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))
69 /* This is a pessimistic view. Scalar must be purely a read-write PV to copy-
73 /* ============================================================================
75 =head1 Allocation and deallocation of SVs.
77 An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
78 sv, av, hv...) contains type and reference count information, and for
79 many types, a pointer to the body (struct xrv, xpv, xpviv...), which
80 contains fields specific to each type. Some types store all they need
81 in the head, so don't have a body.
83 In all but the most memory-paranoid configurations (ex: PURIFY), heads
84 and bodies are allocated out of arenas, which by default are
85 approximately 4K chunks of memory parcelled up into N heads or bodies.
86 Sv-bodies are allocated by their sv-type, guaranteeing size
87 consistency needed to allocate safely from arrays.
89 For SV-heads, the first slot in each arena is reserved, and holds a
90 link to the next arena, some flags, and a note of the number of slots.
91 Snaked through each arena chain is a linked list of free items; when
92 this becomes empty, an extra arena is allocated and divided up into N
93 items which are threaded into the free list.
95 SV-bodies are similar, but they use arena-sets by default, which
96 separate the link and info from the arena itself, and reclaim the 1st
97 slot in the arena. SV-bodies are further described later.
99 The following global variables are associated with arenas:
101 PL_sv_arenaroot pointer to list of SV arenas
102 PL_sv_root pointer to list of free SV structures
104 PL_body_arenas head of linked-list of body arenas
105 PL_body_roots[] array of pointers to list of free bodies of svtype
106 arrays are indexed by the svtype needed
108 A few special SV heads are not allocated from an arena, but are
109 instead directly created in the interpreter structure, eg PL_sv_undef.
110 The size of arenas can be changed from the default by setting
111 PERL_ARENA_SIZE appropriately at compile time.
113 The SV arena serves the secondary purpose of allowing still-live SVs
114 to be located and destroyed during final cleanup.
116 At the lowest level, the macros new_SV() and del_SV() grab and free
117 an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
118 to return the SV to the free list with error checking.) new_SV() calls
119 more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
120 SVs in the free list have their SvTYPE field set to all ones.
122 At the time of very final cleanup, sv_free_arenas() is called from
123 perl_destruct() to physically free all the arenas allocated since the
124 start of the interpreter.
126 The function visit() scans the SV arenas list, and calls a specified
127 function for each SV it finds which is still live - ie which has an SvTYPE
128 other than all 1's, and a non-zero SvREFCNT. visit() is used by the
129 following functions (specified as [function that calls visit()] / [function
130 called by visit() for each SV]):
132 sv_report_used() / do_report_used()
133 dump all remaining SVs (debugging aid)
135 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
136 do_clean_named_io_objs()
137 Attempt to free all objects pointed to by RVs,
138 and try to do the same for all objects indirectly
139 referenced by typeglobs too. 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) Safefree((sv)->sv_debug_file)
186 # define DEBUG_SV_SERIAL(sv) \
187 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) del_SV\n", \
188 PTR2UV(sv), (long)(sv)->sv_debug_serial))
190 # define FREE_SV_DEBUG_FILE(sv)
191 # define DEBUG_SV_SERIAL(sv) NOOP
195 # define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
196 # define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
197 /* Whilst I'd love to do this, it seems that things like to check on
199 # define POSION_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
201 # define POSION_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
202 PoisonNew(&SvREFCNT(sv), 1, U32)
204 # define SvARENA_CHAIN(sv) SvANY(sv)
205 # define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
206 # define POSION_SV_HEAD(sv)
209 /* Mark an SV head as unused, and add to free list.
211 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
212 * its refcount artificially decremented during global destruction, so
213 * there may be dangling pointers to it. The last thing we want in that
214 * case is for it to be reused. */
216 #define plant_SV(p) \
218 const U32 old_flags = SvFLAGS(p); \
219 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
220 DEBUG_SV_SERIAL(p); \
221 FREE_SV_DEBUG_FILE(p); \
223 SvFLAGS(p) = SVTYPEMASK; \
224 if (!(old_flags & SVf_BREAK)) { \
225 SvARENA_CHAIN_SET(p, PL_sv_root); \
231 #define uproot_SV(p) \
234 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
239 /* make some more SVs by adding another arena */
246 char *chunk; /* must use New here to match call to */
247 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
248 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
253 /* new_SV(): return a new, empty SV head */
255 #ifdef DEBUG_LEAKING_SCALARS
256 /* provide a real function for a debugger to play with */
258 S_new_SV(pTHX_ const char *file, int line, const char *func)
265 sv = S_more_sv(aTHX);
269 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
270 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
276 sv->sv_debug_inpad = 0;
277 sv->sv_debug_parent = NULL;
278 sv->sv_debug_file = PL_curcop ? savepv(CopFILE(PL_curcop)): NULL;
280 sv->sv_debug_serial = PL_sv_serial++;
282 MEM_LOG_NEW_SV(sv, file, line, func);
283 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%"UVxf": (%05ld) new_SV (from %s:%d [%s])\n",
284 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
288 # define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
296 (p) = S_more_sv(aTHX); \
300 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
305 /* del_SV(): return an empty SV head to the free list */
318 S_del_sv(pTHX_ SV *p)
322 PERL_ARGS_ASSERT_DEL_SV;
327 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
328 const SV * const sv = sva + 1;
329 const SV * const svend = &sva[SvREFCNT(sva)];
330 if (p >= sv && p < svend) {
336 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
337 "Attempt to free non-arena SV: 0x%"UVxf
338 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
345 #else /* ! DEBUGGING */
347 #define del_SV(p) plant_SV(p)
349 #endif /* DEBUGGING */
353 =head1 SV Manipulation Functions
355 =for apidoc sv_add_arena
357 Given a chunk of memory, link it to the head of the list of arenas,
358 and split it into a list of free SVs.
364 S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
367 SV *const sva = MUTABLE_SV(ptr);
371 PERL_ARGS_ASSERT_SV_ADD_ARENA;
373 /* The first SV in an arena isn't an SV. */
374 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
375 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
376 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
378 PL_sv_arenaroot = sva;
379 PL_sv_root = sva + 1;
381 svend = &sva[SvREFCNT(sva) - 1];
384 SvARENA_CHAIN_SET(sv, (sv + 1));
388 /* Must always set typemask because it's always checked in on cleanup
389 when the arenas are walked looking for objects. */
390 SvFLAGS(sv) = SVTYPEMASK;
393 SvARENA_CHAIN_SET(sv, 0);
397 SvFLAGS(sv) = SVTYPEMASK;
400 /* visit(): call the named function for each non-free SV in the arenas
401 * whose flags field matches the flags/mask args. */
404 S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
410 PERL_ARGS_ASSERT_VISIT;
412 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
413 register const SV * const svend = &sva[SvREFCNT(sva)];
415 for (sv = sva + 1; sv < svend; ++sv) {
416 if (SvTYPE(sv) != (svtype)SVTYPEMASK
417 && (sv->sv_flags & mask) == flags
430 /* called by sv_report_used() for each live SV */
433 do_report_used(pTHX_ SV *const sv)
435 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
436 PerlIO_printf(Perl_debug_log, "****\n");
443 =for apidoc sv_report_used
445 Dump the contents of all SVs not yet freed. (Debugging aid).
451 Perl_sv_report_used(pTHX)
454 visit(do_report_used, 0, 0);
460 /* called by sv_clean_objs() for each live SV */
463 do_clean_objs(pTHX_ SV *const ref)
468 SV * const target = SvRV(ref);
469 if (SvOBJECT(target)) {
470 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
471 if (SvWEAKREF(ref)) {
472 sv_del_backref(target, ref);
478 SvREFCNT_dec(target);
483 /* XXX Might want to check arrays, etc. */
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)));
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)));
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)));
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)));
528 SvREFCNT_dec(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)));
551 SvREFCNT_dec(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_BIND, 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, FALSE, 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 You generally want to use the C<SvUPGRADE> macro wrapper. See also C<svtype>.
1136 Perl_sv_upgrade(pTHX_ register SV *const sv, svtype new_type)
1141 const svtype old_type = SvTYPE(sv);
1142 const struct body_details *new_type_details;
1143 const struct body_details *old_type_details
1144 = bodies_by_type + old_type;
1145 SV *referant = NULL;
1147 PERL_ARGS_ASSERT_SV_UPGRADE;
1149 if (old_type == new_type)
1152 /* This clause was purposefully added ahead of the early return above to
1153 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1154 inference by Nick I-S that it would fix other troublesome cases. See
1155 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1157 Given that shared hash key scalars are no longer PVIV, but PV, there is
1158 no longer need to unshare so as to free up the IVX slot for its proper
1159 purpose. So it's safe to move the early return earlier. */
1161 if (new_type != SVt_PV && SvIsCOW(sv)) {
1162 sv_force_normal_flags(sv, 0);
1165 old_body = SvANY(sv);
1167 /* Copying structures onto other structures that have been neatly zeroed
1168 has a subtle gotcha. Consider XPVMG
1170 +------+------+------+------+------+-------+-------+
1171 | NV | CUR | LEN | IV | MAGIC | STASH |
1172 +------+------+------+------+------+-------+-------+
1173 0 4 8 12 16 20 24 28
1175 where NVs are aligned to 8 bytes, so that sizeof that structure is
1176 actually 32 bytes long, with 4 bytes of padding at the end:
1178 +------+------+------+------+------+-------+-------+------+
1179 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1180 +------+------+------+------+------+-------+-------+------+
1181 0 4 8 12 16 20 24 28 32
1183 so what happens if you allocate memory for this structure:
1185 +------+------+------+------+------+-------+-------+------+------+...
1186 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1187 +------+------+------+------+------+-------+-------+------+------+...
1188 0 4 8 12 16 20 24 28 32 36
1190 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1191 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1192 started out as zero once, but it's quite possible that it isn't. So now,
1193 rather than a nicely zeroed GP, you have it pointing somewhere random.
1196 (In fact, GP ends up pointing at a previous GP structure, because the
1197 principle cause of the padding in XPVMG getting garbage is a copy of
1198 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1199 this happens to be moot because XPVGV has been re-ordered, with GP
1200 no longer after STASH)
1202 So we are careful and work out the size of used parts of all the
1210 referant = SvRV(sv);
1211 old_type_details = &fake_rv;
1212 if (new_type == SVt_NV)
1213 new_type = SVt_PVNV;
1215 if (new_type < SVt_PVIV) {
1216 new_type = (new_type == SVt_NV)
1217 ? SVt_PVNV : SVt_PVIV;
1222 if (new_type < SVt_PVNV) {
1223 new_type = SVt_PVNV;
1227 assert(new_type > SVt_PV);
1228 assert(SVt_IV < SVt_PV);
1229 assert(SVt_NV < SVt_PV);
1236 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1237 there's no way that it can be safely upgraded, because perl.c
1238 expects to Safefree(SvANY(PL_mess_sv)) */
1239 assert(sv != PL_mess_sv);
1240 /* This flag bit is used to mean other things in other scalar types.
1241 Given that it only has meaning inside the pad, it shouldn't be set
1242 on anything that can get upgraded. */
1243 assert(!SvPAD_TYPED(sv));
1246 if (old_type_details->cant_upgrade)
1247 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1248 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1251 if (old_type > new_type)
1252 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1253 (int)old_type, (int)new_type);
1255 new_type_details = bodies_by_type + new_type;
1257 SvFLAGS(sv) &= ~SVTYPEMASK;
1258 SvFLAGS(sv) |= new_type;
1260 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1261 the return statements above will have triggered. */
1262 assert (new_type != SVt_NULL);
1265 assert(old_type == SVt_NULL);
1266 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv));
1270 assert(old_type == SVt_NULL);
1271 SvANY(sv) = new_XNV();
1276 assert(new_type_details->body_size);
1279 assert(new_type_details->arena);
1280 assert(new_type_details->arena_size);
1281 /* This points to the start of the allocated area. */
1282 new_body_inline(new_body, new_type);
1283 Zero(new_body, new_type_details->body_size, char);
1284 new_body = ((char *)new_body) - new_type_details->offset;
1286 /* We always allocated the full length item with PURIFY. To do this
1287 we fake things so that arena is false for all 16 types.. */
1288 new_body = new_NOARENAZ(new_type_details);
1290 SvANY(sv) = new_body;
1291 if (new_type == SVt_PVAV) {
1295 if (old_type_details->body_size) {
1298 /* It will have been zeroed when the new body was allocated.
