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
36 # if __STDC_VERSION__ >= 199901L && !defined(VMS)
47 /* Missing proto on LynxOS */
48 char *gconvert(double, int, int, char *);
51 #ifdef PERL_UTF8_CACHE_ASSERT
52 /* if adding more checks watch out for the following tests:
53 * t/op/index.t t/op/length.t t/op/pat.t t/op/substr.t
54 * lib/utf8.t lib/Unicode/Collate/t/index.t
57 # define ASSERT_UTF8_CACHE(cache) \
58 STMT_START { if (cache) { assert((cache)[0] <= (cache)[1]); \
59 assert((cache)[2] <= (cache)[3]); \
60 assert((cache)[3] <= (cache)[1]);} \
63 # define ASSERT_UTF8_CACHE(cache) NOOP
66 #ifdef PERL_OLD_COPY_ON_WRITE
67 #define SV_COW_NEXT_SV(sv) INT2PTR(SV *,SvUVX(sv))
68 #define SV_COW_NEXT_SV_SET(current,next) SvUV_set(current, PTR2UV(next))
71 /* ============================================================================
73 =head1 Allocation and deallocation of SVs.
75 An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
76 sv, av, hv...) contains type and reference count information, and for
77 many types, a pointer to the body (struct xrv, xpv, xpviv...), which
78 contains fields specific to each type. Some types store all they need
79 in the head, so don't have a body.
81 In all but the most memory-paranoid configurations (ex: PURIFY), heads
82 and bodies are allocated out of arenas, which by default are
83 approximately 4K chunks of memory parcelled up into N heads or bodies.
84 Sv-bodies are allocated by their sv-type, guaranteeing size
85 consistency needed to allocate safely from arrays.
87 For SV-heads, the first slot in each arena is reserved, and holds a
88 link to the next arena, some flags, and a note of the number of slots.
89 Snaked through each arena chain is a linked list of free items; when
90 this becomes empty, an extra arena is allocated and divided up into N
91 items which are threaded into the free list.
93 SV-bodies are similar, but they use arena-sets by default, which
94 separate the link and info from the arena itself, and reclaim the 1st
95 slot in the arena. SV-bodies are further described later.
97 The following global variables are associated with arenas:
99 PL_sv_arenaroot pointer to list of SV arenas
100 PL_sv_root pointer to list of free SV structures
102 PL_body_arenas head of linked-list of body arenas
103 PL_body_roots[] array of pointers to list of free bodies of svtype
104 arrays are indexed by the svtype needed
106 A few special SV heads are not allocated from an arena, but are
107 instead directly created in the interpreter structure, eg PL_sv_undef.
108 The size of arenas can be changed from the default by setting
109 PERL_ARENA_SIZE appropriately at compile time.
111 The SV arena serves the secondary purpose of allowing still-live SVs
112 to be located and destroyed during final cleanup.
114 At the lowest level, the macros new_SV() and del_SV() grab and free
115 an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
116 to return the SV to the free list with error checking.) new_SV() calls
117 more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
118 SVs in the free list have their SvTYPE field set to all ones.
120 At the time of very final cleanup, sv_free_arenas() is called from
121 perl_destruct() to physically free all the arenas allocated since the
122 start of the interpreter.
124 The function visit() scans the SV arenas list, and calls a specified
125 function for each SV it finds which is still live - ie which has an SvTYPE
126 other than all 1's, and a non-zero SvREFCNT. visit() is used by the
127 following functions (specified as [function that calls visit()] / [function
128 called by visit() for each SV]):
130 sv_report_used() / do_report_used()
131 dump all remaining SVs (debugging aid)
133 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
134 do_clean_named_io_objs(),do_curse()
135 Attempt to free all objects pointed to by RVs,
136 try to do the same for all objects indir-
137 ectly referenced by typeglobs too, and
138 then do a final sweep, cursing any
139 objects that remain. Called once from
140 perl_destruct(), prior to calling sv_clean_all()
143 sv_clean_all() / do_clean_all()
144 SvREFCNT_dec(sv) each remaining SV, possibly
145 triggering an sv_free(). It also sets the
146 SVf_BREAK flag on the SV to indicate that the
147 refcnt has been artificially lowered, and thus
148 stopping sv_free() from giving spurious warnings
149 about SVs which unexpectedly have a refcnt
150 of zero. called repeatedly from perl_destruct()
151 until there are no SVs left.
153 =head2 Arena allocator API Summary
155 Private API to rest of sv.c
159 new_XPVNV(), del_XPVGV(),
164 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
168 * ========================================================================= */
171 * "A time to plant, and a time to uproot what was planted..."
175 # define MEM_LOG_NEW_SV(sv, file, line, func) \
176 Perl_mem_log_new_sv(sv, file, line, func)
177 # define MEM_LOG_DEL_SV(sv, file, line, func) \
178 Perl_mem_log_del_sv(sv, file, line, func)
180 # define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
181 # define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
184 #ifdef DEBUG_LEAKING_SCALARS
185 # define FREE_SV_DEBUG_FILE(sv) STMT_START { \
186 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
188 # define DEBUG_SV_SERIAL(sv) \
189 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
190 PTR2UV(sv), (long)(sv)->sv_debug_serial))
192 # define FREE_SV_DEBUG_FILE(sv)
193 # define DEBUG_SV_SERIAL(sv) NOOP
197 # define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
198 # define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
199 /* Whilst I'd love to do this, it seems that things like to check on
201 # define POSION_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
203 # define POSION_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
204 PoisonNew(&SvREFCNT(sv), 1, U32)
206 # define SvARENA_CHAIN(sv) SvANY(sv)
207 # define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
208 # define POSION_SV_HEAD(sv)
211 /* Mark an SV head as unused, and add to free list.
213 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
214 * its refcount artificially decremented during global destruction, so
215 * there may be dangling pointers to it. The last thing we want in that
216 * case is for it to be reused. */
218 #define plant_SV(p) \
220 const U32 old_flags = SvFLAGS(p); \
221 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
222 DEBUG_SV_SERIAL(p); \
223 FREE_SV_DEBUG_FILE(p); \
225 SvFLAGS(p) = SVTYPEMASK; \
226 if (!(old_flags & SVf_BREAK)) { \
227 SvARENA_CHAIN_SET(p, PL_sv_root); \
233 #define uproot_SV(p) \
236 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
241 /* make some more SVs by adding another arena */
248 char *chunk; /* must use New here to match call to */
249 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
250 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
255 /* new_SV(): return a new, empty SV head */
257 #ifdef DEBUG_LEAKING_SCALARS
258 /* provide a real function for a debugger to play with */
260 S_new_SV(pTHX_ const char *file, int line, const char *func)
267 sv = S_more_sv(aTHX);
271 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
272 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
278 sv->sv_debug_inpad = 0;
279 sv->sv_debug_parent = NULL;
280 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
282 sv->sv_debug_serial = PL_sv_serial++;
284 MEM_LOG_NEW_SV(sv, file, line, func);
285 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
286 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
290 # define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
298 (p) = S_more_sv(aTHX); \
302 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
307 /* del_SV(): return an empty SV head to the free list */
320 S_del_sv(pTHX_ SV *p)
324 PERL_ARGS_ASSERT_DEL_SV;
329 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
330 const SV * const sv = sva + 1;
331 const SV * const svend = &sva[SvREFCNT(sva)];
332 if (p >= sv && p < svend) {
338 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
339 "Attempt to free non-arena SV: 0x%"UVxf
340 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
347 #else /* ! DEBUGGING */
349 #define del_SV(p) plant_SV(p)
351 #endif /* DEBUGGING */
355 =head1 SV Manipulation Functions
357 =for apidoc sv_add_arena
359 Given a chunk of memory, link it to the head of the list of arenas,
360 and split it into a list of free SVs.
366 S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
369 SV *const sva = MUTABLE_SV(ptr);
373 PERL_ARGS_ASSERT_SV_ADD_ARENA;
375 /* The first SV in an arena isn't an SV. */
376 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
377 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
378 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
380 PL_sv_arenaroot = sva;
381 PL_sv_root = sva + 1;
383 svend = &sva[SvREFCNT(sva) - 1];
386 SvARENA_CHAIN_SET(sv, (sv + 1));
390 /* Must always set typemask because it's always checked in on cleanup
391 when the arenas are walked looking for objects. */
392 SvFLAGS(sv) = SVTYPEMASK;
395 SvARENA_CHAIN_SET(sv, 0);
399 SvFLAGS(sv) = SVTYPEMASK;
402 /* visit(): call the named function for each non-free SV in the arenas
403 * whose flags field matches the flags/mask args. */
406 S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
412 PERL_ARGS_ASSERT_VISIT;
414 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
415 const SV * const svend = &sva[SvREFCNT(sva)];
417 for (sv = sva + 1; sv < svend; ++sv) {
418 if (SvTYPE(sv) != (svtype)SVTYPEMASK
419 && (sv->sv_flags & mask) == flags
432 /* called by sv_report_used() for each live SV */
435 do_report_used(pTHX_ SV *const sv)
437 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
438 PerlIO_printf(Perl_debug_log, "****\n");
445 =for apidoc sv_report_used
447 Dump the contents of all SVs not yet freed (debugging aid).
453 Perl_sv_report_used(pTHX)
456 visit(do_report_used, 0, 0);
462 /* called by sv_clean_objs() for each live SV */
465 do_clean_objs(pTHX_ SV *const ref)
470 SV * const target = SvRV(ref);
471 if (SvOBJECT(target)) {
472 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
473 if (SvWEAKREF(ref)) {
474 sv_del_backref(target, ref);
480 SvREFCNT_dec_NN(target);
487 /* clear any slots in a GV which hold objects - except IO;
488 * called by sv_clean_objs() for each live GV */
491 do_clean_named_objs(pTHX_ SV *const sv)
495 assert(SvTYPE(sv) == SVt_PVGV);
496 assert(isGV_with_GP(sv));
500 /* freeing GP entries may indirectly free the current GV;
501 * hold onto it while we mess with the GP slots */
504 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
505 DEBUG_D((PerlIO_printf(Perl_debug_log,
506 "Cleaning named glob SV object:\n "), sv_dump(obj)));
508 SvREFCNT_dec_NN(obj);
510 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
511 DEBUG_D((PerlIO_printf(Perl_debug_log,
512 "Cleaning named glob AV object:\n "), sv_dump(obj)));
514 SvREFCNT_dec_NN(obj);
516 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
517 DEBUG_D((PerlIO_printf(Perl_debug_log,
518 "Cleaning named glob HV object:\n "), sv_dump(obj)));
520 SvREFCNT_dec_NN(obj);
522 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
523 DEBUG_D((PerlIO_printf(Perl_debug_log,
524 "Cleaning named glob CV object:\n "), sv_dump(obj)));
526 SvREFCNT_dec_NN(obj);
528 SvREFCNT_dec_NN(sv); /* undo the inc above */
531 /* clear any IO slots in a GV which hold objects (except stderr, defout);
532 * called by sv_clean_objs() for each live GV */
535 do_clean_named_io_objs(pTHX_ SV *const sv)
539 assert(SvTYPE(sv) == SVt_PVGV);
540 assert(isGV_with_GP(sv));
541 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
545 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
546 DEBUG_D((PerlIO_printf(Perl_debug_log,
547 "Cleaning named glob IO object:\n "), sv_dump(obj)));
549 SvREFCNT_dec_NN(obj);
551 SvREFCNT_dec_NN(sv); /* undo the inc above */
554 /* Void wrapper to pass to visit() */
556 do_curse(pTHX_ SV * const sv) {
557 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
558 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
564 =for apidoc sv_clean_objs
566 Attempt to destroy all objects not yet freed.
572 Perl_sv_clean_objs(pTHX)
576 PL_in_clean_objs = TRUE;
577 visit(do_clean_objs, SVf_ROK, SVf_ROK);
578 /* Some barnacles may yet remain, clinging to typeglobs.
579 * Run the non-IO destructors first: they may want to output
580 * error messages, close files etc */
581 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
582 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
583 /* And if there are some very tenacious barnacles clinging to arrays,
584 closures, or what have you.... */
585 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
586 olddef = PL_defoutgv;
587 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
588 if (olddef && isGV_with_GP(olddef))
589 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
590 olderr = PL_stderrgv;
591 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
592 if (olderr && isGV_with_GP(olderr))
593 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
594 SvREFCNT_dec(olddef);
595 PL_in_clean_objs = FALSE;
598 /* called by sv_clean_all() for each live SV */
601 do_clean_all(pTHX_ SV *const sv)
604 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
605 /* don't clean pid table and strtab */
608 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%"UVxf"\n", PTR2UV(sv)) ));
609 SvFLAGS(sv) |= SVf_BREAK;
614 =for apidoc sv_clean_all
616 Decrement the refcnt of each remaining SV, possibly triggering a
617 cleanup. This function may have to be called multiple times to free
618 SVs which are in complex self-referential hierarchies.
624 Perl_sv_clean_all(pTHX)
628 PL_in_clean_all = TRUE;
629 cleaned = visit(do_clean_all, 0,0);
634 ARENASETS: a meta-arena implementation which separates arena-info
635 into struct arena_set, which contains an array of struct
636 arena_descs, each holding info for a single arena. By separating
637 the meta-info from the arena, we recover the 1st slot, formerly
638 borrowed for list management. The arena_set is about the size of an
639 arena, avoiding the needless malloc overhead of a naive linked-list.
641 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
642 memory in the last arena-set (1/2 on average). In trade, we get
643 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
644 smaller types). The recovery of the wasted space allows use of
645 small arenas for large, rare body types, by changing array* fields
646 in body_details_by_type[] below.
649 char *arena; /* the raw storage, allocated aligned */
650 size_t size; /* its size ~4k typ */
651 svtype utype; /* bodytype stored in arena */
656 /* Get the maximum number of elements in set[] such that struct arena_set
657 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
658 therefore likely to be 1 aligned memory page. */
660 #define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
661 - 2 * sizeof(int)) / sizeof (struct arena_desc))
664 struct arena_set* next;
665 unsigned int set_size; /* ie ARENAS_PER_SET */
666 unsigned int curr; /* index of next available arena-desc */
667 struct arena_desc set[ARENAS_PER_SET];
671 =for apidoc sv_free_arenas
673 Deallocate the memory used by all arenas. Note that all the individual SV
674 heads and bodies within the arenas must already have been freed.
