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 It croaks if the SV is already in a more complex form than requested. You
1131 generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1132 before calling C<sv_upgrade>, and hence does not croak. See also
1139 Perl_sv_upgrade(pTHX_ register SV *const sv, svtype new_type)
1144 const svtype old_type = SvTYPE(sv);
1145 const struct body_details *new_type_details;
1146 const struct body_details *old_type_details
1147 = bodies_by_type + old_type;
1148 SV *referant = NULL;
1150 PERL_ARGS_ASSERT_SV_UPGRADE;
1152 if (old_type == new_type)
1155 /* This clause was purposefully added ahead of the early return above to
1156 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1157 inference by Nick I-S that it would fix other troublesome cases. See
1158 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1160 Given that shared hash key scalars are no longer PVIV, but PV, there is
1161 no longer need to unshare so as to free up the IVX slot for its proper
1162 purpose. So it's safe to move the early return earlier. */
1164 if (new_type != SVt_PV && SvIsCOW(sv)) {
1165 sv_force_normal_flags(sv, 0);
1168 old_body = SvANY(sv);
1170 /* Copying structures onto other structures that have been neatly zeroed
1171 has a subtle gotcha. Consider XPVMG
1173 +------+------+------+------+------+-------+-------+
1174 | NV | CUR | LEN | IV | MAGIC | STASH |
1175 +------+------+------+------+------+-------+-------+
1176 0 4 8 12 16 20 24 28
1178 where NVs are aligned to 8 bytes, so that sizeof that structure is
1179 actually 32 bytes long, with 4 bytes of padding at the end:
1181 +------+------+------+------+------+-------+-------+------+
1182 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1183 +------+------+------+------+------+-------+-------+------+
1184 0 4 8 12 16 20 24 28 32
1186 so what happens if you allocate memory for this structure:
1188 +------+------+------+------+------+-------+-------+------+------+...
1189 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1190 +------+------+------+------+------+-------+-------+------+------+...
1191 0 4 8 12 16 20 24 28 32 36
1193 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1194 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1195 started out as zero once, but it's quite possible that it isn't. So now,
1196 rather than a nicely zeroed GP, you have it pointing somewhere random.
1199 (In fact, GP ends up pointing at a previous GP structure, because the
1200 principle cause of the padding in XPVMG getting garbage is a copy of
1201 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1202 this happens to be moot because XPVGV has been re-ordered, with GP
1203 no longer after STASH)
1205 So we are careful and work out the size of used parts of all the
1213 referant = SvRV(sv);
1214 old_type_details = &fake_rv;
1215 if (new_type == SVt_NV)
1216 new_type = SVt_PVNV;
1218 if (new_type < SVt_PVIV) {
1219 new_type = (new_type == SVt_NV)
1220 ? SVt_PVNV : SVt_PVIV;
1225 if (new_type < SVt_PVNV) {
1226 new_type = SVt_PVNV;
1230 assert(new_type > SVt_PV);
1231 assert(SVt_IV < SVt_PV);
1232 assert(SVt_NV < SVt_PV);
1239 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1240 there's no way that it can be safely upgraded, because perl.c
1241 expects to Safefree(SvANY(PL_mess_sv)) */
1242 assert(sv != PL_mess_sv);
1243 /* This flag bit is used to mean other things in other scalar types.
1244 Given that it only has meaning inside the pad, it shouldn't be set
1245 on anything that can get upgraded. */
1246 assert(!SvPAD_TYPED(sv));
1249 if (old_type_details->cant_upgrade)
1250 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1251 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1254 if (old_type > new_type)
1255 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1256 (int)old_type, (int)new_type);
1258 new_type_details = bodies_by_type + new_type;
1260 SvFLAGS(sv) &= ~SVTYPEMASK;
1261 SvFLAGS(sv) |= new_type;
1263 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1264 the return statements above will have triggered. */
1265 assert (new_type != SVt_NULL);
1268 assert(old_type == SVt_NULL);
1269 SvANY(sv) = (XPVIV*)((char*)&(sv->sv_u.svu_iv) - STRUCT_OFFSET(XPVIV, xiv_iv));
1273 assert(old_type == SVt_NULL);
1274 SvANY(sv) = new_XNV();
1279 assert(new_type_details->body_size);
1282 assert(new_type_details->arena);
1283 assert(new_type_details->arena_size);
1284 /* This points to the start of the allocated area. */
1285 new_body_inline(new_body, new_type);
1286 Zero(new_body, new_type_details->body_size, char);
1287 new_body = ((char *)new_body) - new_type_details->offset;
1289 /* We always allocated the full length item with PURIFY. To do this
1290 we fake things so that arena is false for all 16 types.. */
1291 new_body = new_NOARENAZ(new_type_details);
1293 SvANY(sv) = new_body;
1294 if (new_type == SVt_PVAV) {
1298 if (old_type_details->body_size) {
1301 /* It will have been zeroed when the new body was allocated.
1302 Lets not write to it, in case it confuses a write-back
1308 #ifndef NODEFAULT_SHAREKEYS
1309 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1311 HvMAX(sv) = 7; /* (start with 8 buckets) */
1314 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1315 The target created by newSVrv also is, and it can have magic.
1316 However, it never has SvPVX set.
1318 if (old_type == SVt_IV) {
1320 } else if (old_type >= SVt_PV) {
1321 assert(SvPVX_const(sv) == 0);
1324 if (old_type >= SVt_PVMG) {
1325 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1326 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1328 sv->sv_u.svu_array = NULL; /* or svu_hash */
1334 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
1335 sv_force_normal_flags(sv) is called. */
1338 /* XXX Is this still needed? Was it ever needed? Surely as there is
1339 no route from NV to PVIV, NOK can never be true */
1340 assert(!SvNOKp(sv));
1351 assert(new_type_details->body_size);
1352 /* We always allocated the full length item with PURIFY. To do this
1353 we fake things so that arena is false for all 16 types.. */
1354 if(new_type_details->arena) {
1355 /* This points to the start of the allocated area. */
1356 new_body_inline(new_body, new_type);
1357 Zero(new_body, new_type_details->body_size, char);
1358 new_body = ((char *)new_body) - new_type_details->offset;
1360 new_body = new_NOARENAZ(new_type_details);
1362 SvANY(sv) = new_body;
1364 if (old_type_details->copy) {
1365 /* There is now the potential for an upgrade from something without
1366 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1367 int offset = old_type_details->offset;
1368 int length = old_type_details->copy;
1370 if (new_type_details->offset > old_type_details->offset) {
1371 const int difference
1372 = new_type_details->offset - old_type_details->offset;
1373 offset += difference;
1374 length -= difference;
1376 assert (length >= 0);
1378 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1382 #ifndef NV_ZERO_IS_ALLBITS_ZERO
1383 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
1384 * correct 0.0 for us. Otherwise, if the old body didn't have an
1385 * NV slot, but the new one does, then we need to initialise the
1386 * freshly created NV slot with whatever the correct bit pattern is
1388 if (old_type_details->zero_nv && !new_type_details->zero_nv
1389 && !isGV_with_GP(sv))
1393 if (new_type == SVt_PVIO) {
1394 IO * const io = MUTABLE_IO(sv);
1395 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
1398 /* Clear the stashcache because a new IO could overrule a package
1400 hv_clear(PL_stashcache);
1402 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1403 IoPAGE_LEN(sv) = 60;
1405 if (old_type < SVt_PV) {
1406 /* referant will be NULL unless the old type was SVt_IV emulating
1408 sv->sv_u.svu_rv = referant;
1412 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1413 (unsigned long)new_type);
1416 if (old_type > SVt_IV) {
1420 /* Note that there is an assumption that all bodies of types that
1421 can be upgraded came from arenas. Only the more complex non-
1422 upgradable types are allowed to be directly malloc()ed. */
1423 assert(old_type_details->arena);
1424 del_body((void*)((char*)old_body + old_type_details->offset),
1425 &PL_body_roots[old_type]);
1431 =for apidoc sv_backoff
1433 Remove any string offset. You should normally use the C<SvOOK_off> macro
1440 Perl_sv_backoff(pTHX_ register SV *const sv)
1443 const char * const s = SvPVX_const(sv);
1445 PERL_ARGS_ASSERT_SV_BACKOFF;
1446 PERL_UNUSED_CONTEXT;
1449 assert(SvTYPE(sv) != SVt_PVHV);
1450 assert(SvTYPE(sv) != SVt_PVAV);
1452 SvOOK_offset(sv, delta);
1454 SvLEN_set(sv, SvLEN(sv) + delta);
1455 SvPV_set(sv, SvPVX(sv) - delta);
1456 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1457 SvFLAGS(sv) &= ~SVf_OOK;
1464 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1465 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1466 Use the C<SvGROW> wrapper instead.
1472 Perl_sv_grow(pTHX_ register SV *const sv, register STRLEN newlen)
1476 PERL_ARGS_ASSERT_SV_GROW;
1478 if (PL_madskills && newlen >= 0x100000) {
1479 PerlIO_printf(Perl_debug_log,
1480 "Allocation too large: %"UVxf"\n", (UV)newlen);
1482 #ifdef HAS_64K_LIMIT
1483 if (newlen >= 0x10000) {
1484 PerlIO_printf(Perl_debug_log,
1485 "Allocation too large: %"UVxf"\n", (UV)newlen);
1488 #endif /* HAS_64K_LIMIT */
1491 if (SvTYPE(sv) < SVt_PV) {
1492 sv_upgrade(sv, SVt_PV);
1493 s = SvPVX_mutable(sv);
1495 else if (SvOOK(sv)) { /* pv is offset? */
1497 s = SvPVX_mutable(sv);
1498 if (newlen > SvLEN(sv))
1499 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1500 #ifdef HAS_64K_LIMIT
1501 if (newlen >= 0x10000)
1506 s = SvPVX_mutable(sv);
1508 if (newlen > SvLEN(sv)) { /* need more room? */
1509 STRLEN minlen = SvCUR(sv);
1510 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1511 if (newlen < minlen)
1513 #ifndef Perl_safesysmalloc_size
1514 newlen = PERL_STRLEN_ROUNDUP(newlen);
1516 if (SvLEN(sv) && s) {
1517 s = (char*)saferealloc(s, newlen);
1520 s = (char*)safemalloc(newlen);
1521 if (SvPVX_const(sv) && SvCUR(sv)) {
1522 Move(SvPVX_const(sv), s, (newlen < SvCUR(sv)) ? newlen : SvCUR(sv), char);
1526 #ifdef Perl_safesysmalloc_size
1527 /* Do this here, do it once, do it right, and then we will never get
1528 called back into sv_grow() unless there really is some growing
1530 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1532 SvLEN_set(sv, newlen);