1299 Lets not write to it, in case it confuses a write-back
1305 #ifndef NODEFAULT_SHAREKEYS
1306 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1308 HvMAX(sv) = 7; /* (start with 8 buckets) */
1311 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1312 The target created by newSVrv also is, and it can have magic.
1313 However, it never has SvPVX set.
1315 if (old_type == SVt_IV) {
1317 } else if (old_type >= SVt_PV) {
1318 assert(SvPVX_const(sv) == 0);
1321 if (old_type >= SVt_PVMG) {
1322 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1323 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1325 sv->sv_u.svu_array = NULL; /* or svu_hash */
1331 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
1332 sv_force_normal_flags(sv) is called. */
1335 /* XXX Is this still needed? Was it ever needed? Surely as there is
1336 no route from NV to PVIV, NOK can never be true */
1337 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 (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 hv_clear(PL_stashcache);
1399 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1400 IoPAGE_LEN(sv) = 60;
1402 if (old_type < SVt_PV) {
1403 /* referant will be NULL unless the old type was SVt_IV emulating
1405 sv->sv_u.svu_rv = referant;
1409 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1410 (unsigned long)new_type);
1413 if (old_type > SVt_IV) {
1417 /* Note that there is an assumption that all bodies of types that
1418 can be upgraded came from arenas. Only the more complex non-
1419 upgradable types are allowed to be directly malloc()ed. */
1420 assert(old_type_details->arena);
1421 del_body((void*)((char*)old_body + old_type_details->offset),
1422 &PL_body_roots[old_type]);
1428 =for apidoc sv_backoff
1430 Remove any string offset. You should normally use the C<SvOOK_off> macro
1437 Perl_sv_backoff(pTHX_ register SV *const sv)
1440 const char * const s = SvPVX_const(sv);
1442 PERL_ARGS_ASSERT_SV_BACKOFF;
1443 PERL_UNUSED_CONTEXT;
1446 assert(SvTYPE(sv) != SVt_PVHV);
1447 assert(SvTYPE(sv) != SVt_PVAV);
1449 SvOOK_offset(sv, delta);
1451 SvLEN_set(sv, SvLEN(sv) + delta);
1452 SvPV_set(sv, SvPVX(sv) - delta);
1453 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1454 SvFLAGS(sv) &= ~SVf_OOK;
1461 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1462 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1463 Use the C<SvGROW> wrapper instead.
1469 Perl_sv_grow(pTHX_ register SV *const sv, register STRLEN newlen)
1473 PERL_ARGS_ASSERT_SV_GROW;
1475 if (PL_madskills && newlen >= 0x100000) {
1476 PerlIO_printf(Perl_debug_log,
1477 "Allocation too large: %"UVxf"\n", (UV)newlen);
1479 #ifdef HAS_64K_LIMIT
1480 if (newlen >= 0x10000) {
1481 PerlIO_printf(Perl_debug_log,
1482 "Allocation too large: %"UVxf"\n", (UV)newlen);
1485 #endif /* HAS_64K_LIMIT */
1488 if (SvTYPE(sv) < SVt_PV) {
1489 sv_upgrade(sv, SVt_PV);
1490 s = SvPVX_mutable(sv);
1492 else if (SvOOK(sv)) { /* pv is offset? */
1494 s = SvPVX_mutable(sv);
1495 if (newlen > SvLEN(sv))
1496 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1497 #ifdef HAS_64K_LIMIT
1498 if (newlen >= 0x10000)
1503 s = SvPVX_mutable(sv);
1505 if (newlen > SvLEN(sv)) { /* need more room? */
1506 STRLEN minlen = SvCUR(sv);
1507 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1508 if (newlen < minlen)
1510 #ifndef Perl_safesysmalloc_size
1511 newlen = PERL_STRLEN_ROUNDUP(newlen);
1513 if (SvLEN(sv) && s) {
1514 s = (char*)saferealloc(s, newlen);
1517 s = (char*)safemalloc(newlen);
1518 if (SvPVX_const(sv) && SvCUR(sv)) {
1519 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1523 #ifdef Perl_safesysmalloc_size
1524 /* Do this here, do it once, do it right, and then we will never get
1525 called back into sv_grow() unless there really is some growing
1527 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1529 SvLEN_set(sv, newlen);
1536 =for apidoc sv_setiv
1538 Copies an integer into the given SV, upgrading first if necessary.
1539 Does not handle 'set' magic. See also C<sv_setiv_mg>.
1545 Perl_sv_setiv(pTHX_ register SV *const sv, const IV i)
1549 PERL_ARGS_ASSERT_SV_SETIV;
1551 SV_CHECK_THINKFIRST_COW_DROP(sv);
1552 switch (SvTYPE(sv)) {
1555 sv_upgrade(sv, SVt_IV);
1558 sv_upgrade(sv, SVt_PVIV);
1562 if (!isGV_with_GP(sv))
1569 /* diag_listed_as: Can't coerce %s to %s in %s */
1570 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1574 (void)SvIOK_only(sv); /* validate number */
1580 =for apidoc sv_setiv_mg
1582 Like C<sv_setiv>, but also handles 'set' magic.
1588 Perl_sv_setiv_mg(pTHX_ register SV *const sv, const IV i)
1590 PERL_ARGS_ASSERT_SV_SETIV_MG;
1597 =for apidoc sv_setuv
1599 Copies an unsigned integer into the given SV, upgrading first if necessary.
1600 Does not handle 'set' magic. See also C<sv_setuv_mg>.
1606 Perl_sv_setuv(pTHX_ register SV *const sv, const UV u)
1608 PERL_ARGS_ASSERT_SV_SETUV;
1610 /* With these two if statements:
1611 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1614 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1616 If you wish to remove them, please benchmark to see what the effect is
1618 if (u <= (UV)IV_MAX) {
1619 sv_setiv(sv, (IV)u);
1628 =for apidoc sv_setuv_mg
1630 Like C<sv_setuv>, but also handles 'set' magic.
1636 Perl_sv_setuv_mg(pTHX_ register SV *const sv, const UV u)
1638 PERL_ARGS_ASSERT_SV_SETUV_MG;
1645 =for apidoc sv_setnv
1647 Copies a double into the given SV, upgrading first if necessary.
1648 Does not handle 'set' magic. See also C<sv_setnv_mg>.
1654 Perl_sv_setnv(pTHX_ register SV *const sv, const NV num)
1658 PERL_ARGS_ASSERT_SV_SETNV;
1660 SV_CHECK_THINKFIRST_COW_DROP(sv);
1661 switch (SvTYPE(sv)) {
1664 sv_upgrade(sv, SVt_NV);
1668 sv_upgrade(sv, SVt_PVNV);
1672 if (!isGV_with_GP(sv))
1679 /* diag_listed_as: Can't coerce %s to %s in %s */
1680 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1685 (void)SvNOK_only(sv); /* validate number */
1690 =for apidoc sv_setnv_mg
1692 Like C<sv_setnv>, but also handles 'set' magic.
1698 Perl_sv_setnv_mg(pTHX_ register SV *const sv, const NV num)
1700 PERL_ARGS_ASSERT_SV_SETNV_MG;
1706 /* Print an "isn't numeric" warning, using a cleaned-up,
1707 * printable version of the offending string
1711 S_not_a_number(pTHX_ SV *const sv)
1718 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1721 dsv = newSVpvs_flags("", SVs_TEMP);
1722 pv = sv_uni_display(dsv, sv, 10, UNI_DISPLAY_ISPRINT);
1725 const char * const limit = tmpbuf + sizeof(tmpbuf) - 8;
1726 /* each *s can expand to 4 chars + "...\0",
1727 i.e. need room for 8 chars */
1729 const char *s = SvPVX_const(sv);
1730 const char * const end = s + SvCUR(sv);
1731 for ( ; s < end && d < limit; s++ ) {
1733 if (ch & 128 && !isPRINT_LC(ch)) {
1742 else if (ch == '\r') {
1746 else if (ch == '\f') {
1750 else if (ch == '\\') {
1754 else if (ch == '\0') {
1758 else if (isPRINT_LC(ch))
1775 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1776 "Argument \"%s\" isn't numeric in %s", pv,
1779 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1780 "Argument \"%s\" isn't numeric", pv);
1784 =for apidoc looks_like_number
1786 Test if the content of an SV looks like a number (or is a number).
1787 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1788 non-numeric warning), even if your atof() doesn't grok them.
1794 Perl_looks_like_number(pTHX_ SV *const sv)
1796 register const char *sbegin;
1799 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1802 sbegin = SvPVX_const(sv);
1805 else if (SvPOKp(sv))
1806 sbegin = SvPV_const(sv, len);
1808 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1809 return grok_number(sbegin, len, NULL);
1813 S_glob_2number(pTHX_ GV * const gv)
1815 const U32 wasfake = SvFLAGS(gv) & SVf_FAKE;
1816 SV *const buffer = sv_newmortal();
1818 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1820 /* FAKE globs can get coerced, so need to turn this off temporarily if it
1823 gv_efullname3(buffer, gv, "*");
1824 SvFLAGS(gv) |= wasfake;
1826 /* We know that all GVs stringify to something that is not-a-number,
1827 so no need to test that. */
1828 if (ckWARN(WARN_NUMERIC))
1829 not_a_number(buffer);
1830 /* We just want something true to return, so that S_sv_2iuv_common
1831 can tail call us and return true. */
1835 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1836 until proven guilty, assume that things are not that bad... */
1841 As 64 bit platforms often have an NV that doesn't preserve all bits of
1842 an IV (an assumption perl has been based on to date) it becomes necessary
1843 to remove the assumption that the NV always carries enough precision to
1844 recreate the IV whenever needed, and that the NV is the canonical form.
1845 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1846 precision as a side effect of conversion (which would lead to insanity
1847 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1848 1) to distinguish between IV/UV/NV slots that have cached a valid
1849 conversion where precision was lost and IV/UV/NV slots that have a
1850 valid conversion which has lost no precision
1851 2) to ensure that if a numeric conversion to one form is requested that
1852 would lose precision, the precise conversion (or differently
1853 imprecise conversion) is also performed and cached, to prevent
1854 requests for different numeric formats on the same SV causing
1855 lossy conversion chains. (lossless conversion chains are perfectly
1860 SvIOKp is true if the IV slot contains a valid value
1861 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1862 SvNOKp is true if the NV slot contains a valid value
1863 SvNOK is true only if the NV value is accurate
1866 while converting from PV to NV, check to see if converting that NV to an
1867 IV(or UV) would lose accuracy over a direct conversion from PV to
1868 IV(or UV). If it would, cache both conversions, return NV, but mark
1869 SV as IOK NOKp (ie not NOK).
1871 While converting from PV to IV, check to see if converting that IV to an
1872 NV would lose accuracy over a direct conversion from PV to NV. If it
1873 would, cache both conversions, flag similarly.
1875 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
1876 correctly because if IV & NV were set NV *always* overruled.