679 Perl_sv_free_arenas(pTHX)
686 /* Free arenas here, but be careful about fake ones. (We assume
687 contiguity of the fake ones with the corresponding real ones.) */
689 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
690 svanext = MUTABLE_SV(SvANY(sva));
691 while (svanext && SvFAKE(svanext))
692 svanext = MUTABLE_SV(SvANY(svanext));
699 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
702 struct arena_set *current = aroot;
705 assert(aroot->set[i].arena);
706 Safefree(aroot->set[i].arena);
714 i = PERL_ARENA_ROOTS_SIZE;
716 PL_body_roots[i] = 0;
723 Here are mid-level routines that manage the allocation of bodies out
724 of the various arenas. There are 5 kinds of arenas:
726 1. SV-head arenas, which are discussed and handled above
727 2. regular body arenas
728 3. arenas for reduced-size bodies
731 Arena types 2 & 3 are chained by body-type off an array of
732 arena-root pointers, which is indexed by svtype. Some of the
733 larger/less used body types are malloced singly, since a large
734 unused block of them is wasteful. Also, several svtypes dont have
735 bodies; the data fits into the sv-head itself. The arena-root
736 pointer thus has a few unused root-pointers (which may be hijacked
737 later for arena types 4,5)
739 3 differs from 2 as an optimization; some body types have several
740 unused fields in the front of the structure (which are kept in-place
741 for consistency). These bodies can be allocated in smaller chunks,
742 because the leading fields arent accessed. Pointers to such bodies
743 are decremented to point at the unused 'ghost' memory, knowing that
744 the pointers are used with offsets to the real memory.
747 =head1 SV-Body Allocation
749 Allocation of SV-bodies is similar to SV-heads, differing as follows;
750 the allocation mechanism is used for many body types, so is somewhat
751 more complicated, it uses arena-sets, and has no need for still-live
754 At the outermost level, (new|del)_X*V macros return bodies of the
755 appropriate type. These macros call either (new|del)_body_type or
756 (new|del)_body_allocated macro pairs, depending on specifics of the
757 type. Most body types use the former pair, the latter pair is used to
758 allocate body types with "ghost fields".
760 "ghost fields" are fields that are unused in certain types, and
761 consequently don't need to actually exist. They are declared because
762 they're part of a "base type", which allows use of functions as
763 methods. The simplest examples are AVs and HVs, 2 aggregate types
764 which don't use the fields which support SCALAR semantics.
766 For these types, the arenas are carved up into appropriately sized
767 chunks, we thus avoid wasted memory for those unaccessed members.
768 When bodies are allocated, we adjust the pointer back in memory by the
769 size of the part not allocated, so it's as if we allocated the full
770 structure. (But things will all go boom if you write to the part that
771 is "not there", because you'll be overwriting the last members of the
772 preceding structure in memory.)
774 We calculate the correction using the STRUCT_OFFSET macro on the first
775 member present. If the allocated structure is smaller (no initial NV
776 actually allocated) then the net effect is to subtract the size of the NV
777 from the pointer, to return a new pointer as if an initial NV were actually
778 allocated. (We were using structures named *_allocated for this, but
779 this turned out to be a subtle bug, because a structure without an NV
780 could have a lower alignment constraint, but the compiler is allowed to
781 optimised accesses based on the alignment constraint of the actual pointer
782 to the full structure, for example, using a single 64 bit load instruction
783 because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
785 This is the same trick as was used for NV and IV bodies. Ironically it
786 doesn't need to be used for NV bodies any more, because NV is now at
787 the start of the structure. IV bodies don't need it either, because
788 they are no longer allocated.
790 In turn, the new_body_* allocators call S_new_body(), which invokes
791 new_body_inline macro, which takes a lock, and takes a body off the
792 linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
793 necessary to refresh an empty list. Then the lock is released, and
794 the body is returned.
796 Perl_more_bodies allocates a new arena, and carves it up into an array of N
797 bodies, which it strings into a linked list. It looks up arena-size
798 and body-size from the body_details table described below, thus
799 supporting the multiple body-types.
801 If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
802 the (new|del)_X*V macros are mapped directly to malloc/free.
804 For each sv-type, struct body_details bodies_by_type[] carries
805 parameters which control these aspects of SV handling:
807 Arena_size determines whether arenas are used for this body type, and if
808 so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
809 zero, forcing individual mallocs and frees.
811 Body_size determines how big a body is, and therefore how many fit into
812 each arena. Offset carries the body-pointer adjustment needed for
813 "ghost fields", and is used in *_allocated macros.
815 But its main purpose is to parameterize info needed in
816 Perl_sv_upgrade(). The info here dramatically simplifies the function
817 vs the implementation in 5.8.8, making it table-driven. All fields
818 are used for this, except for arena_size.
820 For the sv-types that have no bodies, arenas are not used, so those
821 PL_body_roots[sv_type] are unused, and can be overloaded. In
822 something of a special case, SVt_NULL is borrowed for HE arenas;
823 PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
824 bodies_by_type[SVt_NULL] slot is not used, as the table is not
829 struct body_details {
830 U8 body_size; /* Size to allocate */
831 U8 copy; /* Size of structure to copy (may be shorter) */
833 unsigned int type : 4; /* We have space for a sanity check. */
834 unsigned int cant_upgrade : 1; /* Cannot upgrade this type */
835 unsigned int zero_nv : 1; /* zero the NV when upgrading from this */
836 unsigned int arena : 1; /* Allocated from an arena */
837 size_t arena_size; /* Size of arena to allocate */
845 /* With -DPURFIY we allocate everything directly, and don't use arenas.
846 This seems a rather elegant way to simplify some of the code below. */
847 #define HASARENA FALSE
849 #define HASARENA TRUE
851 #define NOARENA FALSE
853 /* Size the arenas to exactly fit a given number of bodies. A count
854 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
855 simplifying the default. If count > 0, the arena is sized to fit
856 only that many bodies, allowing arenas to be used for large, rare
857 bodies (XPVFM, XPVIO) without undue waste. The arena size is
858 limited by PERL_ARENA_SIZE, so we can safely oversize the
861 #define FIT_ARENA0(body_size) \
862 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
863 #define FIT_ARENAn(count,body_size) \
864 ( count * body_size <= PERL_ARENA_SIZE) \
865 ? count * body_size \
866 : FIT_ARENA0 (body_size)
867 #define FIT_ARENA(count,body_size) \
869 ? FIT_ARENAn (count, body_size) \
870 : FIT_ARENA0 (body_size)
872 /* Calculate the length to copy. Specifically work out the length less any
873 final padding the compiler needed to add. See the comment in sv_upgrade
874 for why copying the padding proved to be a bug. */
876 #define copy_length(type, last_member) \
877 STRUCT_OFFSET(type, last_member) \
878 + sizeof (((type*)SvANY((const SV *)0))->last_member)
880 static const struct body_details bodies_by_type[] = {
881 /* HEs use this offset for their arena. */
882 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
884 /* The bind placeholder pretends to be an RV for now.
885 Also it's marked as "can't upgrade" to stop anyone using it before it's
887 { 0, 0, 0, SVt_DUMMY, TRUE, NONV, NOARENA, 0 },
889 /* IVs are in the head, so the allocation size is 0. */
891 sizeof(IV), /* This is used to copy out the IV body. */
892 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
893 NOARENA /* IVS don't need an arena */, 0
896 { sizeof(NV), sizeof(NV),
897 STRUCT_OFFSET(XPVNV, xnv_u),
898 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
900 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
901 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
902 + STRUCT_OFFSET(XPV, xpv_cur),
903 SVt_PV, FALSE, NONV, HASARENA,
904 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
906 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
907 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
908 + STRUCT_OFFSET(XPV, xpv_cur),
909 SVt_PVIV, FALSE, NONV, HASARENA,
910 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
912 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
913 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
914 + STRUCT_OFFSET(XPV, xpv_cur),
915 SVt_PVNV, FALSE, HADNV, HASARENA,
916 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
918 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
919 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
924 SVt_REGEXP, TRUE, NONV, HASARENA,
925 FIT_ARENA(0, sizeof(regexp))
928 { sizeof(XPVGV), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
929 HASARENA, FIT_ARENA(0, sizeof(XPVGV)) },
931 { sizeof(XPVLV), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
932 HASARENA, FIT_ARENA(0, sizeof(XPVLV)) },
935 copy_length(XPVAV, xav_alloc),
937 SVt_PVAV, TRUE, NONV, HASARENA,
938 FIT_ARENA(0, sizeof(XPVAV)) },
941 copy_length(XPVHV, xhv_max),
943 SVt_PVHV, TRUE, NONV, HASARENA,
944 FIT_ARENA(0, sizeof(XPVHV)) },
949 SVt_PVCV, TRUE, NONV, HASARENA,
950 FIT_ARENA(0, sizeof(XPVCV)) },
955 SVt_PVFM, TRUE, NONV, NOARENA,
956 FIT_ARENA(20, sizeof(XPVFM)) },
961 SVt_PVIO, TRUE, NONV, HASARENA,
962 FIT_ARENA(24, sizeof(XPVIO)) },
965 #define new_body_allocated(sv_type) \
966 (void *)((char *)S_new_body(aTHX_ sv_type) \
967 - bodies_by_type[sv_type].offset)
969 /* return a thing to the free list */
971 #define del_body(thing, root) \
973 void ** const thing_copy = (void **)thing; \
974 *thing_copy = *root; \
975 *root = (void*)thing_copy; \
980 #define new_XNV() safemalloc(sizeof(XPVNV))
981 #define new_XPVNV() safemalloc(sizeof(XPVNV))
982 #define new_XPVMG() safemalloc(sizeof(XPVMG))
984 #define del_XPVGV(p) safefree(p)
988 #define new_XNV() new_body_allocated(SVt_NV)
989 #define new_XPVNV() new_body_allocated(SVt_PVNV)
990 #define new_XPVMG() new_body_allocated(SVt_PVMG)
992 #define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
993 &PL_body_roots[SVt_PVGV])
997 /* no arena for you! */
999 #define new_NOARENA(details) \
1000 safemalloc((details)->body_size + (details)->offset)
1001 #define new_NOARENAZ(details) \
1002 safecalloc((details)->body_size + (details)->offset, 1)
1005 Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1006 const size_t arena_size)
1009 void ** const root = &PL_body_roots[sv_type];
1010 struct arena_desc *adesc;
1011 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1015 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
1016 #if defined(DEBUGGING) && !defined(PERL_GLOBAL_STRUCT_PRIVATE)
1017 static bool done_sanity_check;
1019 /* PERL_GLOBAL_STRUCT_PRIVATE cannot coexist with global
1020 * variables like done_sanity_check. */
1021 if (!done_sanity_check) {
1022 unsigned int i = SVt_LAST;
1024 done_sanity_check = TRUE;
1027 assert (bodies_by_type[i].type == i);
1033 /* may need new arena-set to hold new arena */
1034 if (!aroot || aroot->curr >= aroot->set_size) {
1035 struct arena_set *newroot;
1036 Newxz(newroot, 1, struct arena_set);
1037 newroot->set_size = ARENAS_PER_SET;
1038 newroot->next = aroot;
1040 PL_body_arenas = (void *) newroot;
1041 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1044 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1045 curr = aroot->curr++;
1046 adesc = &(aroot->set[curr]);
1047 assert(!adesc->arena);
1049 Newx(adesc->arena, good_arena_size, char);
1050 adesc->size = good_arena_size;
1051 adesc->utype = sv_type;
1052 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %"UVuf"\n",
1053 curr, (void*)adesc->arena, (UV)good_arena_size));
1055 start = (char *) adesc->arena;
1057 /* Get the address of the byte after the end of the last body we can fit.
1058 Remember, this is integer division: */
1059 end = start + good_arena_size / body_size * body_size;
1061 /* computed count doesn't reflect the 1st slot reservation */
1062 #if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1063 DEBUG_m(PerlIO_printf(Perl_debug_log,
1064 "arena %p end %p arena-size %d (from %d) type %d "
1066 (void*)start, (void*)end, (int)good_arena_size,
1067 (int)arena_size, sv_type, (int)body_size,
1068 (int)good_arena_size / (int)body_size));
1070 DEBUG_m(PerlIO_printf(Perl_debug_log,
1071 "arena %p end %p arena-size %d type %d size %d ct %d\n",
1072 (void*)start, (void*)end,
1073 (int)arena_size, sv_type, (int)body_size,
1074 (int)good_arena_size / (int)body_size));
1076 *root = (void *)start;
1079 /* Where the next body would start: */
1080 char * const next = start + body_size;
1083 /* This is the last body: */
1084 assert(next == end);
1086 *(void **)start = 0;
1090 *(void**) start = (void *)next;
1095 /* grab a new thing from the free list, allocating more if necessary.
1096 The inline version is used for speed in hot routines, and the
1097 function using it serves the rest (unless PURIFY).
1099 #define new_body_inline(xpv, sv_type) \
1101 void ** const r3wt = &PL_body_roots[sv_type]; \
1102 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1103 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
1104 bodies_by_type[sv_type].body_size,\
1105 bodies_by_type[sv_type].arena_size)); \
1106 *(r3wt) = *(void**)(xpv); \
1112 S_new_body(pTHX_ const svtype sv_type)
1116 new_body_inline(xpv, sv_type);
1122 static const struct body_details fake_rv =
1123 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1126 =for apidoc sv_upgrade
1128 Upgrade an SV to a more complex form. Generally adds a new body type to the
1129 SV, then copies across as much information as possible from the old body.
1130 It croaks if the SV is already in a more complex form than requested. You
1131 generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1132 before calling C<sv_upgrade>, and hence does not croak. See also
1139 Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
1144 const svtype old_type = SvTYPE(sv);
1145 const struct body_details *new_type_details;
1146 const struct body_details *old_type_details
1147 = bodies_by_type + old_type;
1148 SV *referant = NULL;
1150 PERL_ARGS_ASSERT_SV_UPGRADE;
1152 if (old_type == new_type)
1155 /* This clause was purposefully added ahead of the early return above to
1156 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1157 inference by Nick I-S that it would fix other troublesome cases. See
1158 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1160 Given that shared hash key scalars are no longer PVIV, but PV, there is
1161 no longer need to unshare so as to free up the IVX slot for its proper
1162 purpose. So it's safe to move the early return earlier. */
1164 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
1165 sv_force_normal_flags(sv, 0);
1168 old_body = SvANY(sv);
1170 /* Copying structures onto other structures that have been neatly zeroed
1171 has a subtle gotcha. Consider XPVMG
1173 +------+------+------+------+------+-------+-------+
1174 | NV | CUR | LEN | IV | MAGIC | STASH |
1175 +------+------+------+------+------+-------+-------+
1176 0 4 8 12 16 20 24 28
1178 where NVs are aligned to 8 bytes, so that sizeof that structure is
1179 actually 32 bytes long, with 4 bytes of padding at the end:
1181 +------+------+------+------+------+-------+-------+------+
1182 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1183 +------+------+------+------+------+-------+-------+------+
1184 0 4 8 12 16 20 24 28 32
1186 so what happens if you allocate memory for this structure:
1188 +------+------+------+------+------+-------+-------+------+------+...
