatever the correct bit pattern is
1477 if (old_type_details->zero_nv && !new_type_details->zero_nv
1478 && !isGV_with_GP(sv))
1482 if (UNLIKELY(new_type == SVt_PVIO)) {
1483 IO * const io = MUTABLE_IO(sv);
1484 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
1487 /* Clear the stashcache because a new IO could overrule a package
1489 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
1490 hv_clear(PL_stashcache);
1492 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1493 IoPAGE_LEN(sv) = 60;
1495 if (UNLIKELY(new_type == SVt_REGEXP))
1496 sv->sv_u.svu_rx = (regexp *)new_body;
1497 else if (old_type < SVt_PV) {
1498 /* referant will be NULL unless the old type was SVt_IV emulating
1500 sv->sv_u.svu_rv = referant;
1504 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1505 (unsigned long)new_type);
1508 /* if this is zero, this is a body-less SVt_NULL, SVt_IV/SVt_RV,
1509 and sometimes SVt_NV */
1510 if (old_type_details->body_size) {
1514 /* Note that there is an assumption that all bodies of types that
1515 can be upgraded came from arenas. Only the more complex non-
1516 upgradable types are allowed to be directly malloc()ed. */
1517 assert(old_type_details->arena);
1518 del_body((void*)((char*)old_body + old_type_details->offset),
1519 &PL_body_roots[old_type]);
1525 =for apidoc sv_backoff
1527 Remove any string offset. You should normally use the C<SvOOK_off> macro
1534 Perl_sv_backoff(SV *const sv)
1537 const char * const s = SvPVX_const(sv);
1539 PERL_ARGS_ASSERT_SV_BACKOFF;
1542 assert(SvTYPE(sv) != SVt_PVHV);
1543 assert(SvTYPE(sv) != SVt_PVAV);
1545 SvOOK_offset(sv, delta);
1547 SvLEN_set(sv, SvLEN(sv) + delta);
1548 SvPV_set(sv, SvPVX(sv) - delta);
1549 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1550 SvFLAGS(sv) &= ~SVf_OOK;
1557 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1558 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1559 Use the C<SvGROW> wrapper instead.
1564 static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1567 Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
1571 PERL_ARGS_ASSERT_SV_GROW;
1575 if (SvTYPE(sv) < SVt_PV) {
1576 sv_upgrade(sv, SVt_PV);
1577 s = SvPVX_mutable(sv);
1579 else if (SvOOK(sv)) { /* pv is offset? */
1581 s = SvPVX_mutable(sv);
1582 if (newlen > SvLEN(sv))
1583 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1587 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
1588 s = SvPVX_mutable(sv);
1591 #ifdef PERL_NEW_COPY_ON_WRITE
1592 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
1593 * to store the COW count. So in general, allocate one more byte than
1594 * asked for, to make it likely this byte is always spare: and thus
1595 * make more strings COW-able.
1596 * If the new size is a big power of two, don't bother: we assume the
1597 * caller wanted a nice 2^N sized block and will be annoyed at getting
1599 * Only increment if the allocation isn't MEM_SIZE_MAX,
1600 * otherwise it will wrap to 0.
1602 if (newlen & 0xff && newlen != MEM_SIZE_MAX)
1606 #if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1607 #define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1610 if (newlen > SvLEN(sv)) { /* need more room? */
1611 STRLEN minlen = SvCUR(sv);
1612 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1613 if (newlen < minlen)
1615 #ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1617 /* Don't round up on the first allocation, as odds are pretty good that
1618 * the initial request is accurate as to what is really needed */
1620 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1621 if (rounded > newlen)
1625 if (SvLEN(sv) && s) {
1626 s = (char*)saferealloc(s, newlen);
1629 s = (char*)safemalloc(newlen);
1630 if (SvPVX_const(sv) && SvCUR(sv)) {
1631 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1635 #ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1636 /* Do this here, do it once, do it right, and then we will never get
1637 called back into sv_grow() unless there really is some growing
1639 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1641 SvLEN_set(sv, newlen);
1648 =for apidoc sv_setiv
1650 Copies an integer into the given SV, upgrading first if necessary.
1651 Does not handle 'set' magic. See also C<sv_setiv_mg>.
1657 Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
1659 PERL_ARGS_ASSERT_SV_SETIV;
1661 SV_CHECK_THINKFIRST_COW_DROP(sv);
1662 switch (SvTYPE(sv)) {
1665 sv_upgrade(sv, SVt_IV);
1668 sv_upgrade(sv, SVt_PVIV);
1672 if (!isGV_with_GP(sv))
1679 /* diag_listed_as: Can't coerce %s to %s in %s */
1680 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<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<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),
1796 (void)SvNOK_only(sv); /* validate number */
1801 =for apidoc sv_setnv_mg
1803 Like C<sv_setnv>, but also handles 'set' magic.
1809 Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
1811 PERL_ARGS_ASSERT_SV_SETNV_MG;
1817 /* Return a cleaned-up, printable version of sv, for non-numeric, or
1818 * not incrementable warning display.
1819 * Originally part of S_not_a_number().
1820 * The return value may be != tmpbuf.
1824 S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1827 PERL_ARGS_ASSERT_SV_DISPLAY;
1830 SV *dsv = newSVpvs_flags("", SVs_TEMP);
1831 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
1834 const char * const limit = tmpbuf + tmpbuf_size - 8;
1835 /* each *s can expand to 4 chars + "...\0",
1836 i.e. need room for 8 chars */
1838 const char *s = SvPVX_const(sv);
1839 const char * const end = s + SvCUR(sv);
1840 for ( ; s < end && d < limit; s++ ) {
1842 if (! isASCII(ch) && !isPRINT_LC(ch)) {
1846 /* Map to ASCII "equivalent" of Latin1 */
1847 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
1853 else if (ch == '\r') {
1857 else if (ch == '\f') {
1861 else if (ch == '\\') {
1865 else if (ch == '\0') {
1869 else if (isPRINT_LC(ch))
1888 /* Print an "isn't numeric" warning, using a cleaned-up,
1889 * printable version of the offending string
1893 S_not_a_number(pTHX_ SV *const sv)
1898 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1900 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1903 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1904 /* diag_listed_as: Argument "%s" isn't numeric%s */
1905 "Argument \"%s\" isn't numeric in %s", pv,
1908 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1909 /* diag_listed_as: Argument "%s" isn't numeric%s */
1910 "Argument \"%s\" isn't numeric", pv);
1914 S_not_incrementable(pTHX_ SV *const sv) {
1918 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1920 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1922 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1923 "Argument \"%s\" treated as 0 in increment (++)", pv);
1927 =for apidoc looks_like_number
1929 Test if the content of an SV looks like a number (or is a number).
1930 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1931 non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1938 Perl_looks_like_number(pTHX_ SV *const sv)
1944 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1946 if (SvPOK(sv) || SvPOKp(sv)) {
1947 sbegin = SvPV_nomg_const(sv, len);
1950 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1951 numtype = grok_number(sbegin, len, NULL);
1952 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
1956 S_glob_2number(pTHX_ GV * const gv)
1958 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1960 /* We know that all GVs stringify to something that is not-a-number,
1961 so no need to test that. */
1962 if (ckWARN(WARN_NUMERIC))
1964 SV *const buffer = sv_newmortal();
1965 gv_efullname3(buffer, gv, "*");
1966 not_a_number(buffer);
1968 /* We just want something true to return, so that S_sv_2iuv_common
1969 can tail call us and return true. */
1973 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1974 until proven guilty, assume that things are not that bad... */
1979 As 64 bit platforms often have an NV that doesn't preserve all bits of
1980 an IV (an assumption perl has been based on to date) it becomes necessary
1981 to remove the assumption that the NV always carries enough precision to
1982 recreate the IV whenever needed, and that the NV is the canonical form.
1983 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1984 precision as a side effect of conversion (which would lead to insanity
1985 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1986 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1987 where precision was lost, and IV/UV/NV slots that have a valid conversion
1988 which has lost no precision
1989 2) to ensure that if a numeric conversion to one form is requested that
1990 would lose precision, the precise conversion (or differently
1991 imprecise conversion) is also performed and cached, to prevent
1992 requests for different numeric formats on the same SV causing
1993 lossy conversion chains. (lossless conversion chains are perfectly
1998 SvIOKp is true if the IV slot contains a valid value
1999 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
2000 SvNOKp is true if the NV slot contains a valid value
2001 SvNOK is true only if the NV value is accurate
2004 while converting from PV to NV, check to see if converting that NV to an
2005 IV(or UV) would lose accuracy over a direct conversion from PV to
2006 IV(or UV). If it would, cache both conversions, return NV, but mark
2007 SV as IOK NOKp (ie not NOK).
2009 While converting from PV to IV, check to see if converting that IV to an
2010 NV would lose accuracy over a direct conversion from PV to NV. If it
2011 would, cache both conversions, flag similarly.
2013 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
2014 correctly because if IV & NV were set NV *always* overruled.
2015 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
2016 changes - now IV and NV together means that the two are interchangeable:
2017 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
2019 The benefit of this is that operations such as pp_add know that if
2020 SvIOK is true for both left and right operands, then integer addition
2021 can be used instead of floating point (for cases where the result won't
2022 overflow). Before, floating point was always used, which could lead to
2023 loss of precision compared with integer addition.