1877 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
1878 changes - now IV and NV together means that the two are interchangeable:
1879 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
1881 The benefit of this is that operations such as pp_add know that if
1882 SvIOK is true for both left and right operands, then integer addition
1883 can be used instead of floating point (for cases where the result won't
1884 overflow). Before, floating point was always used, which could lead to
1885 loss of precision compared with integer addition.
1887 * making IV and NV equal status should make maths accurate on 64 bit
1889 * may speed up maths somewhat if pp_add and friends start to use
1890 integers when possible instead of fp. (Hopefully the overhead in
1891 looking for SvIOK and checking for overflow will not outweigh the
1892 fp to integer speedup)
1893 * will slow down integer operations (callers of SvIV) on "inaccurate"
1894 values, as the change from SvIOK to SvIOKp will cause a call into
1895 sv_2iv each time rather than a macro access direct to the IV slot
1896 * should speed up number->string conversion on integers as IV is
1897 favoured when IV and NV are equally accurate
1899 ####################################################################
1900 You had better be using SvIOK_notUV if you want an IV for arithmetic:
1901 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
1902 On the other hand, SvUOK is true iff UV.
1903 ####################################################################
1905 Your mileage will vary depending your CPU's relative fp to integer
1909 #ifndef NV_PRESERVES_UV
1910 # define IS_NUMBER_UNDERFLOW_IV 1
1911 # define IS_NUMBER_UNDERFLOW_UV 2
1912 # define IS_NUMBER_IV_AND_UV 2
1913 # define IS_NUMBER_OVERFLOW_IV 4
1914 # define IS_NUMBER_OVERFLOW_UV 5
1916 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
1918 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
1920 S_sv_2iuv_non_preserve(pTHX_ register SV *const sv
1928 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
1930 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));
1931 if (SvNVX(sv) < (NV)IV_MIN) {
1932 (void)SvIOKp_on(sv);
1934 SvIV_set(sv, IV_MIN);
1935 return IS_NUMBER_UNDERFLOW_IV;
1937 if (SvNVX(sv) > (NV)UV_MAX) {
1938 (void)SvIOKp_on(sv);
1941 SvUV_set(sv, UV_MAX);
1942 return IS_NUMBER_OVERFLOW_UV;
1944 (void)SvIOKp_on(sv);
1946 /* Can't use strtol etc to convert this string. (See truth table in
1948 if (SvNVX(sv) <= (UV)IV_MAX) {
1949 SvIV_set(sv, I_V(SvNVX(sv)));
1950 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
1951 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
1953 /* Integer is imprecise. NOK, IOKp */
1955 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
1958 SvUV_set(sv, U_V(SvNVX(sv)));
1959 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
1960 if (SvUVX(sv) == UV_MAX) {
1961 /* As we know that NVs don't preserve UVs, UV_MAX cannot
1962 possibly be preserved by NV. Hence, it must be overflow.
1964 return IS_NUMBER_OVERFLOW_UV;
1966 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
1968 /* Integer is imprecise. NOK, IOKp */
1970 return IS_NUMBER_OVERFLOW_IV;
1972 #endif /* !NV_PRESERVES_UV*/
1975 S_sv_2iuv_common(pTHX_ SV *const sv)
1979 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
1982 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
1983 * without also getting a cached IV/UV from it at the same time
1984 * (ie PV->NV conversion should detect loss of accuracy and cache
1985 * IV or UV at same time to avoid this. */
1986 /* IV-over-UV optimisation - choose to cache IV if possible */
1988 if (SvTYPE(sv) == SVt_NV)
1989 sv_upgrade(sv, SVt_PVNV);
1991 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
1992 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
1993 certainly cast into the IV range at IV_MAX, whereas the correct
1994 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
1996 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
1997 if (Perl_isnan(SvNVX(sv))) {
2003 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2004 SvIV_set(sv, I_V(SvNVX(sv)));
2005 if (SvNVX(sv) == (NV) SvIVX(sv)
2006 #ifndef NV_PRESERVES_UV
2007 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2008 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2009 /* Don't flag it as "accurately an integer" if the number
2010 came from a (by definition imprecise) NV operation, and
2011 we're outside the range of NV integer precision */
2015 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2017 /* scalar has trailing garbage, eg "42a" */
2019 DEBUG_c(PerlIO_printf(Perl_debug_log,
2020 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
2026 /* IV not precise. No need to convert from PV, as NV
2027 conversion would already have cached IV if it detected
2028 that PV->IV would be better than PV->NV->IV
2029 flags already correct - don't set public IOK. */
2030 DEBUG_c(PerlIO_printf(Perl_debug_log,
2031 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
2036 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2037 but the cast (NV)IV_MIN rounds to a the value less (more
2038 negative) than IV_MIN which happens to be equal to SvNVX ??
2039 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2040 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2041 (NV)UVX == NVX are both true, but the values differ. :-(
2042 Hopefully for 2s complement IV_MIN is something like
2043 0x8000000000000000 which will be exact. NWC */
2046 SvUV_set(sv, U_V(SvNVX(sv)));
2048 (SvNVX(sv) == (NV) SvUVX(sv))
2049 #ifndef NV_PRESERVES_UV
2050 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2051 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2052 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2053 /* Don't flag it as "accurately an integer" if the number
2054 came from a (by definition imprecise) NV operation, and
2055 we're outside the range of NV integer precision */
2061 DEBUG_c(PerlIO_printf(Perl_debug_log,
2062 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2068 else if (SvPOKp(sv) && SvLEN(sv)) {
2070 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2071 /* We want to avoid a possible problem when we cache an IV/ a UV which
2072 may be later translated to an NV, and the resulting NV is not
2073 the same as the direct translation of the initial string
2074 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2075 be careful to ensure that the value with the .456 is around if the
2076 NV value is requested in the future).
2078 This means that if we cache such an IV/a UV, we need to cache the
2079 NV as well. Moreover, we trade speed for space, and do not
2080 cache the NV if we are sure it's not needed.
2083 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2084 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2085 == IS_NUMBER_IN_UV) {
2086 /* It's definitely an integer, only upgrade to PVIV */
2087 if (SvTYPE(sv) < SVt_PVIV)
2088 sv_upgrade(sv, SVt_PVIV);
2090 } else if (SvTYPE(sv) < SVt_PVNV)
2091 sv_upgrade(sv, SVt_PVNV);
2093 /* If NVs preserve UVs then we only use the UV value if we know that
2094 we aren't going to call atof() below. If NVs don't preserve UVs
2095 then the value returned may have more precision than atof() will
2096 return, even though value isn't perfectly accurate. */
2097 if ((numtype & (IS_NUMBER_IN_UV
2098 #ifdef NV_PRESERVES_UV
2101 )) == IS_NUMBER_IN_UV) {
2102 /* This won't turn off the public IOK flag if it was set above */
2103 (void)SvIOKp_on(sv);
2105 if (!(numtype & IS_NUMBER_NEG)) {
2107 if (value <= (UV)IV_MAX) {
2108 SvIV_set(sv, (IV)value);
2110 /* it didn't overflow, and it was positive. */
2111 SvUV_set(sv, value);
2115 /* 2s complement assumption */
2116 if (value <= (UV)IV_MIN) {
2117 SvIV_set(sv, -(IV)value);
2119 /* Too negative for an IV. This is a double upgrade, but
2120 I'm assuming it will be rare. */
2121 if (SvTYPE(sv) < SVt_PVNV)
2122 sv_upgrade(sv, SVt_PVNV);
2126 SvNV_set(sv, -(NV)value);
2127 SvIV_set(sv, IV_MIN);
2131 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2132 will be in the previous block to set the IV slot, and the next
2133 block to set the NV slot. So no else here. */
2135 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2136 != IS_NUMBER_IN_UV) {
2137 /* It wasn't an (integer that doesn't overflow the UV). */
2138 SvNV_set(sv, Atof(SvPVX_const(sv)));
2140 if (! numtype && ckWARN(WARN_NUMERIC))
2143 #if defined(USE_LONG_DOUBLE)
2144 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" PERL_PRIgldbl ")\n",
2145 PTR2UV(sv), SvNVX(sv)));
2147 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"NVgf")\n",
2148 PTR2UV(sv), SvNVX(sv)));
2151 #ifdef NV_PRESERVES_UV
2152 (void)SvIOKp_on(sv);
2154 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2155 SvIV_set(sv, I_V(SvNVX(sv)));
2156 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2159 NOOP; /* Integer is imprecise. NOK, IOKp */
2161 /* UV will not work better than IV */
2163 if (SvNVX(sv) > (NV)UV_MAX) {
2165 /* Integer is inaccurate. NOK, IOKp, is UV */
2166 SvUV_set(sv, UV_MAX);
2168 SvUV_set(sv, U_V(SvNVX(sv)));
2169 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2170 NV preservse UV so can do correct comparison. */
2171 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2174 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2179 #else /* NV_PRESERVES_UV */
2180 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2181 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2182 /* The IV/UV slot will have been set from value returned by
2183 grok_number above. The NV slot has just been set using
2186 assert (SvIOKp(sv));
2188 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2189 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2190 /* Small enough to preserve all bits. */
2191 (void)SvIOKp_on(sv);
2193 SvIV_set(sv, I_V(SvNVX(sv)));
2194 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2196 /* Assumption: first non-preserved integer is < IV_MAX,
2197 this NV is in the preserved range, therefore: */
2198 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2200 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);
2204 0 0 already failed to read UV.
2205 0 1 already failed to read UV.
2206 1 0 you won't get here in this case. IV/UV
2207 slot set, public IOK, Atof() unneeded.
2208 1 1 already read UV.
2209 so there's no point in sv_2iuv_non_preserve() attempting
2210 to use atol, strtol, strtoul etc. */
2212 sv_2iuv_non_preserve (sv, numtype);
2214 sv_2iuv_non_preserve (sv);
2218 #endif /* NV_PRESERVES_UV */
2219 /* It might be more code efficient to go through the entire logic above
2220 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2221 gets complex and potentially buggy, so more programmer efficient
2222 to do it this way, by turning off the public flags: */
2224 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2228 if (isGV_with_GP(sv))
2229 return glob_2number(MUTABLE_GV(sv));
2231 if (!SvPADTMP(sv)) {
2232 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2235 if (SvTYPE(sv) < SVt_IV)
2236 /* Typically the caller expects that sv_any is not NULL now. */
2237 sv_upgrade(sv, SVt_IV);
2238 /* Return 0 from the caller. */
2245 =for apidoc sv_2iv_flags
2247 Return the integer value of an SV, doing any necessary string
2248 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2249 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2255 Perl_sv_2iv_flags(pTHX_ register SV *const sv, const I32 flags)
2260 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2261 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2262 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2263 In practice they are extremely unlikely to actually get anywhere
2264 accessible by user Perl code - the only way that I'm aware of is when
2265 a constant subroutine which is used as the second argument to index.