1189 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1190 +------+------+------+------+------+-------+-------+------+------+...
1191 0 4 8 12 16 20 24 28 32 36
1193 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1194 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1195 started out as zero once, but it's quite possible that it isn't. So now,
1196 rather than a nicely zeroed GP, you have it pointing somewhere random.
1199 (In fact, GP ends up pointing at a previous GP structure, because the
1200 principle cause of the padding in XPVMG getting garbage is a copy of
1201 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1202 this happens to be moot because XPVGV has been re-ordered, with GP
1203 no longer after STASH)
1205 So we are careful and work out the size of used parts of all the
1213 referant = SvRV(sv);
1214 old_type_details = &fake_rv;
1215 if (new_type == SVt_NV)
1216 new_type = SVt_PVNV;
1218 if (new_type < SVt_PVIV) {
1219 new_type = (new_type == SVt_NV)
1220 ? SVt_PVNV : SVt_PVIV;
1225 if (new_type < SVt_PVNV) {
1226 new_type = SVt_PVNV;
1230 assert(new_type > SVt_PV);
1231 assert(SVt_IV < SVt_PV);
1232 assert(SVt_NV < SVt_PV);
1239 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1240 there's no way that it can be safely upgraded, because perl.c
1241 expects to Safefree(SvANY(PL_mess_sv)) */
1242 assert(sv != PL_mess_sv);
1243 /* This flag bit is used to mean other things in other scalar types.
1244 Given that it only has meaning inside the pad, it shouldn't be set
1245 on anything that can get upgraded. */
1246 assert(!SvPAD_TYPED(sv));
1249 if (UNLIKELY(old_type_details->cant_upgrade))
1250 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1251 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1254 if (UNLIKELY(old_type > new_type))
1255 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1256 (int)old_type, (int)new_type);
1258 new_type_details = bodies_by_type + new_type;
1260 SvFLAGS(sv) &= ~SVTYPEMASK;
1261 SvFLAGS(sv) |= new_type;
1263 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1264 the return statements above will have triggered. */
1265 assert (new_type != SVt_NULL);
1268 assert(old_type == SVt_NULL);
1269 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv));
1273 assert(old_type == SVt_NULL);
1274 SvANY(sv) = new_XNV();
1279 assert(new_type_details->body_size);
1282 assert(new_type_details->arena);
1283 assert(new_type_details->arena_size);
1284 /* This points to the start of the allocated area. */
1285 new_body_inline(new_body, new_type);
1286 Zero(new_body, new_type_details->body_size, char);
1287 new_body = ((char *)new_body) - new_type_details->offset;
1289 /* We always allocated the full length item with PURIFY. To do this
1290 we fake things so that arena is false for all 16 types.. */
1291 new_body = new_NOARENAZ(new_type_details);
1293 SvANY(sv) = new_body;
1294 if (new_type == SVt_PVAV) {
1298 if (old_type_details->body_size) {
1301 /* It will have been zeroed when the new body was allocated.
1302 Lets not write to it, in case it confuses a write-back
1308 #ifndef NODEFAULT_SHAREKEYS
1309 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1311 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1312 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
1315 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1316 The target created by newSVrv also is, and it can have magic.
1317 However, it never has SvPVX set.
1319 if (old_type == SVt_IV) {
1321 } else if (old_type >= SVt_PV) {
1322 assert(SvPVX_const(sv) == 0);
1325 if (old_type >= SVt_PVMG) {
1326 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1327 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1329 sv->sv_u.svu_array = NULL; /* or svu_hash */
1334 /* XXX Is this still needed? Was it ever needed? Surely as there is
1335 no route from NV to PVIV, NOK can never be true */
1336 assert(!SvNOKp(sv));
1348 assert(new_type_details->body_size);
1349 /* We always allocated the full length item with PURIFY. To do this
1350 we fake things so that arena is false for all 16 types.. */
1351 if(new_type_details->arena) {
1352 /* This points to the start of the allocated area. */
1353 new_body_inline(new_body, new_type);
1354 Zero(new_body, new_type_details->body_size, char);
1355 new_body = ((char *)new_body) - new_type_details->offset;
1357 new_body = new_NOARENAZ(new_type_details);
1359 SvANY(sv) = new_body;
1361 if (old_type_details->copy) {
1362 /* There is now the potential for an upgrade from something without
1363 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1364 int offset = old_type_details->offset;
1365 int length = old_type_details->copy;
1367 if (new_type_details->offset > old_type_details->offset) {
1368 const int difference
1369 = new_type_details->offset - old_type_details->offset;
1370 offset += difference;
1371 length -= difference;
1373 assert (length >= 0);
1375 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1379 #ifndef NV_ZERO_IS_ALLBITS_ZERO
1380 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
1381 * correct 0.0 for us. Otherwise, if the old body didn't have an
1382 * NV slot, but the new one does, then we need to initialise the
1383 * freshly created NV slot with whatever the correct bit pattern is
1385 if (old_type_details->zero_nv && !new_type_details->zero_nv
1386 && !isGV_with_GP(sv))
1390 if (UNLIKELY(new_type == SVt_PVIO)) {
1391 IO * const io = MUTABLE_IO(sv);
1392 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
1395 /* Clear the stashcache because a new IO could overrule a package
1397 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
1398 hv_clear(PL_stashcache);
1400 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1401 IoPAGE_LEN(sv) = 60;
1403 if (UNLIKELY(new_type == SVt_REGEXP))
1404 sv->sv_u.svu_rx = (regexp *)new_body;
1405 else if (old_type < SVt_PV) {
1406 /* referant will be NULL unless the old type was SVt_IV emulating
1408 sv->sv_u.svu_rv = referant;
1412 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1413 (unsigned long)new_type);
1416 if (old_type > SVt_IV) {
1420 /* Note that there is an assumption that all bodies of types that
1421 can be upgraded came from arenas. Only the more complex non-
1422 upgradable types are allowed to be directly malloc()ed. */
1423 assert(old_type_details->arena);
1424 del_body((void*)((char*)old_body + old_type_details->offset),
1425 &PL_body_roots[old_type]);
1431 =for apidoc sv_backoff
1433 Remove any string offset. You should normally use the C<SvOOK_off> macro
1440 Perl_sv_backoff(pTHX_ SV *const sv)
1443 const char * const s = SvPVX_const(sv);
1445 PERL_ARGS_ASSERT_SV_BACKOFF;
1446 PERL_UNUSED_CONTEXT;
1449 assert(SvTYPE(sv) != SVt_PVHV);
1450 assert(SvTYPE(sv) != SVt_PVAV);
1452 SvOOK_offset(sv, delta);
1454 SvLEN_set(sv, SvLEN(sv) + delta);
1455 SvPV_set(sv, SvPVX(sv) - delta);
1456 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1457 SvFLAGS(sv) &= ~SVf_OOK;
1464 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1465 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1466 Use the C<SvGROW> wrapper instead.
1472 Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
1476 PERL_ARGS_ASSERT_SV_GROW;
1478 #ifdef HAS_64K_LIMIT
1479 if (newlen >= 0x10000) {
1480 PerlIO_printf(Perl_debug_log,
1481 "Allocation too large: %"UVxf"\n", (UV)newlen);
1484 #endif /* HAS_64K_LIMIT */
1487 if (SvTYPE(sv) < SVt_PV) {
1488 sv_upgrade(sv, SVt_PV);
1489 s = SvPVX_mutable(sv);
1491 else if (SvOOK(sv)) { /* pv is offset? */
1493 s = SvPVX_mutable(sv);
1494 if (newlen > SvLEN(sv))
1495 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1496 #ifdef HAS_64K_LIMIT
1497 if (newlen >= 0x10000)
1503 if (SvIsCOW(sv)) sv_force_normal(sv);
1504 s = SvPVX_mutable(sv);
1507 #ifdef PERL_NEW_COPY_ON_WRITE
1508 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
1509 * to store the COW count. So in general, allocate one more byte than
1510 * asked for, to make it likely this byte is always spare: and thus
1511 * make more strings COW-able.
1512 * If the new size is a big power of two, don't bother: we assume the
1513 * caller wanted a nice 2^N sized block and will be annoyed at getting
1519 if (newlen > SvLEN(sv)) { /* need more room? */
1520 STRLEN minlen = SvCUR(sv);
1521 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1522 if (newlen < minlen)
1524 #ifndef Perl_safesysmalloc_size
1525 newlen = PERL_STRLEN_ROUNDUP(newlen);
1527 if (SvLEN(sv) && s) {
1528 s = (char*)saferealloc(s, newlen);
1531 s = (char*)safemalloc(newlen);
1532 if (SvPVX_const(sv) && SvCUR(sv)) {
1533 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1537 #ifdef Perl_safesysmalloc_size
1538 /* Do this here, do it once, do it right, and then we will never get
1539 called back into sv_grow() unless there really is some growing
1541 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1543 SvLEN_set(sv, newlen);
1550 =for apidoc sv_setiv
1552 Copies an integer into the given SV, upgrading first if necessary.
1553 Does not handle 'set' magic. See also C<sv_setiv_mg>.
1559 Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
1563 PERL_ARGS_ASSERT_SV_SETIV;
1565 SV_CHECK_THINKFIRST_COW_DROP(sv);
1566 switch (SvTYPE(sv)) {
1569 sv_upgrade(sv, SVt_IV);
1572 sv_upgrade(sv, SVt_PVIV);
1576 if (!isGV_with_GP(sv))
1583 /* diag_listed_as: Can't coerce %s to %s in %s */
1584 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1588 (void)SvIOK_only(sv); /* validate number */
1594 =for apidoc sv_setiv_mg
1596 Like C<sv_setiv>, but also handles 'set' magic.
1602 Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
1604 PERL_ARGS_ASSERT_SV_SETIV_MG;
1611 =for apidoc sv_setuv
1613 Copies an unsigned integer into the given SV, upgrading first if necessary.
1614 Does not handle 'set' magic. See also C<sv_setuv_mg>.
1620 Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
1622 PERL_ARGS_ASSERT_SV_SETUV;
1624 /* With the if statement to ensure that integers are stored as IVs whenever
1626 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1629 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1631 If you wish to remove the following if statement, so that this routine
1632 (and its callers) always return UVs, please benchmark to see what the
1633 effect is. Modern CPUs may be different. Or may not :-)
1635 if (u <= (UV)IV_MAX) {
1636 sv_setiv(sv, (IV)u);
1645 =for apidoc sv_setuv_mg
1647 Like C<sv_setuv>, but also handles 'set' magic.
1653 Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
1655 PERL_ARGS_ASSERT_SV_SETUV_MG;
1662 =for apidoc sv_setnv
1664 Copies a double into the given SV, upgrading first if necessary.
1665 Does not handle 'set' magic. See also C<sv_setnv_mg>.
1671 Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
1675 PERL_ARGS_ASSERT_SV_SETNV;
1677 SV_CHECK_THINKFIRST_COW_DROP(sv);
1678 switch (SvTYPE(sv)) {
1681 sv_upgrade(sv, SVt_NV);
1685 sv_upgrade(sv, SVt_PVNV);
1689 if (!isGV_with_GP(sv))
1696 /* diag_listed_as: Can't coerce %s to %s in %s */
1697 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1702 (void)SvNOK_only(sv); /* validate number */
1707 =for apidoc sv_setnv_mg
1709 Like C<sv_setnv>, but also handles 'set' magic.
1715 Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
1717 PERL_ARGS_ASSERT_SV_SETNV_MG;
1723 /* Print an "isn't numeric" warning, using a cleaned-up,
1724 * printable version of the offending string
1728 S_not_a_number(pTHX_ SV *const sv)
1735 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1738 dsv = newSVpvs_flags("", SVs_TEMP);
1739 pv = sv_uni_display(dsv, sv, 10, UNI_DISPLAY_ISPRINT);
1742 const char * const limit = tmpbuf + sizeof(tmpbuf) - 8;
1743 /* each *s can expand to 4 chars + "...\0",
1744 i.e. need room for 8 chars */
1746 const char *s = SvPVX_const(sv);
1747 const char * const end = s + SvCUR(sv);
1748 for ( ; s < end && d < limit; s++ ) {
1750 if (ch & 128 && !isPRINT_LC(ch)) {
1759 else if (ch == '\r') {
1763 else if (ch == '\f') {
1767 else if (ch == '\\') {
1771 else if (ch == '\0') {
1775 else if (isPRINT_LC(ch))
1792 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1793 /* diag_listed_as: Argument "%s" isn't numeric%s */
1794 "Argument \"%s\" isn't numeric in %s", pv,
1797 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1798 /* diag_listed_as: Argument "%s" isn't numeric%s */
1799 "Argument \"%s\" isn't numeric", pv);
1803 =for apidoc looks_like_number
1805 Test if the content of an SV looks like a number (or is a number).
1806 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1807 non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1814 Perl_looks_like_number(pTHX_ SV *const sv)
1819 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1821 if (SvPOK(sv) || SvPOKp(sv)) {
1822 sbegin = SvPV_nomg_const(sv, len);
1825 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1826 return grok_number(sbegin, len, NULL);
1830 S_glob_2number(pTHX_ GV * const gv)
1832 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1834 /* We know that all GVs stringify to something that is not-a-number,
1835 so no need to test that. */
1836 if (ckWARN(WARN_NUMERIC))
1838 SV *const buffer = sv_newmortal();
1839 gv_efullname3(buffer, gv, "*");
1840 not_a_number(buffer);
1842 /* We just want something true to return, so that S_sv_2iuv_common
1843 can tail call us and return true. */
1847 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1848 until proven guilty, assume that things are not that bad... */
1853 As 64 bit platforms often have an NV that doesn't preserve all bits of
1854 an IV (an assumption perl has been based on to date) it becomes necessary
1855 to remove the assumption that the NV always carries enough precision to
1856 recreate the IV whenever needed, and that the NV is the canonical form.
