3 * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
4 * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009 by Larry Wall
7 * You may distribute under the terms of either the GNU General Public
8 * License or the Artistic License, as specified in the README file.
13 * 'I wonder what the Entish is for "yes" and "no",' he thought.
16 * [p.480 of _The Lord of the Rings_, III/iv: "Treebeard"]
22 * This file contains the code that creates, manipulates and destroys
23 * scalar values (SVs). The other types (AV, HV, GV, etc.) reuse the
24 * structure of an SV, so their creation and destruction is handled
25 * here; higher-level functions are in av.c, hv.c, and so on. Opcode
26 * level functions (eg. substr, split, join) for each of the types are
39 /* Missing proto on LynxOS */
40 char *gconvert(double, int, int, char *);
44 # define SNPRINTF_G(nv, buffer, size, ndig) \
45 quadmath_snprintf(buffer, size, "%.*Qg", (int)ndig, (NV)(nv))
47 # define SNPRINTF_G(nv, buffer, size, ndig) \
48 PERL_UNUSED_RESULT(Gconvert((NV)(nv), (int)ndig, 0, buffer))
51 #ifndef SV_COW_THRESHOLD
52 # define SV_COW_THRESHOLD 0 /* COW iff len > K */
54 #ifndef SV_COWBUF_THRESHOLD
55 # define SV_COWBUF_THRESHOLD 1250 /* COW iff len > K */
57 #ifndef SV_COW_MAX_WASTE_THRESHOLD
58 # define SV_COW_MAX_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
60 #ifndef SV_COWBUF_WASTE_THRESHOLD
61 # define SV_COWBUF_WASTE_THRESHOLD 80 /* COW iff (len - cur) < K */
63 #ifndef SV_COW_MAX_WASTE_FACTOR_THRESHOLD
64 # define SV_COW_MAX_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
66 #ifndef SV_COWBUF_WASTE_FACTOR_THRESHOLD
67 # define SV_COWBUF_WASTE_FACTOR_THRESHOLD 2 /* COW iff len < (cur * K) */
69 /* Work around compiler warnings about unsigned >= THRESHOLD when thres-
72 # define GE_COW_THRESHOLD(cur) ((cur) >= SV_COW_THRESHOLD)
74 # define GE_COW_THRESHOLD(cur) 1
76 #if SV_COWBUF_THRESHOLD
77 # define GE_COWBUF_THRESHOLD(cur) ((cur) >= SV_COWBUF_THRESHOLD)
79 # define GE_COWBUF_THRESHOLD(cur) 1
81 #if SV_COW_MAX_WASTE_THRESHOLD
82 # define GE_COW_MAX_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COW_MAX_WASTE_THRESHOLD)
84 # define GE_COW_MAX_WASTE_THRESHOLD(cur,len) 1
86 #if SV_COWBUF_WASTE_THRESHOLD
87 # define GE_COWBUF_WASTE_THRESHOLD(cur,len) (((len)-(cur)) < SV_COWBUF_WASTE_THRESHOLD)
89 # define GE_COWBUF_WASTE_THRESHOLD(cur,len) 1
91 #if SV_COW_MAX_WASTE_FACTOR_THRESHOLD
92 # define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COW_MAX_WASTE_FACTOR_THRESHOLD * (cur))
94 # define GE_COW_MAX_WASTE_FACTOR_THRESHOLD(cur,len) 1
96 #if SV_COWBUF_WASTE_FACTOR_THRESHOLD
97 # define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) ((len) < SV_COWBUF_WASTE_FACTOR_THRESHOLD * (cur))
99 # define GE_COWBUF_WASTE_FACTOR_THRESHOLD(cur,len) 1
102 #define CHECK_COW_THRESHOLD(cur,len) (\
103 GE_COW_THRESHOLD((cur)) && \
104 GE_COW_MAX_WASTE_THRESHOLD((cur),(len)) && \
105 GE_COW_MAX_WASTE_FACTOR_THRESHOLD((cur),(len)) \
107 #define CHECK_COWBUF_THRESHOLD(cur,len) (\
108 GE_COWBUF_THRESHOLD((cur)) && \
109 GE_COWBUF_WASTE_THRESHOLD((cur),(len)) && \
110 GE_COWBUF_WASTE_FACTOR_THRESHOLD((cur),(len)) \
113 #ifdef PERL_UTF8_CACHE_ASSERT
114 /* if adding more checks watch out for the following tests:
115 * t/op/index.t t/op/length.t t/op/pat.t t/op/substr.t
116 * lib/utf8.t lib/Unicode/Collate/t/index.t
119 # define ASSERT_UTF8_CACHE(cache) \
120 STMT_START { if (cache) { assert((cache)[0] <= (cache)[1]); \
121 assert((cache)[2] <= (cache)[3]); \
122 assert((cache)[3] <= (cache)[1]);} \
125 # define ASSERT_UTF8_CACHE(cache) NOOP
128 static const char S_destroy[] = "DESTROY";
129 #define S_destroy_len (sizeof(S_destroy)-1)
131 /* ============================================================================
133 =head1 Allocation and deallocation of SVs.
135 An SV (or AV, HV, etc.) is allocated in two parts: the head (struct
136 sv, av, hv...) contains type and reference count information, and for
137 many types, a pointer to the body (struct xrv, xpv, xpviv...), which
138 contains fields specific to each type. Some types store all they need
139 in the head, so don't have a body.
141 In all but the most memory-paranoid configurations (ex: PURIFY), heads
142 and bodies are allocated out of arenas, which by default are
143 approximately 4K chunks of memory parcelled up into N heads or bodies.
144 Sv-bodies are allocated by their sv-type, guaranteeing size
145 consistency needed to allocate safely from arrays.
147 For SV-heads, the first slot in each arena is reserved, and holds a
148 link to the next arena, some flags, and a note of the number of slots.
149 Snaked through each arena chain is a linked list of free items; when
150 this becomes empty, an extra arena is allocated and divided up into N
151 items which are threaded into the free list.
153 SV-bodies are similar, but they use arena-sets by default, which
154 separate the link and info from the arena itself, and reclaim the 1st
155 slot in the arena. SV-bodies are further described later.
157 The following global variables are associated with arenas:
159 PL_sv_arenaroot pointer to list of SV arenas
160 PL_sv_root pointer to list of free SV structures
162 PL_body_arenas head of linked-list of body arenas
163 PL_body_roots[] array of pointers to list of free bodies of svtype
164 arrays are indexed by the svtype needed
166 A few special SV heads are not allocated from an arena, but are
167 instead directly created in the interpreter structure, eg PL_sv_undef.
168 The size of arenas can be changed from the default by setting
169 PERL_ARENA_SIZE appropriately at compile time.
171 The SV arena serves the secondary purpose of allowing still-live SVs
172 to be located and destroyed during final cleanup.
174 At the lowest level, the macros new_SV() and del_SV() grab and free
175 an SV head. (If debugging with -DD, del_SV() calls the function S_del_sv()
176 to return the SV to the free list with error checking.) new_SV() calls
177 more_sv() / sv_add_arena() to add an extra arena if the free list is empty.
178 SVs in the free list have their SvTYPE field set to all ones.
180 At the time of very final cleanup, sv_free_arenas() is called from
181 perl_destruct() to physically free all the arenas allocated since the
182 start of the interpreter.
184 The function visit() scans the SV arenas list, and calls a specified
185 function for each SV it finds which is still live - ie which has an SvTYPE
186 other than all 1's, and a non-zero SvREFCNT. visit() is used by the
187 following functions (specified as [function that calls visit()] / [function
188 called by visit() for each SV]):
190 sv_report_used() / do_report_used()
191 dump all remaining SVs (debugging aid)
193 sv_clean_objs() / do_clean_objs(),do_clean_named_objs(),
194 do_clean_named_io_objs(),do_curse()
195 Attempt to free all objects pointed to by RVs,
196 try to do the same for all objects indir-
197 ectly referenced by typeglobs too, and
198 then do a final sweep, cursing any
199 objects that remain. Called once from
200 perl_destruct(), prior to calling sv_clean_all()
203 sv_clean_all() / do_clean_all()
204 SvREFCNT_dec(sv) each remaining SV, possibly
205 triggering an sv_free(). It also sets the
206 SVf_BREAK flag on the SV to indicate that the
207 refcnt has been artificially lowered, and thus
208 stopping sv_free() from giving spurious warnings
209 about SVs which unexpectedly have a refcnt
210 of zero. called repeatedly from perl_destruct()
211 until there are no SVs left.
213 =head2 Arena allocator API Summary
215 Private API to rest of sv.c
219 new_XPVNV(), del_XPVGV(),
224 sv_report_used(), sv_clean_objs(), sv_clean_all(), sv_free_arenas()
228 * ========================================================================= */
231 * "A time to plant, and a time to uproot what was planted..."
235 # define MEM_LOG_NEW_SV(sv, file, line, func) \
236 Perl_mem_log_new_sv(sv, file, line, func)
237 # define MEM_LOG_DEL_SV(sv, file, line, func) \
238 Perl_mem_log_del_sv(sv, file, line, func)
240 # define MEM_LOG_NEW_SV(sv, file, line, func) NOOP
241 # define MEM_LOG_DEL_SV(sv, file, line, func) NOOP
244 #ifdef DEBUG_LEAKING_SCALARS
245 # define FREE_SV_DEBUG_FILE(sv) STMT_START { \
246 if ((sv)->sv_debug_file) PerlMemShared_free((sv)->sv_debug_file); \
248 # define DEBUG_SV_SERIAL(sv) \
249 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%" UVxf ": (%05ld) del_SV\n", \
250 PTR2UV(sv), (long)(sv)->sv_debug_serial))
252 # define FREE_SV_DEBUG_FILE(sv)
253 # define DEBUG_SV_SERIAL(sv) NOOP
257 # define SvARENA_CHAIN(sv) ((sv)->sv_u.svu_rv)
258 # define SvARENA_CHAIN_SET(sv,val) (sv)->sv_u.svu_rv = MUTABLE_SV((val))
259 /* Whilst I'd love to do this, it seems that things like to check on
261 # define POISON_SV_HEAD(sv) PoisonNew(sv, 1, struct STRUCT_SV)
263 # define POISON_SV_HEAD(sv) PoisonNew(&SvANY(sv), 1, void *), \
264 PoisonNew(&SvREFCNT(sv), 1, U32)
266 # define SvARENA_CHAIN(sv) SvANY(sv)
267 # define SvARENA_CHAIN_SET(sv,val) SvANY(sv) = (void *)(val)
268 # define POISON_SV_HEAD(sv)
271 /* Mark an SV head as unused, and add to free list.