2267 if (flags & SV_GMAGIC)
2272 return I_V(SvNVX(sv));
2274 if (SvPOKp(sv) && SvLEN(sv)) {
2277 = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2279 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2280 == IS_NUMBER_IN_UV) {
2281 /* It's definitely an integer */
2282 if (numtype & IS_NUMBER_NEG) {
2283 if (value < (UV)IV_MIN)
2286 if (value < (UV)IV_MAX)
2291 if (ckWARN(WARN_NUMERIC))
2294 return I_V(Atof(SvPVX_const(sv)));
2299 assert(SvTYPE(sv) >= SVt_PVMG);
2300 /* This falls through to the report_uninit inside S_sv_2iuv_common. */
2301 } else if (SvTHINKFIRST(sv)) {
2306 if (flags & SV_SKIP_OVERLOAD)
2308 tmpstr = AMG_CALLunary(sv, numer_amg);
2309 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2310 return SvIV(tmpstr);
2313 return PTR2IV(SvRV(sv));
2316 sv_force_normal_flags(sv, 0);
2318 if (SvREADONLY(sv) && !SvOK(sv)) {
2319 if (ckWARN(WARN_UNINITIALIZED))
2325 if (S_sv_2iuv_common(aTHX_ sv))
2328 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2329 PTR2UV(sv),SvIVX(sv)));
2330 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2334 =for apidoc sv_2uv_flags
2336 Return the unsigned integer value of an SV, doing any necessary string
2337 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2338 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2344 Perl_sv_2uv_flags(pTHX_ register SV *const sv, const I32 flags)
2349 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2350 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2351 the same flag bit as SVf_IVisUV, so must not let them cache IVs. */
2352 if (flags & SV_GMAGIC)
2357 return U_V(SvNVX(sv));
2358 if (SvPOKp(sv) && SvLEN(sv)) {
2361 = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2363 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2364 == IS_NUMBER_IN_UV) {
2365 /* It's definitely an integer */
2366 if (!(numtype & IS_NUMBER_NEG))
2370 if (ckWARN(WARN_NUMERIC))
2373 return U_V(Atof(SvPVX_const(sv)));
2378 assert(SvTYPE(sv) >= SVt_PVMG);
2379 /* This falls through to the report_uninit inside S_sv_2iuv_common. */
2380 } else if (SvTHINKFIRST(sv)) {
2385 if (flags & SV_SKIP_OVERLOAD)
2387 tmpstr = AMG_CALLunary(sv, numer_amg);
2388 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2389 return SvUV(tmpstr);
2392 return PTR2UV(SvRV(sv));
2395 sv_force_normal_flags(sv, 0);
2397 if (SvREADONLY(sv) && !SvOK(sv)) {
2398 if (ckWARN(WARN_UNINITIALIZED))
2404 if (S_sv_2iuv_common(aTHX_ sv))
2408 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2409 PTR2UV(sv),SvUVX(sv)));
2410 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2414 =for apidoc sv_2nv_flags
2416 Return the num value of an SV, doing any necessary string or integer
2417 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2418 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2424 Perl_sv_2nv_flags(pTHX_ register SV *const sv, const I32 flags)
2429 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2430 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2431 the same flag bit as SVf_IVisUV, so must not let them cache NVs. */
2432 if (flags & SV_GMAGIC)
2436 if ((SvPOKp(sv) && SvLEN(sv)) && !SvIOKp(sv)) {
2437 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2438 !grok_number(SvPVX_const(sv), SvCUR(sv), NULL))
2440 return Atof(SvPVX_const(sv));
2444 return (NV)SvUVX(sv);
2446 return (NV)SvIVX(sv);
2451 assert(SvTYPE(sv) >= SVt_PVMG);
2452 /* This falls through to the report_uninit near the end of the
2454 } else if (SvTHINKFIRST(sv)) {
2459 if (flags & SV_SKIP_OVERLOAD)
2461 tmpstr = AMG_CALLunary(sv, numer_amg);
2462 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2463 return SvNV(tmpstr);
2466 return PTR2NV(SvRV(sv));
2469 sv_force_normal_flags(sv, 0);
2471 if (SvREADONLY(sv) && !SvOK(sv)) {
2472 if (ckWARN(WARN_UNINITIALIZED))
2477 if (SvTYPE(sv) < SVt_NV) {
2478 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2479 sv_upgrade(sv, SVt_NV);
2480 #ifdef USE_LONG_DOUBLE
2482 STORE_NUMERIC_LOCAL_SET_STANDARD();
2483 PerlIO_printf(Perl_debug_log,
2484 "0x%"UVxf" num(%" PERL_PRIgldbl ")\n",
2485 PTR2UV(sv), SvNVX(sv));
2486 RESTORE_NUMERIC_LOCAL();
2490 STORE_NUMERIC_LOCAL_SET_STANDARD();
2491 PerlIO_printf(Perl_debug_log, "0x%"UVxf" num(%"NVgf")\n",
2492 PTR2UV(sv), SvNVX(sv));
2493 RESTORE_NUMERIC_LOCAL();
2497 else if (SvTYPE(sv) < SVt_PVNV)
2498 sv_upgrade(sv, SVt_PVNV);
2503 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2504 #ifdef NV_PRESERVES_UV
2510 /* Only set the public NV OK flag if this NV preserves the IV */
2511 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2513 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2514 : (SvIVX(sv) == I_V(SvNVX(sv))))
2520 else if (SvPOKp(sv) && SvLEN(sv)) {
2522 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2523 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2525 #ifdef NV_PRESERVES_UV
2526 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2527 == IS_NUMBER_IN_UV) {
2528 /* It's definitely an integer */
2529 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2531 SvNV_set(sv, Atof(SvPVX_const(sv)));
2537 SvNV_set(sv, Atof(SvPVX_const(sv)));
2538 /* Only set the public NV OK flag if this NV preserves the value in
2539 the PV at least as well as an IV/UV would.
2540 Not sure how to do this 100% reliably. */
2541 /* if that shift count is out of range then Configure's test is
2542 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2544 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2545 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2546 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2547 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2548 /* Can't use strtol etc to convert this string, so don't try.
2549 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2552 /* value has been set. It may not be precise. */
2553 if ((numtype & IS_NUMBER_NEG) && (value > (UV)IV_MIN)) {
2554 /* 2s complement assumption for (UV)IV_MIN */
2555 SvNOK_on(sv); /* Integer is too negative. */
2560 if (numtype & IS_NUMBER_NEG) {
2561 SvIV_set(sv, -(IV)value);
2562 } else if (value <= (UV)IV_MAX) {
2563 SvIV_set(sv, (IV)value);
2565 SvUV_set(sv, value);
2569 if (numtype & IS_NUMBER_NOT_INT) {
2570 /* I believe that even if the original PV had decimals,
2571 they are lost beyond the limit of the FP precision.
2572 However, neither is canonical, so both only get p
2573 flags. NWC, 2000/11/25 */
2574 /* Both already have p flags, so do nothing */
2576 const NV nv = SvNVX(sv);
2577 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2578 if (SvIVX(sv) == I_V(nv)) {
2581 /* It had no "." so it must be integer. */
2585 /* between IV_MAX and NV(UV_MAX).
2586 Could be slightly > UV_MAX */
2588 if (numtype & IS_NUMBER_NOT_INT) {
2589 /* UV and NV both imprecise. */
2591 const UV nv_as_uv = U_V(nv);
2593 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2602 /* It might be more code efficient to go through the entire logic above
2603 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2604 gets complex and potentially buggy, so more programmer efficient
2605 to do it this way, by turning off the public flags: */
2607 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2608 #endif /* NV_PRESERVES_UV */
2611 if (isGV_with_GP(sv)) {
2612 glob_2number(MUTABLE_GV(sv));
2616 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2618 assert (SvTYPE(sv) >= SVt_NV);
2619 /* Typically the caller expects that sv_any is not NULL now. */
2620 /* XXX Ilya implies that this is a bug in callers that assume this
2621 and ideally should be fixed. */
2624 #if defined(USE_LONG_DOUBLE)
2626 STORE_NUMERIC_LOCAL_SET_STANDARD();
2627 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" PERL_PRIgldbl ")\n",
2628 PTR2UV(sv), SvNVX(sv));
2629 RESTORE_NUMERIC_LOCAL();
2633 STORE_NUMERIC_LOCAL_SET_STANDARD();
2634 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 1nv(%"NVgf")\n",
2635 PTR2UV(sv), SvNVX(sv));
2636 RESTORE_NUMERIC_LOCAL();
2645 Return an SV with the numeric value of the source SV, doing any necessary
2646 reference or overload conversion. You must use the C<SvNUM(sv)> macro to
2647 access this function.
2653 Perl_sv_2num(pTHX_ register SV *const sv)
2655 PERL_ARGS_ASSERT_SV_2NUM;
2660 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2661 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2662 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2663 return sv_2num(tmpsv);
2665 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2668 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2669 * UV as a string towards the end of buf, and return pointers to start and
2672 * We assume that buf is at least TYPE_CHARS(UV) long.
2676 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2678 char *ptr = buf + TYPE_CHARS(UV);
2679 char * const ebuf = ptr;
2682 PERL_ARGS_ASSERT_UIV_2BUF;
2694 *--ptr = '0' + (char)(uv % 10);
2703 =for apidoc sv_2pv_flags
2705 Returns a pointer to the string value of an SV, and sets *lp to its length.
2706 If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a string
2708 Normally invoked via the C<SvPV_flags> macro. C<sv_2pv()> and C<sv_2pv_nomg>
2709 usually end up here too.