1539 =for apidoc sv_setiv
1541 Copies an integer into the given SV, upgrading first if necessary.
1542 Does not handle 'set' magic. See also C<sv_setiv_mg>.
1548 Perl_sv_setiv(pTHX_ register SV *const sv, const IV i)
1552 PERL_ARGS_ASSERT_SV_SETIV;
1554 SV_CHECK_THINKFIRST_COW_DROP(sv);
1555 switch (SvTYPE(sv)) {
1558 sv_upgrade(sv, SVt_IV);
1561 sv_upgrade(sv, SVt_PVIV);
1565 if (!isGV_with_GP(sv))
1572 /* diag_listed_as: Can't coerce %s to %s in %s */
1573 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1577 (void)SvIOK_only(sv); /* validate number */
1583 =for apidoc sv_setiv_mg
1585 Like C<sv_setiv>, but also handles 'set' magic.
1591 Perl_sv_setiv_mg(pTHX_ register SV *const sv, const IV i)
1593 PERL_ARGS_ASSERT_SV_SETIV_MG;
1600 =for apidoc sv_setuv
1602 Copies an unsigned integer into the given SV, upgrading first if necessary.
1603 Does not handle 'set' magic. See also C<sv_setuv_mg>.
1609 Perl_sv_setuv(pTHX_ register SV *const sv, const UV u)
1611 PERL_ARGS_ASSERT_SV_SETUV;
1613 /* With the if statement to ensure that integers are stored as IVs whenever
1615 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1618 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1620 If you wish to remove the following if statement, so that this routine
1621 (and its callers) always return UVs, please benchmark to see what the
1622 effect is. Modern CPUs may be different. Or may not :-)
1624 if (u <= (UV)IV_MAX) {
1625 sv_setiv(sv, (IV)u);
1634 =for apidoc sv_setuv_mg
1636 Like C<sv_setuv>, but also handles 'set' magic.
1642 Perl_sv_setuv_mg(pTHX_ register SV *const sv, const UV u)
1644 PERL_ARGS_ASSERT_SV_SETUV_MG;
1651 =for apidoc sv_setnv
1653 Copies a double into the given SV, upgrading first if necessary.
1654 Does not handle 'set' magic. See also C<sv_setnv_mg>.
1660 Perl_sv_setnv(pTHX_ register SV *const sv, const NV num)
1664 PERL_ARGS_ASSERT_SV_SETNV;
1666 SV_CHECK_THINKFIRST_COW_DROP(sv);
1667 switch (SvTYPE(sv)) {
1670 sv_upgrade(sv, SVt_NV);
1674 sv_upgrade(sv, SVt_PVNV);
1678 if (!isGV_with_GP(sv))
1685 /* diag_listed_as: Can't coerce %s to %s in %s */
1686 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1691 (void)SvNOK_only(sv); /* validate number */
1696 =for apidoc sv_setnv_mg
1698 Like C<sv_setnv>, but also handles 'set' magic.
1704 Perl_sv_setnv_mg(pTHX_ register SV *const sv, const NV num)
1706 PERL_ARGS_ASSERT_SV_SETNV_MG;
1712 /* Print an "isn't numeric" warning, using a cleaned-up,
1713 * printable version of the offending string
1717 S_not_a_number(pTHX_ SV *const sv)
1724 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1727 dsv = newSVpvs_flags("", SVs_TEMP);
1728 pv = sv_uni_display(dsv, sv, 10, UNI_DISPLAY_ISPRINT);
1731 const char * const limit = tmpbuf + sizeof(tmpbuf) - 8;
1732 /* each *s can expand to 4 chars + "...\0",
1733 i.e. need room for 8 chars */
1735 const char *s = SvPVX_const(sv);
1736 const char * const end = s + SvCUR(sv);
1737 for ( ; s < end && d < limit; s++ ) {
1739 if (ch & 128 && !isPRINT_LC(ch)) {
1748 else if (ch == '\r') {
1752 else if (ch == '\f') {
1756 else if (ch == '\\') {
1760 else if (ch == '\0') {
1764 else if (isPRINT_LC(ch))
1781 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1782 /* diag_listed_as: Argument "%s" isn't numeric%s */
1783 "Argument \"%s\" isn't numeric in %s", pv,
1786 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1787 /* diag_listed_as: Argument "%s" isn't numeric%s */
1788 "Argument \"%s\" isn't numeric", pv);
1792 =for apidoc looks_like_number
1794 Test if the content of an SV looks like a number (or is a number).
1795 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1796 non-numeric warning), even if your atof() doesn't grok them. Get-magic is
1803 Perl_looks_like_number(pTHX_ SV *const sv)
1805 register const char *sbegin;
1808 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1810 if (SvPOK(sv) || SvPOKp(sv)) {
1811 sbegin = SvPV_nomg_const(sv, len);
1814 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1815 return grok_number(sbegin, len, NULL);
1819 S_glob_2number(pTHX_ GV * const gv)
1821 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1823 /* We know that all GVs stringify to something that is not-a-number,
1824 so no need to test that. */
1825 if (ckWARN(WARN_NUMERIC))
1827 SV *const buffer = sv_newmortal();
1828 gv_efullname3(buffer, gv, "*");
1829 not_a_number(buffer);
1831 /* We just want something true to return, so that S_sv_2iuv_common
1832 can tail call us and return true. */
1836 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1837 until proven guilty, assume that things are not that bad... */
1842 As 64 bit platforms often have an NV that doesn't preserve all bits of
1843 an IV (an assumption perl has been based on to date) it becomes necessary
1844 to remove the assumption that the NV always carries enough precision to
1845 recreate the IV whenever needed, and that the NV is the canonical form.
1846 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1847 precision as a side effect of conversion (which would lead to insanity
1848 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1849 1) to distinguish between IV/UV/NV slots that have cached a valid
1850 conversion where precision was lost and IV/UV/NV slots that have a
1851 valid conversion which has lost no precision
1852 2) to ensure that if a numeric conversion to one form is requested that
1853 would lose precision, the precise conversion (or differently
1854 imprecise conversion) is also performed and cached, to prevent
1855 requests for different numeric formats on the same SV causing
1856 lossy conversion chains. (lossless conversion chains are perfectly
1861 SvIOKp is true if the IV slot contains a valid value
1862 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1863 SvNOKp is true if the NV slot contains a valid value
1864 SvNOK is true only if the NV value is accurate
1867 while converting from PV to NV, check to see if converting that NV to an
1868 IV(or UV) would lose accuracy over a direct conversion from PV to
1869 IV(or UV). If it would, cache both conversions, return NV, but mark
1870 SV as IOK NOKp (ie not NOK).
1872 While converting from PV to IV, check to see if converting that IV to an
1873 NV would lose accuracy over a direct conversion from PV to NV. If it
1874 would, cache both conversions, flag similarly.
1876 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
1877 correctly because if IV & NV were set NV *always* overruled.
1878 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
1879 changes - now IV and NV together means that the two are interchangeable:
1880 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
1882 The benefit of this is that operations such as pp_add know that if
1883 SvIOK is true for both left and right operands, then integer addition
1884 can be used instead of floating point (for cases where the result won't
1885 overflow). Before, floating point was always used, which could lead to
1886 loss of precision compared with integer addition.
1888 * making IV and NV equal status should make maths accurate on 64 bit
1890 * may speed up maths somewhat if pp_add and friends start to use
1891 integers when possible instead of fp. (Hopefully the overhead in
1892 looking for SvIOK and checking for overflow will not outweigh the
1893 fp to integer speedup)
1894 * will slow down integer operations (callers of SvIV) on "inaccurate"
1895 values, as the change from SvIOK to SvIOKp will cause a call into
1896 sv_2iv each time rather than a macro access direct to the IV slot
1897 * should speed up number->string conversion on integers as IV is
1898 favoured when IV and NV are equally accurate
1900 ####################################################################
1901 You had better be using SvIOK_notUV if you want an IV for arithmetic:
1902 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
1903 On the other hand, SvUOK is true iff UV.
1904 ####################################################################
1906 Your mileage will vary depending your CPU's relative fp to integer
1910 #ifndef NV_PRESERVES_UV
1911 # define IS_NUMBER_UNDERFLOW_IV 1
1912 # define IS_NUMBER_UNDERFLOW_UV 2
1913 # define IS_NUMBER_IV_AND_UV 2
1914 # define IS_NUMBER_OVERFLOW_IV 4
1915 # define IS_NUMBER_OVERFLOW_UV 5
1917 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
1919 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
1921 S_sv_2iuv_non_preserve(pTHX_ register SV *const sv
1929 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
1931 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));
1932 if (SvNVX(sv) < (NV)IV_MIN) {
1933 (void)SvIOKp_on(sv);
1935 SvIV_set(sv, IV_MIN);
1936 return IS_NUMBER_UNDERFLOW_IV;
1938 if (SvNVX(sv) > (NV)UV_MAX) {
1939 (void)SvIOKp_on(sv);
1942 SvUV_set(sv, UV_MAX);
1943 return IS_NUMBER_OVERFLOW_UV;
1945 (void)SvIOKp_on(sv);
1947 /* Can't use strtol etc to convert this string. (See truth table in
1949 if (SvNVX(sv) <= (UV)IV_MAX) {
1950 SvIV_set(sv, I_V(SvNVX(sv)));
1951 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
1952 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
1954 /* Integer is imprecise. NOK, IOKp */
1956 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
1959 SvUV_set(sv, U_V(SvNVX(sv)));
1960 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
1961 if (SvUVX(sv) == UV_MAX) {
1962 /* As we know that NVs don't preserve UVs, UV_MAX cannot
1963 possibly be preserved by NV. Hence, it must be overflow.
1965 return IS_NUMBER_OVERFLOW_UV;
1967 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
1969 /* Integer is imprecise. NOK, IOKp */
1971 return IS_NUMBER_OVERFLOW_IV;
1973 #endif /* !NV_PRESERVES_UV*/
1976 S_sv_2iuv_common(pTHX_ SV *const sv)
1980 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
1983 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
1984 * without also getting a cached IV/UV from it at the same time
1985 * (ie PV->NV conversion should detect loss of accuracy and cache
1986 * IV or UV at same time to avoid this. */
1987 /* IV-over-UV optimisation - choose to cache IV if possible */
1989 if (SvTYPE(sv) == SVt_NV)
1990 sv_upgrade(sv, SVt_PVNV);
1992 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
1993 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
1994 certainly cast into the IV range at IV_MAX, whereas the correct
1995 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
1997 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
1998 if (Perl_isnan(SvNVX(sv))) {
2004 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2005 SvIV_set(sv, I_V(SvNVX(sv)));
2006 if (SvNVX(sv) == (NV) SvIVX(sv)
2007 #ifndef NV_PRESERVES_UV
2008 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2009 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2010 /* Don't flag it as "accurately an integer" if the number
2011 came from a (by definition imprecise) NV operation, and
2012 we're outside the range of NV integer precision */
2016 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2018 /* scalar has trailing garbage, eg "42a" */
2020 DEBUG_c(PerlIO_printf(Perl_debug_log,
2021 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (precise)\n",
2027 /* IV not precise. No need to convert from PV, as NV
2028 conversion would already have cached IV if it detected
2029 that PV->IV would be better than PV->NV->IV
2030 flags already correct - don't set public IOK. */
2031 DEBUG_c(PerlIO_printf(Perl_debug_log,
2032 "0x%"UVxf" iv(%"NVgf" => %"IVdf") (imprecise)\n",
2037 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2038 but the cast (NV)IV_MIN rounds to a the value less (more
2039 negative) than IV_MIN which happens to be equal to SvNVX ??