273 * If SVf_BREAK is set, skip adding it to the free list, as this SV had
274 * its refcount artificially decremented during global destruction, so
275 * there may be dangling pointers to it. The last thing we want in that
276 * case is for it to be reused. */
278 #define plant_SV(p) \
280 const U32 old_flags = SvFLAGS(p); \
281 MEM_LOG_DEL_SV(p, __FILE__, __LINE__, FUNCTION__); \
282 DEBUG_SV_SERIAL(p); \
283 FREE_SV_DEBUG_FILE(p); \
285 SvFLAGS(p) = SVTYPEMASK; \
286 if (!(old_flags & SVf_BREAK)) { \
287 SvARENA_CHAIN_SET(p, PL_sv_root); \
293 #define uproot_SV(p) \
296 PL_sv_root = MUTABLE_SV(SvARENA_CHAIN(p)); \
301 /* make some more SVs by adding another arena */
307 char *chunk; /* must use New here to match call to */
308 Newx(chunk,PERL_ARENA_SIZE,char); /* Safefree() in sv_free_arenas() */
309 sv_add_arena(chunk, PERL_ARENA_SIZE, 0);
314 /* new_SV(): return a new, empty SV head */
316 #ifdef DEBUG_LEAKING_SCALARS
317 /* provide a real function for a debugger to play with */
319 S_new_SV(pTHX_ const char *file, int line, const char *func)
326 sv = S_more_sv(aTHX);
330 sv->sv_debug_optype = PL_op ? PL_op->op_type : 0;
331 sv->sv_debug_line = (U16) (PL_parser && PL_parser->copline != NOLINE
337 sv->sv_debug_inpad = 0;
338 sv->sv_debug_parent = NULL;
339 sv->sv_debug_file = PL_curcop ? savesharedpv(CopFILE(PL_curcop)): NULL;
341 sv->sv_debug_serial = PL_sv_serial++;
343 MEM_LOG_NEW_SV(sv, file, line, func);
344 DEBUG_m(PerlIO_printf(Perl_debug_log, "0x%" UVxf ": (%05ld) new_SV (from %s:%d [%s])\n",
345 PTR2UV(sv), (long)sv->sv_debug_serial, file, line, func));
349 # define new_SV(p) (p)=S_new_SV(aTHX_ __FILE__, __LINE__, FUNCTION__)
357 (p) = S_more_sv(aTHX); \
361 MEM_LOG_NEW_SV(p, __FILE__, __LINE__, FUNCTION__); \
366 /* del_SV(): return an empty SV head to the free list */
379 S_del_sv(pTHX_ SV *p)
381 PERL_ARGS_ASSERT_DEL_SV;
386 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
387 const SV * const sv = sva + 1;
388 const SV * const svend = &sva[SvREFCNT(sva)];
389 if (p >= sv && p < svend) {
395 Perl_ck_warner_d(aTHX_ packWARN(WARN_INTERNAL),
396 "Attempt to free non-arena SV: 0x%" UVxf
397 pTHX__FORMAT, PTR2UV(p) pTHX__VALUE);
404 #else /* ! DEBUGGING */
406 #define del_SV(p) plant_SV(p)
408 #endif /* DEBUGGING */
412 =head1 SV Manipulation Functions
414 =for apidoc sv_add_arena
416 Given a chunk of memory, link it to the head of the list of arenas,
417 and split it into a list of free SVs.
423 S_sv_add_arena(pTHX_ char *const ptr, const U32 size, const U32 flags)
425 SV *const sva = MUTABLE_SV(ptr);
429 PERL_ARGS_ASSERT_SV_ADD_ARENA;
431 /* The first SV in an arena isn't an SV. */
432 SvANY(sva) = (void *) PL_sv_arenaroot; /* ptr to next arena */
433 SvREFCNT(sva) = size / sizeof(SV); /* number of SV slots */
434 SvFLAGS(sva) = flags; /* FAKE if not to be freed */
436 PL_sv_arenaroot = sva;
437 PL_sv_root = sva + 1;
439 svend = &sva[SvREFCNT(sva) - 1];
442 SvARENA_CHAIN_SET(sv, (sv + 1));
446 /* Must always set typemask because it's always checked in on cleanup
447 when the arenas are walked looking for objects. */
448 SvFLAGS(sv) = SVTYPEMASK;
451 SvARENA_CHAIN_SET(sv, 0);
455 SvFLAGS(sv) = SVTYPEMASK;
458 /* visit(): call the named function for each non-free SV in the arenas
459 * whose flags field matches the flags/mask args. */
462 S_visit(pTHX_ SVFUNC_t f, const U32 flags, const U32 mask)
467 PERL_ARGS_ASSERT_VISIT;
469 for (sva = PL_sv_arenaroot; sva; sva = MUTABLE_SV(SvANY(sva))) {
470 const SV * const svend = &sva[SvREFCNT(sva)];
472 for (sv = sva + 1; sv < svend; ++sv) {
473 if (SvTYPE(sv) != (svtype)SVTYPEMASK
474 && (sv->sv_flags & mask) == flags
487 /* called by sv_report_used() for each live SV */
490 do_report_used(pTHX_ SV *const sv)
492 if (SvTYPE(sv) != (svtype)SVTYPEMASK) {
493 PerlIO_printf(Perl_debug_log, "****\n");
500 =for apidoc sv_report_used
502 Dump the contents of all SVs not yet freed (debugging aid).
508 Perl_sv_report_used(pTHX)
511 visit(do_report_used, 0, 0);
517 /* called by sv_clean_objs() for each live SV */
520 do_clean_objs(pTHX_ SV *const ref)
524 SV * const target = SvRV(ref);
525 if (SvOBJECT(target)) {
526 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning object ref:\n "), sv_dump(ref)));
527 if (SvWEAKREF(ref)) {
528 sv_del_backref(target, ref);
534 SvREFCNT_dec_NN(target);
541 /* clear any slots in a GV which hold objects - except IO;
542 * called by sv_clean_objs() for each live GV */
545 do_clean_named_objs(pTHX_ SV *const sv)
548 assert(SvTYPE(sv) == SVt_PVGV);
549 assert(isGV_with_GP(sv));
553 /* freeing GP entries may indirectly free the current GV;
554 * hold onto it while we mess with the GP slots */
557 if ( ((obj = GvSV(sv) )) && SvOBJECT(obj)) {
558 DEBUG_D((PerlIO_printf(Perl_debug_log,
559 "Cleaning named glob SV object:\n "), sv_dump(obj)));
561 SvREFCNT_dec_NN(obj);
563 if ( ((obj = MUTABLE_SV(GvAV(sv)) )) && SvOBJECT(obj)) {
564 DEBUG_D((PerlIO_printf(Perl_debug_log,
565 "Cleaning named glob AV object:\n "), sv_dump(obj)));
567 SvREFCNT_dec_NN(obj);
569 if ( ((obj = MUTABLE_SV(GvHV(sv)) )) && SvOBJECT(obj)) {
570 DEBUG_D((PerlIO_printf(Perl_debug_log,
571 "Cleaning named glob HV object:\n "), sv_dump(obj)));
573 SvREFCNT_dec_NN(obj);
575 if ( ((obj = MUTABLE_SV(GvCV(sv)) )) && SvOBJECT(obj)) {
576 DEBUG_D((PerlIO_printf(Perl_debug_log,
577 "Cleaning named glob CV object:\n "), sv_dump(obj)));
579 SvREFCNT_dec_NN(obj);
581 SvREFCNT_dec_NN(sv); /* undo the inc above */
584 /* clear any IO slots in a GV which hold objects (except stderr, defout);
585 * called by sv_clean_objs() for each live GV */
588 do_clean_named_io_objs(pTHX_ SV *const sv)
591 assert(SvTYPE(sv) == SVt_PVGV);
592 assert(isGV_with_GP(sv));
593 if (!GvGP(sv) || sv == (SV*)PL_stderrgv || sv == (SV*)PL_defoutgv)
597 if ( ((obj = MUTABLE_SV(GvIO(sv)) )) && SvOBJECT(obj)) {
598 DEBUG_D((PerlIO_printf(Perl_debug_log,
599 "Cleaning named glob IO object:\n "), sv_dump(obj)));
601 SvREFCNT_dec_NN(obj);
603 SvREFCNT_dec_NN(sv); /* undo the inc above */
606 /* Void wrapper to pass to visit() */
608 do_curse(pTHX_ SV * const sv) {
609 if ((PL_stderrgv && GvGP(PL_stderrgv) && (SV*)GvIO(PL_stderrgv) == sv)
610 || (PL_defoutgv && GvGP(PL_defoutgv) && (SV*)GvIO(PL_defoutgv) == sv))
616 =for apidoc sv_clean_objs
618 Attempt to destroy all objects not yet freed.
624 Perl_sv_clean_objs(pTHX)
627 PL_in_clean_objs = TRUE;
628 visit(do_clean_objs, SVf_ROK, SVf_ROK);
629 /* Some barnacles may yet remain, clinging to typeglobs.
630 * Run the non-IO destructors first: they may want to output
631 * error messages, close files etc */
632 visit(do_clean_named_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
633 visit(do_clean_named_io_objs, SVt_PVGV|SVpgv_GP, SVTYPEMASK|SVp_POK|SVpgv_GP);
634 /* And if there are some very tenacious barnacles clinging to arrays,
635 closures, or what have you.... */
636 visit(do_curse, SVs_OBJECT, SVs_OBJECT);
637 olddef = PL_defoutgv;
638 PL_defoutgv = NULL; /* disable skip of PL_defoutgv */
639 if (olddef && isGV_with_GP(olddef))
640 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olddef));
641 olderr = PL_stderrgv;
642 PL_stderrgv = NULL; /* disable skip of PL_stderrgv */
643 if (olderr && isGV_with_GP(olderr))
644 do_clean_named_io_objs(aTHX_ MUTABLE_SV(olderr));
645 SvREFCNT_dec(olddef);
646 PL_in_clean_objs = FALSE;
649 /* called by sv_clean_all() for each live SV */
652 do_clean_all(pTHX_ SV *const sv)
654 if (sv == (const SV *) PL_fdpid || sv == (const SV *)PL_strtab) {
655 /* don't clean pid table and strtab */
658 DEBUG_D((PerlIO_printf(Perl_debug_log, "Cleaning loops: SV at 0x%" UVxf "\n", PTR2UV(sv)) ));
659 SvFLAGS(sv) |= SVf_BREAK;
664 =for apidoc sv_clean_all
666 Decrement the refcnt of each remaining SV, possibly triggering a
667 cleanup. This function may have to be called multiple times to free
668 SVs which are in complex self-referential hierarchies.
674 Perl_sv_clean_all(pTHX)
677 PL_in_clean_all = TRUE;
678 cleaned = visit(do_clean_all, 0,0);
683 ARENASETS: a meta-arena implementation which separates arena-info
684 into struct arena_set, which contains an array of struct
685 arena_descs, each holding info for a single arena. By separating
686 the meta-info from the arena, we recover the 1st slot, formerly
687 borrowed for list management. The arena_set is about the size of an
688 arena, avoiding the needless malloc overhead of a naive linked-list.