1857 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1858 precision as a side effect of conversion (which would lead to insanity
1859 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1860 1) to distinguish between IV/UV/NV slots that have cached a valid
1861 conversion where precision was lost and IV/UV/NV slots that have a
1862 valid conversion which has lost no precision
1863 2) to ensure that if a numeric conversion to one form is requested that
1864 would lose precision, the precise conversion (or differently
1865 imprecise conversion) is also performed and cached, to prevent
1866 requests for different numeric formats on the same SV causing
1867 lossy conversion chains. (lossless conversion chains are perfectly
1872 SvIOKp is true if the IV slot contains a valid value
1873 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1874 SvNOKp is true if the NV slot contains a valid value
1875 SvNOK is true only if the NV value is accurate
1878 while converting from PV to NV, check to see if converting that NV to an
1879 IV(or UV) would lose accuracy over a direct conversion from PV to
1880 IV(or UV). If it would, cache both conversions, return NV, but mark
1881 SV as IOK NOKp (ie not NOK).
1883 While converting from PV to IV, check to see if converting that IV to an
1884 NV would lose accuracy over a direct conversion from PV to NV. If it
1885 would, cache both conversions, flag similarly.
1887 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
1888 correctly because if IV & NV were set NV *always* overruled.
1889 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
1890 changes - now IV and NV together means that the two are interchangeable:
1891 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
1893 The benefit of this is that operations such as pp_add know that if
1894 SvIOK is true for both left and right operands, then integer addition
1895 can be used instead of floating point (for cases where the result won't
1896 overflow). Before, floating point was always used, which could lead to
1897 loss of precision compared with integer addition.
1899 * making IV and NV equal status should make maths accurate on 64 bit
1901 * may speed up maths somewhat if pp_add and friends start to use
1902 integers when possible instead of fp. (Hopefully the overhead in
1903 looking for SvIOK and checking for overflow will not outweigh the
1904 fp to integer speedup)
1905 * will slow down integer operations (callers of SvIV) on "inaccurate"
1906 values, as the change from SvIOK to SvIOKp will cause a call into
1907 sv_2iv each time rather than a macro access direct to the IV slot
1908 * should speed up number->string conversion on integers as IV is
1909 favoured when IV and NV are equally accurate
1911 ####################################################################
1912 You had better be using SvIOK_notUV if you want an IV for arithmetic:
1913 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
1914 On the other hand, SvUOK is true iff UV.
1915 ####################################################################
1917 Your mileage will vary depending your CPU's relative fp to integer
1921 #ifndef NV_PRESERVES_UV
1922 # define IS_NUMBER_UNDERFLOW_IV 1
1923 # define IS_NUMBER_UNDERFLOW_UV 2
1924 # define IS_NUMBER_IV_AND_UV 2
1925 # define IS_NUMBER_OVERFLOW_IV 4
1926 # define IS_NUMBER_OVERFLOW_UV 5
1928 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
1930 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
1932 S_sv_2iuv_non_preserve(pTHX_ SV *const sv
1940 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
1942 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));
1943 if (SvNVX(sv) < (NV)IV_MIN) {
1944 (void)SvIOKp_on(sv);
1946 SvIV_set(sv, IV_MIN);
1947 return IS_NUMBER_UNDERFLOW_IV;
1949 if (SvNVX(sv) > (NV)UV_MAX) {
1950 (void)SvIOKp_on(sv);
1953 SvUV_set(sv, UV_MAX);
1954 return IS_NUMBER_OVERFLOW_UV;
1956 (void)SvIOKp_on(sv);
1958 /* Can't use strtol etc to convert this string. (See truth table in
1960 if (SvNVX(sv) <= (UV)IV_MAX) {
1961 SvIV_set(sv, I_V(SvNVX(sv)));
1962 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
1963 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
1965 /* Integer is imprecise. NOK, IOKp */
1967 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
1970 SvUV_set(sv, U_V(SvNVX(sv)));
1971 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
1972 if (SvUVX(sv) == UV_MAX) {
1973 /* As we know that NVs don't preserve UVs, UV_MAX cannot
1974 possibly be preserved by NV. Hence, it must be overflow.
1976 return IS_NUMBER_OVERFLOW_UV;
1978 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
1980 /* Integer is imprecise. NOK, IOKp */
1982 return IS_NUMBER_OVERFLOW_IV;
1984 #endif /* !NV_PRESERVES_UV*/
1987 S_sv_2iuv_common(pTHX_ SV *const sv)
1991 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
1994 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
1995 * without also getting a cached IV/UV from it at the same time
1996 * (ie PV->NV conversion should detect loss of accuracy and cache
1997 * IV or UV at same time to avoid this. */
1998 /* IV-over-UV optimisation - choose to cache IV if possible */
2000 if (SvTYPE(sv) == SVt_NV)
2001 sv_upgrade(sv, SVt_PVNV);
2003 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2004 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2005 certainly cast into the IV range at IV_MAX, whereas the correct
2006 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2008 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2009 if (Perl_isnan(SvNVX(sv))) {
2015 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2016 SvIV_set(sv, I_V(SvNVX(sv)));
2017 if (SvNVX(sv) == (NV) SvIVX(sv)
2018 #ifndef NV_PRESERVES_UV
2019 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2020 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2021 /* Don't flag it as "accurately an integer" if the number
2022 came from a (by definition imprecise) NV operation, and
2023 we're outside the range of NV integer precision */
2027 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2029 /* scalar has trailing garbage, eg "42a" */
2031 DEBUG_c(PerlIO_printf(Perl_debug_log,
2032 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
2038 /* IV not precise. No need to convert from PV, as NV
2039 conversion would already have cached IV if it detected
2040 that PV->IV would be better than PV->NV->IV
2041 flags already correct - don't set public IOK. */
2042 DEBUG_c(PerlIO_printf(Perl_debug_log,
2043 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
2048 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2049 but the cast (NV)IV_MIN rounds to a the value less (more
2050 negative) than IV_MIN which happens to be equal to SvNVX ??
2051 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2052 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2053 (NV)UVX == NVX are both true, but the values differ. :-(
2054 Hopefully for 2s complement IV_MIN is something like
2055 0x8000000000000000 which will be exact. NWC */
2058 SvUV_set(sv, U_V(SvNVX(sv)));
2060 (SvNVX(sv) == (NV) SvUVX(sv))
2061 #ifndef NV_PRESERVES_UV
2062 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2063 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2064 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2065 /* Don't flag it as "accurately an integer" if the number
2066 came from a (by definition imprecise) NV operation, and
2067 we're outside the range of NV integer precision */
2073 DEBUG_c(PerlIO_printf(Perl_debug_log,
2074 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2080 else if (SvPOKp(sv)) {
2082 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2083 /* We want to avoid a possible problem when we cache an IV/ a UV which
2084 may be later translated to an NV, and the resulting NV is not
2085 the same as the direct translation of the initial string
2086 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2087 be careful to ensure that the value with the .456 is around if the
2088 NV value is requested in the future).
2090 This means that if we cache such an IV/a UV, we need to cache the
2091 NV as well. Moreover, we trade speed for space, and do not
2092 cache the NV if we are sure it's not needed.
2095 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2096 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2097 == IS_NUMBER_IN_UV) {
2098 /* It's definitely an integer, only upgrade to PVIV */
2099 if (SvTYPE(sv) < SVt_PVIV)
2100 sv_upgrade(sv, SVt_PVIV);
2102 } else if (SvTYPE(sv) < SVt_PVNV)
2103 sv_upgrade(sv, SVt_PVNV);
2105 /* If NVs preserve UVs then we only use the UV value if we know that
2106 we aren't going to call atof() below. If NVs don't preserve UVs
2107 then the value returned may have more precision than atof() will
2108 return, even though value isn't perfectly accurate. */
2109 if ((numtype & (IS_NUMBER_IN_UV
2110 #ifdef NV_PRESERVES_UV
2113 )) == IS_NUMBER_IN_UV) {
2114 /* This won't turn off the public IOK flag if it was set above */
2115 (void)SvIOKp_on(sv);
2117 if (!(numtype & IS_NUMBER_NEG)) {
2119 if (value <= (UV)IV_MAX) {
2120 SvIV_set(sv, (IV)value);
2122 /* it didn't overflow, and it was positive. */
2123 SvUV_set(sv, value);
2127 /* 2s complement assumption */
2128 if (value <= (UV)IV_MIN) {
2129 SvIV_set(sv, -(IV)value);
2131 /* Too negative for an IV. This is a double upgrade, but
2132 I'm assuming it will be rare. */
2133 if (SvTYPE(sv) < SVt_PVNV)
2134 sv_upgrade(sv, SVt_PVNV);
2138 SvNV_set(sv, -(NV)value);
2139 SvIV_set(sv, IV_MIN);
2143 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2144 will be in the previous block to set the IV slot, and the next
2145 block to set the NV slot. So no else here. */
2147 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2148 != IS_NUMBER_IN_UV) {
2149 /* It wasn't an (integer that doesn't overflow the UV). */
2150 SvNV_set(sv, Atof(SvPVX_const(sv)));
2152 if (! numtype && ckWARN(WARN_NUMERIC))
2155 #if defined(USE_LONG_DOUBLE)
2156 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" PERL_PRIgldbl ")\n",
2157 PTR2UV(sv), SvNVX(sv)));
2159 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"NVgf")\n",
2160 PTR2UV(sv), SvNVX(sv)));
2163 #ifdef NV_PRESERVES_UV
2164 (void)SvIOKp_on(sv);
2166 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2167 SvIV_set(sv, I_V(SvNVX(sv)));
2168 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2171 NOOP; /* Integer is imprecise. NOK, IOKp */
2173 /* UV will not work better than IV */
2175 if (SvNVX(sv) > (NV)UV_MAX) {
2177 /* Integer is inaccurate. NOK, IOKp, is UV */
2178 SvUV_set(sv, UV_MAX);
2180 SvUV_set(sv, U_V(SvNVX(sv)));
2181 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2182 NV preservse UV so can do correct comparison. */
2183 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2186 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2191 #else /* NV_PRESERVES_UV */
2192 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2193 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2194 /* The IV/UV slot will have been set from value returned by
2195 grok_number above. The NV slot has just been set using
2198 assert (SvIOKp(sv));
2200 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2201 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2202 /* Small enough to preserve all bits. */
2203 (void)SvIOKp_on(sv);
2205 SvIV_set(sv, I_V(SvNVX(sv)));
2206 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2208 /* Assumption: first non-preserved integer is < IV_MAX,
2209 this NV is in the preserved range, therefore: */
2210 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2212 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);
2216 0 0 already failed to read UV.
2217 0 1 already failed to read UV.
2218 1 0 you won't get here in this case. IV/UV
2219 slot set, public IOK, Atof() unneeded.
2220 1 1 already read UV.
2221 so there's no point in sv_2iuv_non_preserve() attempting
2222 to use atol, strtol, strtoul etc. */
2224 sv_2iuv_non_preserve (sv, numtype);
2226 sv_2iuv_non_preserve (sv);
2230 #endif /* NV_PRESERVES_UV */
2231 /* It might be more code efficient to go through the entire logic above
2232 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2233 gets complex and potentially buggy, so more programmer efficient
2234 to do it this way, by turning off the public flags: */
2236 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2240 if (isGV_with_GP(sv))
2241 return glob_2number(MUTABLE_GV(sv));
2243 if (!SvPADTMP(sv)) {
2244 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2247 if (SvTYPE(sv) < SVt_IV)
2248 /* Typically the caller expects that sv_any is not NULL now. */
2249 sv_upgrade(sv, SVt_IV);
2250 /* Return 0 from the caller. */
2257 =for apidoc sv_2iv_flags
2259 Return the integer value of an SV, doing any necessary string
2260 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2261 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2267 Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
2274 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2280 if (flags & SV_SKIP_OVERLOAD)
2282 tmpstr = AMG_CALLunary(sv, numer_amg);
2283 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2284 return SvIV(tmpstr);
2287 return PTR2IV(SvRV(sv));
2290 if (SvVALID(sv) || isREGEXP(sv)) {
2291 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2292 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2293 In practice they are extremely unlikely to actually get anywhere
2294 accessible by user Perl code - the only way that I'm aware of is when
2295 a constant subroutine which is used as the second argument to index.