2025 * making IV and NV equal status should make maths accurate on 64 bit
2027 * may speed up maths somewhat if pp_add and friends start to use
2028 integers when possible instead of fp. (Hopefully the overhead in
2029 looking for SvIOK and checking for overflow will not outweigh the
2030 fp to integer speedup)
2031 * will slow down integer operations (callers of SvIV) on "inaccurate"
2032 values, as the change from SvIOK to SvIOKp will cause a call into
2033 sv_2iv each time rather than a macro access direct to the IV slot
2034 * should speed up number->string conversion on integers as IV is
2035 favoured when IV and NV are equally accurate
2037 ####################################################################
2038 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2039 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2040 On the other hand, SvUOK is true iff UV.
2041 ####################################################################
2043 Your mileage will vary depending your CPU's relative fp to integer
2047 #ifndef NV_PRESERVES_UV
2048 # define IS_NUMBER_UNDERFLOW_IV 1
2049 # define IS_NUMBER_UNDERFLOW_UV 2
2050 # define IS_NUMBER_IV_AND_UV 2
2051 # define IS_NUMBER_OVERFLOW_IV 4
2052 # define IS_NUMBER_OVERFLOW_UV 5
2054 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
2056 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2058 S_sv_2iuv_non_preserve(pTHX_ SV *const sv
2064 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
2065 PERL_UNUSED_CONTEXT;
2067 DEBUG_c(PerlIO_printf(Perl_debug_log,"sv_2iuv_non '%s', IV=0x%"UVxf" NV=%"NVgf" inttype=%"UVXf"\n", SvPVX_const(sv), SvIVX(sv), SvNVX(sv), (UV)numtype));
2068 if (SvNVX(sv) < (NV)IV_MIN) {
2069 (void)SvIOKp_on(sv);
2071 SvIV_set(sv, IV_MIN);
2072 return IS_NUMBER_UNDERFLOW_IV;
2074 if (SvNVX(sv) > (NV)UV_MAX) {
2075 (void)SvIOKp_on(sv);
2078 SvUV_set(sv, UV_MAX);
2079 return IS_NUMBER_OVERFLOW_UV;
2081 (void)SvIOKp_on(sv);
2083 /* Can't use strtol etc to convert this string. (See truth table in
2085 if (SvNVX(sv) <= (UV)IV_MAX) {
2086 SvIV_set(sv, I_V(SvNVX(sv)));
2087 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2088 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2090 /* Integer is imprecise. NOK, IOKp */
2092 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2095 SvUV_set(sv, U_V(SvNVX(sv)));
2096 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2097 if (SvUVX(sv) == UV_MAX) {
2098 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2099 possibly be preserved by NV. Hence, it must be overflow.
2101 return IS_NUMBER_OVERFLOW_UV;
2103 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2105 /* Integer is imprecise. NOK, IOKp */
2107 return IS_NUMBER_OVERFLOW_IV;
2109 #endif /* !NV_PRESERVES_UV*/
2111 /* If numtype is infnan, set the NV of the sv accordingly.
2112 * If numtype is anything else, try setting the NV using Atof(PV). */
2114 # pragma warning(push)
2115 # pragma warning(disable:4756;disable:4056)
2118 S_sv_setnv(pTHX_ SV* sv, int numtype)
2120 bool pok = cBOOL(SvPOK(sv));
2122 if ((numtype & IS_NUMBER_INFINITY)) {
2123 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
2126 else if ((numtype & IS_NUMBER_NAN)) {
2127 SvNV_set(sv, NV_NAN);
2131 SvNV_set(sv, Atof(SvPVX_const(sv)));
2132 /* Purposefully no true nok here, since we don't want to blow
2133 * away the possible IOK/UV of an existing sv. */
2136 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
2138 SvPOK_on(sv); /* PV is okay, though. */
2142 # pragma warning(pop)
2146 S_sv_2iuv_common(pTHX_ SV *const sv)
2148 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2151 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2152 * without also getting a cached IV/UV from it at the same time
2153 * (ie PV->NV conversion should detect loss of accuracy and cache
2154 * IV or UV at same time to avoid this. */
2155 /* IV-over-UV optimisation - choose to cache IV if possible */
2157 if (SvTYPE(sv) == SVt_NV)
2158 sv_upgrade(sv, SVt_PVNV);
2160 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2161 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2162 certainly cast into the IV range at IV_MAX, whereas the correct
2163 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2165 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2166 if (Perl_isnan(SvNVX(sv))) {
2172 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2173 SvIV_set(sv, I_V(SvNVX(sv)));
2174 if (SvNVX(sv) == (NV) SvIVX(sv)
2175 #ifndef NV_PRESERVES_UV
2176 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
2177 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2178 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2179 /* Don't flag it as "accurately an integer" if the number
2180 came from a (by definition imprecise) NV operation, and
2181 we're outside the range of NV integer precision */
2185 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2187 /* scalar has trailing garbage, eg "42a" */
2189 DEBUG_c(PerlIO_printf(Perl_debug_log,
2190 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
2196 /* IV not precise. No need to convert from PV, as NV
2197 conversion would already have cached IV if it detected
2198 that PV->IV would be better than PV->NV->IV
2199 flags already correct - don't set public IOK. */
2200 DEBUG_c(PerlIO_printf(Perl_debug_log,
2201 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
2206 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2207 but the cast (NV)IV_MIN rounds to a the value less (more
2208 negative) than IV_MIN which happens to be equal to SvNVX ??
2209 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2210 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2211 (NV)UVX == NVX are both true, but the values differ. :-(
2212 Hopefully for 2s complement IV_MIN is something like
2213 0x8000000000000000 which will be exact. NWC */
2216 SvUV_set(sv, U_V(SvNVX(sv)));
2218 (SvNVX(sv) == (NV) SvUVX(sv))
2219 #ifndef NV_PRESERVES_UV
2220 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2221 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2222 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2223 /* Don't flag it as "accurately an integer" if the number
2224 came from a (by definition imprecise) NV operation, and
2225 we're outside the range of NV integer precision */
2231 DEBUG_c(PerlIO_printf(Perl_debug_log,
2232 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2238 else if (SvPOKp(sv)) {
2240 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2241 /* We want to avoid a possible problem when we cache an IV/ a UV which
2242 may be later translated to an NV, and the resulting NV is not
2243 the same as the direct translation of the initial string
2244 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2245 be careful to ensure that the value with the .456 is around if the
2246 NV value is requested in the future).
2248 This means that if we cache such an IV/a UV, we need to cache the
2249 NV as well. Moreover, we trade speed for space, and do not
2250 cache the NV if we are sure it's not needed.
2253 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2254 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2255 == IS_NUMBER_IN_UV) {
2256 /* It's definitely an integer, only upgrade to PVIV */
2257 if (SvTYPE(sv) < SVt_PVIV)
2258 sv_upgrade(sv, SVt_PVIV);
2260 } else if (SvTYPE(sv) < SVt_PVNV)
2261 sv_upgrade(sv, SVt_PVNV);
2263 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
2264 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
2266 S_sv_setnv(aTHX_ sv, numtype);
2270 /* If NVs preserve UVs then we only use the UV value if we know that
2271 we aren't going to call atof() below. If NVs don't preserve UVs
2272 then the value returned may have more precision than atof() will
2273 return, even though value isn't perfectly accurate. */
2274 if ((numtype & (IS_NUMBER_IN_UV
2275 #ifdef NV_PRESERVES_UV
2278 )) == IS_NUMBER_IN_UV) {
2279 /* This won't turn off the public IOK flag if it was set above */
2280 (void)SvIOKp_on(sv);
2282 if (!(numtype & IS_NUMBER_NEG)) {
2284 if (value <= (UV)IV_MAX) {
2285 SvIV_set(sv, (IV)value);
2287 /* it didn't overflow, and it was positive. */
2288 SvUV_set(sv, value);
2292 /* 2s complement assumption */
2293 if (value <= (UV)IV_MIN) {
2294 SvIV_set(sv, value == (UV)IV_MIN
2295 ? IV_MIN : -(IV)value);
2297 /* Too negative for an IV. This is a double upgrade, but
2298 I'm assuming it will be rare. */
2299 if (SvTYPE(sv) < SVt_PVNV)
2300 sv_upgrade(sv, SVt_PVNV);
2304 SvNV_set(sv, -(NV)value);
2305 SvIV_set(sv, IV_MIN);
2309 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2310 will be in the previous block to set the IV slot, and the next
2311 block to set the NV slot. So no else here. */
2313 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2314 != IS_NUMBER_IN_UV) {
2315 /* It wasn't an (integer that doesn't overflow the UV). */
2316 S_sv_setnv(aTHX_ sv, numtype);
2318 if (! numtype && ckWARN(WARN_NUMERIC))
2321 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" NVgf ")\n",
2322 PTR2UV(sv), SvNVX(sv)));
2324 #ifdef NV_PRESERVES_UV
2325 (void)SvIOKp_on(sv);
2327 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2328 if (Perl_isnan(SvNVX(sv))) {
2334 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2335 SvIV_set(sv, I_V(SvNVX(sv)));
2336 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2339 NOOP; /* Integer is imprecise. NOK, IOKp */
2341 /* UV will not work better than IV */
2343 if (SvNVX(sv) > (NV)UV_MAX) {
2345 /* Integer is inaccurate. NOK, IOKp, is UV */
2346 SvUV_set(sv, UV_MAX);
2348 SvUV_set(sv, U_V(SvNVX(sv)));
2349 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2350 NV preservse UV so can do correct comparison. */
2351 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2354 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2359 #else /* NV_PRESERVES_UV */
2360 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2361 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2362 /* The IV/UV slot will have been set from value returned by
2363 grok_number above. The NV slot has just been set using
2366 assert (SvIOKp(sv));
2368 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2369 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2370 /* Small enough to preserve all bits. */
2371 (void)SvIOKp_on(sv);
2373 SvIV_set(sv, I_V(SvNVX(sv)));
2374 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2376 /* Assumption: first non-preserved integer is < IV_MAX,
2377 this NV is in the preserved range, therefore: */
2378 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2380 Perl_croak(aTHX_ "sv_2iv assumed (U_V(fabs((double)SvNVX(sv))) < (UV)IV_MAX) but SvNVX(sv)=%"NVgf" U_V is 0x%"UVxf", IV_MAX is 0x%"UVxf"\n", SvNVX(sv), U_V(SvNVX(sv)), (UV)IV_MAX);
2384 0 0 already failed to read UV.