690 The cost is 1 arena-set malloc per ~320 arena-mallocs, + the unused
691 memory in the last arena-set (1/2 on average). In trade, we get
692 back the 1st slot in each arena (ie 1.7% of a CV-arena, less for
693 smaller types). The recovery of the wasted space allows use of
694 small arenas for large, rare body types, by changing array* fields
695 in body_details_by_type[] below.
698 char *arena; /* the raw storage, allocated aligned */
699 size_t size; /* its size ~4k typ */
700 svtype utype; /* bodytype stored in arena */
705 /* Get the maximum number of elements in set[] such that struct arena_set
706 will fit within PERL_ARENA_SIZE, which is probably just under 4K, and
707 therefore likely to be 1 aligned memory page. */
709 #define ARENAS_PER_SET ((PERL_ARENA_SIZE - sizeof(struct arena_set*) \
710 - 2 * sizeof(int)) / sizeof (struct arena_desc))
713 struct arena_set* next;
714 unsigned int set_size; /* ie ARENAS_PER_SET */
715 unsigned int curr; /* index of next available arena-desc */
716 struct arena_desc set[ARENAS_PER_SET];
720 =for apidoc sv_free_arenas
722 Deallocate the memory used by all arenas. Note that all the individual SV
723 heads and bodies within the arenas must already have been freed.
729 Perl_sv_free_arenas(pTHX)
735 /* Free arenas here, but be careful about fake ones. (We assume
736 contiguity of the fake ones with the corresponding real ones.) */
738 for (sva = PL_sv_arenaroot; sva; sva = svanext) {
739 svanext = MUTABLE_SV(SvANY(sva));
740 while (svanext && SvFAKE(svanext))
741 svanext = MUTABLE_SV(SvANY(svanext));
748 struct arena_set *aroot = (struct arena_set*) PL_body_arenas;
751 struct arena_set *current = aroot;
754 assert(aroot->set[i].arena);
755 Safefree(aroot->set[i].arena);
763 i = PERL_ARENA_ROOTS_SIZE;
765 PL_body_roots[i] = 0;
772 Here are mid-level routines that manage the allocation of bodies out
773 of the various arenas. There are 4 kinds of arenas:
775 1. SV-head arenas, which are discussed and handled above
776 2. regular body arenas
777 3. arenas for reduced-size bodies
780 Arena types 2 & 3 are chained by body-type off an array of
781 arena-root pointers, which is indexed by svtype. Some of the
782 larger/less used body types are malloced singly, since a large
783 unused block of them is wasteful. Also, several svtypes dont have
784 bodies; the data fits into the sv-head itself. The arena-root
785 pointer thus has a few unused root-pointers (which may be hijacked
786 later for arena type 4)
788 3 differs from 2 as an optimization; some body types have several
789 unused fields in the front of the structure (which are kept in-place
790 for consistency). These bodies can be allocated in smaller chunks,
791 because the leading fields arent accessed. Pointers to such bodies
792 are decremented to point at the unused 'ghost' memory, knowing that
793 the pointers are used with offsets to the real memory.
795 Allocation of SV-bodies is similar to SV-heads, differing as follows;
796 the allocation mechanism is used for many body types, so is somewhat
797 more complicated, it uses arena-sets, and has no need for still-live
800 At the outermost level, (new|del)_X*V macros return bodies of the
801 appropriate type. These macros call either (new|del)_body_type or
802 (new|del)_body_allocated macro pairs, depending on specifics of the
803 type. Most body types use the former pair, the latter pair is used to
804 allocate body types with "ghost fields".
806 "ghost fields" are fields that are unused in certain types, and
807 consequently don't need to actually exist. They are declared because
808 they're part of a "base type", which allows use of functions as
809 methods. The simplest examples are AVs and HVs, 2 aggregate types
810 which don't use the fields which support SCALAR semantics.
812 For these types, the arenas are carved up into appropriately sized
813 chunks, we thus avoid wasted memory for those unaccessed members.
814 When bodies are allocated, we adjust the pointer back in memory by the
815 size of the part not allocated, so it's as if we allocated the full
816 structure. (But things will all go boom if you write to the part that
817 is "not there", because you'll be overwriting the last members of the
818 preceding structure in memory.)
820 We calculate the correction using the STRUCT_OFFSET macro on the first
821 member present. If the allocated structure is smaller (no initial NV
822 actually allocated) then the net effect is to subtract the size of the NV
823 from the pointer, to return a new pointer as if an initial NV were actually
824 allocated. (We were using structures named *_allocated for this, but
825 this turned out to be a subtle bug, because a structure without an NV
826 could have a lower alignment constraint, but the compiler is allowed to
827 optimised accesses based on the alignment constraint of the actual pointer
828 to the full structure, for example, using a single 64 bit load instruction
829 because it "knows" that two adjacent 32 bit members will be 8-byte aligned.)
831 This is the same trick as was used for NV and IV bodies. Ironically it
832 doesn't need to be used for NV bodies any more, because NV is now at
833 the start of the structure. IV bodies, and also in some builds NV bodies,
834 don't need it either, because they are no longer allocated.
836 In turn, the new_body_* allocators call S_new_body(), which invokes
837 new_body_inline macro, which takes a lock, and takes a body off the
838 linked list at PL_body_roots[sv_type], calling Perl_more_bodies() if
839 necessary to refresh an empty list. Then the lock is released, and
840 the body is returned.
842 Perl_more_bodies allocates a new arena, and carves it up into an array of N
843 bodies, which it strings into a linked list. It looks up arena-size
844 and body-size from the body_details table described below, thus
845 supporting the multiple body-types.
847 If PURIFY is defined, or PERL_ARENA_SIZE=0, arenas are not used, and
848 the (new|del)_X*V macros are mapped directly to malloc/free.
850 For each sv-type, struct body_details bodies_by_type[] carries
851 parameters which control these aspects of SV handling:
853 Arena_size determines whether arenas are used for this body type, and if
854 so, how big they are. PURIFY or PERL_ARENA_SIZE=0 set this field to
855 zero, forcing individual mallocs and frees.
857 Body_size determines how big a body is, and therefore how many fit into
858 each arena. Offset carries the body-pointer adjustment needed for
859 "ghost fields", and is used in *_allocated macros.
861 But its main purpose is to parameterize info needed in
862 Perl_sv_upgrade(). The info here dramatically simplifies the function
863 vs the implementation in 5.8.8, making it table-driven. All fields
864 are used for this, except for arena_size.
866 For the sv-types that have no bodies, arenas are not used, so those
867 PL_body_roots[sv_type] are unused, and can be overloaded. In
868 something of a special case, SVt_NULL is borrowed for HE arenas;
869 PL_body_roots[HE_SVSLOT=SVt_NULL] is filled by S_more_he, but the
870 bodies_by_type[SVt_NULL] slot is not used, as the table is not
875 struct body_details {
876 U8 body_size; /* Size to allocate */
877 U8 copy; /* Size of structure to copy (may be shorter) */
878 U8 offset; /* Size of unalloced ghost fields to first alloced field*/
879 PERL_BITFIELD8 type : 4; /* We have space for a sanity check. */
880 PERL_BITFIELD8 cant_upgrade : 1;/* Cannot upgrade this type */
881 PERL_BITFIELD8 zero_nv : 1; /* zero the NV when upgrading from this */
882 PERL_BITFIELD8 arena : 1; /* Allocated from an arena */
883 U32 arena_size; /* Size of arena to allocate */
886 #define ALIGNED_TYPE_NAME(name) name##_aligned
887 #define ALIGNED_TYPE(name) \
892 } ALIGNED_TYPE_NAME(name);
894 ALIGNED_TYPE(regexp);
908 /* With -DPURFIY we allocate everything directly, and don't use arenas.