2715 Perl_sv_2pv_flags(pTHX_ register SV *const sv, STRLEN *const lp, const I32 flags)
2725 if (SvGMAGICAL(sv)) {
2726 if (flags & SV_GMAGIC)
2731 if (flags & SV_MUTABLE_RETURN)
2732 return SvPVX_mutable(sv);
2733 if (flags & SV_CONST_RETURN)
2734 return (char *)SvPVX_const(sv);
2737 if (SvIOKp(sv) || SvNOKp(sv)) {
2738 char tbuf[64]; /* Must fit sprintf/Gconvert of longest IV/NV */
2743 ? my_snprintf(tbuf, sizeof(tbuf), "%"UVuf, (UV)SvUVX(sv))
2744 : my_snprintf(tbuf, sizeof(tbuf), "%"IVdf, (IV)SvIVX(sv));
2745 } else if(SvNVX(sv) == 0.0) {
2750 Gconvert(SvNVX(sv), NV_DIG, 0, tbuf);
2757 SvUPGRADE(sv, SVt_PV);
2760 s = SvGROW_mutable(sv, len + 1);
2763 return (char*)memcpy(s, tbuf, len + 1);
2769 assert(SvTYPE(sv) >= SVt_PVMG);
2770 /* This falls through to the report_uninit near the end of the
2772 } else if (SvTHINKFIRST(sv)) {
2777 if (flags & SV_SKIP_OVERLOAD)
2779 tmpstr = AMG_CALLunary(sv, string_amg);
2780 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2781 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2783 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2787 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2788 if (flags & SV_CONST_RETURN) {
2789 pv = (char *) SvPVX_const(tmpstr);
2791 pv = (flags & SV_MUTABLE_RETURN)
2792 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2795 *lp = SvCUR(tmpstr);
2797 pv = sv_2pv_flags(tmpstr, lp, flags);
2810 SV *const referent = SvRV(sv);
2814 retval = buffer = savepvn("NULLREF", len);
2815 } else if (SvTYPE(referent) == SVt_REGEXP) {
2816 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2821 /* If the regex is UTF-8 we want the containing scalar to
2822 have an UTF-8 flag too */
2828 if ((seen_evals = RX_SEEN_EVALS(re)))
2829 PL_reginterp_cnt += seen_evals;
2832 *lp = RX_WRAPLEN(re);
2834 return RX_WRAPPED(re);
2836 const char *const typestr = sv_reftype(referent, 0);
2837 const STRLEN typelen = strlen(typestr);
2838 UV addr = PTR2UV(referent);
2839 const char *stashname = NULL;
2840 STRLEN stashnamelen = 0; /* hush, gcc */
2841 const char *buffer_end;
2843 if (SvOBJECT(referent)) {
2844 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
2847 stashname = HEK_KEY(name);
2848 stashnamelen = HEK_LEN(name);
2850 if (HEK_UTF8(name)) {
2856 stashname = "__ANON__";
2859 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
2860 + 2 * sizeof(UV) + 2 /* )\0 */;
2862 len = typelen + 3 /* (0x */
2863 + 2 * sizeof(UV) + 2 /* )\0 */;
2866 Newx(buffer, len, char);
2867 buffer_end = retval = buffer + len;
2869 /* Working backwards */
2873 *--retval = PL_hexdigit[addr & 15];
2874 } while (addr >>= 4);
2880 memcpy(retval, typestr, typelen);
2884 retval -= stashnamelen;
2885 memcpy(retval, stashname, stashnamelen);
2887 /* retval may not necessarily have reached the start of the
2889 assert (retval >= buffer);
2891 len = buffer_end - retval - 1; /* -1 for that \0 */
2899 if (SvREADONLY(sv) && !SvOK(sv)) {
2902 if (flags & SV_UNDEF_RETURNS_NULL)
2904 if (ckWARN(WARN_UNINITIALIZED))
2909 if (SvIOK(sv) || ((SvIOKp(sv) && !SvNOKp(sv)))) {
2910 /* I'm assuming that if both IV and NV are equally valid then
2911 converting the IV is going to be more efficient */
2912 const U32 isUIOK = SvIsUV(sv);
2913 char buf[TYPE_CHARS(UV)];
2917 if (SvTYPE(sv) < SVt_PVIV)
2918 sv_upgrade(sv, SVt_PVIV);
2919 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
2921 /* inlined from sv_setpvn */
2922 s = SvGROW_mutable(sv, len + 1);
2923 Move(ptr, s, len, char);
2927 else if (SvNOKp(sv)) {
2928 if (SvTYPE(sv) < SVt_PVNV)
2929 sv_upgrade(sv, SVt_PVNV);
2930 if (SvNVX(sv) == 0.0) {
2931 s = SvGROW_mutable(sv, 2);
2936 /* The +20 is pure guesswork. Configure test needed. --jhi */
2937 s = SvGROW_mutable(sv, NV_DIG + 20);
2938 /* some Xenix systems wipe out errno here */
2939 Gconvert(SvNVX(sv), NV_DIG, 0, s);
2949 if (isGV_with_GP(sv)) {
2950 GV *const gv = MUTABLE_GV(sv);
2951 const U32 wasfake = SvFLAGS(gv) & SVf_FAKE;
2952 SV *const buffer = sv_newmortal();
2954 /* FAKE globs can get coerced, so need to turn this off temporarily
2957 gv_efullname3(buffer, gv, "*");
2958 SvFLAGS(gv) |= wasfake;
2960 if (SvPOK(buffer)) {
2962 *lp = SvCUR(buffer);
2964 if ( SvUTF8(buffer) ) SvUTF8_on(sv);
2965 return SvPVX(buffer);
2976 if (flags & SV_UNDEF_RETURNS_NULL)
2978 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2980 if (SvTYPE(sv) < SVt_PV)
2981 /* Typically the caller expects that sv_any is not NULL now. */
2982 sv_upgrade(sv, SVt_PV);
2986 const STRLEN len = s - SvPVX_const(sv);
2992 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
2993 PTR2UV(sv),SvPVX_const(sv)));
2994 if (flags & SV_CONST_RETURN)
2995 return (char *)SvPVX_const(sv);
2996 if (flags & SV_MUTABLE_RETURN)
2997 return SvPVX_mutable(sv);
3002 =for apidoc sv_copypv
3004 Copies a stringified representation of the source SV into the
3005 destination SV. Automatically performs any necessary mg_get and
3006 coercion of numeric values into strings. Guaranteed to preserve
3007 UTF8 flag even from overloaded objects. Similar in nature to
3008 sv_2pv[_flags] but operates directly on an SV instead of just the
3009 string. Mostly uses sv_2pv_flags to do its work, except when that
3010 would lose the UTF-8'ness of the PV.
3016 Perl_sv_copypv(pTHX_ SV *const dsv, register SV *const ssv)
3019 const char * const s = SvPV_const(ssv,len);
3021 PERL_ARGS_ASSERT_SV_COPYPV;
3023 sv_setpvn(dsv,s,len);
3031 =for apidoc sv_2pvbyte
3033 Return a pointer to the byte-encoded representation of the SV, and set *lp
3034 to its length. May cause the SV to be downgraded from UTF-8 as a
3037 Usually accessed via the C<SvPVbyte> macro.
3043 Perl_sv_2pvbyte(pTHX_ register SV *const sv, STRLEN *const lp)
3045 PERL_ARGS_ASSERT_SV_2PVBYTE;
3048 sv_utf8_downgrade(sv,0);
3049 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3053 =for apidoc sv_2pvutf8
3055 Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3056 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3058 Usually accessed via the C<SvPVutf8> macro.
3064 Perl_sv_2pvutf8(pTHX_ register SV *const sv, STRLEN *const lp)
3066 PERL_ARGS_ASSERT_SV_2PVUTF8;
3068 sv_utf8_upgrade(sv);
3069 return lp ? SvPV(sv,*lp) : SvPV_nolen(sv);
3074 =for apidoc sv_2bool
3076 This macro is only used by sv_true() or its macro equivalent, and only if
3077 the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3078 It calls sv_2bool_flags with the SV_GMAGIC flag.
3080 =for apidoc sv_2bool_flags
3082 This function is only used by sv_true() and friends, and only if
3083 the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
3084 contain SV_GMAGIC, then it does an mg_get() first.
3091 Perl_sv_2bool_flags(pTHX_ register SV *const sv, const I32 flags)
3095 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3097 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3103 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3104 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
3105 return cBOOL(SvTRUE(tmpsv));
3107 return SvRV(sv) != 0;
3110 register XPV* const Xpvtmp = (XPV*)SvANY(sv);
3112 (*sv->sv_u.svu_pv > '0' ||
3113 Xpvtmp->xpv_cur > 1 ||
3114 (Xpvtmp->xpv_cur && *sv->sv_u.svu_pv != '0')))
3121 return SvIVX(sv) != 0;
3124 return SvNVX(sv) != 0.0;
3126 if (isGV_with_GP(sv))
3136 =for apidoc sv_utf8_upgrade
3138 Converts the PV of an SV to its UTF-8-encoded form.
3139 Forces the SV to string form if it is not already.
3140 Will C<mg_get> on C<sv> if appropriate.
3141 Always sets the SvUTF8 flag to avoid future validity checks even
3142 if the whole string is the same in UTF-8 as not.
3143 Returns the number of bytes in the converted string
3145 This is not as a general purpose byte encoding to Unicode interface:
3146 use the Encode extension for that.
3148 =for apidoc sv_utf8_upgrade_nomg
3150 Like sv_utf8_upgrade, but doesn't do magic on C<sv>
3152 =for apidoc sv_utf8_upgrade_flags
3154 Converts the PV of an SV to its UTF-8-encoded form.
3155 Forces the SV to string form if it is not already.
3156 Always sets the SvUTF8 flag to avoid future validity checks even
3157 if all the bytes are invariant in UTF-8. If C<flags> has C<SV_GMAGIC> bit set,
3158 will C<mg_get> on C<sv> if appropriate, else not.
3159 Returns the number of bytes in the converted string
3160 C<sv_utf8_upgrade> and
3161 C<sv_utf8_upgrade_nomg> are implemented in terms of this function.
3163 This is not as a general purpose byte encoding to Unicode interface:
3164 use the Encode extension for that.
3168 The grow version is currently not externally documented. It adds a parameter,
3169 extra, which is the number of unused bytes the string of 'sv' is guaranteed to
3170 have free after it upon return. This allows the caller to reserve extra space
3171 that it intends to fill, to avoid extra grows.
3173 Also externally undocumented for the moment is the flag SV_FORCE_UTF8_UPGRADE,
3174 which can be used to tell this function to not first check to see if there are
3175 any characters that are different in UTF-8 (variant characters) which would
3176 force it to allocate a new string to sv, but to assume there are. Typically
3177 this flag is used by a routine that has already parsed the string to find that
3178 there are such characters, and passes this information on so that the work
3179 doesn't have to be repeated.
3181 (One might think that the calling routine could pass in the position of the
3182 first such variant, so it wouldn't have to be found again. But that is not the
3183 case, because typically when the caller is likely to use this flag, it won't be
3184 calling this routine unless it finds something that won't fit into a byte.
3185 Otherwise it tries to not upgrade and just use bytes. But some things that
3186 do fit into a byte are variants in utf8, and the caller may not have been
3187 keeping track of these.)
3189 If the routine itself changes the string, it adds a trailing NUL. Such a NUL
3190 isn't guaranteed due to having other routines do the work in some input cases,
3191 or if the input is already flagged as being in utf8.
3193 The speed of this could perhaps be improved for many cases if someone wanted to
3194 write a fast function that counts the number of variant characters in a string,
3195 especially if it could return the position of the first one.
3200 Perl_sv_utf8_upgrade_flags_grow(pTHX_ register SV *const sv, const I32 flags, STRLEN extra)
3204 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3206 if (sv == &PL_sv_undef)
3210 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3211 (void) sv_2pv_flags(sv,&len, flags);
3213 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3217 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3222 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3227 sv_force_normal_flags(sv, 0);
3230 if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING)) {
3231 sv_recode_to_utf8(sv, PL_encoding);
3232 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3236 if (SvCUR(sv) == 0) {
3237 if (extra) SvGROW(sv, extra);
3238 } else { /* Assume Latin-1/EBCDIC */
3239 /* This function could be much more efficient if we
3240 * had a FLAG in SVs to signal if there are any variant
3241 * chars in the PV. Given that there isn't such a flag
3242 * make the loop as fast as possible (although there are certainly ways
3243 * to speed this up, eg. through vectorization) */
3244 U8 * s = (U8 *) SvPVX_const(sv);
3245 U8 * e = (U8 *) SvEND(sv);
3247 STRLEN two_byte_count = 0;
3249 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3251 /* See if really will need to convert to utf8. We mustn't rely on our
3252 * incoming SV being well formed and having a trailing '\0', as certain
3253 * code in pp_formline can send us partially built SVs. */
3257 if (NATIVE_IS_INVARIANT(ch)) continue;
3259 t--; /* t already incremented; re-point to first variant */
3264 /* utf8 conversion not needed because all are invariants. Mark as
3265 * UTF-8 even if no variant - saves scanning loop */
3271 /* Here, the string should be converted to utf8, either because of an
3272 * input flag (two_byte_count = 0), or because a character that
3273 * requires 2 bytes was found (two_byte_count = 1). t points either to
3274 * the beginning of the string (if we didn't examine anything), or to
3275 * the first variant. In either case, everything from s to t - 1 will
3276 * occupy only 1 byte each on output.
3278 * There are two main ways to convert. One is to create a new string
3279 * and go through the input starting from the beginning, appending each
3280 * converted value onto the new string as we go along. It's probably
3281 * best to allocate enough space in the string for the worst possible
3282 * case rather than possibly running out of space and having to
3283 * reallocate and then copy what we've done so far. Since everything
3284 * from s to t - 1 is invariant, the destination can be initialized
3285 * with these using a fast memory copy
3287 * The other way is to figure out exactly how big the string should be
3288 * by parsing the entire input. Then you don't have to make it big
3289 * enough to handle the worst possible case, and more importantly, if
3290 * the string you already have is large enough, you don't have to
3291 * allocate a new string, you can copy the last character in the input
3292 * string to the final position(s) that will be occupied by the
3293 * converted string and go backwards, stopping at t, since everything
3294 * before that is invariant.
3296 * There are advantages and disadvantages to each method.
3298 * In the first method, we can allocate a new string, do the memory
3299 * copy from the s to t - 1, and then proceed through the rest of the
3300 * string byte-by-byte.