2297 Regexps have no SvIVX and SvNVX fields.
2299 assert(isREGEXP(sv) || SvPOKp(sv));
2302 const char * const ptr =
2303 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2305 = grok_number(ptr, SvCUR(sv), &value);
2307 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2308 == IS_NUMBER_IN_UV) {
2309 /* It's definitely an integer */
2310 if (numtype & IS_NUMBER_NEG) {
2311 if (value < (UV)IV_MIN)
2314 if (value < (UV)IV_MAX)
2319 if (ckWARN(WARN_NUMERIC))
2322 return I_V(Atof(ptr));
2326 if (SvTHINKFIRST(sv)) {
2327 #ifdef PERL_OLD_COPY_ON_WRITE
2329 sv_force_normal_flags(sv, 0);
2332 if (SvREADONLY(sv) && !SvOK(sv)) {
2333 if (ckWARN(WARN_UNINITIALIZED))
2340 if (S_sv_2iuv_common(aTHX_ sv))
2344 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2345 PTR2UV(sv),SvIVX(sv)));
2346 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2350 =for apidoc sv_2uv_flags
2352 Return the unsigned integer value of an SV, doing any necessary string
2353 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2354 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2360 Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
2367 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2373 if (flags & SV_SKIP_OVERLOAD)
2375 tmpstr = AMG_CALLunary(sv, numer_amg);
2376 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2377 return SvUV(tmpstr);
2380 return PTR2UV(SvRV(sv));
2383 if (SvVALID(sv) || isREGEXP(sv)) {
2384 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2385 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2386 Regexps have no SvIVX and SvNVX fields. */
2387 assert(isREGEXP(sv) || SvPOKp(sv));
2390 const char * const ptr =
2391 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2393 = grok_number(ptr, SvCUR(sv), &value);
2395 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2396 == IS_NUMBER_IN_UV) {
2397 /* It's definitely an integer */
2398 if (!(numtype & IS_NUMBER_NEG))
2402 if (ckWARN(WARN_NUMERIC))
2405 return U_V(Atof(ptr));
2409 if (SvTHINKFIRST(sv)) {
2410 #ifdef PERL_OLD_COPY_ON_WRITE
2412 sv_force_normal_flags(sv, 0);
2415 if (SvREADONLY(sv) && !SvOK(sv)) {
2416 if (ckWARN(WARN_UNINITIALIZED))
2423 if (S_sv_2iuv_common(aTHX_ sv))
2427 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2428 PTR2UV(sv),SvUVX(sv)));
2429 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2433 =for apidoc sv_2nv_flags
2435 Return the num value of an SV, doing any necessary string or integer
2436 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2437 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2443 Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
2448 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2449 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2450 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2451 Regexps have no SvIVX and SvNVX fields. */
2453 if (flags & SV_GMAGIC)
2457 if (SvPOKp(sv) && !SvIOKp(sv)) {
2458 ptr = SvPVX_const(sv);
2460 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2461 !grok_number(ptr, SvCUR(sv), NULL))
2467 return (NV)SvUVX(sv);
2469 return (NV)SvIVX(sv);
2475 ptr = RX_WRAPPED((REGEXP *)sv);
2478 assert(SvTYPE(sv) >= SVt_PVMG);
2479 /* This falls through to the report_uninit near the end of the
2481 } else if (SvTHINKFIRST(sv)) {
2486 if (flags & SV_SKIP_OVERLOAD)
2488 tmpstr = AMG_CALLunary(sv, numer_amg);
2489 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2490 return SvNV(tmpstr);
2493 return PTR2NV(SvRV(sv));
2495 #ifdef PERL_OLD_COPY_ON_WRITE
2497 sv_force_normal_flags(sv, 0);
2500 if (SvREADONLY(sv) && !SvOK(sv)) {
2501 if (ckWARN(WARN_UNINITIALIZED))
2506 if (SvTYPE(sv) < SVt_NV) {
2507 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2508 sv_upgrade(sv, SVt_NV);
2509 #ifdef USE_LONG_DOUBLE
2511 STORE_NUMERIC_LOCAL_SET_STANDARD();
2512 PerlIO_printf(Perl_debug_log,
2513 "0x%"UVxf" num(%" PERL_PRIgldbl ")\n",
2514 PTR2UV(sv), SvNVX(sv));
2515 RESTORE_NUMERIC_LOCAL();
2519 STORE_NUMERIC_LOCAL_SET_STANDARD();
2520 PerlIO_printf(Perl_debug_log, "0x%"UVxf" num(%"NVgf")\n",
2521 PTR2UV(sv), SvNVX(sv));
2522 RESTORE_NUMERIC_LOCAL();
2526 else if (SvTYPE(sv) < SVt_PVNV)
2527 sv_upgrade(sv, SVt_PVNV);
2532 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2533 #ifdef NV_PRESERVES_UV
2539 /* Only set the public NV OK flag if this NV preserves the IV */
2540 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2542 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2543 : (SvIVX(sv) == I_V(SvNVX(sv))))
2549 else if (SvPOKp(sv)) {
2551 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2552 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2554 #ifdef NV_PRESERVES_UV
2555 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2556 == IS_NUMBER_IN_UV) {
2557 /* It's definitely an integer */
2558 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2560 SvNV_set(sv, Atof(SvPVX_const(sv)));
2566 SvNV_set(sv, Atof(SvPVX_const(sv)));
2567 /* Only set the public NV OK flag if this NV preserves the value in
2568 the PV at least as well as an IV/UV would.
2569 Not sure how to do this 100% reliably. */
2570 /* if that shift count is out of range then Configure's test is
2571 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2573 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2574 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2575 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2576 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2577 /* Can't use strtol etc to convert this string, so don't try.
2578 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2581 /* value has been set. It may not be precise. */
2582 if ((numtype & IS_NUMBER_NEG) && (value > (UV)IV_MIN)) {
2583 /* 2s complement assumption for (UV)IV_MIN */
2584 SvNOK_on(sv); /* Integer is too negative. */
2589 if (numtype & IS_NUMBER_NEG) {
2590 SvIV_set(sv, -(IV)value);
2591 } else if (value <= (UV)IV_MAX) {
2592 SvIV_set(sv, (IV)value);
2594 SvUV_set(sv, value);
2598 if (numtype & IS_NUMBER_NOT_INT) {
2599 /* I believe that even if the original PV had decimals,
2600 they are lost beyond the limit of the FP precision.
2601 However, neither is canonical, so both only get p
2602 flags. NWC, 2000/11/25 */
2603 /* Both already have p flags, so do nothing */
2605 const NV nv = SvNVX(sv);
2606 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2607 if (SvIVX(sv) == I_V(nv)) {
2610 /* It had no "." so it must be integer. */
2614 /* between IV_MAX and NV(UV_MAX).
2615 Could be slightly > UV_MAX */
2617 if (numtype & IS_NUMBER_NOT_INT) {
2618 /* UV and NV both imprecise. */
2620 const UV nv_as_uv = U_V(nv);
2622 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2631 /* It might be more code efficient to go through the entire logic above
2632 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2633 gets complex and potentially buggy, so more programmer efficient
2634 to do it this way, by turning off the public flags: */
2636 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2637 #endif /* NV_PRESERVES_UV */
2640 if (isGV_with_GP(sv)) {
2641 glob_2number(MUTABLE_GV(sv));
2645 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2647 assert (SvTYPE(sv) >= SVt_NV);
2648 /* Typically the caller expects that sv_any is not NULL now. */
2649 /* XXX Ilya implies that this is a bug in callers that assume this
2650 and ideally should be fixed. */
2653 #if defined(USE_LONG_DOUBLE)
2655 STORE_NUMERIC_LOCAL_SET_STANDARD();
2656 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" PERL_PRIgldbl ")\n",
2657 PTR2UV(sv), SvNVX(sv));
2658 RESTORE_NUMERIC_LOCAL();
2662 STORE_NUMERIC_LOCAL_SET_STANDARD();
2663 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 1nv(%"NVgf")\n",
2664 PTR2UV(sv), SvNVX(sv));
2665 RESTORE_NUMERIC_LOCAL();
2674 Return an SV with the numeric value of the source SV, doing any necessary
2675 reference or overload conversion. You must use the C<SvNUM(sv)> macro to
2676 access this function.
2682 Perl_sv_2num(pTHX_ SV *const sv)
2684 PERL_ARGS_ASSERT_SV_2NUM;
2689 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2690 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2691 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2692 return sv_2num(tmpsv);
2694 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2697 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2698 * UV as a string towards the end of buf, and return pointers to start and
2701 * We assume that buf is at least TYPE_CHARS(UV) long.
2705 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2707 char *ptr = buf + TYPE_CHARS(UV);
2708 char * const ebuf = ptr;
2711 PERL_ARGS_ASSERT_UIV_2BUF;
2723 *--ptr = '0' + (char)(uv % 10);
2732 =for apidoc sv_2pv_flags
2734 Returns a pointer to the string value of an SV, and sets *lp to its length.
2735 If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2736 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2737 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2743 Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2753 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2758 if (flags & SV_SKIP_OVERLOAD)
2760 tmpstr = AMG_CALLunary(sv, string_amg);
2761 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2762 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2764 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2768 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2769 if (flags & SV_CONST_RETURN) {
2770 pv = (char *) SvPVX_const(tmpstr);
2772 pv = (flags & SV_MUTABLE_RETURN)
2773 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2776 *lp = SvCUR(tmpstr);
2778 pv = sv_2pv_flags(tmpstr, lp, flags);
2791 SV *const referent = SvRV(sv);
2795 retval = buffer = savepvn("NULLREF", len);
2796 } else if (SvTYPE(referent) == SVt_REGEXP &&
2797 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
2798 amagic_is_enabled(string_amg))) {
2799 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2803 /* If the regex is UTF-8 we want the containing scalar to
2804 have an UTF-8 flag too */
2811 *lp = RX_WRAPLEN(re);
2813 return RX_WRAPPED(re);
2815 const char *const typestr = sv_reftype(referent, 0);
2816 const STRLEN typelen = strlen(typestr);
2817 UV addr = PTR2UV(referent);
2818 const char *stashname = NULL;
2819 STRLEN stashnamelen = 0; /* hush, gcc */
2820 const char *buffer_end;
2822 if (SvOBJECT(referent)) {
2823 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
2826 stashname = HEK_KEY(name);
2827 stashnamelen = HEK_LEN(name);
2829 if (HEK_UTF8(name)) {
2835 stashname = "__ANON__";
2838 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
2839 + 2 * sizeof(UV) + 2 /* )\0 */;
2841 len = typelen + 3 /* (0x */
2842 + 2 * sizeof(UV) + 2 /* )\0 */;
2845 Newx(buffer, len, char);
2846 buffer_end = retval = buffer + len;
2848 /* Working backwards */
2852 *--retval = PL_hexdigit[addr & 15];
2853 } while (addr >>= 4);
2859 memcpy(retval, typestr, typelen);
2863 retval -= stashnamelen;
2864 memcpy(retval, stashname, stashnamelen);
2866 /* retval may not necessarily have reached the start of the
2868 assert (retval >= buffer);
2870 len = buffer_end - retval - 1; /* -1 for that \0 */
2882 if (flags & SV_MUTABLE_RETURN)
2883 return SvPVX_mutable(sv);
2884 if (flags & SV_CONST_RETURN)
2885 return (char *)SvPVX_const(sv);
2890 /* I'm assuming that if both IV and NV are equally valid then
2891 converting the IV is going to be more efficient */
2892 const U32 isUIOK = SvIsUV(sv);
2893 char buf[TYPE_CHARS(UV)];
2897 if (SvTYPE(sv) < SVt_PVIV)
2898 sv_upgrade(sv, SVt_PVIV);
2899 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
2901 /* inlined from sv_setpvn */
2902 s = SvGROW_mutable(sv, len + 1);
2903 Move(ptr, s, len, char);
2908 else if (SvNOK(sv)) {
2909 if (SvTYPE(sv) < SVt_PVNV)
2910 sv_upgrade(sv, SVt_PVNV);
2911 if (SvNVX(sv) == 0.0) {
2912 s = SvGROW_mutable(sv, 2);
2917 /* The +20 is pure guesswork. Configure test needed. --jhi */
2918 s = SvGROW_mutable(sv, NV_DIG + 20);
2919 /* some Xenix systems wipe out errno here */
2921 #ifndef USE_LOCALE_NUMERIC
2922 Gconvert(SvNVX(sv), NV_DIG, 0, s);
2925 /* Gconvert always uses the current locale. That's the right thing
2926 * to do if we're supposed to be using locales. But otherwise, we
2927 * want the result to be based on the C locale, so we need to
2928 * change to the C locale during the Gconvert and then change back.
2929 * But if we're already in the C locale (PL_numeric_standard is
2930 * TRUE in that case), no need to do any changing */
2931 if (PL_numeric_standard || IN_LOCALE_RUNTIME) {
2932 Gconvert(SvNVX(sv), NV_DIG, 0, s);
2935 char *loc = savepv(setlocale(LC_NUMERIC, NULL));
2936 setlocale(LC_NUMERIC, "C");
2937 Gconvert(SvNVX(sv), NV_DIG, 0, s);
2938 setlocale(LC_NUMERIC, loc);
2942 /* We don't call SvPOK_on(), because it may come to pass that the
2943 * locale changes so that the stringification we just did is no
2944 * longer correct. We will have to re-stringify every time it is
2955 else if (isGV_with_GP(sv)) {
2956 GV *const gv = MUTABLE_GV(sv);
2957 SV *const buffer = sv_newmortal();
2959 gv_efullname3(buffer, gv, "*");
2961 assert(SvPOK(buffer));
2965 *lp = SvCUR(buffer);
2966 return SvPVX(buffer);
2968 else if (isREGEXP(sv)) {
2969 if (lp) *lp = RX_WRAPLEN((REGEXP *)sv);
2970 return RX_WRAPPED((REGEXP *)sv);
2975 if (flags & SV_UNDEF_RETURNS_NULL)
2977 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2979 /* Typically the caller expects that sv_any is not NULL now. */
2980 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
2981 sv_upgrade(sv, SVt_PV);
2986 const STRLEN len = s - SvPVX_const(sv);
2991 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
2992 PTR2UV(sv),SvPVX_const(sv)));
2993 if (flags & SV_CONST_RETURN)
2994 return (char *)SvPVX_const(sv);
2995 if (flags & SV_MUTABLE_RETURN)
2996 return SvPVX_mutable(sv);
3001 =for apidoc sv_copypv
3003 Copies a stringified representation of the source SV into the
3004 destination SV. Automatically performs any necessary mg_get and
3005 coercion of numeric values into strings. Guaranteed to preserve
3006 UTF8 flag even from overloaded objects. Similar in nature to
3007 sv_2pv[_flags] but operates directly on an SV instead of just the
3008 string. Mostly uses sv_2pv_flags to do its work, except when that
3009 would lose the UTF-8'ness of the PV.
3011 =for apidoc sv_copypv_nomg
3013 Like sv_copypv, but doesn't invoke get magic first.
3015 =for apidoc sv_copypv_flags
3017 Implementation of sv_copypv and sv_copypv_nomg. Calls get magic iff flags
3024 Perl_sv_copypv(pTHX_ SV *const dsv, SV *const ssv)
3026 PERL_ARGS_ASSERT_SV_COPYPV;
3028 sv_copypv_flags(dsv, ssv, 0);
3032 Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
3037 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
3039 if ((flags & SV_GMAGIC) && SvGMAGICAL(ssv))
3041 s = SvPV_nomg_const(ssv,len);
3042 sv_setpvn(dsv,s,len);
3050 =for apidoc sv_2pvbyte
3052 Return a pointer to the byte-encoded representation of the SV, and set *lp
3053 to its length. May cause the SV to be downgraded from UTF-8 as a
3056 Usually accessed via the C<SvPVbyte> macro.
3062 Perl_sv_2pvbyte(pTHX_ SV *sv, STRLEN *const lp)
3064 PERL_ARGS_ASSERT_SV_2PVBYTE;
3066 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3067 || isGV_with_GP(sv) || SvROK(sv)) {
3068 SV *sv2 = sv_newmortal();
3072 else SvGETMAGIC(sv);
3073 sv_utf8_downgrade(sv,0);
3074 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3078 =for apidoc sv_2pvutf8
3080 Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3081 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3083 Usually accessed via the C<SvPVutf8> macro.