909 This seems a rather elegant way to simplify some of the code below. */
910 #define HASARENA FALSE
912 #define HASARENA TRUE
914 #define NOARENA FALSE
916 /* Size the arenas to exactly fit a given number of bodies. A count
917 of 0 fits the max number bodies into a PERL_ARENA_SIZE.block,
918 simplifying the default. If count > 0, the arena is sized to fit
919 only that many bodies, allowing arenas to be used for large, rare
920 bodies (XPVFM, XPVIO) without undue waste. The arena size is
921 limited by PERL_ARENA_SIZE, so we can safely oversize the
924 #define FIT_ARENA0(body_size) \
925 ((size_t)(PERL_ARENA_SIZE / body_size) * body_size)
926 #define FIT_ARENAn(count,body_size) \
927 ( count * body_size <= PERL_ARENA_SIZE) \
928 ? count * body_size \
929 : FIT_ARENA0 (body_size)
930 #define FIT_ARENA(count,body_size) \
932 ? FIT_ARENAn (count, body_size) \
933 : FIT_ARENA0 (body_size))
935 /* Calculate the length to copy. Specifically work out the length less any
936 final padding the compiler needed to add. See the comment in sv_upgrade
937 for why copying the padding proved to be a bug. */
939 #define copy_length(type, last_member) \
940 STRUCT_OFFSET(type, last_member) \
941 + sizeof (((type*)SvANY((const SV *)0))->last_member)
943 static const struct body_details bodies_by_type[] = {
944 /* HEs use this offset for their arena. */
945 { 0, 0, 0, SVt_NULL, FALSE, NONV, NOARENA, 0 },
947 /* IVs are in the head, so the allocation size is 0. */
949 sizeof(IV), /* This is used to copy out the IV body. */
950 STRUCT_OFFSET(XPVIV, xiv_iv), SVt_IV, FALSE, NONV,
951 NOARENA /* IVS don't need an arena */, 0
956 STRUCT_OFFSET(XPVNV, xnv_u),
957 SVt_NV, FALSE, HADNV, NOARENA, 0 },
959 { sizeof(NV), sizeof(NV),
960 STRUCT_OFFSET(XPVNV, xnv_u),
961 SVt_NV, FALSE, HADNV, HASARENA, FIT_ARENA(0, sizeof(NV)) },
964 { sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur),
965 copy_length(XPV, xpv_len) - STRUCT_OFFSET(XPV, xpv_cur),
966 + STRUCT_OFFSET(XPV, xpv_cur),
967 SVt_PV, FALSE, NONV, HASARENA,
968 FIT_ARENA(0, sizeof(XPV) - STRUCT_OFFSET(XPV, xpv_cur)) },
970 { sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur),
971 copy_length(XINVLIST, is_offset) - STRUCT_OFFSET(XPV, xpv_cur),
972 + STRUCT_OFFSET(XPV, xpv_cur),
973 SVt_INVLIST, TRUE, NONV, HASARENA,
974 FIT_ARENA(0, sizeof(XINVLIST) - STRUCT_OFFSET(XPV, xpv_cur)) },
976 { sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur),
977 copy_length(XPVIV, xiv_u) - STRUCT_OFFSET(XPV, xpv_cur),
978 + STRUCT_OFFSET(XPV, xpv_cur),
979 SVt_PVIV, FALSE, NONV, HASARENA,
980 FIT_ARENA(0, sizeof(XPVIV) - STRUCT_OFFSET(XPV, xpv_cur)) },
982 { sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur),
983 copy_length(XPVNV, xnv_u) - STRUCT_OFFSET(XPV, xpv_cur),
984 + STRUCT_OFFSET(XPV, xpv_cur),
985 SVt_PVNV, FALSE, HADNV, HASARENA,
986 FIT_ARENA(0, sizeof(XPVNV) - STRUCT_OFFSET(XPV, xpv_cur)) },
988 { sizeof(XPVMG), copy_length(XPVMG, xnv_u), 0, SVt_PVMG, FALSE, HADNV,
989 HASARENA, FIT_ARENA(0, sizeof(XPVMG)) },
991 { sizeof(ALIGNED_TYPE_NAME(regexp)),
994 SVt_REGEXP, TRUE, NONV, HASARENA,
995 FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(regexp)))
998 { sizeof(ALIGNED_TYPE_NAME(XPVGV)), sizeof(XPVGV), 0, SVt_PVGV, TRUE, HADNV,
999 HASARENA, FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(XPVGV))) },
1001 { sizeof(ALIGNED_TYPE_NAME(XPVLV)), sizeof(XPVLV), 0, SVt_PVLV, TRUE, HADNV,
1002 HASARENA, FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(XPVLV))) },
1004 { sizeof(ALIGNED_TYPE_NAME(XPVAV)),
1005 copy_length(XPVAV, xav_alloc),
1007 SVt_PVAV, TRUE, NONV, HASARENA,
1008 FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(XPVAV))) },
1010 { sizeof(ALIGNED_TYPE_NAME(XPVHV)),
1011 copy_length(XPVHV, xhv_max),
1013 SVt_PVHV, TRUE, NONV, HASARENA,
1014 FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(XPVHV))) },
1016 { sizeof(ALIGNED_TYPE_NAME(XPVCV)),
1019 SVt_PVCV, TRUE, NONV, HASARENA,
1020 FIT_ARENA(0, sizeof(ALIGNED_TYPE_NAME(XPVCV))) },
1022 { sizeof(ALIGNED_TYPE_NAME(XPVFM)),
1025 SVt_PVFM, TRUE, NONV, NOARENA,
1026 FIT_ARENA(20, sizeof(ALIGNED_TYPE_NAME(XPVFM))) },
1028 { sizeof(ALIGNED_TYPE_NAME(XPVIO)),
1031 SVt_PVIO, TRUE, NONV, HASARENA,
1032 FIT_ARENA(24, sizeof(ALIGNED_TYPE_NAME(XPVIO))) },
1035 #define new_body_allocated(sv_type) \
1036 (void *)((char *)S_new_body(aTHX_ sv_type) \
1037 - bodies_by_type[sv_type].offset)
1039 /* return a thing to the free list */
1041 #define del_body(thing, root) \
1043 void ** const thing_copy = (void **)thing; \
1044 *thing_copy = *root; \
1045 *root = (void*)thing_copy; \
1049 #if !(NVSIZE <= IVSIZE)
1050 # define new_XNV() safemalloc(sizeof(XPVNV))
1052 #define new_XPVNV() safemalloc(sizeof(XPVNV))
1053 #define new_XPVMG() safemalloc(sizeof(XPVMG))
1055 #define del_XPVGV(p) safefree(p)
1059 #if !(NVSIZE <= IVSIZE)
1060 # define new_XNV() new_body_allocated(SVt_NV)
1062 #define new_XPVNV() new_body_allocated(SVt_PVNV)
1063 #define new_XPVMG() new_body_allocated(SVt_PVMG)
1065 #define del_XPVGV(p) del_body(p + bodies_by_type[SVt_PVGV].offset, \
1066 &PL_body_roots[SVt_PVGV])
1070 /* no arena for you! */
1072 #define new_NOARENA(details) \
1073 safemalloc((details)->body_size + (details)->offset)
1074 #define new_NOARENAZ(details) \
1075 safecalloc((details)->body_size + (details)->offset, 1)
1078 Perl_more_bodies (pTHX_ const svtype sv_type, const size_t body_size,
1079 const size_t arena_size)
1081 void ** const root = &PL_body_roots[sv_type];
1082 struct arena_desc *adesc;
1083 struct arena_set *aroot = (struct arena_set *) PL_body_arenas;
1087 const size_t good_arena_size = Perl_malloc_good_size(arena_size);
1088 #if defined(DEBUGGING)
1089 static bool done_sanity_check;
1091 if (!done_sanity_check) {
1092 unsigned int i = SVt_LAST;
1094 done_sanity_check = TRUE;
1097 assert (bodies_by_type[i].type == i);
1103 /* may need new arena-set to hold new arena */
1104 if (!aroot || aroot->curr >= aroot->set_size) {
1105 struct arena_set *newroot;
1106 Newxz(newroot, 1, struct arena_set);
1107 newroot->set_size = ARENAS_PER_SET;
1108 newroot->next = aroot;
1110 PL_body_arenas = (void *) newroot;
1111 DEBUG_m(PerlIO_printf(Perl_debug_log, "new arenaset %p\n", (void*)aroot));
1114 /* ok, now have arena-set with at least 1 empty/available arena-desc */
1115 curr = aroot->curr++;
1116 adesc = &(aroot->set[curr]);
1117 assert(!adesc->arena);
1119 Newx(adesc->arena, good_arena_size, char);
1120 adesc->size = good_arena_size;
1121 adesc->utype = sv_type;
1122 DEBUG_m(PerlIO_printf(Perl_debug_log, "arena %d added: %p size %" UVuf "\n",
1123 curr, (void*)adesc->arena, (UV)good_arena_size));
1125 start = (char *) adesc->arena;
1127 /* Get the address of the byte after the end of the last body we can fit.
1128 Remember, this is integer division: */
1129 end = start + good_arena_size / body_size * body_size;
1131 /* computed count doesn't reflect the 1st slot reservation */
1132 #if defined(MYMALLOC) || defined(HAS_MALLOC_GOOD_SIZE)
1133 DEBUG_m(PerlIO_printf(Perl_debug_log,
1134 "arena %p end %p arena-size %d (from %d) type %d "
1136 (void*)start, (void*)end, (int)good_arena_size,
1137 (int)arena_size, sv_type, (int)body_size,
1138 (int)good_arena_size / (int)body_size));
1140 DEBUG_m(PerlIO_printf(Perl_debug_log,
1141 "arena %p end %p arena-size %d type %d size %d ct %d\n",
1142 (void*)start, (void*)end,
1143 (int)arena_size, sv_type, (int)body_size,
1144 (int)good_arena_size / (int)body_size));
1146 *root = (void *)start;
1149 /* Where the next body would start: */
1150 char * const next = start + body_size;
1153 /* This is the last body: */
1154 assert(next == end);
1156 *(void **)start = 0;
1160 *(void**) start = (void *)next;
1165 /* grab a new thing from the free list, allocating more if necessary.
1166 The inline version is used for speed in hot routines, and the
1167 function using it serves the rest (unless PURIFY).
1169 #define new_body_inline(xpv, sv_type) \
1171 void ** const r3wt = &PL_body_roots[sv_type]; \
1172 xpv = (PTR_TBL_ENT_t*) (*((void **)(r3wt)) \
1173 ? *((void **)(r3wt)) : Perl_more_bodies(aTHX_ sv_type, \
1174 bodies_by_type[sv_type].body_size,\
1175 bodies_by_type[sv_type].arena_size)); \
1176 *(r3wt) = *(void**)(xpv); \
1182 S_new_body(pTHX_ const svtype sv_type)
1185 new_body_inline(xpv, sv_type);
1191 static const struct body_details fake_rv =
1192 { 0, 0, 0, SVt_IV, FALSE, NONV, NOARENA, 0 };
1195 =for apidoc sv_upgrade
1197 Upgrade an SV to a more complex form. Generally adds a new body type to the
1198 SV, then copies across as much information as possible from the old body.
1199 It croaks if the SV is already in a more complex form than requested. You
1200 generally want to use the C<SvUPGRADE> macro wrapper, which checks the type
1201 before calling C<sv_upgrade>, and hence does not croak. See also
1208 Perl_sv_upgrade(pTHX_ SV *const sv, svtype new_type)
1212 const svtype old_type = SvTYPE(sv);
1213 const struct body_details *new_type_details;
1214 const struct body_details *old_type_details
1215 = bodies_by_type + old_type;
1216 SV *referent = NULL;
1218 PERL_ARGS_ASSERT_SV_UPGRADE;
1220 if (old_type == new_type)
1223 /* This clause was purposefully added ahead of the early return above to
1224 the shared string hackery for (sort {$a <=> $b} keys %hash), with the
1225 inference by Nick I-S that it would fix other troublesome cases. See
1226 changes 7162, 7163 (f130fd4589cf5fbb24149cd4db4137c8326f49c1 and parent)
1228 Given that shared hash key scalars are no longer PVIV, but PV, there is
1229 no longer need to unshare so as to free up the IVX slot for its proper
1230 purpose. So it's safe to move the early return earlier. */
1232 if (new_type > SVt_PVMG && SvIsCOW(sv)) {
1233 sv_force_normal_flags(sv, 0);
1236 old_body = SvANY(sv);
1238 /* Copying structures onto other structures that have been neatly zeroed
1239 has a subtle gotcha. Consider XPVMG
1241 +------+------+------+------+------+-------+-------+
1242 | NV | CUR | LEN | IV | MAGIC | STASH |
1243 +------+------+------+------+------+-------+-------+
1244 0 4 8 12 16 20 24 28
1246 where NVs are aligned to 8 bytes, so that sizeof that structure is
1247 actually 32 bytes long, with 4 bytes of padding at the end:
1249 +------+------+------+------+------+-------+-------+------+
1250 | NV | CUR | LEN | IV | MAGIC | STASH | ??? |
1251 +------+------+------+------+------+-------+-------+------+
1252 0 4 8 12 16 20 24 28 32
1254 so what happens if you allocate memory for this structure:
1256 +------+------+------+------+------+-------+-------+------+------+...