3302 * In the second method, we proceed through the rest of the input
3303 * string just calculating how big the converted string will be. Then
3304 * there are two cases:
3305 * 1) if the string has enough extra space to handle the converted
3306 * value. We go backwards through the string, converting until we
3307 * get to the position we are at now, and then stop. If this
3308 * position is far enough along in the string, this method is
3309 * faster than the other method. If the memory copy were the same
3310 * speed as the byte-by-byte loop, that position would be about
3311 * half-way, as at the half-way mark, parsing to the end and back
3312 * is one complete string's parse, the same amount as starting
3313 * over and going all the way through. Actually, it would be
3314 * somewhat less than half-way, as it's faster to just count bytes
3315 * than to also copy, and we don't have the overhead of allocating
3316 * a new string, changing the scalar to use it, and freeing the
3317 * existing one. But if the memory copy is fast, the break-even
3318 * point is somewhere after half way. The counting loop could be
3319 * sped up by vectorization, etc, to move the break-even point
3320 * further towards the beginning.
3321 * 2) if the string doesn't have enough space to handle the converted
3322 * value. A new string will have to be allocated, and one might
3323 * as well, given that, start from the beginning doing the first
3324 * method. We've spent extra time parsing the string and in
3325 * exchange all we've gotten is that we know precisely how big to
3326 * make the new one. Perl is more optimized for time than space,
3327 * so this case is a loser.
3328 * So what I've decided to do is not use the 2nd method unless it is
3329 * guaranteed that a new string won't have to be allocated, assuming
3330 * the worst case. I also decided not to put any more conditions on it
3331 * than this, for now. It seems likely that, since the worst case is
3332 * twice as big as the unknown portion of the string (plus 1), we won't
3333 * be guaranteed enough space, causing us to go to the first method,
3334 * unless the string is short, or the first variant character is near
3335 * the end of it. In either of these cases, it seems best to use the
3336 * 2nd method. The only circumstance I can think of where this would
3337 * be really slower is if the string had once had much more data in it
3338 * than it does now, but there is still a substantial amount in it */
3341 STRLEN invariant_head = t - s;
3342 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3343 if (SvLEN(sv) < size) {
3345 /* Here, have decided to allocate a new string */
3350 Newx(dst, size, U8);
3352 /* If no known invariants at the beginning of the input string,
3353 * set so starts from there. Otherwise, can use memory copy to
3354 * get up to where we are now, and then start from here */
3356 if (invariant_head <= 0) {
3359 Copy(s, dst, invariant_head, char);
3360 d = dst + invariant_head;
3364 const UV uv = NATIVE8_TO_UNI(*t++);
3365 if (UNI_IS_INVARIANT(uv))
3366 *d++ = (U8)UNI_TO_NATIVE(uv);
3368 *d++ = (U8)UTF8_EIGHT_BIT_HI(uv);
3369 *d++ = (U8)UTF8_EIGHT_BIT_LO(uv);
3373 SvPV_free(sv); /* No longer using pre-existing string */
3374 SvPV_set(sv, (char*)dst);
3375 SvCUR_set(sv, d - dst);
3376 SvLEN_set(sv, size);
3379 /* Here, have decided to get the exact size of the string.
3380 * Currently this happens only when we know that there is
3381 * guaranteed enough space to fit the converted string, so
3382 * don't have to worry about growing. If two_byte_count is 0,
3383 * then t points to the first byte of the string which hasn't
3384 * been examined yet. Otherwise two_byte_count is 1, and t
3385 * points to the first byte in the string that will expand to
3386 * two. Depending on this, start examining at t or 1 after t.
3389 U8 *d = t + two_byte_count;
3392 /* Count up the remaining bytes that expand to two */
3395 const U8 chr = *d++;
3396 if (! NATIVE_IS_INVARIANT(chr)) two_byte_count++;
3399 /* The string will expand by just the number of bytes that
3400 * occupy two positions. But we are one afterwards because of
3401 * the increment just above. This is the place to put the
3402 * trailing NUL, and to set the length before we decrement */
3404 d += two_byte_count;
3405 SvCUR_set(sv, d - s);
3409 /* Having decremented d, it points to the position to put the
3410 * very last byte of the expanded string. Go backwards through
3411 * the string, copying and expanding as we go, stopping when we
3412 * get to the part that is invariant the rest of the way down */
3416 const U8 ch = NATIVE8_TO_UNI(*e--);
3417 if (UNI_IS_INVARIANT(ch)) {
3418 *d-- = UNI_TO_NATIVE(ch);
3420 *d-- = (U8)UTF8_EIGHT_BIT_LO(ch);
3421 *d-- = (U8)UTF8_EIGHT_BIT_HI(ch);
3426 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3427 /* Update pos. We do it at the end rather than during
3428 * the upgrade, to avoid slowing down the common case
3429 * (upgrade without pos) */
3430 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3432 I32 pos = mg->mg_len;
3433 if (pos > 0 && (U32)pos > invariant_head) {
3434 U8 *d = (U8*) SvPVX(sv) + invariant_head;
3435 STRLEN n = (U32)pos - invariant_head;
3437 if (UTF8_IS_START(*d))
3442 mg->mg_len = d - (U8*)SvPVX(sv);
3445 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3446 magic_setutf8(sv,mg); /* clear UTF8 cache */
3451 /* Mark as UTF-8 even if no variant - saves scanning loop */
3457 =for apidoc sv_utf8_downgrade
3459 Attempts to convert the PV of an SV from characters to bytes.
3460 If the PV contains a character that cannot fit
3461 in a byte, this conversion will fail;
3462 in this case, either returns false or, if C<fail_ok> is not
3465 This is not as a general purpose Unicode to byte encoding interface:
3466 use the Encode extension for that.
3472 Perl_sv_utf8_downgrade(pTHX_ register SV *const sv, const bool fail_ok)
3476 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3478 if (SvPOKp(sv) && SvUTF8(sv)) {
3482 int mg_flags = SV_GMAGIC;
3485 sv_force_normal_flags(sv, 0);
3487 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3489 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3491 I32 pos = mg->mg_len;
3493 sv_pos_b2u(sv, &pos);
3494 mg_flags = 0; /* sv_pos_b2u does get magic */
3498 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3499 magic_setutf8(sv,mg); /* clear UTF8 cache */
3502 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3504 if (!utf8_to_bytes(s, &len)) {
3509 Perl_croak(aTHX_ "Wide character in %s",
3512 Perl_croak(aTHX_ "Wide character");
3523 =for apidoc sv_utf8_encode
3525 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3526 flag off so that it looks like octets again.
3532 Perl_sv_utf8_encode(pTHX_ register SV *const sv)
3534 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3537 sv_force_normal_flags(sv, 0);
3539 if (SvREADONLY(sv)) {
3540 Perl_croak_no_modify(aTHX);
3542 (void) sv_utf8_upgrade(sv);
3547 =for apidoc sv_utf8_decode
3549 If the PV of the SV is an octet sequence in UTF-8
3550 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3551 so that it looks like a character. If the PV contains only single-byte
3552 characters, the C<SvUTF8> flag stays off.
3553 Scans PV for validity and returns false if the PV is invalid UTF-8.
3559 Perl_sv_utf8_decode(pTHX_ register SV *const sv)
3561 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3564 const U8 *start, *c;
3567 /* The octets may have got themselves encoded - get them back as
3570 if (!sv_utf8_downgrade(sv, TRUE))
3573 /* it is actually just a matter of turning the utf8 flag on, but
3574 * we want to make sure everything inside is valid utf8 first.
3576 c = start = (const U8 *) SvPVX_const(sv);
3577 if (!is_utf8_string(c, SvCUR(sv)+1))
3579 e = (const U8 *) SvEND(sv);
3582 if (!UTF8_IS_INVARIANT(ch)) {
3587 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3588 /* adjust pos to the start of a UTF8 char sequence */
3589 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3591 I32 pos = mg->mg_len;
3593 for (c = start + pos; c > start; c--) {
3594 if (UTF8_IS_START(*c))
3597 mg->mg_len = c - start;
3600 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3601 magic_setutf8(sv,mg); /* clear UTF8 cache */
3608 =for apidoc sv_setsv
3610 Copies the contents of the source SV C<ssv> into the destination SV
3611 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3612 function if the source SV needs to be reused. Does not handle 'set' magic.
3613 Loosely speaking, it performs a copy-by-value, obliterating any previous
3614 content of the destination.
3616 You probably want to use one of the assortment of wrappers, such as
3617 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3618 C<SvSetMagicSV_nosteal>.
3620 =for apidoc sv_setsv_flags
3622 Copies the contents of the source SV C<ssv> into the destination SV
3623 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3624 function if the source SV needs to be reused. Does not handle 'set' magic.
3625 Loosely speaking, it performs a copy-by-value, obliterating any previous
3626 content of the destination.
3627 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3628 C<ssv> if appropriate, else not. If the C<flags> parameter has the
3629 C<NOSTEAL> bit set then the buffers of temps will not be stolen. <sv_setsv>
3630 and C<sv_setsv_nomg> are implemented in terms of this function.
3632 You probably want to use one of the assortment of wrappers, such as
3633 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3634 C<SvSetMagicSV_nosteal>.
3636 This is the primary function for copying scalars, and most other
3637 copy-ish functions and macros use this underneath.
3643 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3645 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3646 HV *old_stash = NULL;
3648 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3650 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3651 const char * const name = GvNAME(sstr);
3652 const STRLEN len = GvNAMELEN(sstr);
3654 if (dtype >= SVt_PV) {
3660 SvUPGRADE(dstr, SVt_PVGV);
3661 (void)SvOK_off(dstr);
3662 /* FIXME - why are we doing this, then turning it off and on again
3664 isGV_with_GP_on(dstr);
3666 GvSTASH(dstr) = GvSTASH(sstr);
3668 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3669 gv_name_set(MUTABLE_GV(dstr), name, len,
3670 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3671 SvFAKE_on(dstr); /* can coerce to non-glob */
3674 if(GvGP(MUTABLE_GV(sstr))) {
3675 /* If source has method cache entry, clear it */
3677 SvREFCNT_dec(GvCV(sstr));
3678 GvCV_set(sstr, NULL);
3681 /* If source has a real method, then a method is
3684 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3690 /* If dest already had a real method, that's a change as well */
3692 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3693 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3698 /* We don’t need to check the name of the destination if it was not a
3699 glob to begin with. */
3700 if(dtype == SVt_PVGV) {
3701 const char * const name = GvNAME((const GV *)dstr);
3704 /* The stash may have been detached from the symbol table, so
3706 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3707 && GvAV((const GV *)sstr)
3711 const STRLEN len = GvNAMELEN(dstr);
3712 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3713 || (len == 1 && name[0] == ':')) {
3716 /* Set aside the old stash, so we can reset isa caches on
3718 if((old_stash = GvHV(dstr)))
3719 /* Make sure we do not lose it early. */
3720 SvREFCNT_inc_simple_void_NN(
3721 sv_2mortal((SV *)old_stash)
3727 gp_free(MUTABLE_GV(dstr));
3728 isGV_with_GP_off(dstr);
3729 (void)SvOK_off(dstr);
3730 isGV_with_GP_on(dstr);
3731 GvINTRO_off(dstr); /* one-shot flag */
3732 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3733 if (SvTAINTED(sstr))
3735 if (GvIMPORTED(dstr) != GVf_IMPORTED
3736 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3738 GvIMPORTED_on(dstr);
3741 if(mro_changes == 2) {
3743 SV * const sref = (SV *)GvAV((const GV *)dstr);
3744 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3745 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3746 AV * const ary = newAV();
3747 av_push(ary, mg->mg_obj); /* takes the refcount */
3748 mg->mg_obj = (SV *)ary;
3750 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3752 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3753 mro_isa_changed_in(GvSTASH(dstr));
3755 else if(mro_changes == 3) {
3756 HV * const stash = GvHV(dstr);
3757 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
3763 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
3768 S_glob_assign_ref(pTHX_ SV *const dstr, SV *const sstr)
3770 SV * const sref = SvREFCNT_inc(SvRV(sstr));
3772 const int intro = GvINTRO(dstr);
3775 const U32 stype = SvTYPE(sref);
3777 PERL_ARGS_ASSERT_GLOB_ASSIGN_REF;
3780 GvINTRO_off(dstr); /* one-shot flag */
3781 GvLINE(dstr) = CopLINE(PL_curcop);
3782 GvEGV(dstr) = MUTABLE_GV(dstr);
3787 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
3788 import_flag = GVf_IMPORTED_CV;
3791 location = (SV **) &GvHV(dstr);
3792 import_flag = GVf_IMPORTED_HV;
3795 location = (SV **) &GvAV(dstr);
3796 import_flag = GVf_IMPORTED_AV;
3799 location = (SV **) &GvIOp(dstr);
3802 location = (SV **) &GvFORM(dstr);
3805 location = &GvSV(dstr);
3806 import_flag = GVf_IMPORTED_SV;
3809 if (stype == SVt_PVCV) {
3810 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
3811 if (GvCVGEN(dstr)) {
3812 SvREFCNT_dec(GvCV(dstr));
3813 GvCV_set(dstr, NULL);
3814 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3817 SAVEGENERICSV(*location);
3821 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
3822 CV* const cv = MUTABLE_CV(*location);
3824 if (!GvCVGEN((const GV *)dstr) &&
3825 (CvROOT(cv) || CvXSUB(cv)))
3827 /* Redefining a sub - warning is mandatory if
3828 it was a const and its value changed. */
3829 if (CvCONST(cv) && CvCONST((const CV *)sref)
3831 == cv_const_sv((const CV *)sref)) {
3833 /* They are 2 constant subroutines generated from
3834 the same constant. This probably means that
3835 they are really the "same" proxy subroutine
3836 instantiated in 2 places. Most likely this is
3837 when a constant is exported twice. Don't warn.