2385 0 1 already failed to read UV.
2386 1 0 you won't get here in this case. IV/UV
2387 slot set, public IOK, Atof() unneeded.
2388 1 1 already read UV.
2389 so there's no point in sv_2iuv_non_preserve() attempting
2390 to use atol, strtol, strtoul etc. */
2392 sv_2iuv_non_preserve (sv, numtype);
2394 sv_2iuv_non_preserve (sv);
2398 #endif /* NV_PRESERVES_UV */
2399 /* It might be more code efficient to go through the entire logic above
2400 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2401 gets complex and potentially buggy, so more programmer efficient
2402 to do it this way, by turning off the public flags: */
2404 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2408 if (isGV_with_GP(sv))
2409 return glob_2number(MUTABLE_GV(sv));
2411 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2413 if (SvTYPE(sv) < SVt_IV)
2414 /* Typically the caller expects that sv_any is not NULL now. */
2415 sv_upgrade(sv, SVt_IV);
2416 /* Return 0 from the caller. */
2423 =for apidoc sv_2iv_flags
2425 Return the integer value of an SV, doing any necessary string
2426 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2427 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2433 Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
2435 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
2437 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2438 && SvTYPE(sv) != SVt_PVFM);
2440 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2446 if (flags & SV_SKIP_OVERLOAD)
2448 tmpstr = AMG_CALLunary(sv, numer_amg);
2449 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2450 return SvIV(tmpstr);
2453 return PTR2IV(SvRV(sv));
2456 if (SvVALID(sv) || isREGEXP(sv)) {
2457 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2458 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2459 In practice they are extremely unlikely to actually get anywhere
2460 accessible by user Perl code - the only way that I'm aware of is when
2461 a constant subroutine which is used as the second argument to index.
2463 Regexps have no SvIVX and SvNVX fields.
2465 assert(isREGEXP(sv) || SvPOKp(sv));
2468 const char * const ptr =
2469 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2471 = grok_number(ptr, SvCUR(sv), &value);
2473 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2474 == IS_NUMBER_IN_UV) {
2475 /* It's definitely an integer */
2476 if (numtype & IS_NUMBER_NEG) {
2477 if (value < (UV)IV_MIN)
2480 if (value < (UV)IV_MAX)
2485 /* Quite wrong but no good choices. */
2486 if ((numtype & IS_NUMBER_INFINITY)) {
2487 return (numtype & IS_NUMBER_NEG) ? IV_MIN : IV_MAX;
2488 } else if ((numtype & IS_NUMBER_NAN)) {
2489 return 0; /* So wrong. */
2493 if (ckWARN(WARN_NUMERIC))
2496 return I_V(Atof(ptr));
2500 if (SvTHINKFIRST(sv)) {
2501 #ifdef PERL_OLD_COPY_ON_WRITE
2503 sv_force_normal_flags(sv, 0);
2506 if (SvREADONLY(sv) && !SvOK(sv)) {
2507 if (ckWARN(WARN_UNINITIALIZED))
2514 if (S_sv_2iuv_common(aTHX_ sv))
2518 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2519 PTR2UV(sv),SvIVX(sv)));
2520 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2524 =for apidoc sv_2uv_flags
2526 Return the unsigned integer value of an SV, doing any necessary string
2527 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2528 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2534 Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
2536 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
2538 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2544 if (flags & SV_SKIP_OVERLOAD)
2546 tmpstr = AMG_CALLunary(sv, numer_amg);
2547 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2548 return SvUV(tmpstr);
2551 return PTR2UV(SvRV(sv));
2554 if (SvVALID(sv) || isREGEXP(sv)) {
2555 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2556 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2557 Regexps have no SvIVX and SvNVX fields. */
2558 assert(isREGEXP(sv) || SvPOKp(sv));
2561 const char * const ptr =
2562 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2564 = grok_number(ptr, SvCUR(sv), &value);
2566 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2567 == IS_NUMBER_IN_UV) {
2568 /* It's definitely an integer */
2569 if (!(numtype & IS_NUMBER_NEG))
2573 /* Quite wrong but no good choices. */
2574 if ((numtype & IS_NUMBER_INFINITY)) {
2575 return UV_MAX; /* So wrong. */
2576 } else if ((numtype & IS_NUMBER_NAN)) {
2577 return 0; /* So wrong. */
2581 if (ckWARN(WARN_NUMERIC))
2584 return U_V(Atof(ptr));
2588 if (SvTHINKFIRST(sv)) {
2589 #ifdef PERL_OLD_COPY_ON_WRITE
2591 sv_force_normal_flags(sv, 0);
2594 if (SvREADONLY(sv) && !SvOK(sv)) {
2595 if (ckWARN(WARN_UNINITIALIZED))
2602 if (S_sv_2iuv_common(aTHX_ sv))
2606 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2607 PTR2UV(sv),SvUVX(sv)));
2608 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2612 =for apidoc sv_2nv_flags
2614 Return the num value of an SV, doing any necessary string or integer
2615 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2616 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2622 Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
2624 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2626 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2627 && SvTYPE(sv) != SVt_PVFM);
2628 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2629 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2630 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2631 Regexps have no SvIVX and SvNVX fields. */
2633 if (flags & SV_GMAGIC)
2637 if (SvPOKp(sv) && !SvIOKp(sv)) {
2638 ptr = SvPVX_const(sv);
2640 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2641 !grok_number(ptr, SvCUR(sv), NULL))
2647 return (NV)SvUVX(sv);
2649 return (NV)SvIVX(sv);
2655 ptr = RX_WRAPPED((REGEXP *)sv);
2658 assert(SvTYPE(sv) >= SVt_PVMG);
2659 /* This falls through to the report_uninit near the end of the
2661 } else if (SvTHINKFIRST(sv)) {
2666 if (flags & SV_SKIP_OVERLOAD)
2668 tmpstr = AMG_CALLunary(sv, numer_amg);
2669 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2670 return SvNV(tmpstr);
2673 return PTR2NV(SvRV(sv));
2675 #ifdef PERL_OLD_COPY_ON_WRITE
2677 sv_force_normal_flags(sv, 0);
2680 if (SvREADONLY(sv) && !SvOK(sv)) {
2681 if (ckWARN(WARN_UNINITIALIZED))
2686 if (SvTYPE(sv) < SVt_NV) {
2687 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2688 sv_upgrade(sv, SVt_NV);
2690 STORE_NUMERIC_LOCAL_SET_STANDARD();
2691 PerlIO_printf(Perl_debug_log,
2692 "0x%"UVxf" num(%" NVgf ")\n",
2693 PTR2UV(sv), SvNVX(sv));
2694 RESTORE_NUMERIC_LOCAL();
2697 else if (SvTYPE(sv) < SVt_PVNV)
2698 sv_upgrade(sv, SVt_PVNV);
2703 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2704 #ifdef NV_PRESERVES_UV
2710 /* Only set the public NV OK flag if this NV preserves the IV */
2711 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2713 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2714 : (SvIVX(sv) == I_V(SvNVX(sv))))
2720 else if (SvPOKp(sv)) {
2722 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2723 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2725 #ifdef NV_PRESERVES_UV
2726 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2727 == IS_NUMBER_IN_UV) {
2728 /* It's definitely an integer */
2729 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2731 S_sv_setnv(aTHX_ sv, numtype);
2738 SvNV_set(sv, Atof(SvPVX_const(sv)));
2739 /* Only set the public NV OK flag if this NV preserves the value in
2740 the PV at least as well as an IV/UV would.
2741 Not sure how to do this 100% reliably. */
2742 /* if that shift count is out of range then Configure's test is
2743 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2745 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2746 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2747 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2748 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2749 /* Can't use strtol etc to convert this string, so don't try.