1257 | NV | CUR | LEN | IV | MAGIC | STASH | GP | NAME |
1258 +------+------+------+------+------+-------+-------+------+------+...
1259 0 4 8 12 16 20 24 28 32 36
1261 zero it, then copy sizeof(XPVMG) bytes on top of it? Not quite what you
1262 expect, because you copy the area marked ??? onto GP. Now, ??? may have
1263 started out as zero once, but it's quite possible that it isn't. So now,
1264 rather than a nicely zeroed GP, you have it pointing somewhere random.
1267 (In fact, GP ends up pointing at a previous GP structure, because the
1268 principle cause of the padding in XPVMG getting garbage is a copy of
1269 sizeof(XPVMG) bytes from a XPVGV structure in sv_unglob. Right now
1270 this happens to be moot because XPVGV has been re-ordered, with GP
1271 no longer after STASH)
1273 So we are careful and work out the size of used parts of all the
1281 referent = SvRV(sv);
1282 old_type_details = &fake_rv;
1283 if (new_type == SVt_NV)
1284 new_type = SVt_PVNV;
1286 if (new_type < SVt_PVIV) {
1287 new_type = (new_type == SVt_NV)
1288 ? SVt_PVNV : SVt_PVIV;
1293 if (new_type < SVt_PVNV) {
1294 new_type = SVt_PVNV;
1298 assert(new_type > SVt_PV);
1299 STATIC_ASSERT_STMT(SVt_IV < SVt_PV);
1300 STATIC_ASSERT_STMT(SVt_NV < SVt_PV);
1307 /* Because the XPVMG of PL_mess_sv isn't allocated from the arena,
1308 there's no way that it can be safely upgraded, because perl.c
1309 expects to Safefree(SvANY(PL_mess_sv)) */
1310 assert(sv != PL_mess_sv);
1313 if (UNLIKELY(old_type_details->cant_upgrade))
1314 Perl_croak(aTHX_ "Can't upgrade %s (%" UVuf ") to %" UVuf,
1315 sv_reftype(sv, 0), (UV) old_type, (UV) new_type);
1318 if (UNLIKELY(old_type > new_type))
1319 Perl_croak(aTHX_ "sv_upgrade from type %d down to type %d",
1320 (int)old_type, (int)new_type);
1322 new_type_details = bodies_by_type + new_type;
1324 SvFLAGS(sv) &= ~SVTYPEMASK;
1325 SvFLAGS(sv) |= new_type;
1327 /* This can't happen, as SVt_NULL is <= all values of new_type, so one of
1328 the return statements above will have triggered. */
1329 assert (new_type != SVt_NULL);
1332 assert(old_type == SVt_NULL);
1333 SET_SVANY_FOR_BODYLESS_IV(sv);
1337 assert(old_type == SVt_NULL);
1338 #if NVSIZE <= IVSIZE
1339 SET_SVANY_FOR_BODYLESS_NV(sv);
1341 SvANY(sv) = new_XNV();
1347 assert(new_type_details->body_size);
1350 assert(new_type_details->arena);
1351 assert(new_type_details->arena_size);
1352 /* This points to the start of the allocated area. */
1353 new_body_inline(new_body, new_type);
1354 Zero(new_body, new_type_details->body_size, char);
1355 new_body = ((char *)new_body) - new_type_details->offset;
1357 /* We always allocated the full length item with PURIFY. To do this
1358 we fake things so that arena is false for all 16 types.. */
1359 new_body = new_NOARENAZ(new_type_details);
1361 SvANY(sv) = new_body;
1362 if (new_type == SVt_PVAV) {
1366 if (old_type_details->body_size) {
1369 /* It will have been zeroed when the new body was allocated.
1370 Lets not write to it, in case it confuses a write-back
1376 #ifndef NODEFAULT_SHAREKEYS
1377 HvSHAREKEYS_on(sv); /* key-sharing on by default */
1379 /* start with PERL_HASH_DEFAULT_HvMAX+1 buckets: */
1380 HvMAX(sv) = PERL_HASH_DEFAULT_HvMAX;
1383 /* SVt_NULL isn't the only thing upgraded to AV or HV.
1384 The target created by newSVrv also is, and it can have magic.
1385 However, it never has SvPVX set.
1387 if (old_type == SVt_IV) {
1389 } else if (old_type >= SVt_PV) {
1390 assert(SvPVX_const(sv) == 0);
1393 if (old_type >= SVt_PVMG) {
1394 SvMAGIC_set(sv, ((XPVMG*)old_body)->xmg_u.xmg_magic);
1395 SvSTASH_set(sv, ((XPVMG*)old_body)->xmg_stash);
1397 sv->sv_u.svu_array = NULL; /* or svu_hash */
1402 /* XXX Is this still needed? Was it ever needed? Surely as there is
1403 no route from NV to PVIV, NOK can never be true */
1404 assert(!SvNOKp(sv));
1418 assert(new_type_details->body_size);
1419 /* We always allocated the full length item with PURIFY. To do this
1420 we fake things so that arena is false for all 16 types.. */
1421 if(new_type_details->arena) {
1422 /* This points to the start of the allocated area. */
1423 new_body_inline(new_body, new_type);
1424 Zero(new_body, new_type_details->body_size, char);
1425 new_body = ((char *)new_body) - new_type_details->offset;
1427 new_body = new_NOARENAZ(new_type_details);
1429 SvANY(sv) = new_body;
1431 if (old_type_details->copy) {
1432 /* There is now the potential for an upgrade from something without
1433 an offset (PVNV or PVMG) to something with one (PVCV, PVFM) */
1434 int offset = old_type_details->offset;
1435 int length = old_type_details->copy;
1437 if (new_type_details->offset > old_type_details->offset) {
1438 const int difference
1439 = new_type_details->offset - old_type_details->offset;
1440 offset += difference;
1441 length -= difference;
1443 assert (length >= 0);
1445 Copy((char *)old_body + offset, (char *)new_body + offset, length,
1449 #ifndef NV_ZERO_IS_ALLBITS_ZERO
1450 /* If NV 0.0 is stores as all bits 0 then Zero() already creates a
1451 * correct 0.0 for us. Otherwise, if the old body didn't have an
1452 * NV slot, but the new one does, then we need to initialise the
1453 * freshly created NV slot with whatever the correct bit pattern is
1455 if (old_type_details->zero_nv && !new_type_details->zero_nv
1456 && !isGV_with_GP(sv))
1460 if (UNLIKELY(new_type == SVt_PVIO)) {
1461 IO * const io = MUTABLE_IO(sv);
1462 GV *iogv = gv_fetchpvs("IO::File::", GV_ADD, SVt_PVHV);
1465 /* Clear the stashcache because a new IO could overrule a package
1467 DEBUG_o(Perl_deb(aTHX_ "sv_upgrade clearing PL_stashcache\n"));
1468 hv_clear(PL_stashcache);
1470 SvSTASH_set(io, MUTABLE_HV(SvREFCNT_inc(GvHV(iogv))));
1471 IoPAGE_LEN(sv) = 60;
1473 if (old_type < SVt_PV) {
1474 /* referent will be NULL unless the old type was SVt_IV emulating
1476 sv->sv_u.svu_rv = referent;
1480 Perl_croak(aTHX_ "panic: sv_upgrade to unknown type %lu",
1481 (unsigned long)new_type);
1484 /* if this is zero, this is a body-less SVt_NULL, SVt_IV/SVt_RV,
1485 and sometimes SVt_NV */
1486 if (old_type_details->body_size) {
1490 /* Note that there is an assumption that all bodies of types that
1491 can be upgraded came from arenas. Only the more complex non-
1492 upgradable types are allowed to be directly malloc()ed. */
1493 assert(old_type_details->arena);
1494 del_body((void*)((char*)old_body + old_type_details->offset),
1495 &PL_body_roots[old_type]);
1501 =for apidoc sv_backoff
1503 Remove any string offset. You should normally use the C<SvOOK_off> macro
1509 /* prior to 5.000 stable, this function returned the new OOK-less SvFLAGS
1510 prior to 5.23.4 this function always returned 0
1514 Perl_sv_backoff(SV *const sv)
1517 const char * const s = SvPVX_const(sv);
1519 PERL_ARGS_ASSERT_SV_BACKOFF;
1522 assert(SvTYPE(sv) != SVt_PVHV);
1523 assert(SvTYPE(sv) != SVt_PVAV);
1525 SvOOK_offset(sv, delta);
1527 SvLEN_set(sv, SvLEN(sv) + delta);
1528 SvPV_set(sv, SvPVX(sv) - delta);
1529 SvFLAGS(sv) &= ~SVf_OOK;
1530 Move(s, SvPVX(sv), SvCUR(sv)+1, char);
1535 /* forward declaration */
1536 static void S_sv_uncow(pTHX_ SV * const sv, const U32 flags);
1542 Expands the character buffer in the SV. If necessary, uses C<sv_unref> and
1543 upgrades the SV to C<SVt_PV>. Returns a pointer to the character buffer.
1544 Use the C<SvGROW> wrapper instead.
1551 Perl_sv_grow(pTHX_ SV *const sv, STRLEN newlen)
1555 PERL_ARGS_ASSERT_SV_GROW;
1559 if (SvTYPE(sv) < SVt_PV) {
1560 sv_upgrade(sv, SVt_PV);
1561 s = SvPVX_mutable(sv);
1563 else if (SvOOK(sv)) { /* pv is offset? */
1565 s = SvPVX_mutable(sv);
1566 if (newlen > SvLEN(sv))
1567 newlen += 10 * (newlen - SvCUR(sv)); /* avoid copy each time */
1571 if (SvIsCOW(sv)) S_sv_uncow(aTHX_ sv, 0);
1572 s = SvPVX_mutable(sv);
1575 #ifdef PERL_COPY_ON_WRITE
1576 /* the new COW scheme uses SvPVX(sv)[SvLEN(sv)-1] (if spare)
1577 * to store the COW count. So in general, allocate one more byte than
1578 * asked for, to make it likely this byte is always spare: and thus
1579 * make more strings COW-able.