3089 Perl_sv_2pvutf8(pTHX_ SV *sv, STRLEN *const lp)
3091 PERL_ARGS_ASSERT_SV_2PVUTF8;
3093 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3094 || isGV_with_GP(sv) || SvROK(sv))
3095 sv = sv_mortalcopy(sv);
3098 sv_utf8_upgrade_nomg(sv);
3099 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3104 =for apidoc sv_2bool
3106 This macro is only used by sv_true() or its macro equivalent, and only if
3107 the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3108 It calls sv_2bool_flags with the SV_GMAGIC flag.
3110 =for apidoc sv_2bool_flags
3112 This function is only used by sv_true() and friends, and only if
3113 the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
3114 contain SV_GMAGIC, then it does an mg_get() first.
3121 Perl_sv_2bool_flags(pTHX_ SV *const sv, const I32 flags)
3125 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3127 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3133 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3134 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
3135 return cBOOL(SvTRUE(tmpsv));
3137 return SvRV(sv) != 0;
3139 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
3143 =for apidoc sv_utf8_upgrade
3145 Converts the PV of an SV to its UTF-8-encoded form.
3146 Forces the SV to string form if it is not already.
3147 Will C<mg_get> on C<sv> if appropriate.
3148 Always sets the SvUTF8 flag to avoid future validity checks even
3149 if the whole string is the same in UTF-8 as not.
3150 Returns the number of bytes in the converted string
3152 This is not a general purpose byte encoding to Unicode interface:
3153 use the Encode extension for that.
3155 =for apidoc sv_utf8_upgrade_nomg
3157 Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
3159 =for apidoc sv_utf8_upgrade_flags
3161 Converts the PV of an SV to its UTF-8-encoded form.
3162 Forces the SV to string form if it is not already.
3163 Always sets the SvUTF8 flag to avoid future validity checks even
3164 if all the bytes are invariant in UTF-8.
3165 If C<flags> has C<SV_GMAGIC> bit set,
3166 will C<mg_get> on C<sv> if appropriate, else not.
3167 Returns the number of bytes in the converted string
3168 C<sv_utf8_upgrade> and
3169 C<sv_utf8_upgrade_nomg> are implemented in terms of this function.
3171 This is not a general purpose byte encoding to Unicode interface:
3172 use the Encode extension for that.
3176 The grow version is currently not externally documented. It adds a parameter,
3177 extra, which is the number of unused bytes the string of 'sv' is guaranteed to
3178 have free after it upon return. This allows the caller to reserve extra space
3179 that it intends to fill, to avoid extra grows.
3181 Also externally undocumented for the moment is the flag SV_FORCE_UTF8_UPGRADE,
3182 which can be used to tell this function to not first check to see if there are
3183 any characters that are different in UTF-8 (variant characters) which would
3184 force it to allocate a new string to sv, but to assume there are. Typically
3185 this flag is used by a routine that has already parsed the string to find that
3186 there are such characters, and passes this information on so that the work
3187 doesn't have to be repeated.
3189 (One might think that the calling routine could pass in the position of the
3190 first such variant, so it wouldn't have to be found again. But that is not the
3191 case, because typically when the caller is likely to use this flag, it won't be
3192 calling this routine unless it finds something that won't fit into a byte.
3193 Otherwise it tries to not upgrade and just use bytes. But some things that
3194 do fit into a byte are variants in utf8, and the caller may not have been
3195 keeping track of these.)
3197 If the routine itself changes the string, it adds a trailing NUL. Such a NUL
3198 isn't guaranteed due to having other routines do the work in some input cases,
3199 or if the input is already flagged as being in utf8.
3201 The speed of this could perhaps be improved for many cases if someone wanted to
3202 write a fast function that counts the number of variant characters in a string,
3203 especially if it could return the position of the first one.
3208 Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
3212 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3214 if (sv == &PL_sv_undef)
3216 if (!SvPOK_nog(sv)) {
3218 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3219 (void) sv_2pv_flags(sv,&len, flags);
3221 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3225 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3230 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3235 sv_force_normal_flags(sv, 0);
3238 if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING)) {
3239 sv_recode_to_utf8(sv, PL_encoding);
3240 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3244 if (SvCUR(sv) == 0) {
3245 if (extra) SvGROW(sv, extra);
3246 } else { /* Assume Latin-1/EBCDIC */
3247 /* This function could be much more efficient if we
3248 * had a FLAG in SVs to signal if there are any variant
3249 * chars in the PV. Given that there isn't such a flag
3250 * make the loop as fast as possible (although there are certainly ways
3251 * to speed this up, eg. through vectorization) */
3252 U8 * s = (U8 *) SvPVX_const(sv);
3253 U8 * e = (U8 *) SvEND(sv);
3255 STRLEN two_byte_count = 0;
3257 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3259 /* See if really will need to convert to utf8. We mustn't rely on our
3260 * incoming SV being well formed and having a trailing '\0', as certain
3261 * code in pp_formline can send us partially built SVs. */
3265 if (NATIVE_IS_INVARIANT(ch)) continue;
3267 t--; /* t already incremented; re-point to first variant */
3272 /* utf8 conversion not needed because all are invariants. Mark as
3273 * UTF-8 even if no variant - saves scanning loop */
3275 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3280 /* Here, the string should be converted to utf8, either because of an
3281 * input flag (two_byte_count = 0), or because a character that
3282 * requires 2 bytes was found (two_byte_count = 1). t points either to
3283 * the beginning of the string (if we didn't examine anything), or to
3284 * the first variant. In either case, everything from s to t - 1 will
3285 * occupy only 1 byte each on output.
3287 * There are two main ways to convert. One is to create a new string
3288 * and go through the input starting from the beginning, appending each
3289 * converted value onto the new string as we go along. It's probably
3290 * best to allocate enough space in the string for the worst possible
3291 * case rather than possibly running out of space and having to
3292 * reallocate and then copy what we've done so far. Since everything
3293 * from s to t - 1 is invariant, the destination can be initialized
3294 * with these using a fast memory copy
3296 * The other way is to figure out exactly how big the string should be
3297 * by parsing the entire input. Then you don't have to make it big
3298 * enough to handle the worst possible case, and more importantly, if
3299 * the string you already have is large enough, you don't have to
3300 * allocate a new string, you can copy the last character in the input
3301 * string to the final position(s) that will be occupied by the
3302 * converted string and go backwards, stopping at t, since everything
3303 * before that is invariant.
3305 * There are advantages and disadvantages to each method.
3307 * In the first method, we can allocate a new string, do the memory
3308 * copy from the s to t - 1, and then proceed through the rest of the
3309 * string byte-by-byte.
3311 * In the second method, we proceed through the rest of the input
3312 * string just calculating how big the converted string will be. Then
3313 * there are two cases:
3314 * 1) if the string has enough extra space to handle the converted
3315 * value. We go backwards through the string, converting until we
3316 * get to the position we are at now, and then stop. If this
3317 * position is far enough along in the string, this method is
3318 * faster than the other method. If the memory copy were the same
3319 * speed as the byte-by-byte loop, that position would be about
3320 * half-way, as at the half-way mark, parsing to the end and back
3321 * is one complete string's parse, the same amount as starting
3322 * over and going all the way through. Actually, it would be
3323 * somewhat less than half-way, as it's faster to just count bytes
3324 * than to also copy, and we don't have the overhead of allocating
3325 * a new string, changing the scalar to use it, and freeing the
3326 * existing one. But if the memory copy is fast, the break-even
3327 * point is somewhere after half way. The counting loop could be
3328 * sped up by vectorization, etc, to move the break-even point
3329 * further towards the beginning.
3330 * 2) if the string doesn't have enough space to handle the converted
3331 * value. A new string will have to be allocated, and one might
3332 * as well, given that, start from the beginning doing the first
3333 * method. We've spent extra time parsing the string and in
3334 * exchange all we've gotten is that we know precisely how big to
3335 * make the new one. Perl is more optimized for time than space,
3336 * so this case is a loser.
3337 * So what I've decided to do is not use the 2nd method unless it is
3338 * guaranteed that a new string won't have to be allocated, assuming
3339 * the worst case. I also decided not to put any more conditions on it
3340 * than this, for now. It seems likely that, since the worst case is
3341 * twice as big as the unknown portion of the string (plus 1), we won't
3342 * be guaranteed enough space, causing us to go to the first method,
3343 * unless the string is short, or the first variant character is near
3344 * the end of it. In either of these cases, it seems best to use the
3345 * 2nd method. The only circumstance I can think of where this would
3346 * be really slower is if the string had once had much more data in it
3347 * than it does now, but there is still a substantial amount in it */
3350 STRLEN invariant_head = t - s;
3351 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3352 if (SvLEN(sv) < size) {
3354 /* Here, have decided to allocate a new string */
3359 Newx(dst, size, U8);
3361 /* If no known invariants at the beginning of the input string,
3362 * set so starts from there. Otherwise, can use memory copy to
3363 * get up to where we are now, and then start from here */
3365 if (invariant_head <= 0) {
3368 Copy(s, dst, invariant_head, char);
3369 d = dst + invariant_head;
3373 const UV uv = NATIVE8_TO_UNI(*t++);
3374 if (UNI_IS_INVARIANT(uv))
3375 *d++ = (U8)UNI_TO_NATIVE(uv);
3377 *d++ = (U8)UTF8_EIGHT_BIT_HI(uv);
3378 *d++ = (U8)UTF8_EIGHT_BIT_LO(uv);
3382 SvPV_free(sv); /* No longer using pre-existing string */
3383 SvPV_set(sv, (char*)dst);
3384 SvCUR_set(sv, d - dst);
3385 SvLEN_set(sv, size);
3388 /* Here, have decided to get the exact size of the string.
3389 * Currently this happens only when we know that there is
3390 * guaranteed enough space to fit the converted string, so
3391 * don't have to worry about growing. If two_byte_count is 0,
3392 * then t points to the first byte of the string which hasn't
3393 * been examined yet. Otherwise two_byte_count is 1, and t
3394 * points to the first byte in the string that will expand to
3395 * two. Depending on this, start examining at t or 1 after t.
3398 U8 *d = t + two_byte_count;
3401 /* Count up the remaining bytes that expand to two */
3404 const U8 chr = *d++;
3405 if (! NATIVE_IS_INVARIANT(chr)) two_byte_count++;
3408 /* The string will expand by just the number of bytes that
3409 * occupy two positions. But we are one afterwards because of
3410 * the increment just above. This is the place to put the
3411 * trailing NUL, and to set the length before we decrement */
3413 d += two_byte_count;
3414 SvCUR_set(sv, d - s);
3418 /* Having decremented d, it points to the position to put the
3419 * very last byte of the expanded string. Go backwards through
3420 * the string, copying and expanding as we go, stopping when we
3421 * get to the part that is invariant the rest of the way down */
3425 const U8 ch = NATIVE8_TO_UNI(*e--);
3426 if (UNI_IS_INVARIANT(ch)) {
3427 *d-- = UNI_TO_NATIVE(ch);
3429 *d-- = (U8)UTF8_EIGHT_BIT_LO(ch);
3430 *d-- = (U8)UTF8_EIGHT_BIT_HI(ch);
3435 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3436 /* Update pos. We do it at the end rather than during
3437 * the upgrade, to avoid slowing down the common case
3438 * (upgrade without pos) */
3439 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3441 I32 pos = mg->mg_len;
3442 if (pos > 0 && (U32)pos > invariant_head) {
3443 U8 *d = (U8*) SvPVX(sv) + invariant_head;
3444 STRLEN n = (U32)pos - invariant_head;
3446 if (UTF8_IS_START(*d))
3451 mg->mg_len = d - (U8*)SvPVX(sv);
3454 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3455 magic_setutf8(sv,mg); /* clear UTF8 cache */
3460 /* Mark as UTF-8 even if no variant - saves scanning loop */
3466 =for apidoc sv_utf8_downgrade
3468 Attempts to convert the PV of an SV from characters to bytes.
3469 If the PV contains a character that cannot fit
3470 in a byte, this conversion will fail;
3471 in this case, either returns false or, if C<fail_ok> is not
3474 This is not a general purpose Unicode to byte encoding interface:
3475 use the Encode extension for that.
3481 Perl_sv_utf8_downgrade(pTHX_ SV *const sv, const bool fail_ok)
3485 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3487 if (SvPOKp(sv) && SvUTF8(sv)) {
3491 int mg_flags = SV_GMAGIC;
3494 sv_force_normal_flags(sv, 0);
3496 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3498 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3500 I32 pos = mg->mg_len;
3502 sv_pos_b2u(sv, &pos);
3503 mg_flags = 0; /* sv_pos_b2u does get magic */
3507 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3508 magic_setutf8(sv,mg); /* clear UTF8 cache */
3511 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3513 if (!utf8_to_bytes(s, &len)) {
3518 Perl_croak(aTHX_ "Wide character in %s",
3521 Perl_croak(aTHX_ "Wide character");
3532 =for apidoc sv_utf8_encode
3534 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3535 flag off so that it looks like octets again.
3541 Perl_sv_utf8_encode(pTHX_ SV *const sv)
3543 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3545 if (SvREADONLY(sv)) {
3546 sv_force_normal_flags(sv, 0);
3548 (void) sv_utf8_upgrade(sv);
3553 =for apidoc sv_utf8_decode
3555 If the PV of the SV is an octet sequence in UTF-8
3556 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3557 so that it looks like a character. If the PV contains only single-byte
3558 characters, the C<SvUTF8> flag stays off.
3559 Scans PV for validity and returns false if the PV is invalid UTF-8.
3565 Perl_sv_utf8_decode(pTHX_ SV *const sv)
3567 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3570 const U8 *start, *c;
3573 /* The octets may have got themselves encoded - get them back as
3576 if (!sv_utf8_downgrade(sv, TRUE))
3579 /* it is actually just a matter of turning the utf8 flag on, but
3580 * we want to make sure everything inside is valid utf8 first.
3582 c = start = (const U8 *) SvPVX_const(sv);
3583 if (!is_utf8_string(c, SvCUR(sv)))
3585 e = (const U8 *) SvEND(sv);
3588 if (!UTF8_IS_INVARIANT(ch)) {
3593 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3594 /* adjust pos to the start of a UTF8 char sequence */
3595 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3597 I32 pos = mg->mg_len;
3599 for (c = start + pos; c > start; c--) {
3600 if (UTF8_IS_START(*c))
3603 mg->mg_len = c - start;
3606 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3607 magic_setutf8(sv,mg); /* clear UTF8 cache */
3614 =for apidoc sv_setsv
3616 Copies the contents of the source SV C<ssv> into the destination SV
3617 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3618 function if the source SV needs to be reused. Does not handle 'set' magic.