2750 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2753 /* value has been set. It may not be precise. */
2754 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
2755 /* 2s complement assumption for (UV)IV_MIN */
2756 SvNOK_on(sv); /* Integer is too negative. */
2761 if (numtype & IS_NUMBER_NEG) {
2762 /* -IV_MIN is undefined, but we should never reach
2763 * this point with both IS_NUMBER_NEG and value ==
2765 assert(value != (UV)IV_MIN);
2766 SvIV_set(sv, -(IV)value);
2767 } else if (value <= (UV)IV_MAX) {
2768 SvIV_set(sv, (IV)value);
2770 SvUV_set(sv, value);
2774 if (numtype & IS_NUMBER_NOT_INT) {
2775 /* I believe that even if the original PV had decimals,
2776 they are lost beyond the limit of the FP precision.
2777 However, neither is canonical, so both only get p
2778 flags. NWC, 2000/11/25 */
2779 /* Both already have p flags, so do nothing */
2781 const NV nv = SvNVX(sv);
2782 /* XXX should this spot have NAN_COMPARE_BROKEN, too? */
2783 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2784 if (SvIVX(sv) == I_V(nv)) {
2787 /* It had no "." so it must be integer. */
2791 /* between IV_MAX and NV(UV_MAX).
2792 Could be slightly > UV_MAX */
2794 if (numtype & IS_NUMBER_NOT_INT) {
2795 /* UV and NV both imprecise. */
2797 const UV nv_as_uv = U_V(nv);
2799 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2808 /* It might be more code efficient to go through the entire logic above
2809 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2810 gets complex and potentially buggy, so more programmer efficient
2811 to do it this way, by turning off the public flags: */
2813 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2814 #endif /* NV_PRESERVES_UV */
2817 if (isGV_with_GP(sv)) {
2818 glob_2number(MUTABLE_GV(sv));
2822 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2824 assert (SvTYPE(sv) >= SVt_NV);
2825 /* Typically the caller expects that sv_any is not NULL now. */
2826 /* XXX Ilya implies that this is a bug in callers that assume this
2827 and ideally should be fixed. */
2831 STORE_NUMERIC_LOCAL_SET_STANDARD();
2832 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" NVgf ")\n",
2833 PTR2UV(sv), SvNVX(sv));
2834 RESTORE_NUMERIC_LOCAL();
2842 Return an SV with the numeric value of the source SV, doing any necessary
2843 reference or overload conversion. The caller is expected to have handled
2850 Perl_sv_2num(pTHX_ SV *const sv)
2852 PERL_ARGS_ASSERT_SV_2NUM;
2857 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2858 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2859 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2860 return sv_2num(tmpsv);
2862 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2865 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2866 * UV as a string towards the end of buf, and return pointers to start and
2869 * We assume that buf is at least TYPE_CHARS(UV) long.
2873 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2875 char *ptr = buf + TYPE_CHARS(UV);
2876 char * const ebuf = ptr;
2879 PERL_ARGS_ASSERT_UIV_2BUF;
2887 uv = (iv == IV_MIN) ? (UV)iv : (UV)(-iv);
2891 *--ptr = '0' + (char)(uv % 10);
2899 /* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
2900 * infinity or a not-a-number, writes the appropriate strings to the
2901 * buffer, including a zero byte. On success returns the written length,
2902 * excluding the zero byte, on failure (not an infinity, not a nan, or the
2903 * maxlen too small) returns zero.
2905 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2906 * shared string constants we point to, instead of generating a new
2907 * string for each instance. */
2909 S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
2910 assert(maxlen >= 4);
2911 if (maxlen < 4) /* "Inf\0", "NaN\0" */
2915 if (Perl_isinf(nv)) {
2917 if (maxlen < 5) /* "-Inf\0" */
2926 } else if (Perl_isnan(nv)) {
2930 /* XXX optionally output the payload mantissa bits as
2931 * "(unsigned)" (to match the nan("...") C99 function,
2932 * or maybe as "(0xhhh...)" would make more sense...
2933 * provide a format string so that the user can decide?
2934 * NOTE: would affect the maxlen and assert() logic.*/
2939 assert((s == buffer + 3) || (s == buffer + 4));
2941 return s - buffer - 1; /* -1: excluding the zero byte */
2946 =for apidoc sv_2pv_flags
2948 Returns a pointer to the string value of an SV, and sets *lp to its length.
2949 If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2950 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2951 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2957 Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2961 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2963 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2964 && SvTYPE(sv) != SVt_PVFM);
2965 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2970 if (flags & SV_SKIP_OVERLOAD)
2972 tmpstr = AMG_CALLunary(sv, string_amg);
2973 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2974 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2976 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2980 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2981 if (flags & SV_CONST_RETURN) {
2982 pv = (char *) SvPVX_const(tmpstr);
2984 pv = (flags & SV_MUTABLE_RETURN)
2985 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2988 *lp = SvCUR(tmpstr);
2990 pv = sv_2pv_flags(tmpstr, lp, flags);
3003 SV *const referent = SvRV(sv);
3007 retval = buffer = savepvn("NULLREF", len);
3008 } else if (SvTYPE(referent) == SVt_REGEXP &&
3009 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
3010 amagic_is_enabled(string_amg))) {
3011 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
3015 /* If the regex is UTF-8 we want the containing scalar to
3016 have an UTF-8 flag too */
3023 *lp = RX_WRAPLEN(re);
3025 return RX_WRAPPED(re);
3027 const char *const typestr = sv_reftype(referent, 0);
3028 const STRLEN typelen = strlen(typestr);
3029 UV addr = PTR2UV(referent);
3030 const char *stashname = NULL;
3031 STRLEN stashnamelen = 0; /* hush, gcc */
3032 const char *buffer_end;
3034 if (SvOBJECT(referent)) {
3035 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3038 stashname = HEK_KEY(name);
3039 stashnamelen = HEK_LEN(name);
3041 if (HEK_UTF8(name)) {
3047 stashname = "__ANON__";
3050 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3051 + 2 * sizeof(UV) + 2 /* )\0 */;
3053 len = typelen + 3 /* (0x */
3054 + 2 * sizeof(UV) + 2 /* )\0 */;
3057 Newx(buffer, len, char);
3058 buffer_end = retval = buffer + len;
3060 /* Working backwards */
3064 *--retval = PL_hexdigit[addr & 15];
3065 } while (addr >>= 4);
3071 memcpy(retval, typestr, typelen);
3075 retval -= stashnamelen;
3076 memcpy(retval, stashname, stashnamelen);
3078 /* retval may not necessarily have reached the start of the
3080 assert (retval >= buffer);
3082 len = buffer_end - retval - 1; /* -1 for that \0 */
3094 if (flags & SV_MUTABLE_RETURN)
3095 return SvPVX_mutable(sv);
3096 if (flags & SV_CONST_RETURN)
3097 return (char *)SvPVX_const(sv);
3102 /* I'm assuming that if both IV and NV are equally valid then
3103 converting the IV is going to be more efficient */
3104 const U32 isUIOK = SvIsUV(sv);
3105 char buf[TYPE_CHARS(UV)];
3109 if (SvTYPE(sv) < SVt_PVIV)
3110 sv_upgrade(sv, SVt_PVIV);
3111 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
3113 /* inlined from sv_setpvn */
3114 s = SvGROW_mutable(sv, len + 1);
3115 Move(ptr, s, len, char);
3120 else if (SvNOK(sv)) {
3121 if (SvTYPE(sv) < SVt_PVNV)
3122 sv_upgrade(sv, SVt_PVNV);
3123 if (SvNVX(sv) == 0.0
3124 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3125 && !Perl_isnan(SvNVX(sv))
3128 s = SvGROW_mutable(sv, 2);
3133 STRLEN size = 5; /* "-Inf\0" */
3135 s = SvGROW_mutable(sv, size);
3136 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
3142 /* some Xenix systems wipe out errno here */
3151 5 + /* exponent digits */
3155 s = SvGROW_mutable(sv, size);
3156 #ifndef USE_LOCALE_NUMERIC
3157 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3163 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3164 STORE_LC_NUMERIC_SET_TO_NEEDED();
3168 PL_numeric_radix_sv &&
3169 SvUTF8(PL_numeric_radix_sv);
3170 if (local_radix && SvLEN(PL_numeric_radix_sv) > 1) {
3171 size += SvLEN(PL_numeric_radix_sv) - 1;
3172 s = SvGROW_mutable(sv, size);
3175 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3177 /* If the radix character is UTF-8, and actually is in the
3178 * output, turn on the UTF-8 flag for the scalar */
3180 instr(s, SvPVX_const(PL_numeric_radix_sv))) {
3184 RESTORE_LC_NUMERIC();
3187 /* We don't call SvPOK_on(), because it may come to
3188 * pass that the locale changes so that the
3189 * stringification we just did is no longer correct. We
3190 * will have to re-stringify every time it is needed */
3197 else if (isGV_with_GP(sv)) {
3198 GV *const gv = MUTABLE_GV(sv);
3199 SV *const buffer = sv_newmortal();
3201 gv_efullname3(buffer, gv, "*");
3203 assert(SvPOK(buffer));
3207 *lp = SvCUR(buffer);
3208 return SvPVX(buffer);
3210 else if (isREGEXP(sv)) {
3211 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
3212 return RX_WRAPPED((REGEXP *)sv);
3217 if (flags & SV_UNDEF_RETURNS_NULL)
3219 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
3221 /* Typically the caller expects that sv_any is not NULL now. */
3222 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
3223 sv_upgrade(sv, SVt_PV);
3228 const STRLEN len = s - SvPVX_const(sv);
3233 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3234 PTR2UV(sv),SvPVX_const(sv)));
3235 if (flags & SV_CONST_RETURN)
3236 return (char *)SvPVX_const(sv);
3237 if (flags & SV_MUTABLE_RETURN)
3238 return SvPVX_mutable(sv);
3243 =for apidoc sv_copypv
3245 Copies a stringified representation of the source SV into the
3246 destination SV. Automatically performs any necessary mg_get and
3247 coercion of numeric values into strings. Guaranteed to preserve
3248 UTF8 flag even from overloaded objects. Similar in nature to
3249 sv_2pv[_flags] but operates directly on an SV instead of just the
3250 string. Mostly uses sv_2pv_flags to do its work, except when that
3251 would lose the UTF-8'ness of the PV.