1581 * Only increment if the allocation isn't MEM_SIZE_MAX,
1582 * otherwise it will wrap to 0.
1584 if ( newlen != MEM_SIZE_MAX )
1588 #if defined(PERL_USE_MALLOC_SIZE) && defined(Perl_safesysmalloc_size)
1589 #define PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1592 if (newlen > SvLEN(sv)) { /* need more room? */
1593 STRLEN minlen = SvCUR(sv);
1594 minlen += (minlen >> PERL_STRLEN_EXPAND_SHIFT) + 10;
1595 if (newlen < minlen)
1597 #ifndef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1599 /* Don't round up on the first allocation, as odds are pretty good that
1600 * the initial request is accurate as to what is really needed */
1602 STRLEN rounded = PERL_STRLEN_ROUNDUP(newlen);
1603 if (rounded > newlen)
1607 if (SvLEN(sv) && s) {
1608 s = (char*)saferealloc(s, newlen);
1611 s = (char*)safemalloc(newlen);
1612 if (SvPVX_const(sv) && SvCUR(sv)) {
1613 Move(SvPVX_const(sv), s, SvCUR(sv), char);
1617 #ifdef PERL_UNWARANTED_CHUMMINESS_WITH_MALLOC
1618 /* Do this here, do it once, do it right, and then we will never get
1619 called back into sv_grow() unless there really is some growing
1621 SvLEN_set(sv, Perl_safesysmalloc_size(s));
1623 SvLEN_set(sv, newlen);
1630 =for apidoc sv_setiv
1632 Copies an integer into the given SV, upgrading first if necessary.
1633 Does not handle 'set' magic. See also C<L</sv_setiv_mg>>.
1639 Perl_sv_setiv(pTHX_ SV *const sv, const IV i)
1641 PERL_ARGS_ASSERT_SV_SETIV;
1643 SV_CHECK_THINKFIRST_COW_DROP(sv);
1644 switch (SvTYPE(sv)) {
1647 sv_upgrade(sv, SVt_IV);
1650 sv_upgrade(sv, SVt_PVIV);
1654 if (!isGV_with_GP(sv))
1662 /* diag_listed_as: Can't coerce %s to %s in %s */
1663 Perl_croak(aTHX_ "Can't coerce %s to integer in %s", sv_reftype(sv,0),
1665 NOT_REACHED; /* NOTREACHED */
1669 (void)SvIOK_only(sv); /* validate number */
1675 =for apidoc sv_setiv_mg
1677 Like C<sv_setiv>, but also handles 'set' magic.
1683 Perl_sv_setiv_mg(pTHX_ SV *const sv, const IV i)
1685 PERL_ARGS_ASSERT_SV_SETIV_MG;
1692 =for apidoc sv_setuv
1694 Copies an unsigned integer into the given SV, upgrading first if necessary.
1695 Does not handle 'set' magic. See also C<L</sv_setuv_mg>>.
1701 Perl_sv_setuv(pTHX_ SV *const sv, const UV u)
1703 PERL_ARGS_ASSERT_SV_SETUV;
1705 /* With the if statement to ensure that integers are stored as IVs whenever
1707 u=1.49 s=0.52 cu=72.49 cs=10.64 scripts=270 tests=20865
1710 u=1.35 s=0.47 cu=73.45 cs=11.43 scripts=270 tests=20865
1712 If you wish to remove the following if statement, so that this routine
1713 (and its callers) always return UVs, please benchmark to see what the
1714 effect is. Modern CPUs may be different. Or may not :-)
1716 if (u <= (UV)IV_MAX) {
1717 sv_setiv(sv, (IV)u);
1726 =for apidoc sv_setuv_mg
1728 Like C<sv_setuv>, but also handles 'set' magic.
1734 Perl_sv_setuv_mg(pTHX_ SV *const sv, const UV u)
1736 PERL_ARGS_ASSERT_SV_SETUV_MG;
1743 =for apidoc sv_setnv
1745 Copies a double into the given SV, upgrading first if necessary.
1746 Does not handle 'set' magic. See also C<L</sv_setnv_mg>>.
1752 Perl_sv_setnv(pTHX_ SV *const sv, const NV num)
1754 PERL_ARGS_ASSERT_SV_SETNV;
1756 SV_CHECK_THINKFIRST_COW_DROP(sv);
1757 switch (SvTYPE(sv)) {
1760 sv_upgrade(sv, SVt_NV);
1764 sv_upgrade(sv, SVt_PVNV);
1768 if (!isGV_with_GP(sv))
1776 /* diag_listed_as: Can't coerce %s to %s in %s */
1777 Perl_croak(aTHX_ "Can't coerce %s to number in %s", sv_reftype(sv,0),
1779 NOT_REACHED; /* NOTREACHED */
1784 (void)SvNOK_only(sv); /* validate number */
1789 =for apidoc sv_setnv_mg
1791 Like C<sv_setnv>, but also handles 'set' magic.
1797 Perl_sv_setnv_mg(pTHX_ SV *const sv, const NV num)
1799 PERL_ARGS_ASSERT_SV_SETNV_MG;
1805 /* Return a cleaned-up, printable version of sv, for non-numeric, or
1806 * not incrementable warning display.
1807 * Originally part of S_not_a_number().
1808 * The return value may be != tmpbuf.
1812 S_sv_display(pTHX_ SV *const sv, char *tmpbuf, STRLEN tmpbuf_size) {
1815 PERL_ARGS_ASSERT_SV_DISPLAY;
1818 SV *dsv = newSVpvs_flags("", SVs_TEMP);
1819 pv = sv_uni_display(dsv, sv, 32, UNI_DISPLAY_ISPRINT);
1822 const char * const limit = tmpbuf + tmpbuf_size - 8;
1823 /* each *s can expand to 4 chars + "...\0",
1824 i.e. need room for 8 chars */
1826 const char *s = SvPVX_const(sv);
1827 const char * const end = s + SvCUR(sv);
1828 for ( ; s < end && d < limit; s++ ) {
1830 if (! isASCII(ch) && !isPRINT_LC(ch)) {
1834 /* Map to ASCII "equivalent" of Latin1 */
1835 ch = LATIN1_TO_NATIVE(NATIVE_TO_LATIN1(ch) & 127);
1841 else if (ch == '\r') {
1845 else if (ch == '\f') {
1849 else if (ch == '\\') {
1853 else if (ch == '\0') {
1857 else if (isPRINT_LC(ch))
1876 /* Print an "isn't numeric" warning, using a cleaned-up,
1877 * printable version of the offending string
1881 S_not_a_number(pTHX_ SV *const sv)
1886 PERL_ARGS_ASSERT_NOT_A_NUMBER;
1888 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1891 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1892 /* diag_listed_as: Argument "%s" isn't numeric%s */
1893 "Argument \"%s\" isn't numeric in %s", pv,
1896 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1897 /* diag_listed_as: Argument "%s" isn't numeric%s */
1898 "Argument \"%s\" isn't numeric", pv);
1902 S_not_incrementable(pTHX_ SV *const sv) {
1906 PERL_ARGS_ASSERT_NOT_INCREMENTABLE;
1908 pv = sv_display(sv, tmpbuf, sizeof(tmpbuf));
1910 Perl_warner(aTHX_ packWARN(WARN_NUMERIC),
1911 "Argument \"%s\" treated as 0 in increment (++)", pv);
1915 =for apidoc looks_like_number
1917 Test if the content of an SV looks like a number (or is a number).
1918 C<Inf> and C<Infinity> are treated as numbers (so will not issue a
1919 non-numeric warning), even if your C<atof()> doesn't grok them. Get-magic is
1926 Perl_looks_like_number(pTHX_ SV *const sv)
1932 PERL_ARGS_ASSERT_LOOKS_LIKE_NUMBER;
1934 if (SvPOK(sv) || SvPOKp(sv)) {
1935 sbegin = SvPV_nomg_const(sv, len);
1938 return SvFLAGS(sv) & (SVf_NOK|SVp_NOK|SVf_IOK|SVp_IOK);
1939 numtype = grok_number(sbegin, len, NULL);
1940 return ((numtype & IS_NUMBER_TRAILING)) ? 0 : numtype;
1944 S_glob_2number(pTHX_ GV * const gv)
1946 PERL_ARGS_ASSERT_GLOB_2NUMBER;
1948 /* We know that all GVs stringify to something that is not-a-number,
1949 so no need to test that. */
1950 if (ckWARN(WARN_NUMERIC))
1952 SV *const buffer = sv_newmortal();
1953 gv_efullname3(buffer, gv, "*");
1954 not_a_number(buffer);
1956 /* We just want something true to return, so that S_sv_2iuv_common
1957 can tail call us and return true. */
1961 /* Actually, ISO C leaves conversion of UV to IV undefined, but
1962 until proven guilty, assume that things are not that bad... */
1967 As 64 bit platforms often have an NV that doesn't preserve all bits of
1968 an IV (an assumption perl has been based on to date) it becomes necessary
1969 to remove the assumption that the NV always carries enough precision to
1970 recreate the IV whenever needed, and that the NV is the canonical form.
1971 Instead, IV/UV and NV need to be given equal rights. So as to not lose
1972 precision as a side effect of conversion (which would lead to insanity
1973 and the dragon(s) in t/op/numconvert.t getting very angry) the intent is
1974 1) to distinguish between IV/UV/NV slots that have a valid conversion cached
1975 where precision was lost, and IV/UV/NV slots that have a valid conversion
1976 which has lost no precision
1977 2) to ensure that if a numeric conversion to one form is requested that
1978 would lose precision, the precise conversion (or differently
1979 imprecise conversion) is also performed and cached, to prevent
1980 requests for different numeric formats on the same SV causing
1981 lossy conversion chains. (lossless conversion chains are perfectly
1986 SvIOKp is true if the IV slot contains a valid value
1987 SvIOK is true only if the IV value is accurate (UV if SvIOK_UV true)
1988 SvNOKp is true if the NV slot contains a valid value
1989 SvNOK is true only if the NV value is accurate
1992 while converting from PV to NV, check to see if converting that NV to an
1993 IV(or UV) would lose accuracy over a direct conversion from PV to
1994 IV(or UV). If it would, cache both conversions, return NV, but mark
1995 SV as IOK NOKp (ie not NOK).
1997 While converting from PV to IV, check to see if converting that IV to an
1998 NV would lose accuracy over a direct conversion from PV to NV. If it
1999 would, cache both conversions, flag similarly.