3619 Loosely speaking, it performs a copy-by-value, obliterating any previous
3620 content of the destination.
3622 You probably want to use one of the assortment of wrappers, such as
3623 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3624 C<SvSetMagicSV_nosteal>.
3626 =for apidoc sv_setsv_flags
3628 Copies the contents of the source SV C<ssv> into the destination SV
3629 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3630 function if the source SV needs to be reused. Does not handle 'set' magic.
3631 Loosely speaking, it performs a copy-by-value, obliterating any previous
3632 content of the destination.
3633 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3634 C<ssv> if appropriate, else not. If the C<flags>
3635 parameter has the C<NOSTEAL> bit set then the
3636 buffers of temps will not be stolen. <sv_setsv>
3637 and C<sv_setsv_nomg> are implemented in terms of this function.
3639 You probably want to use one of the assortment of wrappers, such as
3640 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3641 C<SvSetMagicSV_nosteal>.
3643 This is the primary function for copying scalars, and most other
3644 copy-ish functions and macros use this underneath.
3650 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3652 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3653 HV *old_stash = NULL;
3655 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3657 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3658 const char * const name = GvNAME(sstr);
3659 const STRLEN len = GvNAMELEN(sstr);
3661 if (dtype >= SVt_PV) {
3667 SvUPGRADE(dstr, SVt_PVGV);
3668 (void)SvOK_off(dstr);
3669 /* We have to turn this on here, even though we turn it off
3670 below, as GvSTASH will fail an assertion otherwise. */
3671 isGV_with_GP_on(dstr);
3673 GvSTASH(dstr) = GvSTASH(sstr);
3675 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3676 gv_name_set(MUTABLE_GV(dstr), name, len,
3677 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3678 SvFAKE_on(dstr); /* can coerce to non-glob */
3681 if(GvGP(MUTABLE_GV(sstr))) {
3682 /* If source has method cache entry, clear it */
3684 SvREFCNT_dec(GvCV(sstr));
3685 GvCV_set(sstr, NULL);
3688 /* If source has a real method, then a method is
3691 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3697 /* If dest already had a real method, that's a change as well */
3699 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3700 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3705 /* We don't need to check the name of the destination if it was not a
3706 glob to begin with. */
3707 if(dtype == SVt_PVGV) {
3708 const char * const name = GvNAME((const GV *)dstr);
3711 /* The stash may have been detached from the symbol table, so
3713 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3717 const STRLEN len = GvNAMELEN(dstr);
3718 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3719 || (len == 1 && name[0] == ':')) {
3722 /* Set aside the old stash, so we can reset isa caches on
3724 if((old_stash = GvHV(dstr)))
3725 /* Make sure we do not lose it early. */
3726 SvREFCNT_inc_simple_void_NN(
3727 sv_2mortal((SV *)old_stash)
3733 gp_free(MUTABLE_GV(dstr));
3734 isGV_with_GP_off(dstr); /* SvOK_off does not like globs. */
3735 (void)SvOK_off(dstr);
3736 isGV_with_GP_on(dstr);
3737 GvINTRO_off(dstr); /* one-shot flag */
3738 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3739 if (SvTAINTED(sstr))
3741 if (GvIMPORTED(dstr) != GVf_IMPORTED
3742 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3744 GvIMPORTED_on(dstr);
3747 if(mro_changes == 2) {
3748 if (GvAV((const GV *)sstr)) {
3750 SV * const sref = (SV *)GvAV((const GV *)dstr);
3751 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3752 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3753 AV * const ary = newAV();
3754 av_push(ary, mg->mg_obj); /* takes the refcount */
3755 mg->mg_obj = (SV *)ary;
3757 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3759 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3761 mro_isa_changed_in(GvSTASH(dstr));
3763 else if(mro_changes == 3) {
3764 HV * const stash = GvHV(dstr);
3765 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
3771 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
3772 if (GvIO(dstr) && dtype == SVt_PVGV) {
3773 DEBUG_o(Perl_deb(aTHX_
3774 "glob_assign_glob clearing PL_stashcache\n"));
3775 /* It's a cache. It will rebuild itself quite happily.
3776 It's a lot of effort to work out exactly which key (or keys)
3777 might be invalidated by the creation of the this file handle.
3779 hv_clear(PL_stashcache);
3785 S_glob_assign_ref(pTHX_ SV *const dstr, SV *const sstr)
3787 SV * const sref = SvRV(sstr);
3789 const int intro = GvINTRO(dstr);
3792 const U32 stype = SvTYPE(sref);
3794 PERL_ARGS_ASSERT_GLOB_ASSIGN_REF;
3797 GvINTRO_off(dstr); /* one-shot flag */
3798 GvLINE(dstr) = CopLINE(PL_curcop);
3799 GvEGV(dstr) = MUTABLE_GV(dstr);
3804 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
3805 import_flag = GVf_IMPORTED_CV;
3808 location = (SV **) &GvHV(dstr);
3809 import_flag = GVf_IMPORTED_HV;
3812 location = (SV **) &GvAV(dstr);
3813 import_flag = GVf_IMPORTED_AV;
3816 location = (SV **) &GvIOp(dstr);
3819 location = (SV **) &GvFORM(dstr);
3822 location = &GvSV(dstr);
3823 import_flag = GVf_IMPORTED_SV;
3826 if (stype == SVt_PVCV) {
3827 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
3828 if (GvCVGEN(dstr)) {
3829 SvREFCNT_dec(GvCV(dstr));
3830 GvCV_set(dstr, NULL);
3831 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3834 /* SAVEt_GVSLOT takes more room on the savestack and has more
3835 overhead in leave_scope than SAVEt_GENERIC_SV. But for CVs
3836 leave_scope needs access to the GV so it can reset method
3837 caches. We must use SAVEt_GVSLOT whenever the type is
3838 SVt_PVCV, even if the stash is anonymous, as the stash may
3839 gain a name somehow before leave_scope. */
3840 if (stype == SVt_PVCV) {
3841 /* There is no save_pushptrptrptr. Creating it for this
3842 one call site would be overkill. So inline the ss add
3846 SS_ADD_PTR(location);
3847 SS_ADD_PTR(SvREFCNT_inc(*location));
3848 SS_ADD_UV(SAVEt_GVSLOT);
3851 else SAVEGENERICSV(*location);
3854 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
3855 CV* const cv = MUTABLE_CV(*location);
3857 if (!GvCVGEN((const GV *)dstr) &&
3858 (CvROOT(cv) || CvXSUB(cv)) &&
3859 /* redundant check that avoids creating the extra SV
3860 most of the time: */
3861 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
3863 SV * const new_const_sv =
3864 CvCONST((const CV *)sref)
3865 ? cv_const_sv((const CV *)sref)
3867 report_redefined_cv(
3868 sv_2mortal(Perl_newSVpvf(aTHX_
3871 HvNAME_HEK(GvSTASH((const GV *)dstr))
3873 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr)))
3876 CvCONST((const CV *)sref) ? &new_const_sv : NULL
3880 cv_ckproto_len_flags(cv, (const GV *)dstr,
3881 SvPOK(sref) ? CvPROTO(sref) : NULL,
3882 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
3883 SvPOK(sref) ? SvUTF8(sref) : 0);
3885 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3886 GvASSUMECV_on(dstr);
3887 if(GvSTASH(dstr)) gv_method_changed(dstr); /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
3889 *location = SvREFCNT_inc_simple_NN(sref);
3890 if (import_flag && !(GvFLAGS(dstr) & import_flag)
3891 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
3892 GvFLAGS(dstr) |= import_flag;
3894 if (stype == SVt_PVHV) {
3895 const char * const name = GvNAME((GV*)dstr);
3896 const STRLEN len = GvNAMELEN(dstr);
3899 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
3900 || (len == 1 && name[0] == ':')
3902 && (!dref || HvENAME_get(dref))
3905 (HV *)sref, (HV *)dref,
3911 stype == SVt_PVAV && sref != dref
3912 && strEQ(GvNAME((GV*)dstr), "ISA")
3913 /* The stash may have been detached from the symbol table, so
3914 check its name before doing anything. */
3915 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3918 MAGIC * const omg = dref && SvSMAGICAL(dref)
3919 ? mg_find(dref, PERL_MAGIC_isa)
3921 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3922 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3923 AV * const ary = newAV();
3924 av_push(ary, mg->mg_obj); /* takes the refcount */
3925 mg->mg_obj = (SV *)ary;
3928 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
3929 SV **svp = AvARRAY((AV *)omg->mg_obj);
3930 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
3934 SvREFCNT_inc_simple_NN(*svp++)
3940 SvREFCNT_inc_simple_NN(omg->mg_obj)
3944 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
3949 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
3951 mg = mg_find(sref, PERL_MAGIC_isa);
3953 /* Since the *ISA assignment could have affected more than
3954 one stash, don't call mro_isa_changed_in directly, but let
3955 magic_clearisa do it for us, as it already has the logic for
3956 dealing with globs vs arrays of globs. */
3958 Perl_magic_clearisa(aTHX_ NULL, mg);
3960 else if (stype == SVt_PVIO) {
3961 DEBUG_o(Perl_deb(aTHX_ "glob_assign_ref clearing PL_stashcache\n"));
3962 /* It's a cache. It will rebuild itself quite happily.
3963 It's a lot of effort to work out exactly which key (or keys)
3964 might be invalidated by the creation of the this file handle.
3966 hv_clear(PL_stashcache);
3970 if (!intro) SvREFCNT_dec(dref);
3971 if (SvTAINTED(sstr))
3976 /* Work around compiler warnings about unsigned >= THRESHOLD when thres-
3978 #if SV_COW_THRESHOLD
3979 # define GE_COW_THRESHOLD(len) ((len) >= SV_COW_THRESHOLD)
3981 # define GE_COW_THRESHOLD(len) 1
3983 #if SV_COWBUF_THRESHOLD
3984 # define GE_COWBUF_THRESHOLD(len) ((len) >= SV_COWBUF_THRESHOLD)
3986 # define GE_COWBUF_THRESHOLD(len) 1
3990 Perl_sv_setsv_flags(pTHX_ SV *dstr, SV* sstr, const I32 flags)
3997 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
4002 if (SvIS_FREED(dstr)) {
4003 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
4004 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
4006 SV_CHECK_THINKFIRST_COW_DROP(dstr);
4008 sstr = &PL_sv_undef;
4009 if (SvIS_FREED(sstr)) {
4010 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
4011 (void*)sstr, (void*)dstr);
4013 stype = SvTYPE(sstr);
4014 dtype = SvTYPE(dstr);
4016 /* There's a lot of redundancy below but we're going for speed here */
4021 if (dtype != SVt_PVGV && dtype != SVt_PVLV) {
4022 (void)SvOK_off(dstr);
4030 sv_upgrade(dstr, SVt_IV);
4034 sv_upgrade(dstr, SVt_PVIV);
4038 goto end_of_first_switch;
4040 (void)SvIOK_only(dstr);
4041 SvIV_set(dstr, SvIVX(sstr));
4044 /* SvTAINTED can only be true if the SV has taint magic, which in
4045 turn means that the SV type is PVMG (or greater). This is the
4046 case statement for SVt_IV, so this cannot be true (whatever gcov
4048 assert(!SvTAINTED(sstr));
4053 if (dtype < SVt_PV && dtype != SVt_IV)
4054 sv_upgrade(dstr, SVt_IV);
4062 sv_upgrade(dstr, SVt_NV);
4066 sv_upgrade(dstr, SVt_PVNV);
4070 goto end_of_first_switch;
4072 SvNV_set(dstr, SvNVX(sstr));
4073 (void)SvNOK_only(dstr);
4074 /* SvTAINTED can only be true if the SV has taint magic, which in
4075 turn means that the SV type is PVMG (or greater). This is the
4076 case statement for SVt_NV, so this cannot be true (whatever gcov
4078 assert(!SvTAINTED(sstr));
4085 sv_upgrade(dstr, SVt_PV);
4088 if (dtype < SVt_PVIV)
4089 sv_upgrade(dstr, SVt_PVIV);
4092 if (dtype < SVt_PVNV)
4093 sv_upgrade(dstr, SVt_PVNV);
4097 const char * const type = sv_reftype(sstr,0);
4099 /* diag_listed_as: Bizarre copy of %s */
4100 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4102 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4108 if (dtype < SVt_REGEXP)
4110 if (dtype >= SVt_PV) {
4116 sv_upgrade(dstr, SVt_REGEXP);
4120 /* case SVt_DUMMY: */
4124 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4126 if (SvTYPE(sstr) != stype)
4127 stype = SvTYPE(sstr);
4129 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4130 glob_assign_glob(dstr, sstr, dtype);
4133 if (stype == SVt_PVLV)
4135 if (isREGEXP(sstr)) goto upgregexp;
4136 SvUPGRADE(dstr, SVt_PVNV);
4139 SvUPGRADE(dstr, (svtype)stype);
4141 end_of_first_switch:
4143 /* dstr may have been upgraded. */
4144 dtype = SvTYPE(dstr);
4145 sflags = SvFLAGS(sstr);
4147 if (dtype == SVt_PVCV) {
4148 /* Assigning to a subroutine sets the prototype. */
4151 const char *const ptr = SvPV_const(sstr, len);
4153 SvGROW(dstr, len + 1);
4154 Copy(ptr, SvPVX(dstr), len + 1, char);
4155 SvCUR_set(dstr, len);
4157 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4158 CvAUTOLOAD_off(dstr);
4163 else if (dtype == SVt_PVAV || dtype == SVt_PVHV || dtype == SVt_PVFM) {
4164 const char * const type = sv_reftype(dstr,0);
4166 /* diag_listed_as: Cannot copy to %s */
4167 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4169 Perl_croak(aTHX_ "Cannot copy to %s", type);
4170 } else if (sflags & SVf_ROK) {
4171 if (isGV_with_GP(dstr)
4172 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4175 if (GvIMPORTED(dstr) != GVf_IMPORTED
4176 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4178 GvIMPORTED_on(dstr);
4183 glob_assign_glob(dstr, sstr, dtype);
4187 if (dtype >= SVt_PV) {
4188 if (isGV_with_GP(dstr)) {
4189 glob_assign_ref(dstr, sstr);
4192 if (SvPVX_const(dstr)) {
4198 (void)SvOK_off(dstr);
4199 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4200 SvFLAGS(dstr) |= sflags & SVf_ROK;
4201 assert(!(sflags & SVp_NOK));
4202 assert(!(sflags & SVp_IOK));
4203 assert(!(sflags & SVf_NOK));
4204 assert(!(sflags & SVf_IOK));
4206 else if (isGV_with_GP(dstr)) {
4207 if (!(sflags & SVf_OK)) {
4208 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4209 "Undefined value assigned to typeglob");
4212 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4213 if (dstr != (const SV *)gv) {
4214 const char * const name = GvNAME((const GV *)dstr);
4215 const STRLEN len = GvNAMELEN(dstr);
4216 HV *old_stash = NULL;
4217 bool reset_isa = FALSE;
4218 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4219 || (len == 1 && name[0] == ':')) {
4220 /* Set aside the old stash, so we can reset isa caches
4221 on its subclasses. */
4222 if((old_stash = GvHV(dstr))) {
4223 /* Make sure we do not lose it early. */
4224 SvREFCNT_inc_simple_void_NN(
4225 sv_2mortal((SV *)old_stash)
4232 gp_free(MUTABLE_GV(dstr));
4233 GvGP_set(dstr, gp_ref(GvGP(gv)));
4236 HV * const stash = GvHV(dstr);
4238 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4248 else if ((dtype == SVt_REGEXP || dtype == SVt_PVLV)
4249 && (stype == SVt_REGEXP || isREGEXP(sstr))) {
4250 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4252 else if (sflags & SVp_POK) {
4254 const STRLEN cur = SvCUR(sstr);
4255 const STRLEN len = SvLEN(sstr);
4258 * Check to see if we can just swipe the string. If so, it's a
4259 * possible small lose on short strings, but a big win on long ones.