2040 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2041 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2042 (NV)UVX == NVX are both true, but the values differ. :-(
2043 Hopefully for 2s complement IV_MIN is something like
2044 0x8000000000000000 which will be exact. NWC */
2047 SvUV_set(sv, U_V(SvNVX(sv)));
2049 (SvNVX(sv) == (NV) SvUVX(sv))
2050 #ifndef NV_PRESERVES_UV
2051 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2052 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2053 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2054 /* Don't flag it as "accurately an integer" if the number
2055 came from a (by definition imprecise) NV operation, and
2056 we're outside the range of NV integer precision */
2062 DEBUG_c(PerlIO_printf(Perl_debug_log,
2063 "0x%"UVxf" 2iv(%"UVuf" => %"IVdf") (as unsigned)\n",
2069 else if (SvPOKp(sv) && SvLEN(sv)) {
2071 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2072 /* We want to avoid a possible problem when we cache an IV/ a UV which
2073 may be later translated to an NV, and the resulting NV is not
2074 the same as the direct translation of the initial string
2075 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2076 be careful to ensure that the value with the .456 is around if the
2077 NV value is requested in the future).
2079 This means that if we cache such an IV/a UV, we need to cache the
2080 NV as well. Moreover, we trade speed for space, and do not
2081 cache the NV if we are sure it's not needed.
2084 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2085 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2086 == IS_NUMBER_IN_UV) {
2087 /* It's definitely an integer, only upgrade to PVIV */
2088 if (SvTYPE(sv) < SVt_PVIV)
2089 sv_upgrade(sv, SVt_PVIV);
2091 } else if (SvTYPE(sv) < SVt_PVNV)
2092 sv_upgrade(sv, SVt_PVNV);
2094 /* If NVs preserve UVs then we only use the UV value if we know that
2095 we aren't going to call atof() below. If NVs don't preserve UVs
2096 then the value returned may have more precision than atof() will
2097 return, even though value isn't perfectly accurate. */
2098 if ((numtype & (IS_NUMBER_IN_UV
2099 #ifdef NV_PRESERVES_UV
2102 )) == IS_NUMBER_IN_UV) {
2103 /* This won't turn off the public IOK flag if it was set above */
2104 (void)SvIOKp_on(sv);
2106 if (!(numtype & IS_NUMBER_NEG)) {
2108 if (value <= (UV)IV_MAX) {
2109 SvIV_set(sv, (IV)value);
2111 /* it didn't overflow, and it was positive. */
2112 SvUV_set(sv, value);
2116 /* 2s complement assumption */
2117 if (value <= (UV)IV_MIN) {
2118 SvIV_set(sv, -(IV)value);
2120 /* Too negative for an IV. This is a double upgrade, but
2121 I'm assuming it will be rare. */
2122 if (SvTYPE(sv) < SVt_PVNV)
2123 sv_upgrade(sv, SVt_PVNV);
2127 SvNV_set(sv, -(NV)value);
2128 SvIV_set(sv, IV_MIN);
2132 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2133 will be in the previous block to set the IV slot, and the next
2134 block to set the NV slot. So no else here. */
2136 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2137 != IS_NUMBER_IN_UV) {
2138 /* It wasn't an (integer that doesn't overflow the UV). */
2139 SvNV_set(sv, Atof(SvPVX_const(sv)));
2141 if (! numtype && ckWARN(WARN_NUMERIC))
2144 #if defined(USE_LONG_DOUBLE)
2145 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%" PERL_PRIgldbl ")\n",
2146 PTR2UV(sv), SvNVX(sv)));
2148 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"NVgf")\n",
2149 PTR2UV(sv), SvNVX(sv)));
2152 #ifdef NV_PRESERVES_UV
2153 (void)SvIOKp_on(sv);
2155 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2156 SvIV_set(sv, I_V(SvNVX(sv)));
2157 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2160 NOOP; /* Integer is imprecise. NOK, IOKp */
2162 /* UV will not work better than IV */
2164 if (SvNVX(sv) > (NV)UV_MAX) {
2166 /* Integer is inaccurate. NOK, IOKp, is UV */
2167 SvUV_set(sv, UV_MAX);
2169 SvUV_set(sv, U_V(SvNVX(sv)));
2170 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2171 NV preservse UV so can do correct comparison. */
2172 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2175 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2180 #else /* NV_PRESERVES_UV */
2181 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2182 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2183 /* The IV/UV slot will have been set from value returned by
2184 grok_number above. The NV slot has just been set using
2187 assert (SvIOKp(sv));
2189 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2190 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2191 /* Small enough to preserve all bits. */
2192 (void)SvIOKp_on(sv);
2194 SvIV_set(sv, I_V(SvNVX(sv)));
2195 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2197 /* Assumption: first non-preserved integer is < IV_MAX,
2198 this NV is in the preserved range, therefore: */
2199 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2201 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);
2205 0 0 already failed to read UV.
2206 0 1 already failed to read UV.
2207 1 0 you won't get here in this case. IV/UV
2208 slot set, public IOK, Atof() unneeded.
2209 1 1 already read UV.
2210 so there's no point in sv_2iuv_non_preserve() attempting
2211 to use atol, strtol, strtoul etc. */
2213 sv_2iuv_non_preserve (sv, numtype);
2215 sv_2iuv_non_preserve (sv);
2219 #endif /* NV_PRESERVES_UV */
2220 /* It might be more code efficient to go through the entire logic above
2221 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2222 gets complex and potentially buggy, so more programmer efficient
2223 to do it this way, by turning off the public flags: */
2225 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2229 if (isGV_with_GP(sv))
2230 return glob_2number(MUTABLE_GV(sv));
2232 if (!SvPADTMP(sv)) {
2233 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2236 if (SvTYPE(sv) < SVt_IV)
2237 /* Typically the caller expects that sv_any is not NULL now. */
2238 sv_upgrade(sv, SVt_IV);
2239 /* Return 0 from the caller. */
2246 =for apidoc sv_2iv_flags
2248 Return the integer value of an SV, doing any necessary string
2249 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2250 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2256 Perl_sv_2iv_flags(pTHX_ register SV *const sv, const I32 flags)
2261 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2262 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2263 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2264 In practice they are extremely unlikely to actually get anywhere
2265 accessible by user Perl code - the only way that I'm aware of is when
2266 a constant subroutine which is used as the second argument to index.
2268 if (flags & SV_GMAGIC)
2273 return I_V(SvNVX(sv));
2275 if (SvPOKp(sv) && SvLEN(sv)) {
2278 = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2280 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2281 == IS_NUMBER_IN_UV) {
2282 /* It's definitely an integer */
2283 if (numtype & IS_NUMBER_NEG) {
2284 if (value < (UV)IV_MIN)
2287 if (value < (UV)IV_MAX)
2292 if (ckWARN(WARN_NUMERIC))
2295 return I_V(Atof(SvPVX_const(sv)));
2300 assert(SvTYPE(sv) >= SVt_PVMG);
2301 /* This falls through to the report_uninit inside S_sv_2iuv_common. */
2302 } else if (SvTHINKFIRST(sv)) {
2307 if (flags & SV_SKIP_OVERLOAD)
2309 tmpstr = AMG_CALLunary(sv, numer_amg);
2310 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2311 return SvIV(tmpstr);
2314 return PTR2IV(SvRV(sv));
2317 sv_force_normal_flags(sv, 0);
2319 if (SvREADONLY(sv) && !SvOK(sv)) {
2320 if (ckWARN(WARN_UNINITIALIZED))
2326 if (S_sv_2iuv_common(aTHX_ sv))
2329 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2iv(%"IVdf")\n",
2330 PTR2UV(sv),SvIVX(sv)));
2331 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2335 =for apidoc sv_gmagical_2iv_please
2337 Used internally by C<SvIV_please_nomg>, this function sets the C<SvIVX>
2338 slot if C<sv_2iv> would have made the scalar C<SvIOK> had it not been
2339 magical. In that case it returns true.
2345 Perl_sv_gmagical_2iv_please(pTHX_ register SV *sv)
2348 PERL_ARGS_ASSERT_SV_GMAGICAL_2IV_PLEASE;
2349 assert(SvGMAGICAL(sv) && !SvIOKp(sv) && (SvNOKp(sv) || SvPOKp(sv)));
2350 if (S_sv_2iuv_common(aTHX_ sv)) { SvNIOK_off(sv); return 0; }
2351 has_int = !!SvIOK(sv);
2352 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2357 =for apidoc sv_2uv_flags
2359 Return the unsigned integer value of an SV, doing any necessary string
2360 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2361 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2367 Perl_sv_2uv_flags(pTHX_ register SV *const sv, const I32 flags)
2372 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2373 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2374 the same flag bit as SVf_IVisUV, so must not let them cache IVs. */
2375 if (flags & SV_GMAGIC)
2380 return U_V(SvNVX(sv));
2381 if (SvPOKp(sv) && SvLEN(sv)) {
2384 = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2386 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2387 == IS_NUMBER_IN_UV) {
2388 /* It's definitely an integer */
2389 if (!(numtype & IS_NUMBER_NEG))
2393 if (ckWARN(WARN_NUMERIC))
2396 return U_V(Atof(SvPVX_const(sv)));
2401 assert(SvTYPE(sv) >= SVt_PVMG);
2402 /* This falls through to the report_uninit inside S_sv_2iuv_common. */
2403 } else if (SvTHINKFIRST(sv)) {
2408 if (flags & SV_SKIP_OVERLOAD)
2410 tmpstr = AMG_CALLunary(sv, numer_amg);
2411 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2412 return SvUV(tmpstr);
2415 return PTR2UV(SvRV(sv));
2418 sv_force_normal_flags(sv, 0);
2420 if (SvREADONLY(sv) && !SvOK(sv)) {
2421 if (ckWARN(WARN_UNINITIALIZED))
2427 if (S_sv_2iuv_common(aTHX_ sv))
2431 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2uv(%"UVuf")\n",
2432 PTR2UV(sv),SvUVX(sv)));
2433 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2437 =for apidoc sv_2nv_flags
2439 Return the num value of an SV, doing any necessary string or integer
2440 conversion. If flags includes SV_GMAGIC, does an mg_get() first.