2001 Before, the SV value "3.2" could become NV=3.2 IV=3 NOK, IOK quite
2002 correctly because if IV & NV were set NV *always* overruled.
2003 Now, "3.2" will become NV=3.2 IV=3 NOK, IOKp, because the flag's meaning
2004 changes - now IV and NV together means that the two are interchangeable:
2005 SvIVX == (IV) SvNVX && SvNVX == (NV) SvIVX;
2007 The benefit of this is that operations such as pp_add know that if
2008 SvIOK is true for both left and right operands, then integer addition
2009 can be used instead of floating point (for cases where the result won't
2010 overflow). Before, floating point was always used, which could lead to
2011 loss of precision compared with integer addition.
2013 * making IV and NV equal status should make maths accurate on 64 bit
2015 * may speed up maths somewhat if pp_add and friends start to use
2016 integers when possible instead of fp. (Hopefully the overhead in
2017 looking for SvIOK and checking for overflow will not outweigh the
2018 fp to integer speedup)
2019 * will slow down integer operations (callers of SvIV) on "inaccurate"
2020 values, as the change from SvIOK to SvIOKp will cause a call into
2021 sv_2iv each time rather than a macro access direct to the IV slot
2022 * should speed up number->string conversion on integers as IV is
2023 favoured when IV and NV are equally accurate
2025 ####################################################################
2026 You had better be using SvIOK_notUV if you want an IV for arithmetic:
2027 SvIOK is true if (IV or UV), so you might be getting (IV)SvUV.
2028 On the other hand, SvUOK is true iff UV.
2029 ####################################################################
2031 Your mileage will vary depending your CPU's relative fp to integer
2035 #ifndef NV_PRESERVES_UV
2036 # define IS_NUMBER_UNDERFLOW_IV 1
2037 # define IS_NUMBER_UNDERFLOW_UV 2
2038 # define IS_NUMBER_IV_AND_UV 2
2039 # define IS_NUMBER_OVERFLOW_IV 4
2040 # define IS_NUMBER_OVERFLOW_UV 5
2042 /* sv_2iuv_non_preserve(): private routine for use by sv_2iv() and sv_2uv() */
2044 /* For sv_2nv these three cases are "SvNOK and don't bother casting" */
2046 S_sv_2iuv_non_preserve(pTHX_ SV *const sv
2052 PERL_ARGS_ASSERT_SV_2IUV_NON_PRESERVE;
2053 PERL_UNUSED_CONTEXT;
2055 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));
2056 if (SvNVX(sv) < (NV)IV_MIN) {
2057 (void)SvIOKp_on(sv);
2059 SvIV_set(sv, IV_MIN);
2060 return IS_NUMBER_UNDERFLOW_IV;
2062 if (SvNVX(sv) > (NV)UV_MAX) {
2063 (void)SvIOKp_on(sv);
2066 SvUV_set(sv, UV_MAX);
2067 return IS_NUMBER_OVERFLOW_UV;
2069 (void)SvIOKp_on(sv);
2071 /* Can't use strtol etc to convert this string. (See truth table in
2073 if (SvNVX(sv) <= (UV)IV_MAX) {
2074 SvIV_set(sv, I_V(SvNVX(sv)));
2075 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2076 SvIOK_on(sv); /* Integer is precise. NOK, IOK */
2078 /* Integer is imprecise. NOK, IOKp */
2080 return SvNVX(sv) < 0 ? IS_NUMBER_UNDERFLOW_UV : IS_NUMBER_IV_AND_UV;
2083 SvUV_set(sv, U_V(SvNVX(sv)));
2084 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2085 if (SvUVX(sv) == UV_MAX) {
2086 /* As we know that NVs don't preserve UVs, UV_MAX cannot
2087 possibly be preserved by NV. Hence, it must be overflow.
2089 return IS_NUMBER_OVERFLOW_UV;
2091 SvIOK_on(sv); /* Integer is precise. NOK, UOK */
2093 /* Integer is imprecise. NOK, IOKp */
2095 return IS_NUMBER_OVERFLOW_IV;
2097 #endif /* !NV_PRESERVES_UV*/
2099 /* If numtype is infnan, set the NV of the sv accordingly.
2100 * If numtype is anything else, try setting the NV using Atof(PV). */
2102 S_sv_setnv(pTHX_ SV* sv, int numtype)
2104 bool pok = cBOOL(SvPOK(sv));
2107 if ((numtype & IS_NUMBER_INFINITY)) {
2108 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF);
2113 if ((numtype & IS_NUMBER_NAN)) {
2114 SvNV_set(sv, NV_NAN);
2119 SvNV_set(sv, Atof(SvPVX_const(sv)));
2120 /* Purposefully no true nok here, since we don't want to blow
2121 * away the possible IOK/UV of an existing sv. */
2124 SvNOK_only(sv); /* No IV or UV please, this is pure infnan. */
2126 SvPOK_on(sv); /* PV is okay, though. */
2131 S_sv_2iuv_common(pTHX_ SV *const sv)
2133 PERL_ARGS_ASSERT_SV_2IUV_COMMON;
2136 /* erm. not sure. *should* never get NOKp (without NOK) from sv_2nv
2137 * without also getting a cached IV/UV from it at the same time
2138 * (ie PV->NV conversion should detect loss of accuracy and cache
2139 * IV or UV at same time to avoid this. */
2140 /* IV-over-UV optimisation - choose to cache IV if possible */
2142 if (SvTYPE(sv) == SVt_NV)
2143 sv_upgrade(sv, SVt_PVNV);
2145 (void)SvIOKp_on(sv); /* Must do this first, to clear any SvOOK */
2146 /* < not <= as for NV doesn't preserve UV, ((NV)IV_MAX+1) will almost
2147 certainly cast into the IV range at IV_MAX, whereas the correct
2148 answer is the UV IV_MAX +1. Hence < ensures that dodgy boundary
2150 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2151 if (Perl_isnan(SvNVX(sv))) {
2157 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2158 SvIV_set(sv, I_V(SvNVX(sv)));
2159 if (SvNVX(sv) == (NV) SvIVX(sv)
2160 #ifndef NV_PRESERVES_UV
2161 && SvIVX(sv) != IV_MIN /* avoid negating IV_MIN below */
2162 && (((UV)1 << NV_PRESERVES_UV_BITS) >
2163 (UV)(SvIVX(sv) > 0 ? SvIVX(sv) : -SvIVX(sv)))
2164 /* Don't flag it as "accurately an integer" if the number
2165 came from a (by definition imprecise) NV operation, and
2166 we're outside the range of NV integer precision */
2170 SvIOK_on(sv); /* Can this go wrong with rounding? NWC */
2172 /* scalar has trailing garbage, eg "42a" */
2174 DEBUG_c(PerlIO_printf(Perl_debug_log,
2175 "0x%" UVxf " iv(%" NVgf " => %" IVdf ") (precise)\n",
2181 /* IV not precise. No need to convert from PV, as NV
2182 conversion would already have cached IV if it detected
2183 that PV->IV would be better than PV->NV->IV
2184 flags already correct - don't set public IOK. */
2185 DEBUG_c(PerlIO_printf(Perl_debug_log,
2186 "0x%" UVxf " iv(%" NVgf " => %" IVdf ") (imprecise)\n",
2191 /* Can the above go wrong if SvIVX == IV_MIN and SvNVX < IV_MIN,
2192 but the cast (NV)IV_MIN rounds to a the value less (more
2193 negative) than IV_MIN which happens to be equal to SvNVX ??
2194 Analogous to 0xFFFFFFFFFFFFFFFF rounding up to NV (2**64) and
2195 NV rounding back to 0xFFFFFFFFFFFFFFFF, so UVX == UV(NVX) and
2196 (NV)UVX == NVX are both true, but the values differ. :-(
2197 Hopefully for 2s complement IV_MIN is something like
2198 0x8000000000000000 which will be exact. NWC */
2201 SvUV_set(sv, U_V(SvNVX(sv)));
2203 (SvNVX(sv) == (NV) SvUVX(sv))
2204 #ifndef NV_PRESERVES_UV
2205 /* Make sure it's not 0xFFFFFFFFFFFFFFFF */
2206 /*&& (SvUVX(sv) != UV_MAX) irrelevant with code below */
2207 && (((UV)1 << NV_PRESERVES_UV_BITS) > SvUVX(sv))
2208 /* Don't flag it as "accurately an integer" if the number
2209 came from a (by definition imprecise) NV operation, and
2210 we're outside the range of NV integer precision */
2216 DEBUG_c(PerlIO_printf(Perl_debug_log,
2217 "0x%" UVxf " 2iv(%" UVuf " => %" IVdf ") (as unsigned)\n",
2223 else if (SvPOKp(sv)) {
2226 const char *s = SvPVX_const(sv);
2227 const STRLEN cur = SvCUR(sv);
2229 /* short-cut for a single digit string like "1" */
2234 if (SvTYPE(sv) < SVt_PVIV)
2235 sv_upgrade(sv, SVt_PVIV);
2237 SvIV_set(sv, (IV)(c - '0'));
2242 numtype = grok_number(s, cur, &value);
2243 /* We want to avoid a possible problem when we cache an IV/ a UV which
2244 may be later translated to an NV, and the resulting NV is not
2245 the same as the direct translation of the initial string
2246 (eg 123.456 can shortcut to the IV 123 with atol(), but we must
2247 be careful to ensure that the value with the .456 is around if the
2248 NV value is requested in the future).
2250 This means that if we cache such an IV/a UV, we need to cache the
2251 NV as well. Moreover, we trade speed for space, and do not
2252 cache the NV if we are sure it's not needed.