4260 * It might even be a win on short strings if SvPVX_const(dstr)
4261 * has to be allocated and SvPVX_const(sstr) has to be freed.
4262 * Likewise if we can set up COW rather than doing an actual copy, we
4263 * drop to the else clause, as the swipe code and the COW setup code
4264 * have much in common.
4267 /* Whichever path we take through the next code, we want this true,
4268 and doing it now facilitates the COW check. */
4269 (void)SvPOK_only(dstr);
4272 /* If we're already COW then this clause is not true, and if COW
4273 is allowed then we drop down to the else and make dest COW
4274 with us. If caller hasn't said that we're allowed to COW
4275 shared hash keys then we don't do the COW setup, even if the
4276 source scalar is a shared hash key scalar. */
4277 (((flags & SV_COW_SHARED_HASH_KEYS)
4278 ? !(sflags & SVf_IsCOW)
4279 #ifdef PERL_NEW_COPY_ON_WRITE
4281 ((!GE_COWBUF_THRESHOLD(cur) && SvLEN(dstr) > cur)
4282 /* If this is a regular (non-hek) COW, only so many COW
4283 "copies" are possible. */
4284 || CowREFCNT(sstr) == SV_COW_REFCNT_MAX))
4286 : 1 /* If making a COW copy is forbidden then the behaviour we
4287 desire is as if the source SV isn't actually already
4288 COW, even if it is. So we act as if the source flags
4289 are not COW, rather than actually testing them. */
4291 #ifndef PERL_ANY_COW
4292 /* The change that added SV_COW_SHARED_HASH_KEYS makes the logic
4293 when PERL_OLD_COPY_ON_WRITE is defined a little wrong.
4294 Conceptually PERL_OLD_COPY_ON_WRITE being defined should
4295 override SV_COW_SHARED_HASH_KEYS, because it means "always COW"
4296 but in turn, it's somewhat dead code, never expected to go
4297 live, but more kept as a placeholder on how to do it better
4298 in a newer implementation. */
4299 /* If we are COW and dstr is a suitable target then we drop down
4300 into the else and make dest a COW of us. */
4301 || (SvFLAGS(dstr) & SVf_BREAK)
4306 #ifdef PERL_NEW_COPY_ON_WRITE
4307 /* slated for free anyway (and not COW)? */
4308 (sflags & (SVs_TEMP|SVf_IsCOW)) == SVs_TEMP &&
4310 (sflags & SVs_TEMP) && /* slated for free anyway? */
4312 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4313 (!(flags & SV_NOSTEAL)) &&
4314 /* and we're allowed to steal temps */
4315 SvREFCNT(sstr) == 1 && /* and no other references to it? */
4316 len) /* and really is a string */
4318 && ((flags & SV_COW_SHARED_HASH_KEYS)
4319 ? (!((sflags & CAN_COW_MASK) == CAN_COW_FLAGS
4320 # ifdef PERL_OLD_COPY_ON_WRITE
4321 && (SvFLAGS(dstr) & CAN_COW_MASK) == CAN_COW_FLAGS
4322 && SvTYPE(sstr) >= SVt_PVIV
4324 && !(SvFLAGS(dstr) & SVf_BREAK)
4325 && !(sflags & SVf_IsCOW)
4326 && GE_COW_THRESHOLD(cur) && cur+1 < len
4327 && (GE_COWBUF_THRESHOLD(cur) || SvLEN(dstr) < cur+1)
4333 /* Failed the swipe test, and it's not a shared hash key either.
4334 Have to copy the string. */
4335 SvGROW(dstr, cur + 1); /* inlined from sv_setpvn */
4336 Move(SvPVX_const(sstr),SvPVX(dstr),cur,char);
4337 SvCUR_set(dstr, cur);
4338 *SvEND(dstr) = '\0';
4340 /* If PERL_OLD_COPY_ON_WRITE is not defined, then isSwipe will always
4342 /* Either it's a shared hash key, or it's suitable for
4343 copy-on-write or we can swipe the string. */
4345 PerlIO_printf(Perl_debug_log, "Copy on write: sstr --> dstr\n");
4351 if (!(sflags & SVf_IsCOW)) {
4353 # ifdef PERL_OLD_COPY_ON_WRITE
4354 /* Make the source SV into a loop of 1.
4355 (about to become 2) */
4356 SV_COW_NEXT_SV_SET(sstr, sstr);
4358 CowREFCNT(sstr) = 0;
4363 /* Initial code is common. */
4364 if (SvPVX_const(dstr)) { /* we know that dtype >= SVt_PV */
4369 /* making another shared SV. */
4372 # ifdef PERL_OLD_COPY_ON_WRITE
4373 assert (SvTYPE(dstr) >= SVt_PVIV);
4374 /* SvIsCOW_normal */
4375 /* splice us in between source and next-after-source. */
4376 SV_COW_NEXT_SV_SET(dstr, SV_COW_NEXT_SV(sstr));
4377 SV_COW_NEXT_SV_SET(sstr, dstr);
4381 SvPV_set(dstr, SvPVX_mutable(sstr));
4385 /* SvIsCOW_shared_hash */
4386 DEBUG_C(PerlIO_printf(Perl_debug_log,
4387 "Copy on write: Sharing hash\n"));
4389 assert (SvTYPE(dstr) >= SVt_PV);
4391 HEK_KEY(share_hek_hek(SvSHARED_HEK_FROM_PV(SvPVX_const(sstr)))));
4393 SvLEN_set(dstr, len);
4394 SvCUR_set(dstr, cur);
4398 { /* Passes the swipe test. */
4399 SvPV_set(dstr, SvPVX_mutable(sstr));
4400 SvLEN_set(dstr, SvLEN(sstr));
4401 SvCUR_set(dstr, SvCUR(sstr));
4404 (void)SvOK_off(sstr); /* NOTE: nukes most SvFLAGS on sstr */
4405 SvPV_set(sstr, NULL);
4411 if (sflags & SVp_NOK) {
4412 SvNV_set(dstr, SvNVX(sstr));
4414 if (sflags & SVp_IOK) {
4415 SvIV_set(dstr, SvIVX(sstr));
4416 /* Must do this otherwise some other overloaded use of 0x80000000
4417 gets confused. I guess SVpbm_VALID */
4418 if (sflags & SVf_IVisUV)
4421 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_NOK|SVp_NOK|SVf_UTF8);
4423 const MAGIC * const smg = SvVSTRING_mg(sstr);
4425 sv_magic(dstr, NULL, PERL_MAGIC_vstring,
4426 smg->mg_ptr, smg->mg_len);
4427 SvRMAGICAL_on(dstr);
4431 else if (sflags & (SVp_IOK|SVp_NOK)) {
4432 (void)SvOK_off(dstr);
4433 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_IVisUV|SVf_NOK|SVp_NOK);
4434 if (sflags & SVp_IOK) {
4435 /* XXXX Do we want to set IsUV for IV(ROK)? Be extra safe... */
4436 SvIV_set(dstr, SvIVX(sstr));
4438 if (sflags & SVp_NOK) {
4439 SvNV_set(dstr, SvNVX(sstr));
4443 if (isGV_with_GP(sstr)) {
4444 gv_efullname3(dstr, MUTABLE_GV(sstr), "*");
4447 (void)SvOK_off(dstr);
4449 if (SvTAINTED(sstr))
4454 =for apidoc sv_setsv_mg
4456 Like C<sv_setsv>, but also handles 'set' magic.
4462 Perl_sv_setsv_mg(pTHX_ SV *const dstr, SV *const sstr)
4464 PERL_ARGS_ASSERT_SV_SETSV_MG;
4466 sv_setsv(dstr,sstr);
4471 # ifdef PERL_OLD_COPY_ON_WRITE
4472 # define SVt_COW SVt_PVIV
4474 # define SVt_COW SVt_PV
4477 Perl_sv_setsv_cow(pTHX_ SV *dstr, SV *sstr)
4479 STRLEN cur = SvCUR(sstr);
4480 STRLEN len = SvLEN(sstr);
4483 PERL_ARGS_ASSERT_SV_SETSV_COW;
4486 PerlIO_printf(Perl_debug_log, "Fast copy on write: %p -> %p\n",
4487 (void*)sstr, (void*)dstr);
4494 if (SvTHINKFIRST(dstr))
4495 sv_force_normal_flags(dstr, SV_COW_DROP_PV);
4496 else if (SvPVX_const(dstr))
4497 Safefree(SvPVX_mutable(dstr));
4501 SvUPGRADE(dstr, SVt_COW);
4503 assert (SvPOK(sstr));
4504 assert (SvPOKp(sstr));
4505 # ifdef PERL_OLD_COPY_ON_WRITE
4506 assert (!SvIOK(sstr));
4507 assert (!SvIOKp(sstr));
4508 assert (!SvNOK(sstr));
4509 assert (!SvNOKp(sstr));
4512 if (SvIsCOW(sstr)) {
4514 if (SvLEN(sstr) == 0) {
4515 /* source is a COW shared hash key. */
4516 DEBUG_C(PerlIO_printf(Perl_debug_log,
4517 "Fast copy on write: Sharing hash\n"));
4518 new_pv = HEK_KEY(share_hek_hek(SvSHARED_HEK_FROM_PV(SvPVX_const(sstr))));
4521 # ifdef PERL_OLD_COPY_ON_WRITE
4522 SV_COW_NEXT_SV_SET(dstr, SV_COW_NEXT_SV(sstr));
4524 assert(SvCUR(sstr)+1 < SvLEN(sstr));
4525 assert(CowREFCNT(sstr) < SV_COW_REFCNT_MAX);
4528 assert ((SvFLAGS(sstr) & CAN_COW_MASK) == CAN_COW_FLAGS);
4529 SvUPGRADE(sstr, SVt_COW);
4531 DEBUG_C(PerlIO_printf(Perl_debug_log,
4532 "Fast copy on write: Converting sstr to COW\n"));
4533 # ifdef PERL_OLD_COPY_ON_WRITE
4534 SV_COW_NEXT_SV_SET(dstr, sstr);
4536 CowREFCNT(sstr) = 0;
4539 # ifdef PERL_OLD_COPY_ON_WRITE
4540 SV_COW_NEXT_SV_SET(sstr, dstr);
4544 new_pv = SvPVX_mutable(sstr);
4547 SvPV_set(dstr, new_pv);
4548 SvFLAGS(dstr) = (SVt_COW|SVf_POK|SVp_POK|SVf_IsCOW);
4551 SvLEN_set(dstr, len);
4552 SvCUR_set(dstr, cur);
4561 =for apidoc sv_setpvn
4563 Copies a string into an SV. The C<len> parameter indicates the number of
4564 bytes to be copied. If the C<ptr> argument is NULL the SV will become
4565 undefined. Does not handle 'set' magic. See C<sv_setpvn_mg>.
4571 Perl_sv_setpvn(pTHX_ SV *const sv, const char *const ptr, const STRLEN len)
4576 PERL_ARGS_ASSERT_SV_SETPVN;
4578 SV_CHECK_THINKFIRST_COW_DROP(sv);
4584 /* len is STRLEN which is unsigned, need to copy to signed */
4587 Perl_croak(aTHX_ "panic: sv_setpvn called with negative strlen %"
4590 SvUPGRADE(sv, SVt_PV);
4592 dptr = SvGROW(sv, len + 1);
4593 Move(ptr,dptr,len,char);
4596 (void)SvPOK_only_UTF8(sv); /* validate pointer */
4598 if (SvTYPE(sv) == SVt_PVCV) CvAUTOLOAD_off(sv);
4602 =for apidoc sv_setpvn_mg
4604 Like C<sv_setpvn>, but also handles 'set' magic.
4610 Perl_sv_setpvn_mg(pTHX_ SV *const sv, const char *const ptr, const STRLEN len)
4612 PERL_ARGS_ASSERT_SV_SETPVN_MG;
4614 sv_setpvn(sv,ptr,len);
4619 =for apidoc sv_setpv
4621 Copies a string into an SV. The string must be null-terminated. Does not
4622 handle 'set' magic. See C<sv_setpv_mg>.
4628 Perl_sv_setpv(pTHX_ SV *const sv, const char *const ptr)
4633 PERL_ARGS_ASSERT_SV_SETPV;
4635 SV_CHECK_THINKFIRST_COW_DROP(sv);
4641 SvUPGRADE(sv, SVt_PV);
4643 SvGROW(sv, len + 1);
4644 Move(ptr,SvPVX(sv),len+1,char);
4646 (void)SvPOK_only_UTF8(sv); /* validate pointer */