3253 =for apidoc sv_copypv_nomg
3255 Like sv_copypv, but doesn't invoke get magic first.
3257 =for apidoc sv_copypv_flags
3259 Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3266 Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
3271 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
3273 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
3274 sv_setpvn(dsv,s,len);
3282 =for apidoc sv_2pvbyte
3284 Return a pointer to the byte-encoded representation of the SV, and set *lp
3285 to its length. May cause the SV to be downgraded from UTF-8 as a
3288 Usually accessed via the C<SvPVbyte> macro.
3294 Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
3296 PERL_ARGS_ASSERT_SV_2PVBYTE;
3299 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3300 || isGV_with_GP(sv) || SvROK(sv)) {
3301 SV *sv2 = sv_newmortal();
3302 sv_copypv_nomg(sv2,sv);
3305 sv_utf8_downgrade(sv,0);
3306 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3310 =for apidoc sv_2pvutf8
3312 Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3313 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3315 Usually accessed via the C<SvPVutf8> macro.
3321 Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
3323 PERL_ARGS_ASSERT_SV_2PVUTF8;
3325 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3326 || isGV_with_GP(sv) || SvROK(sv))
3327 sv = sv_mortalcopy(sv);
3330 sv_utf8_upgrade_nomg(sv);
3331 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3336 =for apidoc sv_2bool
3338 This macro is only used by sv_true() or its macro equivalent, and only if
3339 the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3340 It calls sv_2bool_flags with the SV_GMAGIC flag.
3342 =for apidoc sv_2bool_flags
3344 This function is only used by sv_true() and friends, and only if
3345 the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
3346 contain SV_GMAGIC, then it does an mg_get() first.
3353 Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
3355 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3358 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3364 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3365 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3368 if(SvGMAGICAL(sv)) {
3370 goto restart; /* call sv_2bool */
3372 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3373 else if(!SvOK(sv)) {
3376 else if(SvPOK(sv)) {
3377 svb = SvPVXtrue(sv);
3379 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3380 svb = (SvIOK(sv) && SvIVX(sv) != 0)
3381 || (SvNOK(sv) && SvNVX(sv) != 0.0);
3385 goto restart; /* call sv_2bool_nomg */
3390 return SvRV(sv) != 0;
3394 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
3395 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
3399 =for apidoc sv_utf8_upgrade
3401 Converts the PV of an SV to its UTF-8-encoded form.
3402 Forces the SV to string form if it is not already.
3403 Will C<mg_get> on C<sv> if appropriate.
3404 Always sets the SvUTF8 flag to avoid future validity checks even
3405 if the whole string is the same in UTF-8 as not.
3406 Returns the number of bytes in the converted string
3408 This is not a general purpose byte encoding to Unicode interface:
3409 use the Encode extension for that.
3411 =for apidoc sv_utf8_upgrade_nomg
3413 Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
3415 =for apidoc sv_utf8_upgrade_flags
3417 Converts the PV of an SV to its UTF-8-encoded form.
3418 Forces the SV to string form if it is not already.
3419 Always sets the SvUTF8 flag to avoid future validity checks even
3420 if all the bytes are invariant in UTF-8.
3421 If C<flags> has C<SV_GMAGIC> bit set,
3422 will C<mg_get> on C<sv> if appropriate, else not.
3424 If C<flags> has SV_FORCE_UTF8_UPGRADE set, this function assumes that the PV
3425 will expand when converted to UTF-8, and skips the extra work of checking for
3426 that. Typically this flag is used by a routine that has already parsed the
3427 string and found such characters, and passes this information on so that the
3428 work doesn't have to be repeated.
3430 Returns the number of bytes in the converted string.
3432 This is not a general purpose byte encoding to Unicode interface:
3433 use the Encode extension for that.
3435 =for apidoc sv_utf8_upgrade_flags_grow
3437 Like sv_utf8_upgrade_flags, but has an additional parameter C<extra>, which is
3438 the number of unused bytes the string of 'sv' is guaranteed to have free after
3439 it upon return. This allows the caller to reserve extra space that it intends
3440 to fill, to avoid extra grows.
3442 C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3443 are implemented in terms of this function.
3445 Returns the number of bytes in the converted string (not including the spares).
3449 (One might think that the calling routine could pass in the position of the
3450 first variant character when it has set SV_FORCE_UTF8_UPGRADE, so it wouldn't
3451 have to be found again. But that is not the case, because typically when the
3452 caller is likely to use this flag, it won't be calling this routine unless it
3453 finds something that won't fit into a byte. Otherwise it tries to not upgrade
3454 and just use bytes. But some things that do fit into a byte are variants in
3455 utf8, and the caller may not have been keeping track of these.)
3457 If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3458 C<NUL> isn't guaranteed due to having other routines do the work in some input
3459 cases, or if the input is already flagged as being in utf8.
3461 The speed of this could perhaps be improved for many cases if someone wanted to
3462 write a fast function that counts the number of variant characters in a string,
3463 especially if it could return the position of the first one.
3468 Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
3470 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3472 if (sv == &PL_sv_undef)
3474 if (!SvPOK_nog(sv)) {
3476 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3477 (void) sv_2pv_flags(sv,&len, flags);
3479 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3483 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3488 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3493 S_sv_uncow(aTHX_ sv, 0);
3496 if (IN_ENCODING && !(flags & SV_UTF8_NO_ENCODING)) {
3497 sv_recode_to_utf8(sv, _get_encoding());
3498 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3502 if (SvCUR(sv) == 0) {
3503 if (extra) SvGROW(sv, extra);
3504 } else { /* Assume Latin-1/EBCDIC */
3505 /* This function could be much more efficient if we
3506 * had a FLAG in SVs to signal if there are any variant
3507 * chars in the PV. Given that there isn't such a flag
3508 * make the loop as fast as possible (although there are certainly ways
3509 * to speed this up, eg. through vectorization) */
3510 U8 * s = (U8 *) SvPVX_const(sv);
3511 U8 * e = (U8 *) SvEND(sv);
3513 STRLEN two_byte_count = 0;
3515 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3517 /* See if really will need to convert to utf8. We mustn't rely on our
3518 * incoming SV being well formed and having a trailing '\0', as certain
3519 * code in pp_formline can send us partially built SVs. */
3523 if (NATIVE_BYTE_IS_INVARIANT(ch)) continue;
3525 t--; /* t already incremented; re-point to first variant */
3530 /* utf8 conversion not needed because all are invariants. Mark as
3531 * UTF-8 even if no variant - saves scanning loop */
3533 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3538 /* Here, the string should be converted to utf8, either because of an
3539 * input flag (two_byte_count = 0), or because a character that
3540 * requires 2 bytes was found (two_byte_count = 1). t points either to
3541 * the beginning of the string (if we didn't examine anything), or to
3542 * the first variant. In either case, everything from s to t - 1 will
3543 * occupy only 1 byte each on output.
3545 * There are two main ways to convert. One is to create a new string
3546 * and go through the input starting from the beginning, appending each
3547 * converted value onto the new string as we go along. It's probably
3548 * best to allocate enough space in the string for the worst possible
3549 * case rather than possibly running out of space and having to
3550 * reallocate and then copy what we've done so far. Since everything
3551 * from s to t - 1 is invariant, the destination can be initialized
3552 * with these using a fast memory copy
3554 * The other way is to figure out exactly how big the string should be
3555 * by parsing the entire input. Then you don't have to make it big
3556 * enough to handle the worst possible case, and more importantly, if
3557 * the string you already have is large enough, you don't have to
3558 * allocate a new string, you can copy the last character in the input
3559 * string to the final position(s) that will be occupied by the
3560 * converted string and go backwards, stopping at t, since everything
3561 * before that is invariant.
3563 * There are advantages and disadvantages to each method.
3565 * In the first method, we can allocate a new string, do the memory
3566 * copy from the s to t - 1, and then proceed through the rest of the
3567 * string byte-by-byte.
3569 * In the second method, we proceed through the rest of the input
3570 * string just calculating how big the converted string will be. Then
3571 * there are two cases:
3572 * 1) if the string has enough extra space to handle the converted
3573 * value. We go backwards through the string, converting until we
3574 * get to the position we are at now, and then stop. If this
3575 * position is far enough along in the string, this method is
3576 * faster than the other method. If the memory copy were the same
3577 * speed as the byte-by-byte loop, that position would be about
3578 * half-way, as at the half-way mark, parsing to the end and back
3579 * is one complete string's parse, the same amount as starting
3580 * over and going all the way through. Actually, it would be
3581 * somewhat less than half-way, as it's faster to just count bytes
3582 * than to also copy, and we don't have the overhead of allocating
3583 * a new string, changing the scalar to use it, and freeing the
3584 * existing one. But if the memory copy is fast, the break-even
3585 * point is somewhere after half way. The counting loop could be
3586 * sped up by vectorization, etc, to move the break-even point
3587 * further towards the beginning.