2441 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2447 Perl_sv_2nv_flags(pTHX_ register SV *const sv, const I32 flags)
2452 if (SvGMAGICAL(sv) || SvVALID(sv)) {
2453 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2454 the same flag bit as SVf_IVisUV, so must not let them cache NVs. */
2455 if (flags & SV_GMAGIC)
2459 if ((SvPOKp(sv) && SvLEN(sv)) && !SvIOKp(sv)) {
2460 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2461 !grok_number(SvPVX_const(sv), SvCUR(sv), NULL))
2463 return Atof(SvPVX_const(sv));
2467 return (NV)SvUVX(sv);
2469 return (NV)SvIVX(sv);
2474 assert(SvTYPE(sv) >= SVt_PVMG);
2475 /* This falls through to the report_uninit near the end of the
2477 } else if (SvTHINKFIRST(sv)) {
2482 if (flags & SV_SKIP_OVERLOAD)
2484 tmpstr = AMG_CALLunary(sv, numer_amg);
2485 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2486 return SvNV(tmpstr);
2489 return PTR2NV(SvRV(sv));
2492 sv_force_normal_flags(sv, 0);
2494 if (SvREADONLY(sv) && !SvOK(sv)) {
2495 if (ckWARN(WARN_UNINITIALIZED))
2500 if (SvTYPE(sv) < SVt_NV) {
2501 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2502 sv_upgrade(sv, SVt_NV);
2503 #ifdef USE_LONG_DOUBLE
2505 STORE_NUMERIC_LOCAL_SET_STANDARD();
2506 PerlIO_printf(Perl_debug_log,
2507 "0x%"UVxf" num(%" PERL_PRIgldbl ")\n",
2508 PTR2UV(sv), SvNVX(sv));
2509 RESTORE_NUMERIC_LOCAL();
2513 STORE_NUMERIC_LOCAL_SET_STANDARD();
2514 PerlIO_printf(Perl_debug_log, "0x%"UVxf" num(%"NVgf")\n",
2515 PTR2UV(sv), SvNVX(sv));
2516 RESTORE_NUMERIC_LOCAL();
2520 else if (SvTYPE(sv) < SVt_PVNV)
2521 sv_upgrade(sv, SVt_PVNV);
2526 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2527 #ifdef NV_PRESERVES_UV
2533 /* Only set the public NV OK flag if this NV preserves the IV */
2534 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2536 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2537 : (SvIVX(sv) == I_V(SvNVX(sv))))
2543 else if (SvPOKp(sv) && SvLEN(sv)) {
2545 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2546 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2548 #ifdef NV_PRESERVES_UV
2549 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2550 == IS_NUMBER_IN_UV) {
2551 /* It's definitely an integer */
2552 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2554 SvNV_set(sv, Atof(SvPVX_const(sv)));
2560 SvNV_set(sv, Atof(SvPVX_const(sv)));
2561 /* Only set the public NV OK flag if this NV preserves the value in
2562 the PV at least as well as an IV/UV would.
2563 Not sure how to do this 100% reliably. */
2564 /* if that shift count is out of range then Configure's test is
2565 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2567 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2568 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2569 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2570 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2571 /* Can't use strtol etc to convert this string, so don't try.
2572 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2575 /* value has been set. It may not be precise. */
2576 if ((numtype & IS_NUMBER_NEG) && (value > (UV)IV_MIN)) {
2577 /* 2s complement assumption for (UV)IV_MIN */
2578 SvNOK_on(sv); /* Integer is too negative. */
2583 if (numtype & IS_NUMBER_NEG) {
2584 SvIV_set(sv, -(IV)value);
2585 } else if (value <= (UV)IV_MAX) {
2586 SvIV_set(sv, (IV)value);
2588 SvUV_set(sv, value);
2592 if (numtype & IS_NUMBER_NOT_INT) {
2593 /* I believe that even if the original PV had decimals,
2594 they are lost beyond the limit of the FP precision.
2595 However, neither is canonical, so both only get p
2596 flags. NWC, 2000/11/25 */
2597 /* Both already have p flags, so do nothing */
2599 const NV nv = SvNVX(sv);
2600 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2601 if (SvIVX(sv) == I_V(nv)) {
2604 /* It had no "." so it must be integer. */
2608 /* between IV_MAX and NV(UV_MAX).
2609 Could be slightly > UV_MAX */
2611 if (numtype & IS_NUMBER_NOT_INT) {
2612 /* UV and NV both imprecise. */
2614 const UV nv_as_uv = U_V(nv);
2616 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2625 /* It might be more code efficient to go through the entire logic above
2626 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2627 gets complex and potentially buggy, so more programmer efficient
2628 to do it this way, by turning off the public flags: */
2630 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2631 #endif /* NV_PRESERVES_UV */
2634 if (isGV_with_GP(sv)) {
2635 glob_2number(MUTABLE_GV(sv));
2639 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2641 assert (SvTYPE(sv) >= SVt_NV);
2642 /* Typically the caller expects that sv_any is not NULL now. */
2643 /* XXX Ilya implies that this is a bug in callers that assume this
2644 and ideally should be fixed. */
2647 #if defined(USE_LONG_DOUBLE)
2649 STORE_NUMERIC_LOCAL_SET_STANDARD();
2650 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2nv(%" PERL_PRIgldbl ")\n",
2651 PTR2UV(sv), SvNVX(sv));
2652 RESTORE_NUMERIC_LOCAL();
2656 STORE_NUMERIC_LOCAL_SET_STANDARD();
2657 PerlIO_printf(Perl_debug_log, "0x%"UVxf" 1nv(%"NVgf")\n",
2658 PTR2UV(sv), SvNVX(sv));
2659 RESTORE_NUMERIC_LOCAL();
2668 Return an SV with the numeric value of the source SV, doing any necessary
2669 reference or overload conversion. You must use the C<SvNUM(sv)> macro to
2670 access this function.
2676 Perl_sv_2num(pTHX_ register SV *const sv)
2678 PERL_ARGS_ASSERT_SV_2NUM;
2683 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2684 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2685 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2686 return sv_2num(tmpsv);
2688 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2691 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2692 * UV as a string towards the end of buf, and return pointers to start and
2695 * We assume that buf is at least TYPE_CHARS(UV) long.
2699 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2701 char *ptr = buf + TYPE_CHARS(UV);
2702 char * const ebuf = ptr;
2705 PERL_ARGS_ASSERT_UIV_2BUF;
2717 *--ptr = '0' + (char)(uv % 10);
2726 =for apidoc sv_2pv_flags
2728 Returns a pointer to the string value of an SV, and sets *lp to its length.
2729 If flags includes SV_GMAGIC, does an mg_get() first. Coerces sv to a
2730 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2731 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2737 Perl_sv_2pv_flags(pTHX_ register SV *const sv, STRLEN *const lp, const I32 flags)
2747 if (SvGMAGICAL(sv)) {
2748 if (flags & SV_GMAGIC)
2753 if (flags & SV_MUTABLE_RETURN)
2754 return SvPVX_mutable(sv);
2755 if (flags & SV_CONST_RETURN)
2756 return (char *)SvPVX_const(sv);
2759 if (SvIOKp(sv) || SvNOKp(sv)) {
2760 char tbuf[64]; /* Must fit sprintf/Gconvert of longest IV/NV */
2765 ? my_snprintf(tbuf, sizeof(tbuf), "%"UVuf, (UV)SvUVX(sv))
2766 : my_snprintf(tbuf, sizeof(tbuf), "%"IVdf, (IV)SvIVX(sv));
2767 } else if(SvNVX(sv) == 0.0) {
2772 Gconvert(SvNVX(sv), NV_DIG, 0, tbuf);
2779 SvUPGRADE(sv, SVt_PV);
2782 s = SvGROW_mutable(sv, len + 1);
2785 return (char*)memcpy(s, tbuf, len + 1);
2791 assert(SvTYPE(sv) >= SVt_PVMG);
2792 /* This falls through to the report_uninit near the end of the
2794 } else if (SvTHINKFIRST(sv)) {
2799 if (flags & SV_SKIP_OVERLOAD)
2801 tmpstr = AMG_CALLunary(sv, string_amg);
2802 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
2803 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2805 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
2809 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
2810 if (flags & SV_CONST_RETURN) {
2811 pv = (char *) SvPVX_const(tmpstr);
2813 pv = (flags & SV_MUTABLE_RETURN)
2814 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
2817 *lp = SvCUR(tmpstr);
2819 pv = sv_2pv_flags(tmpstr, lp, flags);
2832 SV *const referent = SvRV(sv);
2836 retval = buffer = savepvn("NULLREF", len);
2837 } else if (SvTYPE(referent) == SVt_REGEXP && (
2838 !(PL_curcop->cop_hints & HINT_NO_AMAGIC)
2839 || amagic_is_enabled(string_amg)
2841 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
2845 /* If the regex is UTF-8 we want the containing scalar to
2846 have an UTF-8 flag too */
2853 *lp = RX_WRAPLEN(re);
2855 return RX_WRAPPED(re);
2857 const char *const typestr = sv_reftype(referent, 0);
2858 const STRLEN typelen = strlen(typestr);
2859 UV addr = PTR2UV(referent);
2860 const char *stashname = NULL;
2861 STRLEN stashnamelen = 0; /* hush, gcc */
2862 const char *buffer_end;
2864 if (SvOBJECT(referent)) {
2865 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
2868 stashname = HEK_KEY(name);
2869 stashnamelen = HEK_LEN(name);
2871 if (HEK_UTF8(name)) {
2877 stashname = "__ANON__";
2880 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
2881 + 2 * sizeof(UV) + 2 /* )\0 */;
2883 len = typelen + 3 /* (0x */
2884 + 2 * sizeof(UV) + 2 /* )\0 */;
2887 Newx(buffer, len, char);
2888 buffer_end = retval = buffer + len;
2890 /* Working backwards */
2894 *--retval = PL_hexdigit[addr & 15];
2895 } while (addr >>= 4);
2901 memcpy(retval, typestr, typelen);
2905 retval -= stashnamelen;
2906 memcpy(retval, stashname, stashnamelen);
2908 /* retval may not necessarily have reached the start of the
2910 assert (retval >= buffer);
2912 len = buffer_end - retval - 1; /* -1 for that \0 */
2920 if (SvREADONLY(sv) && !SvOK(sv)) {
2923 if (flags & SV_UNDEF_RETURNS_NULL)
2925 if (ckWARN(WARN_UNINITIALIZED))
2930 if (SvIOK(sv) || ((SvIOKp(sv) && !SvNOKp(sv)))) {
2931 /* I'm assuming that if both IV and NV are equally valid then
2932 converting the IV is going to be more efficient */
2933 const U32 isUIOK = SvIsUV(sv);
2934 char buf[TYPE_CHARS(UV)];
2938 if (SvTYPE(sv) < SVt_PVIV)
2939 sv_upgrade(sv, SVt_PVIV);
2940 ptr = uiv_2buf(buf, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
2942 /* inlined from sv_setpvn */
2943 s = SvGROW_mutable(sv, len + 1);
2944 Move(ptr, s, len, char);
2948 else if (SvNOKp(sv)) {
2949 if (SvTYPE(sv) < SVt_PVNV)
2950 sv_upgrade(sv, SVt_PVNV);
2951 if (SvNVX(sv) == 0.0) {
2952 s = SvGROW_mutable(sv, 2);
2957 /* The +20 is pure guesswork. Configure test needed. --jhi */
2958 s = SvGROW_mutable(sv, NV_DIG + 20);
2959 /* some Xenix systems wipe out errno here */
2960 Gconvert(SvNVX(sv), NV_DIG, 0, s);
2970 if (isGV_with_GP(sv)) {
2971 GV *const gv = MUTABLE_GV(sv);
2972 SV *const buffer = sv_newmortal();
2974 gv_efullname3(buffer, gv, "*");
2976 assert(SvPOK(buffer));
2978 *lp = SvCUR(buffer);
2980 if ( SvUTF8(buffer) ) SvUTF8_on(sv);
2981 return SvPVX(buffer);
2986 if (flags & SV_UNDEF_RETURNS_NULL)
2988 if (!PL_localizing && !SvPADTMP(sv) && ckWARN(WARN_UNINITIALIZED))
2990 if (SvTYPE(sv) < SVt_PV)
2991 /* Typically the caller expects that sv_any is not NULL now. */
2992 sv_upgrade(sv, SVt_PV);
2996 const STRLEN len = s - SvPVX_const(sv);
3002 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%"UVxf" 2pv(%s)\n",
3003 PTR2UV(sv),SvPVX_const(sv)));
3004 if (flags & SV_CONST_RETURN)
3005 return (char *)SvPVX_const(sv);
3006 if (flags & SV_MUTABLE_RETURN)
3007 return SvPVX_mutable(sv);
3012 =for apidoc sv_copypv
3014 Copies a stringified representation of the source SV into the
3015 destination SV. Automatically performs any necessary mg_get and
3016 coercion of numeric values into strings. Guaranteed to preserve
3017 UTF8 flag even from overloaded objects. Similar in nature to
3018 sv_2pv[_flags] but operates directly on an SV instead of just the
3019 string. Mostly uses sv_2pv_flags to do its work, except when that
3020 would lose the UTF-8'ness of the PV.