3840 else if (ckWARN(WARN_REDEFINE)
3842 && (!CvCONST((const CV *)sref)
3843 || sv_cmp(cv_const_sv(cv),
3844 cv_const_sv((const CV *)
3846 Perl_warner(aTHX_ packWARN(WARN_REDEFINE),
3849 ? "Constant subroutine %"HEKf
3850 "::%"HEKf" redefined"
3851 : "Subroutine %"HEKf"::%"HEKf
3854 HvNAME_HEK(GvSTASH((const GV *)dstr))
3856 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr))));
3860 cv_ckproto_len_flags(cv, (const GV *)dstr,
3861 SvPOK(sref) ? SvPVX_const(sref) : NULL,
3862 SvPOK(sref) ? SvCUR(sref) : 0,
3863 SvPOK(sref) ? SvUTF8(sref) : 0);
3865 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3866 GvASSUMECV_on(dstr);
3867 if(GvSTASH(dstr)) mro_method_changed_in(GvSTASH(dstr)); /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
3870 if (import_flag && !(GvFLAGS(dstr) & import_flag)
3871 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
3872 GvFLAGS(dstr) |= import_flag;
3874 if (stype == SVt_PVHV) {
3875 const char * const name = GvNAME((GV*)dstr);
3876 const STRLEN len = GvNAMELEN(dstr);
3879 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
3880 || (len == 1 && name[0] == ':')
3882 && (!dref || HvENAME_get(dref))
3885 (HV *)sref, (HV *)dref,
3891 stype == SVt_PVAV && sref != dref
3892 && strEQ(GvNAME((GV*)dstr), "ISA")
3893 /* The stash may have been detached from the symbol table, so
3894 check its name before doing anything. */
3895 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3898 MAGIC * const omg = dref && SvSMAGICAL(dref)
3899 ? mg_find(dref, PERL_MAGIC_isa)
3901 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3902 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3903 AV * const ary = newAV();
3904 av_push(ary, mg->mg_obj); /* takes the refcount */
3905 mg->mg_obj = (SV *)ary;
3908 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
3909 SV **svp = AvARRAY((AV *)omg->mg_obj);
3910 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
3914 SvREFCNT_inc_simple_NN(*svp++)
3920 SvREFCNT_inc_simple_NN(omg->mg_obj)
3924 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
3929 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
3931 mg = mg_find(sref, PERL_MAGIC_isa);
3933 /* Since the *ISA assignment could have affected more than
3934 one stash, don’t call mro_isa_changed_in directly, but let
3935 magic_clearisa do it for us, as it already has the logic for
3936 dealing with globs vs arrays of globs. */
3938 Perl_magic_clearisa(aTHX_ NULL, mg);
3943 if (SvTAINTED(sstr))
3949 Perl_sv_setsv_flags(pTHX_ SV *dstr, register SV* sstr, const I32 flags)
3952 register U32 sflags;
3954 register svtype stype;
3956 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
3961 if (SvIS_FREED(dstr)) {
3962 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
3963 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
3965 SV_CHECK_THINKFIRST_COW_DROP(dstr);
3967 sstr = &PL_sv_undef;
3968 if (SvIS_FREED(sstr)) {
3969 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
3970 (void*)sstr, (void*)dstr);
3972 stype = SvTYPE(sstr);
3973 dtype = SvTYPE(dstr);
3975 (void)SvAMAGIC_off(dstr);
3978 /* need to nuke the magic */
3982 /* There's a lot of redundancy below but we're going for speed here */
3987 if (dtype != SVt_PVGV && dtype != SVt_PVLV) {
3988 (void)SvOK_off(dstr);
3996 sv_upgrade(dstr, SVt_IV);
4000 sv_upgrade(dstr, SVt_PVIV);
4004 goto end_of_first_switch;
4006 (void)SvIOK_only(dstr);
4007 SvIV_set(dstr, SvIVX(sstr));
4010 /* SvTAINTED can only be true if the SV has taint magic, which in
4011 turn means that the SV type is PVMG (or greater). This is the
4012 case statement for SVt_IV, so this cannot be true (whatever gcov
4014 assert(!SvTAINTED(sstr));
4019 if (dtype < SVt_PV && dtype != SVt_IV)
4020 sv_upgrade(dstr, SVt_IV);
4028 sv_upgrade(dstr, SVt_NV);
4032 sv_upgrade(dstr, SVt_PVNV);
4036 goto end_of_first_switch;
4038 SvNV_set(dstr, SvNVX(sstr));
4039 (void)SvNOK_only(dstr);
4040 /* SvTAINTED can only be true if the SV has taint magic, which in
4041 turn means that the SV type is PVMG (or greater). This is the
4042 case statement for SVt_NV, so this cannot be true (whatever gcov
4044 assert(!SvTAINTED(sstr));
4050 #ifdef PERL_OLD_COPY_ON_WRITE
4051 if ((SvFLAGS(sstr) & CAN_COW_MASK) == CAN_COW_FLAGS) {
4052 if (dtype < SVt_PVIV)
4053 sv_upgrade(dstr, SVt_PVIV);
4060 sv_upgrade(dstr, SVt_PV);
4063 if (dtype < SVt_PVIV)
4064 sv_upgrade(dstr, SVt_PVIV);
4067 if (dtype < SVt_PVNV)
4068 sv_upgrade(dstr, SVt_PVNV);
4072 const char * const type = sv_reftype(sstr,0);
4074 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4076 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4081 if (dtype < SVt_REGEXP)
4082 sv_upgrade(dstr, SVt_REGEXP);
4085 /* case SVt_BIND: */
4089 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4091 if (SvTYPE(sstr) != stype)
4092 stype = SvTYPE(sstr);
4094 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4095 glob_assign_glob(dstr, sstr, dtype);
4098 if (stype == SVt_PVLV)
4099 SvUPGRADE(dstr, SVt_PVNV);
4101 SvUPGRADE(dstr, (svtype)stype);
4103 end_of_first_switch:
4105 /* dstr may have been upgraded. */
4106 dtype = SvTYPE(dstr);
4107 sflags = SvFLAGS(sstr);
4109 if (dtype == SVt_PVCV || dtype == SVt_PVFM) {
4110 /* Assigning to a subroutine sets the prototype. */
4113 const char *const ptr = SvPV_const(sstr, len);
4115 SvGROW(dstr, len + 1);
4116 Copy(ptr, SvPVX(dstr), len + 1, char);
4117 SvCUR_set(dstr, len);
4119 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4123 } else if (dtype == SVt_PVAV || dtype == SVt_PVHV) {
4124 const char * const type = sv_reftype(dstr,0);
4126 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4128 Perl_croak(aTHX_ "Cannot copy to %s", type);
4129 } else if (sflags & SVf_ROK) {
4130 if (isGV_with_GP(dstr)
4131 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4134 if (GvIMPORTED(dstr) != GVf_IMPORTED
4135 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4137 GvIMPORTED_on(dstr);
4142 glob_assign_glob(dstr, sstr, dtype);
4146 if (dtype >= SVt_PV) {
4147 if (isGV_with_GP(dstr)) {
4148 glob_assign_ref(dstr, sstr);
4151 if (SvPVX_const(dstr)) {
4157 (void)SvOK_off(dstr);
4158 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4159 SvFLAGS(dstr) |= sflags & SVf_ROK;
4160 assert(!(sflags & SVp_NOK));
4161 assert(!(sflags & SVp_IOK));
4162 assert(!(sflags & SVf_NOK));
4163 assert(!(sflags & SVf_IOK));
4165 else if (isGV_with_GP(dstr)) {
4166 if (!(sflags & SVf_OK)) {
4167 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4168 "Undefined value assigned to typeglob");
4171 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4172 if (dstr != (const SV *)gv) {
4173 const char * const name = GvNAME((const GV *)dstr);
4174 const STRLEN len = GvNAMELEN(dstr);
4175 HV *old_stash = NULL;
4176 bool reset_isa = FALSE;
4177 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4178 || (len == 1 && name[0] == ':')) {
4179 /* Set aside the old stash, so we can reset isa caches
4180 on its subclasses. */
4181 if((old_stash = GvHV(dstr))) {
4182 /* Make sure we do not lose it early. */
4183 SvREFCNT_inc_simple_void_NN(
4184 sv_2mortal((SV *)old_stash)
4191 gp_free(MUTABLE_GV(dstr));
4192 GvGP_set(dstr, gp_ref(GvGP(gv)));
4195 HV * const stash = GvHV(dstr);
4197 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4207 else if (dtype == SVt_REGEXP && stype == SVt_REGEXP) {
4208 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4210 else if (sflags & SVp_POK) {
4214 * Check to see if we can just swipe the string. If so, it's a
4215 * possible small lose on short strings, but a big win on long ones.
4216 * It might even be a win on short strings if SvPVX_const(dstr)
4217 * has to be allocated and SvPVX_const(sstr) has to be freed.
4218 * Likewise if we can set up COW rather than doing an actual copy, we
4219 * drop to the else clause, as the swipe code and the COW setup code
4220 * have much in common.
4223 /* Whichever path we take through the next code, we want this true,
4224 and doing it now facilitates the COW check. */
4225 (void)SvPOK_only(dstr);
4228 /* If we're already COW then this clause is not true, and if COW
4229 is allowed then we drop down to the else and make dest COW
4230 with us. If caller hasn't said that we're allowed to COW
4231 shared hash keys then we don't do the COW setup, even if the
4232 source scalar is a shared hash key scalar. */
4233 (((flags & SV_COW_SHARED_HASH_KEYS)
4234 ? (sflags & (SVf_FAKE|SVf_READONLY)) != (SVf_FAKE|SVf_READONLY)
4235 : 1 /* If making a COW copy is forbidden then the behaviour we
4236 desire is as if the source SV isn't actually already
4237 COW, even if it is. So we act as if the source flags
4238 are not COW, rather than actually testing them. */
4240 #ifndef PERL_OLD_COPY_ON_WRITE
4241 /* The change that added SV_COW_SHARED_HASH_KEYS makes the logic
4242 when PERL_OLD_COPY_ON_WRITE is defined a little wrong.