3588 * 2) if the string doesn't have enough space to handle the converted
3589 * value. A new string will have to be allocated, and one might
3590 * as well, given that, start from the beginning doing the first
3591 * method. We've spent extra time parsing the string and in
3592 * exchange all we've gotten is that we know precisely how big to
3593 * make the new one. Perl is more optimized for time than space,
3594 * so this case is a loser.
3595 * So what I've decided to do is not use the 2nd method unless it is
3596 * guaranteed that a new string won't have to be allocated, assuming
3597 * the worst case. I also decided not to put any more conditions on it
3598 * than this, for now. It seems likely that, since the worst case is
3599 * twice as big as the unknown portion of the string (plus 1), we won't
3600 * be guaranteed enough space, causing us to go to the first method,
3601 * unless the string is short, or the first variant character is near
3602 * the end of it. In either of these cases, it seems best to use the
3603 * 2nd method. The only circumstance I can think of where this would
3604 * be really slower is if the string had once had much more data in it
3605 * than it does now, but there is still a substantial amount in it */
3608 STRLEN invariant_head = t - s;
3609 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3610 if (SvLEN(sv) < size) {
3612 /* Here, have decided to allocate a new string */
3617 Newx(dst, size, U8);
3619 /* If no known invariants at the beginning of the input string,
3620 * set so starts from there. Otherwise, can use memory copy to
3621 * get up to where we are now, and then start from here */
3623 if (invariant_head == 0) {
3626 Copy(s, dst, invariant_head, char);
3627 d = dst + invariant_head;
3631 append_utf8_from_native_byte(*t, &d);
3635 SvPV_free(sv); /* No longer using pre-existing string */
3636 SvPV_set(sv, (char*)dst);
3637 SvCUR_set(sv, d - dst);
3638 SvLEN_set(sv, size);
3641 /* Here, have decided to get the exact size of the string.
3642 * Currently this happens only when we know that there is
3643 * guaranteed enough space to fit the converted string, so
3644 * don't have to worry about growing. If two_byte_count is 0,
3645 * then t points to the first byte of the string which hasn't
3646 * been examined yet. Otherwise two_byte_count is 1, and t
3647 * points to the first byte in the string that will expand to
3648 * two. Depending on this, start examining at t or 1 after t.
3651 U8 *d = t + two_byte_count;
3654 /* Count up the remaining bytes that expand to two */
3657 const U8 chr = *d++;
3658 if (! NATIVE_BYTE_IS_INVARIANT(chr)) two_byte_count++;
3661 /* The string will expand by just the number of bytes that
3662 * occupy two positions. But we are one afterwards because of
3663 * the increment just above. This is the place to put the
3664 * trailing NUL, and to set the length before we decrement */
3666 d += two_byte_count;
3667 SvCUR_set(sv, d - s);
3671 /* Having decremented d, it points to the position to put the
3672 * very last byte of the expanded string. Go backwards through
3673 * the string, copying and expanding as we go, stopping when we
3674 * get to the part that is invariant the rest of the way down */
3678 if (NATIVE_BYTE_IS_INVARIANT(*e)) {
3681 *d-- = UTF8_EIGHT_BIT_LO(*e);
3682 *d-- = UTF8_EIGHT_BIT_HI(*e);
3688 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3689 /* Update pos. We do it at the end rather than during
3690 * the upgrade, to avoid slowing down the common case
3691 * (upgrade without pos).
3692 * pos can be stored as either bytes or characters. Since
3693 * this was previously a byte string we can just turn off
3694 * the bytes flag. */
3695 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3697 mg->mg_flags &= ~MGf_BYTES;
3699 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3700 magic_setutf8(sv,mg); /* clear UTF8 cache */
3705 /* Mark as UTF-8 even if no variant - saves scanning loop */
3711 =for apidoc sv_utf8_downgrade
3713 Attempts to convert the PV of an SV from characters to bytes.
3714 If the PV contains a character that cannot fit
3715 in a byte, this conversion will fail;
3716 in this case, either returns false or, if C<fail_ok> is not
3719 This is not a general purpose Unicode to byte encoding interface:
3720 use the Encode extension for that.
3726 Perl_sv_utf8_downgrade(pTHX_ SV *const sv, const bool fail_ok)
3728 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3730 if (SvPOKp(sv) && SvUTF8(sv)) {
3734 int mg_flags = SV_GMAGIC;
3737 S_sv_uncow(aTHX_ sv, 0);
3739 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3741 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3742 if (mg && mg->mg_len > 0 && mg->mg_flags & MGf_BYTES) {
3743 mg->mg_len = sv_pos_b2u_flags(sv, mg->mg_len,
3744 SV_GMAGIC|SV_CONST_RETURN);
3745 mg_flags = 0; /* sv_pos_b2u does get magic */
3747 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3748 magic_setutf8(sv,mg); /* clear UTF8 cache */
3751 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3753 if (!utf8_to_bytes(s, &len)) {
3758 Perl_croak(aTHX_ "Wide character in %s",
3761 Perl_croak(aTHX_ "Wide character");
3772 =for apidoc sv_utf8_encode
3774 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3775 flag off so that it looks like octets again.
3781 Perl_sv_utf8_encode(pTHX_ SV *const sv)
3783 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3785 if (SvREADONLY(sv)) {
3786 sv_force_normal_flags(sv, 0);
3788 (void) sv_utf8_upgrade(sv);
3793 =for apidoc sv_utf8_decode
3795 If the PV of the SV is an octet sequence in UTF-8
3796 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3797 so that it looks like a character. If the PV contains only single-byte
3798 characters, the C<SvUTF8> flag stays off.
3799 Scans PV for validity and returns false if the PV is invalid UTF-8.
3805 Perl_sv_utf8_decode(pTHX_ SV *const sv)
3807 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3810 const U8 *start, *c;
3813 /* The octets may have got themselves encoded - get them back as
3816 if (!sv_utf8_downgrade(sv, TRUE))
3819 /* it is actually just a matter of turning the utf8 flag on, but
3820 * we want to make sure everything inside is valid utf8 first.
3822 c = start = (const U8 *) SvPVX_const(sv);
3823 if (!is_utf8_string(c, SvCUR(sv)))
3825 e = (const U8 *) SvEND(sv);
3828 if (!UTF8_IS_INVARIANT(ch)) {
3833 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3834 /* XXX Is this dead code? XS_utf8_decode calls SvSETMAGIC
3835 after this, clearing pos. Does anything on CPAN
3837 /* adjust pos to the start of a UTF8 char sequence */
3838 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3840 I32 pos = mg->mg_len;
3842 for (c = start + pos; c > start; c--) {
3843 if (UTF8_IS_START(*c))
3846 mg->mg_len = c - start;
3849 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3850 magic_setutf8(sv,mg); /* clear UTF8 cache */
3857 =for apidoc sv_setsv
3859 Copies the contents of the source SV C<ssv> into the destination SV
3860 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3861 function if the source SV needs to be reused. Does not handle 'set' magic on
3862 destination SV. Calls 'get' magic on source SV. Loosely speaking, it
3863 performs a copy-by-value, obliterating any previous content of the
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 =for apidoc sv_setsv_flags
3872 Copies the contents of the source SV C<ssv> into the destination SV
3873 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3874 function if the source SV needs to be reused. Does not handle 'set' magic.
3875 Loosely speaking, it performs a copy-by-value, obliterating any previous
3876 content of the destination.
3877 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3878 C<ssv> if appropriate, else not. If the C<flags>
3879 parameter has the C<SV_NOSTEAL> bit set then the
3880 buffers of temps will not be stolen. <sv_setsv>
3881 and C<sv_setsv_nomg> are implemented in terms of this function.
3883 You probably want to use one of the assortment of wrappers, such as
3884 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3885 C<SvSetMagicSV_nosteal>.
3887 This is the primary function for copying scalars, and most other
3888 copy-ish functions and macros use this underneath.