3026 Perl_sv_copypv(pTHX_ SV *const dsv, register SV *const ssv)
3029 const char * const s = SvPV_const(ssv,len);
3031 PERL_ARGS_ASSERT_SV_COPYPV;
3033 sv_setpvn(dsv,s,len);
3041 =for apidoc sv_2pvbyte
3043 Return a pointer to the byte-encoded representation of the SV, and set *lp
3044 to its length. May cause the SV to be downgraded from UTF-8 as a
3047 Usually accessed via the C<SvPVbyte> macro.
3053 Perl_sv_2pvbyte(pTHX_ register SV *sv, STRLEN *const lp)
3055 PERL_ARGS_ASSERT_SV_2PVBYTE;
3057 if ((SvTHINKFIRST(sv) && !SvIsCOW(sv)) || isGV_with_GP(sv)) {
3058 SV *sv2 = sv_newmortal();
3062 else SvGETMAGIC(sv);
3063 sv_utf8_downgrade(sv,0);
3064 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3068 =for apidoc sv_2pvutf8
3070 Return a pointer to the UTF-8-encoded representation of the SV, and set *lp
3071 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3073 Usually accessed via the C<SvPVutf8> macro.
3079 Perl_sv_2pvutf8(pTHX_ register SV *sv, STRLEN *const lp)
3081 PERL_ARGS_ASSERT_SV_2PVUTF8;
3083 if ((SvTHINKFIRST(sv) && !SvIsCOW(sv)) || isGV_with_GP(sv))
3084 sv = sv_mortalcopy(sv);
3085 sv_utf8_upgrade(sv);
3086 if (SvGMAGICAL(sv)) SvFLAGS(sv) &= ~SVf_POK;
3088 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3093 =for apidoc sv_2bool
3095 This macro is only used by sv_true() or its macro equivalent, and only if
3096 the latter's argument is neither SvPOK, SvIOK nor SvNOK.
3097 It calls sv_2bool_flags with the SV_GMAGIC flag.
3099 =for apidoc sv_2bool_flags
3101 This function is only used by sv_true() and friends, and only if
3102 the latter's argument is neither SvPOK, SvIOK nor SvNOK. If the flags
3103 contain SV_GMAGIC, then it does an mg_get() first.
3110 Perl_sv_2bool_flags(pTHX_ register SV *const sv, const I32 flags)
3114 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3116 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3122 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3123 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
3124 return cBOOL(SvTRUE(tmpsv));
3126 return SvRV(sv) != 0;
3129 register XPV* const Xpvtmp = (XPV*)SvANY(sv);
3131 (*sv->sv_u.svu_pv > '0' ||
3132 Xpvtmp->xpv_cur > 1 ||
3133 (Xpvtmp->xpv_cur && *sv->sv_u.svu_pv != '0')))
3140 return SvIVX(sv) != 0;
3143 return SvNVX(sv) != 0.0;
3145 if (isGV_with_GP(sv))
3155 =for apidoc sv_utf8_upgrade
3157 Converts the PV of an SV to its UTF-8-encoded form.
3158 Forces the SV to string form if it is not already.
3159 Will C<mg_get> on C<sv> if appropriate.
3160 Always sets the SvUTF8 flag to avoid future validity checks even
3161 if the whole string is the same in UTF-8 as not.
3162 Returns the number of bytes in the converted string
3164 This is not as a general purpose byte encoding to Unicode interface:
3165 use the Encode extension for that.
3167 =for apidoc sv_utf8_upgrade_nomg
3169 Like sv_utf8_upgrade, but doesn't do magic on C<sv>.
3171 =for apidoc sv_utf8_upgrade_flags
3173 Converts the PV of an SV to its UTF-8-encoded form.
3174 Forces the SV to string form if it is not already.
3175 Always sets the SvUTF8 flag to avoid future validity checks even
3176 if all the bytes are invariant in UTF-8.
3177 If C<flags> has C<SV_GMAGIC> bit set,
3178 will C<mg_get> on C<sv> if appropriate, else not.
3179 Returns the number of bytes in the converted string
3180 C<sv_utf8_upgrade> and
3181 C<sv_utf8_upgrade_nomg> are implemented in terms of this function.
3183 This is not as a general purpose byte encoding to Unicode interface:
3184 use the Encode extension for that.
3188 The grow version is currently not externally documented. It adds a parameter,
3189 extra, which is the number of unused bytes the string of 'sv' is guaranteed to
3190 have free after it upon return. This allows the caller to reserve extra space
3191 that it intends to fill, to avoid extra grows.
3193 Also externally undocumented for the moment is the flag SV_FORCE_UTF8_UPGRADE,
3194 which can be used to tell this function to not first check to see if there are
3195 any characters that are different in UTF-8 (variant characters) which would
3196 force it to allocate a new string to sv, but to assume there are. Typically
3197 this flag is used by a routine that has already parsed the string to find that
3198 there are such characters, and passes this information on so that the work
3199 doesn't have to be repeated.
3201 (One might think that the calling routine could pass in the position of the
3202 first such variant, so it wouldn't have to be found again. But that is not the
3203 case, because typically when the caller is likely to use this flag, it won't be
3204 calling this routine unless it finds something that won't fit into a byte.
3205 Otherwise it tries to not upgrade and just use bytes. But some things that
3206 do fit into a byte are variants in utf8, and the caller may not have been
3207 keeping track of these.)
3209 If the routine itself changes the string, it adds a trailing NUL. Such a NUL
3210 isn't guaranteed due to having other routines do the work in some input cases,
3211 or if the input is already flagged as being in utf8.
3213 The speed of this could perhaps be improved for many cases if someone wanted to
3214 write a fast function that counts the number of variant characters in a string,
3215 especially if it could return the position of the first one.
3220 Perl_sv_utf8_upgrade_flags_grow(pTHX_ register SV *const sv, const I32 flags, STRLEN extra)
3224 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3226 if (sv == &PL_sv_undef)
3230 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3231 (void) sv_2pv_flags(sv,&len, flags);
3233 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3237 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3242 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3247 sv_force_normal_flags(sv, 0);
3250 if (PL_encoding && !(flags & SV_UTF8_NO_ENCODING)) {
3251 sv_recode_to_utf8(sv, PL_encoding);
3252 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3256 if (SvCUR(sv) == 0) {
3257 if (extra) SvGROW(sv, extra);
3258 } else { /* Assume Latin-1/EBCDIC */
3259 /* This function could be much more efficient if we
3260 * had a FLAG in SVs to signal if there are any variant
3261 * chars in the PV. Given that there isn't such a flag
3262 * make the loop as fast as possible (although there are certainly ways
3263 * to speed this up, eg. through vectorization) */
3264 U8 * s = (U8 *) SvPVX_const(sv);
3265 U8 * e = (U8 *) SvEND(sv);
3267 STRLEN two_byte_count = 0;
3269 if (flags & SV_FORCE_UTF8_UPGRADE) goto must_be_utf8;
3271 /* See if really will need to convert to utf8. We mustn't rely on our
3272 * incoming SV being well formed and having a trailing '\0', as certain
3273 * code in pp_formline can send us partially built SVs. */
3277 if (NATIVE_IS_INVARIANT(ch)) continue;
3279 t--; /* t already incremented; re-point to first variant */
3284 /* utf8 conversion not needed because all are invariants. Mark as
3285 * UTF-8 even if no variant - saves scanning loop */
3287 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3292 /* Here, the string should be converted to utf8, either because of an
3293 * input flag (two_byte_count = 0), or because a character that
3294 * requires 2 bytes was found (two_byte_count = 1). t points either to
3295 * the beginning of the string (if we didn't examine anything), or to
3296 * the first variant. In either case, everything from s to t - 1 will
3297 * occupy only 1 byte each on output.
3299 * There are two main ways to convert. One is to create a new string
3300 * and go through the input starting from the beginning, appending each
3301 * converted value onto the new string as we go along. It's probably
3302 * best to allocate enough space in the string for the worst possible
3303 * case rather than possibly running out of space and having to
3304 * reallocate and then copy what we've done so far. Since everything
3305 * from s to t - 1 is invariant, the destination can be initialized
3306 * with these using a fast memory copy
3308 * The other way is to figure out exactly how big the string should be
3309 * by parsing the entire input. Then you don't have to make it big
3310 * enough to handle the worst possible case, and more importantly, if
3311 * the string you already have is large enough, you don't have to
3312 * allocate a new string, you can copy the last character in the input
3313 * string to the final position(s) that will be occupied by the
3314 * converted string and go backwards, stopping at t, since everything
3315 * before that is invariant.