4243 Conceptually PERL_OLD_COPY_ON_WRITE being defined should
4244 override SV_COW_SHARED_HASH_KEYS, because it means "always COW"
4245 but in turn, it's somewhat dead code, never expected to go
4246 live, but more kept as a placeholder on how to do it better
4247 in a newer implementation. */
4248 /* If we are COW and dstr is a suitable target then we drop down
4249 into the else and make dest a COW of us. */
4250 || (SvFLAGS(dstr) & CAN_COW_MASK) != CAN_COW_FLAGS
4255 (sflags & SVs_TEMP) && /* slated for free anyway? */
4256 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4257 (!(flags & SV_NOSTEAL)) &&
4258 /* and we're allowed to steal temps */
4259 SvREFCNT(sstr) == 1 && /* and no other references to it? */
4260 SvLEN(sstr)) /* and really is a string */
4261 #ifdef PERL_OLD_COPY_ON_WRITE
4262 && ((flags & SV_COW_SHARED_HASH_KEYS)
4263 ? (!((sflags & CAN_COW_MASK) == CAN_COW_FLAGS
4264 && (SvFLAGS(dstr) & CAN_COW_MASK) == CAN_COW_FLAGS
4265 && SvTYPE(sstr) >= SVt_PVIV && SvTYPE(sstr) != SVt_PVFM))
4269 /* Failed the swipe test, and it's not a shared hash key either.
4270 Have to copy the string. */
4271 STRLEN len = SvCUR(sstr);
4272 SvGROW(dstr, len + 1); /* inlined from sv_setpvn */
4273 Move(SvPVX_const(sstr),SvPVX(dstr),len,char);
4274 SvCUR_set(dstr, len);
4275 *SvEND(dstr) = '\0';
4277 /* If PERL_OLD_COPY_ON_WRITE is not defined, then isSwipe will always
4279 /* Either it's a shared hash key, or it's suitable for
4280 copy-on-write or we can swipe the string. */
4282 PerlIO_printf(Perl_debug_log, "Copy on write: sstr --> dstr\n");
4286 #ifdef PERL_OLD_COPY_ON_WRITE
4288 if ((sflags & (SVf_FAKE | SVf_READONLY))
4289 != (SVf_FAKE | SVf_READONLY)) {
4290 SvREADONLY_on(sstr);
4292 /* Make the source SV into a loop of 1.
4293 (about to become 2) */
4294 SV_COW_NEXT_SV_SET(sstr, sstr);
4298 /* Initial code is common. */
4299 if (SvPVX_const(dstr)) { /* we know that dtype >= SVt_PV */
4304 /* making another shared SV. */
4305 STRLEN cur = SvCUR(sstr);
4306 STRLEN len = SvLEN(sstr);
4307 #ifdef PERL_OLD_COPY_ON_WRITE
4309 assert (SvTYPE(dstr) >= SVt_PVIV);
4310 /* SvIsCOW_normal */
4311 /* splice us in between source and next-after-source. */
4312 SV_COW_NEXT_SV_SET(dstr, SV_COW_NEXT_SV(sstr));
4313 SV_COW_NEXT_SV_SET(sstr, dstr);
4314 SvPV_set(dstr, SvPVX_mutable(sstr));
4318 /* SvIsCOW_shared_hash */
4319 DEBUG_C(PerlIO_printf(Perl_debug_log,
4320 "Copy on write: Sharing hash\n"));
4322 assert (SvTYPE(dstr) >= SVt_PV);
4324 HEK_KEY(share_hek_hek(SvSHARED_HEK_FROM_PV(SvPVX_const(sstr)))));
4326 SvLEN_set(dstr, len);
4327 SvCUR_set(dstr, cur);
4328 SvREADONLY_on(dstr);
4332 { /* Passes the swipe test. */
4333 SvPV_set(dstr, SvPVX_mutable(sstr));
4334 SvLEN_set(dstr, SvLEN(sstr));
4335 SvCUR_set(dstr, SvCUR(sstr));
4338 (void)SvOK_off(sstr); /* NOTE: nukes most SvFLAGS on sstr */
4339 SvPV_set(sstr, NULL);
4345 if (sflags & SVp_NOK) {
4346 SvNV_set(dstr, SvNVX(sstr));
4348 if (sflags & SVp_IOK) {
4349 SvIV_set(dstr, SvIVX(sstr));
4350 /* Must do this otherwise some other overloaded use of 0x80000000
4351 gets confused. I guess SVpbm_VALID */
4352 if (sflags & SVf_IVisUV)
4355 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_NOK|SVp_NOK|SVf_UTF8);
4357 const MAGIC * const smg = SvVSTRING_mg(sstr);
4359 sv_magic(dstr, NULL, PERL_MAGIC_vstring,
4360 smg->mg_ptr, smg->mg_len);
4361 SvRMAGICAL_on(dstr);
4365 else if (sflags & (SVp_IOK|SVp_NOK)) {
4366 (void)SvOK_off(dstr);
4367 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_IVisUV|SVf_NOK|SVp_NOK);
4368 if (sflags & SVp_IOK) {
4369 /* XXXX Do we want to set IsUV for IV(ROK)? Be extra safe... */
4370 SvIV_set(dstr, SvIVX(sstr));
4372 if (sflags & SVp_NOK) {
4373 SvNV_set(dstr, SvNVX(sstr));
4377 if (isGV_with_GP(sstr)) {
4378 /* This stringification rule for globs is spread in 3 places.
4379 This feels bad. FIXME. */
4380 const U32 wasfake = sflags & SVf_FAKE;
4382 /* FAKE globs can get coerced, so need to turn this off
4383 temporarily if it is on. */
4385 gv_efullname3(dstr, MUTABLE_GV(sstr), "*");
4386 SvFLAGS(sstr) |= wasfake;
4389 (void)SvOK_off(dstr);
4391 if (SvTAINTED(sstr))
4396 =for apidoc sv_setsv_mg
4398 Like C<sv_setsv>, but also handles 'set' magic.
4404 Perl_sv_setsv_mg(pTHX_ SV *const dstr, register SV *const sstr)
4406 PERL_ARGS_ASSERT_SV_SETSV_MG;
4408 sv_setsv(dstr,sstr);
4412 #ifdef PERL_OLD_COPY_ON_WRITE
4414 Perl_sv_setsv_cow(pTHX_ SV *dstr, SV *sstr)
4416 STRLEN cur = SvCUR(sstr);
4417 STRLEN len = SvLEN(sstr);
4418 register char *new_pv;
4420 PERL_ARGS_ASSERT_SV_SETSV_COW;
4423 PerlIO_printf(Perl_debug_log, "Fast copy on write: %p -> %p\n",
4424 (void*)sstr, (void*)dstr);
4431 if (SvTHINKFIRST(dstr))
4432 sv_force_normal_flags(dstr, SV_COW_DROP_PV);
4433 else if (SvPVX_const(dstr))
4434 Safefree(SvPVX_const(dstr));
4438 SvUPGRADE(dstr, SVt_PVIV);
4440 assert (SvPOK(sstr));
4441 assert (SvPOKp(sstr));
4442 assert (!SvIOK(sstr));
4443 assert (!SvIOKp(sstr));
4444 assert (!SvNOK(sstr));
4445 assert (!SvNOKp(sstr));
4447 if (SvIsCOW(sstr)) {
4449 if (SvLEN(sstr) == 0) {
4450 /* source is a COW shared hash key. */
4451 DEBUG_C(PerlIO_printf(Perl_debug_log,
4452 "Fast copy on write: Sharing hash\n"));
4453 new_pv = HEK_KEY(share_hek_hek(SvSHARED_HEK_FROM_PV(SvPVX_const(sstr))));
4456 SV_COW_NEXT_SV_SET(dstr, SV_COW_NEXT_SV(sstr));
4458 assert ((SvFLAGS(sstr) & CAN_COW_MASK) == CAN_COW_FLAGS);
4459 SvUPGRADE(sstr, SVt_PVIV);
4460 SvREADONLY_on(sstr);
4462 DEBUG_C(PerlIO_printf(Perl_debug_log,
4463 "Fast copy on write: Converting sstr to COW\n"));
4464 SV_COW_NEXT_SV_SET(dstr, sstr);
4466 SV_COW_NEXT_SV_SET(sstr, dstr);
4467 new_pv = SvPVX_mutable(sstr);
4470 SvPV_set(dstr, new_pv);
4471 SvFLAGS(dstr) = (SVt_PVIV|SVf_POK|SVp_POK|SVf_FAKE|SVf_READONLY);
4474 SvLEN_set(dstr, len);
4475 SvCUR_set(dstr, cur);
4484 =for apidoc sv_setpvn
4486 Copies a string into an SV. The C<len> parameter indicates the number of
4487 bytes to be copied. If the C<ptr> argument is NULL the SV will become
4488 undefined. Does not handle 'set' magic. See C<sv_setpvn_mg>.
4494 Perl_sv_setpvn(pTHX_ register SV *const sv, register const char *const ptr, register const STRLEN len)
4497 register char *dptr;
4499 PERL_ARGS_ASSERT_SV_SETPVN;
4501 SV_CHECK_THINKFIRST_COW_DROP(sv);
4507 /* len is STRLEN which is unsigned, need to copy to signed */
4510 Perl_croak(aTHX_ "panic: sv_setpvn called with negative strlen");
4512 SvUPGRADE(sv, SVt_PV);
4514 dptr = SvGROW(sv, len + 1);
4515 Move(ptr,dptr,len,char);
4518 (void)SvPOK_only_UTF8(sv); /* validate pointer */
4523 =for apidoc sv_setpvn_mg
4525 Like C<sv_setpvn>, but also handles 'set' magic.
4531 Perl_sv_setpvn_mg(pTHX_ register SV *const sv, register const char *const ptr, register const STRLEN len)
4533 PERL_ARGS_ASSERT_SV_SETPVN_MG;
4535 sv_setpvn(sv,ptr,len);
4540 =for apidoc sv_setpv
4542 Copies a string into an SV. The string must be null-terminated. Does not
4543 handle 'set' magic. See C<sv_setpv_mg>.
4549 Perl_sv_setpv(pTHX_ register SV *const sv, register const char *const ptr)
4552 register STRLEN len;
4554 PERL_ARGS_ASSERT_SV_SETPV;
4556 SV_CHECK_THINKFIRST_COW_DROP(sv);
4562 SvUPGRADE(sv, SVt_PV);
4564 SvGROW(sv, len + 1);
4565 Move(ptr,SvPVX(sv),len+1,char);
4567 (void)SvPOK_only_UTF8(sv); /* validate pointer */
4572 =for apidoc sv_setpv_mg
4574 Like C<sv_setpv>, but also handles 'set' magic.
4580 Perl_sv_setpv_mg(pTHX_ register SV *const sv, register const char *const ptr)
4582 PERL_ARGS_ASSERT_SV_SETPV_MG;
4589 Perl_sv_sethek(pTHX_ register SV *const sv, const HEK *const hek)
4593 PERL_ARGS_ASSERT_SV_SETHEK;
4599 if (HEK_LEN(hek) == HEf_SVKEY) {
4600 sv_setsv(sv, *(SV**)HEK_KEY(hek));
4603 const int flags = HEK_FLAGS(hek);
4604 if (flags & HVhek_WASUTF8) {
4605 STRLEN utf8_len = HEK_LEN(hek);
4606 char *as_utf8 = (char *)bytes_to_utf8((U8*)HEK_KEY(hek), &utf8_len);
4607 sv_usepvn_flags(sv, as_utf8, utf8_len, SV_HAS_TRAILING_NUL);
4610 } else if (flags & (HVhek_REHASH|HVhek_UNSHARED)) {
4611 sv_setpvn(sv, HEK_KEY(hek), HEK_LEN(hek));