3894 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3896 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3897 HV *old_stash = NULL;
3899 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3901 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3902 const char * const name = GvNAME(sstr);
3903 const STRLEN len = GvNAMELEN(sstr);
3905 if (dtype >= SVt_PV) {
3911 SvUPGRADE(dstr, SVt_PVGV);
3912 (void)SvOK_off(dstr);
3913 isGV_with_GP_on(dstr);
3915 GvSTASH(dstr) = GvSTASH(sstr);
3917 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3918 gv_name_set(MUTABLE_GV(dstr), name, len,
3919 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3920 SvFAKE_on(dstr); /* can coerce to non-glob */
3923 if(GvGP(MUTABLE_GV(sstr))) {
3924 /* If source has method cache entry, clear it */
3926 SvREFCNT_dec(GvCV(sstr));
3927 GvCV_set(sstr, NULL);
3930 /* If source has a real method, then a method is
3933 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3939 /* If dest already had a real method, that's a change as well */
3941 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3942 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3947 /* We don't need to check the name of the destination if it was not a
3948 glob to begin with. */
3949 if(dtype == SVt_PVGV) {
3950 const char * const name = GvNAME((const GV *)dstr);
3953 /* The stash may have been detached from the symbol table, so
3955 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3959 const STRLEN len = GvNAMELEN(dstr);
3960 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3961 || (len == 1 && name[0] == ':')) {
3964 /* Set aside the old stash, so we can reset isa caches on
3966 if((old_stash = GvHV(dstr)))
3967 /* Make sure we do not lose it early. */
3968 SvREFCNT_inc_simple_void_NN(
3969 sv_2mortal((SV *)old_stash)
3974 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
3977 gp_free(MUTABLE_GV(dstr));
3978 GvINTRO_off(dstr); /* one-shot flag */
3979 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3980 if (SvTAINTED(sstr))
3982 if (GvIMPORTED(dstr) != GVf_IMPORTED
3983 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3985 GvIMPORTED_on(dstr);
3988 if(mro_changes == 2) {
3989 if (GvAV((const GV *)sstr)) {
3991 SV * const sref = (SV *)GvAV((const GV *)dstr);
3992 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3993 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3994 AV * const ary = newAV();
3995 av_push(ary, mg->mg_obj); /* takes the refcount */
3996 mg->mg_obj = (SV *)ary;
3998 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
4000 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
4002 mro_isa_changed_in(GvSTASH(dstr));
4004 else if(mro_changes == 3) {
4005 HV * const stash = GvHV(dstr);
4006 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
4012 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
4013 if (GvIO(dstr) && dtype == SVt_PVGV) {
4014 DEBUG_o(Perl_deb(aTHX_
4015 "glob_assign_glob clearing PL_stashcache\n"));
4016 /* It's a cache. It will rebuild itself quite happily.
4017 It's a lot of effort to work out exactly which key (or keys)
4018 might be invalidated by the creation of the this file handle.
4020 hv_clear(PL_stashcache);
4026 Perl_gv_setref(pTHX_ SV *const dstr, SV *const sstr)
4028 SV * const sref = SvRV(sstr);
4030 const int intro = GvINTRO(dstr);
4033 const U32 stype = SvTYPE(sref);
4035 PERL_ARGS_ASSERT_GV_SETREF;
4038 GvINTRO_off(dstr); /* one-shot flag */
4039 GvLINE(dstr) = CopLINE(PL_curcop);
4040 GvEGV(dstr) = MUTABLE_GV(dstr);
4045 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
4046 import_flag = GVf_IMPORTED_CV;
4049 location = (SV **) &GvHV(dstr);
4050 import_flag = GVf_IMPORTED_HV;
4053 location = (SV **) &GvAV(dstr);
4054 import_flag = GVf_IMPORTED_AV;
4057 location = (SV **) &GvIOp(dstr);
4060 location = (SV **) &GvFORM(dstr);
4063 location = &GvSV(dstr);
4064 import_flag = GVf_IMPORTED_SV;
4067 if (stype == SVt_PVCV) {
4068 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
4069 if (GvCVGEN(dstr)) {
4070 SvREFCNT_dec(GvCV(dstr));
4071 GvCV_set(dstr, NULL);
4072 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4075 /* SAVEt_GVSLOT takes more room on the savestack and has more
4076 overhead in leave_scope than SAVEt_GENERIC_SV. But for CVs
4077 leave_scope needs access to the GV so it can reset method
4078 caches. We must use SAVEt_GVSLOT whenever the type is
4079 SVt_PVCV, even if the stash is anonymous, as the stash may
4080 gain a name somehow before leave_scope. */
4081 if (stype == SVt_PVCV) {
4082 /* There is no save_pushptrptrptr. Creating it for this
4083 one call site would be overkill. So inline the ss add
4087 SS_ADD_PTR(location);
4088 SS_ADD_PTR(SvREFCNT_inc(*location));
4089 SS_ADD_UV(SAVEt_GVSLOT);
4092 else SAVEGENERICSV(*location);
4095 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
4096 CV* const cv = MUTABLE_CV(*location);
4098 if (!GvCVGEN((const GV *)dstr) &&
4099 (CvROOT(cv) || CvXSUB(cv)) &&
4100 /* redundant check that avoids creating the extra SV
4101 most of the time: */
4102 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
4104 SV * const new_const_sv =
4105 CvCONST((const CV *)sref)
4106 ? cv_const_sv((const CV *)sref)
4108 report_redefined_cv(
4109 sv_2mortal(Perl_newSVpvf(aTHX_
4112 HvNAME_HEK(GvSTASH((const GV *)dstr))
4114 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr)))
4117 CvCONST((const CV *)sref) ? &new_const_sv : NULL
4121 cv_ckproto_len_flags(cv, (const GV *)dstr,
4122 SvPOK(sref) ? CvPROTO(sref) : NULL,
4123 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
4124 SvPOK(sref) ? SvUTF8(sref) : 0);
4126 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4127 GvASSUMECV_on(dstr);
4128 if(GvSTASH(dstr)) { /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
4129 if (intro && GvREFCNT(dstr) > 1) {
4130 /* temporary remove extra savestack's ref */
4132 gv_method_changed(dstr);
4135 else gv_method_changed(dstr);
4138 *location = SvREFCNT_inc_simple_NN(sref);
4139 if (import_flag && !(GvFLAGS(dstr) & import_flag)
4140 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
4141 GvFLAGS(dstr) |= import_flag;
4143 if (import_flag == GVf_IMPORTED_SV) {
4145 save_aliased_sv((GV *)dstr);
4147 /* Turn off the flag if sref is not referenced elsewhere,
4148 even by weak refs. (SvRMAGICAL is a pessimistic check for
4150 if (SvREFCNT(sref) <= 2 && !SvRMAGICAL(sref))
4151 GvALIASED_SV_off(dstr);
4153 GvALIASED_SV_on(dstr);
4155 if (stype == SVt_PVHV) {
4156 const char * const name = GvNAME((GV*)dstr);
4157 const STRLEN len = GvNAMELEN(dstr);
4160 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
4161 || (len == 1 && name[0] == ':')
4163 && (!dref || HvENAME_get(dref))
4166 (HV *)sref, (HV *)dref,
4172 stype == SVt_PVAV && sref != dref
4173 && strEQ(GvNAME((GV*)dstr), "ISA")
4174 /* The stash may have been detached from the symbol table, so
4175 check its name before doing anything. */
4176 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
4179 MAGIC * const omg = dref && SvSMAGICAL(dref)
4180 ? mg_find(dref, PERL_MAGIC_isa)
4182 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
4183 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
4184 AV * const ary = newAV();
4185 av_push(ary, mg->mg_obj); /* takes the refcount */
4186 mg->mg_obj = (SV *)ary;
4189 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
4190 SV **svp = AvARRAY((AV *)omg->mg_obj);
4191 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
4195 SvREFCNT_inc_simple_NN(*svp++)
4201 SvREFCNT_inc_simple_NN(omg->mg_obj)
4205 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
4210 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
4212 mg = mg_find(sref, PERL_MAGIC_isa);
4214 /* Since the *ISA assignment could have affected more than
4215 one stash, don't call mro_isa_changed_in directly, but let
4216 magic_clearisa do it for us, as it already has the logic for
4217 dealing with globs vs arrays of globs. */
4219 Perl_magic_clearisa(aTHX_ NULL, mg);
4221 else if (stype == SVt_PVIO) {
4222 DEBUG_o(Perl_deb(aTHX_ "gv_setref clearing PL_stashcache\n"));
4223 /* It's a cache. It will rebuild itself quite happily.
4224 It's a lot of effort to work out exactly which key (or keys)
4225 might be invalidated by the creation of the this file handle.