3317 * There are advantages and disadvantages to each method.
3319 * In the first method, we can allocate a new string, do the memory
3320 * copy from the s to t - 1, and then proceed through the rest of the
3321 * string byte-by-byte.
3323 * In the second method, we proceed through the rest of the input
3324 * string just calculating how big the converted string will be. Then
3325 * there are two cases:
3326 * 1) if the string has enough extra space to handle the converted
3327 * value. We go backwards through the string, converting until we
3328 * get to the position we are at now, and then stop. If this
3329 * position is far enough along in the string, this method is
3330 * faster than the other method. If the memory copy were the same
3331 * speed as the byte-by-byte loop, that position would be about
3332 * half-way, as at the half-way mark, parsing to the end and back
3333 * is one complete string's parse, the same amount as starting
3334 * over and going all the way through. Actually, it would be
3335 * somewhat less than half-way, as it's faster to just count bytes
3336 * than to also copy, and we don't have the overhead of allocating
3337 * a new string, changing the scalar to use it, and freeing the
3338 * existing one. But if the memory copy is fast, the break-even
3339 * point is somewhere after half way. The counting loop could be
3340 * sped up by vectorization, etc, to move the break-even point
3341 * further towards the beginning.
3342 * 2) if the string doesn't have enough space to handle the converted
3343 * value. A new string will have to be allocated, and one might
3344 * as well, given that, start from the beginning doing the first
3345 * method. We've spent extra time parsing the string and in
3346 * exchange all we've gotten is that we know precisely how big to
3347 * make the new one. Perl is more optimized for time than space,
3348 * so this case is a loser.
3349 * So what I've decided to do is not use the 2nd method unless it is
3350 * guaranteed that a new string won't have to be allocated, assuming
3351 * the worst case. I also decided not to put any more conditions on it
3352 * than this, for now. It seems likely that, since the worst case is
3353 * twice as big as the unknown portion of the string (plus 1), we won't
3354 * be guaranteed enough space, causing us to go to the first method,
3355 * unless the string is short, or the first variant character is near
3356 * the end of it. In either of these cases, it seems best to use the
3357 * 2nd method. The only circumstance I can think of where this would
3358 * be really slower is if the string had once had much more data in it
3359 * than it does now, but there is still a substantial amount in it */
3362 STRLEN invariant_head = t - s;
3363 STRLEN size = invariant_head + (e - t) * 2 + 1 + extra;
3364 if (SvLEN(sv) < size) {
3366 /* Here, have decided to allocate a new string */
3371 Newx(dst, size, U8);
3373 /* If no known invariants at the beginning of the input string,
3374 * set so starts from there. Otherwise, can use memory copy to
3375 * get up to where we are now, and then start from here */
3377 if (invariant_head <= 0) {
3380 Copy(s, dst, invariant_head, char);
3381 d = dst + invariant_head;
3385 const UV uv = NATIVE8_TO_UNI(*t++);
3386 if (UNI_IS_INVARIANT(uv))
3387 *d++ = (U8)UNI_TO_NATIVE(uv);
3389 *d++ = (U8)UTF8_EIGHT_BIT_HI(uv);
3390 *d++ = (U8)UTF8_EIGHT_BIT_LO(uv);
3394 SvPV_free(sv); /* No longer using pre-existing string */
3395 SvPV_set(sv, (char*)dst);
3396 SvCUR_set(sv, d - dst);
3397 SvLEN_set(sv, size);
3400 /* Here, have decided to get the exact size of the string.
3401 * Currently this happens only when we know that there is
3402 * guaranteed enough space to fit the converted string, so
3403 * don't have to worry about growing. If two_byte_count is 0,
3404 * then t points to the first byte of the string which hasn't
3405 * been examined yet. Otherwise two_byte_count is 1, and t
3406 * points to the first byte in the string that will expand to
3407 * two. Depending on this, start examining at t or 1 after t.
3410 U8 *d = t + two_byte_count;
3413 /* Count up the remaining bytes that expand to two */
3416 const U8 chr = *d++;
3417 if (! NATIVE_IS_INVARIANT(chr)) two_byte_count++;
3420 /* The string will expand by just the number of bytes that
3421 * occupy two positions. But we are one afterwards because of
3422 * the increment just above. This is the place to put the
3423 * trailing NUL, and to set the length before we decrement */
3425 d += two_byte_count;
3426 SvCUR_set(sv, d - s);
3430 /* Having decremented d, it points to the position to put the
3431 * very last byte of the expanded string. Go backwards through
3432 * the string, copying and expanding as we go, stopping when we
3433 * get to the part that is invariant the rest of the way down */
3437 const U8 ch = NATIVE8_TO_UNI(*e--);
3438 if (UNI_IS_INVARIANT(ch)) {
3439 *d-- = UNI_TO_NATIVE(ch);
3441 *d-- = (U8)UTF8_EIGHT_BIT_LO(ch);
3442 *d-- = (U8)UTF8_EIGHT_BIT_HI(ch);
3447 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3448 /* Update pos. We do it at the end rather than during
3449 * the upgrade, to avoid slowing down the common case
3450 * (upgrade without pos) */
3451 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3453 I32 pos = mg->mg_len;
3454 if (pos > 0 && (U32)pos > invariant_head) {
3455 U8 *d = (U8*) SvPVX(sv) + invariant_head;
3456 STRLEN n = (U32)pos - invariant_head;
3458 if (UTF8_IS_START(*d))
3463 mg->mg_len = d - (U8*)SvPVX(sv);
3466 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3467 magic_setutf8(sv,mg); /* clear UTF8 cache */
3472 /* Mark as UTF-8 even if no variant - saves scanning loop */
3478 =for apidoc sv_utf8_downgrade
3480 Attempts to convert the PV of an SV from characters to bytes.
3481 If the PV contains a character that cannot fit
3482 in a byte, this conversion will fail;
3483 in this case, either returns false or, if C<fail_ok> is not
3486 This is not as a general purpose Unicode to byte encoding interface:
3487 use the Encode extension for that.
3493 Perl_sv_utf8_downgrade(pTHX_ register SV *const sv, const bool fail_ok)
3497 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE;
3499 if (SvPOKp(sv) && SvUTF8(sv)) {
3503 int mg_flags = SV_GMAGIC;
3506 sv_force_normal_flags(sv, 0);
3508 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3510 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3512 I32 pos = mg->mg_len;
3514 sv_pos_b2u(sv, &pos);
3515 mg_flags = 0; /* sv_pos_b2u does get magic */
3519 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3520 magic_setutf8(sv,mg); /* clear UTF8 cache */
3523 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3525 if (!utf8_to_bytes(s, &len)) {
3530 Perl_croak(aTHX_ "Wide character in %s",
3533 Perl_croak(aTHX_ "Wide character");
3544 =for apidoc sv_utf8_encode
3546 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3547 flag off so that it looks like octets again.
3553 Perl_sv_utf8_encode(pTHX_ register SV *const sv)
3555 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3557 if (SvREADONLY(sv)) {
3558 sv_force_normal_flags(sv, 0);
3560 (void) sv_utf8_upgrade(sv);
3565 =for apidoc sv_utf8_decode
3567 If the PV of the SV is an octet sequence in UTF-8
3568 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3569 so that it looks like a character. If the PV contains only single-byte
3570 characters, the C<SvUTF8> flag stays off.
3571 Scans PV for validity and returns false if the PV is invalid UTF-8.
3577 Perl_sv_utf8_decode(pTHX_ register SV *const sv)
3579 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3582 const U8 *start, *c;
3585 /* The octets may have got themselves encoded - get them back as
3588 if (!sv_utf8_downgrade(sv, TRUE))
3591 /* it is actually just a matter of turning the utf8 flag on, but
3592 * we want to make sure everything inside is valid utf8 first.
3594 c = start = (const U8 *) SvPVX_const(sv);
3595 if (!is_utf8_string(c, SvCUR(sv)))
3597 e = (const U8 *) SvEND(sv);
3600 if (!UTF8_IS_INVARIANT(ch)) {
3605 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3606 /* adjust pos to the start of a UTF8 char sequence */
3607 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3609 I32 pos = mg->mg_len;
3611 for (c = start + pos; c > start; c--) {
3612 if (UTF8_IS_START(*c))
3615 mg->mg_len = c - start;
3618 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3619 magic_setutf8(sv,mg); /* clear UTF8 cache */
3626 =for apidoc sv_setsv
3628 Copies the contents of the source SV C<ssv> into the destination SV
3629 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3630 function if the source SV needs to be reused. Does not handle 'set' magic.
3631 Loosely speaking, it performs a copy-by-value, obliterating any previous
3632 content of the destination.
3634 You probably want to use one of the assortment of wrappers, such as
3635 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3636 C<SvSetMagicSV_nosteal>.
3638 =for apidoc sv_setsv_flags
3640 Copies the contents of the source SV C<ssv> into the destination SV
3641 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3642 function if the source SV needs to be reused. Does not handle 'set' magic.
3643 Loosely speaking, it performs a copy-by-value, obliterating any previous
3644 content of the destination.
3645 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3646 C<ssv> if appropriate, else not. If the C<flags>
3647 parameter has the C<NOSTEAL> bit set then the
3648 buffers of temps will not be stolen. <sv_setsv>
3649 and C<sv_setsv_nomg> are implemented in terms of this function.
3651 You probably want to use one of the assortment of wrappers, such as
3652 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3653 C<SvSetMagicSV_nosteal>.
3655 This is the primary function for copying scalars, and most other
3656 copy-ish functions and macros use this underneath.