4227 hv_clear(PL_stashcache);
4231 if (!intro) SvREFCNT_dec(dref);
4232 if (SvTAINTED(sstr))
4240 #ifdef PERL_DEBUG_READONLY_COW
4241 # include <sys/mman.h>
4243 # ifndef PERL_MEMORY_DEBUG_HEADER_SIZE
4244 # define PERL_MEMORY_DEBUG_HEADER_SIZE 0
4248 Perl_sv_buf_to_ro(pTHX_ SV *sv)
4250 struct perl_memory_debug_header * const header =
4251 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4252 const MEM_SIZE len = header->size;
4253 PERL_ARGS_ASSERT_SV_BUF_TO_RO;
4254 # ifdef PERL_TRACK_MEMPOOL
4255 if (!header->readonly) header->readonly = 1;
4257 if (mprotect(header, len, PROT_READ))
4258 Perl_warn(aTHX_ "mprotect RW for COW string %p %lu failed with %d",
4259 header, len, errno);
4263 S_sv_buf_to_rw(pTHX_ SV *sv)
4265 struct perl_memory_debug_header * const header =
4266 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4267 const MEM_SIZE len = header->size;
4268 PERL_ARGS_ASSERT_SV_BUF_TO_RW;
4269 if (mprotect(header, len, PROT_READ|PROT_WRITE))
4270 Perl_warn(aTHX_ "mprotect for COW string %p %lu failed with %d",
4271 header, len, errno);
4272 # ifdef PERL_TRACK_MEMPOOL
4273 header->readonly = 0;
4278 # define sv_buf_to_ro(sv) NOOP
4279 # define sv_buf_to_rw(sv) NOOP
4283 Perl_sv_setsv_flags(pTHX_ SV *dstr, SV* sstr, const I32 flags)
4289 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
4291 if (UNLIKELY( sstr == dstr ))
4294 if (SvIS_FREED(dstr)) {
4295 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
4296 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
4298 SV_CHECK_THINKFIRST_COW_DROP(dstr);
4299 if (UNLIKELY( !sstr ))
4300 sstr = &PL_sv_undef;
4301 if (SvIS_FREED(sstr)) {
4302 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
4303 (void*)sstr, (void*)dstr);
4305 stype = SvTYPE(sstr);
4306 dtype = SvTYPE(dstr);
4308 /* There's a lot of redundancy below but we're going for speed here */
4313 if (LIKELY( dtype != SVt_PVGV && dtype != SVt_PVLV )) {
4314 (void)SvOK_off(dstr);
4322 /* For performance, we inline promoting to type SVt_IV. */
4323 /* We're starting from SVt_NULL, so provided that define is
4324 * actual 0, we don't have to unset any SV type flags
4325 * to promote to SVt_IV. */
4326 STATIC_ASSERT_STMT(SVt_NULL == 0);
4327 SET_SVANY_FOR_BODYLESS_IV(dstr);
4328 SvFLAGS(dstr) |= SVt_IV;
4332 sv_upgrade(dstr, SVt_PVIV);
4336 goto end_of_first_switch;
4338 (void)SvIOK_only(dstr);
4339 SvIV_set(dstr, SvIVX(sstr));
4342 /* SvTAINTED can only be true if the SV has taint magic, which in
4343 turn means that the SV type is PVMG (or greater). This is the
4344 case statement for SVt_IV, so this cannot be true (whatever gcov
4346 assert(!SvTAINTED(sstr));
4351 if (dtype < SVt_PV && dtype != SVt_IV)
4352 sv_upgrade(dstr, SVt_IV);
4356 if (LIKELY( SvNOK(sstr) )) {
4360 sv_upgrade(dstr, SVt_NV);
4364 sv_upgrade(dstr, SVt_PVNV);
4368 goto end_of_first_switch;
4370 SvNV_set(dstr, SvNVX(sstr));
4371 (void)SvNOK_only(dstr);
4372 /* SvTAINTED can only be true if the SV has taint magic, which in
4373 turn means that the SV type is PVMG (or greater). This is the
4374 case statement for SVt_NV, so this cannot be true (whatever gcov
4376 assert(!SvTAINTED(sstr));
4383 sv_upgrade(dstr, SVt_PV);
4386 if (dtype < SVt_PVIV)
4387 sv_upgrade(dstr, SVt_PVIV);
4390 if (dtype < SVt_PVNV)
4391 sv_upgrade(dstr, SVt_PVNV);
4395 const char * const type = sv_reftype(sstr,0);
4397 /* diag_listed_as: Bizarre copy of %s */
4398 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4400 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4402 NOT_REACHED; /* NOTREACHED */
4406 if (dtype < SVt_REGEXP)
4408 if (dtype >= SVt_PV) {
4414 sv_upgrade(dstr, SVt_REGEXP);
4422 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4424 if (SvTYPE(sstr) != stype)
4425 stype = SvTYPE(sstr);
4427 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4428 glob_assign_glob(dstr, sstr, dtype);
4431 if (stype == SVt_PVLV)
4433 if (isREGEXP(sstr)) goto upgregexp;
4434 SvUPGRADE(dstr, SVt_PVNV);
4437 SvUPGRADE(dstr, (svtype)stype);
4439 end_of_first_switch:
4441 /* dstr may have been upgraded. */
4442 dtype = SvTYPE(dstr);
4443 sflags = SvFLAGS(sstr);
4445 if (UNLIKELY( dtype == SVt_PVCV )) {
4446 /* Assigning to a subroutine sets the prototype. */
4449 const char *const ptr = SvPV_const(sstr, len);
4451 SvGROW(dstr, len + 1);
4452 Copy(ptr, SvPVX(dstr), len + 1, char);
4453 SvCUR_set(dstr, len);
4455 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4456 CvAUTOLOAD_off(dstr);
4461 else if (UNLIKELY(dtype == SVt_PVAV || dtype == SVt_PVHV
4462 || dtype == SVt_PVFM))
4464 const char * const type = sv_reftype(dstr,0);
4466 /* diag_listed_as: Cannot copy to %s */
4467 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4469 Perl_croak(aTHX_ "Cannot copy to %s", type);
4470 } else if (sflags & SVf_ROK) {
4471 if (isGV_with_GP(dstr)
4472 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4475 if (GvIMPORTED(dstr) != GVf_IMPORTED
4476 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4478 GvIMPORTED_on(dstr);
4483 glob_assign_glob(dstr, sstr, dtype);
4487 if (dtype >= SVt_PV) {
4488 if (isGV_with_GP(dstr)) {
4489 gv_setref(dstr, sstr);
4492 if (SvPVX_const(dstr)) {
4498 (void)SvOK_off(dstr);
4499 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4500 SvFLAGS(dstr) |= sflags & SVf_ROK;
4501 assert(!(sflags & SVp_NOK));
4502 assert(!(sflags & SVp_IOK));
4503 assert(!(sflags & SVf_NOK));
4504 assert(!(sflags & SVf_IOK));
4506 else if (isGV_with_GP(dstr)) {
4507 if (!(sflags & SVf_OK)) {
4508 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4509 "Undefined value assigned to typeglob");
4512 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4513 if (dstr != (const SV *)gv) {
4514 const char * const name = GvNAME((const GV *)dstr);
4515 const STRLEN len = GvNAMELEN(dstr);
4516 HV *old_stash = NULL;
4517 bool reset_isa = FALSE;
4518 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4519 || (len == 1 && name[0] == ':')) {
4520 /* Set aside the old stash, so we can reset isa caches
4521 on its subclasses. */
4522 if((old_stash = GvHV(dstr))) {
4523 /* Make sure we do not lose it early. */
4524 SvREFCNT_inc_simple_void_NN(
4525 sv_2mortal((SV *)old_stash)
4532 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
4533 gp_free(MUTABLE_GV(dstr));
4535 GvGP_set(dstr, gp_ref(GvGP(gv)));
4538 HV * const stash = GvHV(dstr);
4540 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4550 else if ((dtype == SVt_REGEXP || dtype == SVt_PVLV)
4551 && (stype == SVt_REGEXP || isREGEXP(sstr))) {
4552 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4554 else if (sflags & SVp_POK) {
4555 const STRLEN cur = SvCUR(sstr);
4556 const STRLEN len = SvLEN(sstr);
4559 * We have three basic ways to copy the string:
4565 * Which we choose is based on various factors. The following
4566 * things are listed in order of speed, fastest to slowest:
4568 * - Copying a short string
4569 * - Copy-on-write bookkeeping
4571 * - Copying a long string
4573 * We swipe the string (steal the string buffer) if the SV on the
4574 * rhs is about to be freed anyway (TEMP and refcnt==1). This is a
4575 * big win on long strings. It should be a win on short strings if
4576 * SvPVX_const(dstr) has to be allocated. If not, it should not
4577 * slow things down, as SvPVX_const(sstr) would have been freed
4580 * We also steal the buffer from a PADTMP (operator target) if it
4581 * is ‘long enough’. For short strings, a swipe does not help
4582 * here, as it causes more malloc calls the next time the target
4583 * is used. Benchmarks show that even if SvPVX_const(dstr) has to
4584 * be allocated it is still not worth swiping PADTMPs for short
4585 * strings, as the savings here are small.
4587 * If swiping is not an option, then we see whether it is
4588 * worth using copy-on-write. If the lhs already has a buf-
4589 * fer big enough and the string is short, we skip it and fall back
4590 * to method 3, since memcpy is faster for short strings than the
4591 * later bookkeeping overhead that copy-on-write entails.
4593 * If the rhs is not a copy-on-write string yet, then we also
4594 * consider whether the buffer is too large relative to the string
4595 * it holds. Some operations such as readline allocate a large
4596 * buffer in the expectation of reusing it. But turning such into
4597 * a COW buffer is counter-productive because it increases memory
4598 * usage by making readline allocate a new large buffer the sec-
4599 * ond time round. So, if the buffer is too large, again, we use
4602 * Finally, if there is no buffer on the left, or the buffer is too
4603 * small, then we use copy-on-write and make both SVs share the
4608 /* Whichever path we take through the next code, we want this true,
4609 and doing it now facilitates the COW check. */
4610 (void)SvPOK_only(dstr);
4614 /* slated for free anyway (and not COW)? */
4615 (sflags & (SVs_TEMP|SVf_IsCOW)) == SVs_TEMP
4616 /* or a swipable TARG */
4618 (SVs_PADTMP|SVf_READONLY|SVf_PROTECT|SVf_IsCOW))
4620 /* whose buffer is worth stealing */
4621 && CHECK_COWBUF_THRESHOLD(cur,len)
4624 !(sflags & SVf_OOK) &&
https://perl5.git.perl.org / perl5.git / blob