3662 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3664 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3665 HV *old_stash = NULL;
3667 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3669 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3670 const char * const name = GvNAME(sstr);
3671 const STRLEN len = GvNAMELEN(sstr);
3673 if (dtype >= SVt_PV) {
3679 SvUPGRADE(dstr, SVt_PVGV);
3680 (void)SvOK_off(dstr);
3681 /* We have to turn this on here, even though we turn it off
3682 below, as GvSTASH will fail an assertion otherwise. */
3683 isGV_with_GP_on(dstr);
3685 GvSTASH(dstr) = GvSTASH(sstr);
3687 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3688 gv_name_set(MUTABLE_GV(dstr), name, len,
3689 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3690 SvFAKE_on(dstr); /* can coerce to non-glob */
3693 if(GvGP(MUTABLE_GV(sstr))) {
3694 /* If source has method cache entry, clear it */
3696 SvREFCNT_dec(GvCV(sstr));
3697 GvCV_set(sstr, NULL);
3700 /* If source has a real method, then a method is
3703 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3709 /* If dest already had a real method, that's a change as well */
3711 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3712 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3717 /* We don't need to check the name of the destination if it was not a
3718 glob to begin with. */
3719 if(dtype == SVt_PVGV) {
3720 const char * const name = GvNAME((const GV *)dstr);
3723 /* The stash may have been detached from the symbol table, so
3725 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3726 && GvAV((const GV *)sstr)
3730 const STRLEN len = GvNAMELEN(dstr);
3731 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3732 || (len == 1 && name[0] == ':')) {
3735 /* Set aside the old stash, so we can reset isa caches on
3737 if((old_stash = GvHV(dstr)))
3738 /* Make sure we do not lose it early. */
3739 SvREFCNT_inc_simple_void_NN(
3740 sv_2mortal((SV *)old_stash)
3746 gp_free(MUTABLE_GV(dstr));
3747 isGV_with_GP_off(dstr); /* SvOK_off does not like globs. */
3748 (void)SvOK_off(dstr);
3749 isGV_with_GP_on(dstr);
3750 GvINTRO_off(dstr); /* one-shot flag */
3751 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3752 if (SvTAINTED(sstr))
3754 if (GvIMPORTED(dstr) != GVf_IMPORTED
3755 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3757 GvIMPORTED_on(dstr);
3760 if(mro_changes == 2) {
3762 SV * const sref = (SV *)GvAV((const GV *)dstr);
3763 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3764 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3765 AV * const ary = newAV();
3766 av_push(ary, mg->mg_obj); /* takes the refcount */
3767 mg->mg_obj = (SV *)ary;
3769 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3771 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3772 mro_isa_changed_in(GvSTASH(dstr));
3774 else if(mro_changes == 3) {
3775 HV * const stash = GvHV(dstr);
3776 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
3782 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
3787 S_glob_assign_ref(pTHX_ SV *const dstr, SV *const sstr)
3789 SV * const sref = SvREFCNT_inc(SvRV(sstr));
3791 const int intro = GvINTRO(dstr);
3794 const U32 stype = SvTYPE(sref);
3796 PERL_ARGS_ASSERT_GLOB_ASSIGN_REF;
3799 GvINTRO_off(dstr); /* one-shot flag */
3800 GvLINE(dstr) = CopLINE(PL_curcop);
3801 GvEGV(dstr) = MUTABLE_GV(dstr);
3806 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
3807 import_flag = GVf_IMPORTED_CV;
3810 location = (SV **) &GvHV(dstr);
3811 import_flag = GVf_IMPORTED_HV;
3814 location = (SV **) &GvAV(dstr);
3815 import_flag = GVf_IMPORTED_AV;
3818 location = (SV **) &GvIOp(dstr);
3821 location = (SV **) &GvFORM(dstr);
3824 location = &GvSV(dstr);
3825 import_flag = GVf_IMPORTED_SV;
3828 if (stype == SVt_PVCV) {
3829 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
3830 if (GvCVGEN(dstr)) {
3831 SvREFCNT_dec(GvCV(dstr));
3832 GvCV_set(dstr, NULL);
3833 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3836 SAVEGENERICSV(*location);
3840 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
3841 CV* const cv = MUTABLE_CV(*location);
3843 if (!GvCVGEN((const GV *)dstr) &&
3844 (CvROOT(cv) || CvXSUB(cv)) &&
3845 /* redundant check that avoids creating the extra SV
3846 most of the time: */
3847 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
3849 SV * const new_const_sv =
3850 CvCONST((const CV *)sref)
3851 ? cv_const_sv((const CV *)sref)
3853 report_redefined_cv(
3854 sv_2mortal(Perl_newSVpvf(aTHX_
3857 HvNAME_HEK(GvSTASH((const GV *)dstr))
3859 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr)))
3862 CvCONST((const CV *)sref) ? &new_const_sv : NULL
3866 cv_ckproto_len_flags(cv, (const GV *)dstr,
3867 SvPOK(sref) ? CvPROTO(sref) : NULL,
3868 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
3869 SvPOK(sref) ? SvUTF8(sref) : 0);
3871 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
3872 GvASSUMECV_on(dstr);
3873 if(GvSTASH(dstr)) mro_method_changed_in(GvSTASH(dstr)); /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
3876 if (import_flag && !(GvFLAGS(dstr) & import_flag)
3877 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
3878 GvFLAGS(dstr) |= import_flag;
3880 if (stype == SVt_PVHV) {
3881 const char * const name = GvNAME((GV*)dstr);
3882 const STRLEN len = GvNAMELEN(dstr);
3885 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
3886 || (len == 1 && name[0] == ':')
3888 && (!dref || HvENAME_get(dref))
3891 (HV *)sref, (HV *)dref,
3897 stype == SVt_PVAV && sref != dref
3898 && strEQ(GvNAME((GV*)dstr), "ISA")
3899 /* The stash may have been detached from the symbol table, so
3900 check its name before doing anything. */
3901 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3904 MAGIC * const omg = dref && SvSMAGICAL(dref)
3905 ? mg_find(dref, PERL_MAGIC_isa)
3907 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3908 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3909 AV * const ary = newAV();
3910 av_push(ary, mg->mg_obj); /* takes the refcount */
3911 mg->mg_obj = (SV *)ary;
3914 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
3915 SV **svp = AvARRAY((AV *)omg->mg_obj);
3916 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
3920 SvREFCNT_inc_simple_NN(*svp++)
3926 SvREFCNT_inc_simple_NN(omg->mg_obj)
3930 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
3935 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
3937 mg = mg_find(sref, PERL_MAGIC_isa);
3939 /* Since the *ISA assignment could have affected more than
3940 one stash, don't call mro_isa_changed_in directly, but let
3941 magic_clearisa do it for us, as it already has the logic for
3942 dealing with globs vs arrays of globs. */
3944 Perl_magic_clearisa(aTHX_ NULL, mg);
3949 if (SvTAINTED(sstr))
3955 Perl_sv_setsv_flags(pTHX_ SV *dstr, register SV* sstr, const I32 flags)
3958 register U32 sflags;
3960 register svtype stype;
3962 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
3967 if (SvIS_FREED(dstr)) {
3968 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
3969 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
3971 SV_CHECK_THINKFIRST_COW_DROP(dstr);
3973 sstr = &PL_sv_undef;
3974 if (SvIS_FREED(sstr)) {
3975 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
3976 (void*)sstr, (void*)dstr);
3978 stype = SvTYPE(sstr);
3979 dtype = SvTYPE(dstr);
3983 /* need to nuke the magic */
3984 sv_unmagic(dstr, PERL_MAGIC_vstring);
3987 /* There's a lot of redundancy below but we're going for speed here */
3992 if (dtype != SVt_PVGV && dtype != SVt_PVLV) {
3993 (void)SvOK_off(dstr);
4001 sv_upgrade(dstr, SVt_IV);
4005 sv_upgrade(dstr, SVt_PVIV);
4009 goto end_of_first_switch;
4011 (void)SvIOK_only(dstr);
4012 SvIV_set(dstr, SvIVX(sstr));
4015 /* SvTAINTED can only be true if the SV has taint magic, which in
4016 turn means that the SV type is PVMG (or greater). This is the
4017 case statement for SVt_IV, so this cannot be true (whatever gcov
4019 assert(!SvTAINTED(sstr));
4024 if (dtype < SVt_PV && dtype != SVt_IV)
4025 sv_upgrade(dstr, SVt_IV);
4033 sv_upgrade(dstr, SVt_NV);
4037 sv_upgrade(dstr, SVt_PVNV);
4041 goto end_of_first_switch;
4043 SvNV_set(dstr, SvNVX(sstr));
4044 (void)SvNOK_only(dstr);
4045 /* SvTAINTED can only be true if the SV has taint magic, which in
4046 turn means that the SV type is PVMG (or greater). This is the
4047 case statement for SVt_NV, so this cannot be true (whatever gcov
4049 assert(!SvTAINTED(sstr));
4055 #ifdef PERL_OLD_COPY_ON_WRITE
4056 if ((SvFLAGS(sstr) & CAN_COW_MASK) == CAN_COW_FLAGS) {
4057 if (dtype < SVt_PVIV)
4058 sv_upgrade(dstr, SVt_PVIV);
4065 sv_upgrade(dstr, SVt_PV);
4068 if (dtype < SVt_PVIV)
4069 sv_upgrade(dstr, SVt_PVIV);
4072 if (dtype < SVt_PVNV)
4073 sv_upgrade(dstr, SVt_PVNV);
4077 const char * const type = sv_reftype(sstr,0);
4079 /* diag_listed_as: Bizarre copy of %s */
4080 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4082 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4087 if (dtype < SVt_REGEXP)
4088 sv_upgrade(dstr, SVt_REGEXP);
4091 /* case SVt_BIND: */
4095 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4097 if (SvTYPE(sstr) != stype)
4098 stype = SvTYPE(sstr);
4100 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4101 glob_assign_glob(dstr, sstr, dtype);
4104 if (stype == SVt_PVLV)
4105 SvUPGRADE(dstr, SVt_PVNV);
4107 SvUPGRADE(dstr, (svtype)stype);
4109 end_of_first_switch:
4111 /* dstr may have been upgraded. */
4112 dtype = SvTYPE(dstr);
4113 sflags = SvFLAGS(sstr);
4115 if (dtype == SVt_PVCV || dtype == SVt_PVFM) {
4116 /* Assigning to a subroutine sets the prototype. */
4119 const char *const ptr = SvPV_const(sstr, len);
4121 SvGROW(dstr, len + 1);
4122 Copy(ptr, SvPVX(dstr), len + 1, char);
4123 SvCUR_set(dstr, len);
4125 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4126 CvAUTOLOAD_off(dstr);
4130 } else if (dtype == SVt_PVAV || dtype == SVt_PVHV) {
4131 const char * const type = sv_reftype(dstr,0);
4133 /* diag_listed_as: Cannot copy to %s */
4134 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4136 Perl_croak(aTHX_ "Cannot copy to %s", type);
4137 } else if (sflags & SVf_ROK) {
4138 if (isGV_with_GP(dstr)
4139 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4142 if (GvIMPORTED(dstr) != GVf_IMPORTED
4143 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4145 GvIMPORTED_on(dstr);
4150 glob_assign_glob(dstr, sstr, dtype);
4154 if (dtype >= SVt_PV) {
4155 if (isGV_with_GP(dstr)) {
4156 glob_assign_ref(dstr, sstr);
4159 if (SvPVX_const(dstr)) {
4165 (void)SvOK_off(dstr);
4166 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4167 SvFLAGS(dstr) |= sflags & SVf_ROK;
4168 assert(!(sflags & SVp_NOK));
4169 assert(!(sflags & SVp_IOK));
4170 assert(!(sflags & SVf_NOK));
4171 assert(!(sflags & SVf_IOK));
4173 else if (isGV_with_GP(dstr)) {
4174 if (!(sflags & SVf_OK)) {
4175 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4176 "Undefined value assigned to typeglob");
4179 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4180 if (dstr != (const SV *)gv) {
4181 const char * const name = GvNAME((const GV *)dstr);
4182 const STRLEN len = GvNAMELEN(dstr);
4183 HV *old_stash = NULL;
4184 bool reset_isa = FALSE;
4185 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4186 || (len == 1 && name[0] == ':')) {
4187 /* Set aside the old stash, so we can reset isa caches
4188 on its subclasses. */
4189 if((old_stash = GvHV(dstr))) {
4190 /* Make sure we do not lose it early. */
4191 SvREFCNT_inc_simple_void_NN(
4192 sv_2mortal((SV *)old_stash)
4199 gp_free(MUTABLE_GV(dstr));
4200 GvGP_set(dstr, gp_ref(GvGP(gv)));
4203 HV * const stash = GvHV(dstr);
4205 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4215 else if (dtype == SVt_REGEXP && stype == SVt_REGEXP) {
4216 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4218 else if (sflags & SVp_POK) {
4222 * Check to see if we can just swipe the string. If so, it's a
4223 * possible small lose on short strings, but a big win on long ones.
4224 * It might even be a win on short strings if SvPVX_const(dstr)
4225 * has to be allocated and SvPVX_const(sstr) has to be freed.
4226 * Likewise if we can set up COW rather than doing an actual copy, we
4227 * drop to the else clause, as the swipe code and the COW setup code
4228 * have much in common.
4231 /* Whichever path we take through the next code, we want this true,
4232 and doing it now facilitates the COW check. */
4233 (void)SvPOK_only(dstr);
4236 /* If we're already COW then this clause is not true, and if COW
4237 is allowed then we drop down to the else and make dest COW
4238 with us. If caller hasn't said that we're allowed to COW
4239 shared hash keys then we don't do the COW setup, even if the
4240 source scalar is a shared hash key scalar. */
4241 (((flags & SV_COW_SHARED_HASH_KEYS)
4242 ? (sflags & (SVf_FAKE|SVf_READONLY)) != (SVf_FAKE|SVf_READONLY)
4243 : 1 /* If making a COW copy is forbidden then the behaviour we
4244 desire is as if the source SV isn't actually already
4245 COW, even if it is. So we act as if the source flags
4246 are not COW, rather than actually testing them. */
4248 #ifndef PERL_OLD_COPY_ON_WRITE
4249 /* The change that added SV_COW_SHARED_HASH_KEYS makes the logic
4250 when PERL_OLD_COPY_ON_WRITE is defined a little wrong.
4251 Conceptually PERL_OLD_COPY_ON_WRITE being defined should
4252 override SV_COW_SHARED_HASH_KEYS, because it means "always COW"
4253 but in turn, it's somewhat dead code, never expected to go
4254 live, but more kept as a placeholder on how to do it better
4255 in a newer implementation. */
4256 /* If we are COW and dstr is a suitable target then we drop down
4257 into the else and make dest a COW of us. */
4258 || (SvFLAGS(dstr) & CAN_COW_MASK) != CAN_COW_FLAGS
4263 (sflags & SVs_TEMP) && /* slated for free anyway? */
4264 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4265 (!(flags & SV_NOSTEAL)) &&
4266 /* and we're allowed to steal temps */
4267 SvREFCNT(sstr) == 1 && /* and no other references to it? */
4268 SvLEN(sstr)) /* and really is a string */
4269 #ifdef PERL_OLD_COPY_ON_WRITE
4270 && ((flags & SV_COW_SHARED_HASH_KEYS)
4271 ? (!((sflags & CAN_COW_MASK) == CAN_COW_FLAGS
4272 && (SvFLAGS(dstr) & CAN_COW_MASK) == CAN_COW_FLAGS
4273 && SvTYPE(sstr) >= SVt_PVIV && SvTYPE(sstr) != SVt_PVFM))
4277 /* Failed the swipe test, and it's not a shared hash key either.
4278 Have to copy the string. */
4279 STRLEN len = SvCUR(sstr);
4280 SvGROW(dstr, len + 1); /* inlined from sv_setpvn */
4281 Move(SvPVX_const(sstr),SvPVX(dstr),len,char);
4282 SvCUR_set(dstr, len);
4283 *SvEND(dstr) = '\0';
4285 /* If PERL_OLD_COPY_ON_WRITE is not defined, then isSwipe will always
4287 /* Either it's a shared hash key, or it's suitable for
4288 copy-on-write or we can swipe the string. */
4290 PerlIO_printf(Perl_debug_log, "Copy on write: sstr --> dstr\n");
4294 #ifdef PERL_OLD_COPY_ON_WRITE
4296 if ((sflags & (SVf_FAKE | SVf_READONLY))
4297 != (SVf_FAKE | SVf_READONLY)) {
4298 SvREADONLY_on(sstr);
4300 /* Make the source SV into a loop of 1.
4301 (about to become 2) */
4302 SV_COW_NEXT_SV_SET(sstr, sstr);
4306 /* Initial code is common. */
4307 if (SvPVX_const(dstr)) { /* we know that dtype >= SVt_PV */
4312 /* making another shared SV. */
4313 STRLEN cur = SvCUR(sstr);
4314 STRLEN len = SvLEN(sstr);
4315 #ifdef PERL_OLD_COPY_ON_WRITE
4317 assert (SvTYPE(dstr) >= SVt_PVIV);
4318 /* SvIsCOW_normal */
4319 /* splice us in between source and next-after-source. */
4320 SV_COW_NEXT_SV_SET(dstr, SV_COW_NEXT_SV(sstr));
4321 SV_COW_NEXT_SV_SET(sstr, dstr);
4322 SvPV_set(dstr, SvPVX_mutable(sstr));
4326 /* SvIsCOW_shared_hash */
4327 DEBUG_C(PerlIO_printf(Perl_debug_log,
4328 "Copy on write: Sharing hash\n"));
4330 assert (SvTYPE(dstr) >= SVt_PV);
4332 HEK_KEY(share_hek_hek(SvSHARED_HEK_FROM_PV(SvPVX_const(sstr)))));
4334 SvLEN_set(dstr, len);
4335 SvCUR_set(dstr, cur);
4336 SvREADONLY_on(dstr);
4340 { /* Passes the swipe test. */
4341 SvPV_set(dstr, SvPVX_mutable(sstr));
4342 SvLEN_set(dstr, SvLEN(sstr));
4343 SvCUR_set(dstr, SvCUR(sstr));
4346 (void)SvOK_off(sstr); /* NOTE: nukes most SvFLAGS on sstr */
4347 SvPV_set(sstr, NULL);
4353 if (sflags & SVp_NOK) {
4354 SvNV_set(dstr, SvNVX(sstr));
4356 if (sflags & SVp_IOK) {
4357 SvIV_set(dstr, SvIVX(sstr));
4358 /* Must do this otherwise some other overloaded use of 0x80000000
4359 gets confused. I guess SVpbm_VALID */
4360 if (sflags & SVf_IVisUV)
4363 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_NOK|SVp_NOK|SVf_UTF8);
4365 const MAGIC * const smg = SvVSTRING_mg(sstr);
4367 sv_magic(dstr, NULL, PERL_MAGIC_vstring,
4368 smg->mg_ptr, smg->mg_len);
4369 SvRMAGICAL_on(dstr);
4373 else if (sflags & (SVp_IOK|SVp_NOK)) {
4374 (void)SvOK_off(dstr);
4375 SvFLAGS(dstr) |= sflags & (SVf_IOK|SVp_IOK|SVf_IVisUV|SVf_NOK|SVp_NOK);
4376 if (sflags & SVp_IOK) {
4377 /* XXXX Do we want to set IsUV for IV(ROK)? Be extra safe... */
4378 SvIV_set(dstr, SvIVX(sstr));
4380 if (sflags & SVp_NOK) {
4381 SvNV_set(dstr, SvNVX(sstr));
4385 if (isGV_with_GP(sstr)) {
4386 gv_efullname3(dstr, MUTABLE_GV(sstr), "*");
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 %"
4513 SvUPGRADE(sv, SVt_PV);
4515 dptr = SvGROW(sv, len + 1);
4516 Move(ptr,dptr,len,char);
4519 (void)SvPOK_only_UTF8(sv); /* validate pointer */
4521 if (SvTYPE(sv) == SVt_PVCV) CvAUTOLOAD_off(sv);
4525 =for apidoc sv_setpvn_mg
4527 Like C<sv_setpvn>, but also handles 'set' magic.
4533 Perl_sv_setpvn_mg(pTHX_ register SV *const sv, register const char *const ptr, register const STRLEN len)
4535 PERL_ARGS_ASSERT_SV_SETPVN_MG;
4537 sv_setpvn(sv,ptr,len);
4542 =for apidoc sv_setpv
4544 Copies a string into an SV. The string must be null-terminated. Does not
4545 handle 'set' magic. See C<sv_setpv_mg>.
4551 Perl_sv_setpv(pTHX_ register SV *const sv, register const char *const ptr)
4554 register STRLEN len;
4556 PERL_ARGS_ASSERT_SV_SETPV;
4558 SV_CHECK_THINKFIRST_COW_DROP(sv);
4564 SvUPGRADE(sv, SVt_PV);
4566 SvGROW(sv, len + 1);
4567 Move(ptr,SvPVX(sv),len+1,char);
4569 (void)SvPOK_only_UTF8(sv); /* validate pointer */
4571 if (SvTYPE(sv) == SVt_PVCV) CvAUTOLOAD_off(sv);
4575 =for apidoc sv_setpv_mg
4577 Like C<sv_setpv>, but also handles 'set' magic.
4583 Perl_sv_setpv_mg(pTHX_ register SV *const sv, register const char *const ptr)
4585 PERL_ARGS_ASSERT_SV_SETPV_MG;
4592 Perl_sv_sethek(pTHX_ register SV *const sv, const HEK *const hek)
4596 PERL_ARGS_ASSERT_SV_SETHEK;
4602 if (HEK_LEN(hek) == HEf_SVKEY) {
4603 sv_setsv(sv, *(SV**)HEK_KEY(hek));
4606 const int flags = HEK_FLAGS(hek);
4607 if (flags & HVhek_WASUTF8) {
4608 STRLEN utf8_len = HEK_LEN(hek);
4609 char *as_utf8 = (char *)bytes_to_utf8((U8*)HEK_KEY(hek), &utf8_len);
4610 sv_usepvn_flags(sv, as_utf8, utf8_len, SV_HAS_TRAILING_NUL);
4613 } else if (flags & (HVhek_REHASH|HVhek_UNSHARED)) {