2255 /* SVt_PVNV is one higher than SVt_PVIV, hence this order */
2256 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2257 == IS_NUMBER_IN_UV) {
2258 /* It's definitely an integer, only upgrade to PVIV */
2259 if (SvTYPE(sv) < SVt_PVIV)
2260 sv_upgrade(sv, SVt_PVIV);
2262 } else if (SvTYPE(sv) < SVt_PVNV)
2263 sv_upgrade(sv, SVt_PVNV);
2265 if ((numtype & (IS_NUMBER_INFINITY | IS_NUMBER_NAN))) {
2266 if (ckWARN(WARN_NUMERIC) && ((numtype & IS_NUMBER_TRAILING)))
2268 S_sv_setnv(aTHX_ sv, numtype);
2272 /* If NVs preserve UVs then we only use the UV value if we know that
2273 we aren't going to call atof() below. If NVs don't preserve UVs
2274 then the value returned may have more precision than atof() will
2275 return, even though value isn't perfectly accurate. */
2276 if ((numtype & (IS_NUMBER_IN_UV
2277 #ifdef NV_PRESERVES_UV
2280 )) == IS_NUMBER_IN_UV) {
2281 /* This won't turn off the public IOK flag if it was set above */
2282 (void)SvIOKp_on(sv);
2284 if (!(numtype & IS_NUMBER_NEG)) {
2286 if (value <= (UV)IV_MAX) {
2287 SvIV_set(sv, (IV)value);
2289 /* it didn't overflow, and it was positive. */
2290 SvUV_set(sv, value);
2294 /* 2s complement assumption */
2295 if (value <= (UV)IV_MIN) {
2296 SvIV_set(sv, value == (UV)IV_MIN
2297 ? IV_MIN : -(IV)value);
2299 /* Too negative for an IV. This is a double upgrade, but
2300 I'm assuming it will be rare. */
2301 if (SvTYPE(sv) < SVt_PVNV)
2302 sv_upgrade(sv, SVt_PVNV);
2306 SvNV_set(sv, -(NV)value);
2307 SvIV_set(sv, IV_MIN);
2311 /* For !NV_PRESERVES_UV and IS_NUMBER_IN_UV and IS_NUMBER_NOT_INT we
2312 will be in the previous block to set the IV slot, and the next
2313 block to set the NV slot. So no else here. */
2315 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2316 != IS_NUMBER_IN_UV) {
2317 /* It wasn't an (integer that doesn't overflow the UV). */
2318 S_sv_setnv(aTHX_ sv, numtype);
2320 if (! numtype && ckWARN(WARN_NUMERIC))
2323 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%" UVxf " 2iv(%" NVgf ")\n",
2324 PTR2UV(sv), SvNVX(sv)));
2326 #ifdef NV_PRESERVES_UV
2327 (void)SvIOKp_on(sv);
2329 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
2330 if (Perl_isnan(SvNVX(sv))) {
2336 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2337 SvIV_set(sv, I_V(SvNVX(sv)));
2338 if ((NV)(SvIVX(sv)) == SvNVX(sv)) {
2341 NOOP; /* Integer is imprecise. NOK, IOKp */
2343 /* UV will not work better than IV */
2345 if (SvNVX(sv) > (NV)UV_MAX) {
2347 /* Integer is inaccurate. NOK, IOKp, is UV */
2348 SvUV_set(sv, UV_MAX);
2350 SvUV_set(sv, U_V(SvNVX(sv)));
2351 /* 0xFFFFFFFFFFFFFFFF not an issue in here, NVs
2352 NV preservse UV so can do correct comparison. */
2353 if ((NV)(SvUVX(sv)) == SvNVX(sv)) {
2356 NOOP; /* Integer is imprecise. NOK, IOKp, is UV */
2361 #else /* NV_PRESERVES_UV */
2362 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2363 == (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT)) {
2364 /* The IV/UV slot will have been set from value returned by
2365 grok_number above. The NV slot has just been set using
2368 assert (SvIOKp(sv));
2370 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2371 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2372 /* Small enough to preserve all bits. */
2373 (void)SvIOKp_on(sv);
2375 SvIV_set(sv, I_V(SvNVX(sv)));
2376 if ((NV)(SvIVX(sv)) == SvNVX(sv))
2378 /* Assumption: first non-preserved integer is < IV_MAX,
2379 this NV is in the preserved range, therefore: */
2380 if (!(U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))
2382 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);
2386 0 0 already failed to read UV.
2387 0 1 already failed to read UV.
2388 1 0 you won't get here in this case. IV/UV
2389 slot set, public IOK, Atof() unneeded.
2390 1 1 already read UV.
2391 so there's no point in sv_2iuv_non_preserve() attempting
2392 to use atol, strtol, strtoul etc. */
2394 sv_2iuv_non_preserve (sv, numtype);
2396 sv_2iuv_non_preserve (sv);
2400 #endif /* NV_PRESERVES_UV */
2401 /* It might be more code efficient to go through the entire logic above
2402 and conditionally set with SvIOKp_on() rather than SvIOK(), but it
2403 gets complex and potentially buggy, so more programmer efficient
2404 to do it this way, by turning off the public flags: */
2406 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2410 if (isGV_with_GP(sv))
2411 return glob_2number(MUTABLE_GV(sv));
2413 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2415 if (SvTYPE(sv) < SVt_IV)
2416 /* Typically the caller expects that sv_any is not NULL now. */
2417 sv_upgrade(sv, SVt_IV);
2418 /* Return 0 from the caller. */
2425 =for apidoc sv_2iv_flags
2427 Return the integer value of an SV, doing any necessary string
2428 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2429 Normally used via the C<SvIV(sv)> and C<SvIVx(sv)> macros.
2435 Perl_sv_2iv_flags(pTHX_ SV *const sv, const I32 flags)
2437 PERL_ARGS_ASSERT_SV_2IV_FLAGS;
2439 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2440 && SvTYPE(sv) != SVt_PVFM);
2442 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2448 if (flags & SV_SKIP_OVERLOAD)
2450 tmpstr = AMG_CALLunary(sv, numer_amg);
2451 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2452 return SvIV(tmpstr);
2455 return PTR2IV(SvRV(sv));
2458 if (SvVALID(sv) || isREGEXP(sv)) {
2459 /* FBMs use the space for SvIVX and SvNVX for other purposes, so
2460 must not let them cache IVs.
2461 In practice they are extremely unlikely to actually get anywhere
2462 accessible by user Perl code - the only way that I'm aware of is when
2463 a constant subroutine which is used as the second argument to index.
2465 Regexps have no SvIVX and SvNVX fields.
2470 const char * const ptr =
2471 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2473 = grok_number(ptr, SvCUR(sv), &value);
2475 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2476 == IS_NUMBER_IN_UV) {
2477 /* It's definitely an integer */
2478 if (numtype & IS_NUMBER_NEG) {
2479 if (value < (UV)IV_MIN)
2482 if (value < (UV)IV_MAX)
2487 /* Quite wrong but no good choices. */
2488 if ((numtype & IS_NUMBER_INFINITY)) {
2489 return (numtype & IS_NUMBER_NEG) ? IV_MIN : IV_MAX;
2490 } else if ((numtype & IS_NUMBER_NAN)) {
2491 return 0; /* So wrong. */
2495 if (ckWARN(WARN_NUMERIC))
2498 return I_V(Atof(ptr));
2502 if (SvTHINKFIRST(sv)) {
2503 if (SvREADONLY(sv) && !SvOK(sv)) {
2504 if (ckWARN(WARN_UNINITIALIZED))
2511 if (S_sv_2iuv_common(aTHX_ sv))
2515 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%" UVxf " 2iv(%" IVdf ")\n",
2516 PTR2UV(sv),SvIVX(sv)));
2517 return SvIsUV(sv) ? (IV)SvUVX(sv) : SvIVX(sv);
2521 =for apidoc sv_2uv_flags
2523 Return the unsigned integer value of an SV, doing any necessary string
2524 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2525 Normally used via the C<SvUV(sv)> and C<SvUVx(sv)> macros.
2527 =for apidoc Amnh||SV_GMAGIC
2533 Perl_sv_2uv_flags(pTHX_ SV *const sv, const I32 flags)
2535 PERL_ARGS_ASSERT_SV_2UV_FLAGS;
2537 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
2543 if (flags & SV_SKIP_OVERLOAD)
2545 tmpstr = AMG_CALLunary(sv, numer_amg);
2546 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2547 return SvUV(tmpstr);
2550 return PTR2UV(SvRV(sv));
2553 if (SvVALID(sv) || isREGEXP(sv)) {
2554 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2555 the same flag bit as SVf_IVisUV, so must not let them cache IVs.
2556 Regexps have no SvIVX and SvNVX fields. */
2560 const char * const ptr =
2561 isREGEXP(sv) ? RX_WRAPPED((REGEXP*)sv) : SvPVX_const(sv);
2563 = grok_number(ptr, SvCUR(sv), &value);
2565 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2566 == IS_NUMBER_IN_UV) {
2567 /* It's definitely an integer */
2568 if (!(numtype & IS_NUMBER_NEG))
2572 /* Quite wrong but no good choices. */
2573 if ((numtype & IS_NUMBER_INFINITY)) {
2574 return UV_MAX; /* So wrong. */
2575 } else if ((numtype & IS_NUMBER_NAN)) {
2576 return 0; /* So wrong. */
2580 if (ckWARN(WARN_NUMERIC))
2583 return U_V(Atof(ptr));
2587 if (SvTHINKFIRST(sv)) {
2588 if (SvREADONLY(sv) && !SvOK(sv)) {
2589 if (ckWARN(WARN_UNINITIALIZED))
2596 if (S_sv_2iuv_common(aTHX_ sv))
2600 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%" UVxf " 2uv(%" UVuf ")\n",
2601 PTR2UV(sv),SvUVX(sv)));
2602 return SvIsUV(sv) ? SvUVX(sv) : (UV)SvIVX(sv);
2606 =for apidoc sv_2nv_flags
2608 Return the num value of an SV, doing any necessary string or integer
2609 conversion. If C<flags> has the C<SV_GMAGIC> bit set, does an C<mg_get()> first.
2610 Normally used via the C<SvNV(sv)> and C<SvNVx(sv)> macros.
2616 Perl_sv_2nv_flags(pTHX_ SV *const sv, const I32 flags)
2618 PERL_ARGS_ASSERT_SV_2NV_FLAGS;
2620 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
2621 && SvTYPE(sv) != SVt_PVFM);
2622 if (SvGMAGICAL(sv) || SvVALID(sv) || isREGEXP(sv)) {
2623 /* FBMs use the space for SvIVX and SvNVX for other purposes, and use
2624 the same flag bit as SVf_IVisUV, so must not let them cache NVs.
2625 Regexps have no SvIVX and SvNVX fields. */
2627 if (flags & SV_GMAGIC)
2631 if (SvPOKp(sv) && !SvIOKp(sv)) {
2632 ptr = SvPVX_const(sv);
2633 if (!SvIOKp(sv) && ckWARN(WARN_NUMERIC) &&
2634 !grok_number(ptr, SvCUR(sv), NULL))
2640 return (NV)SvUVX(sv);
2642 return (NV)SvIVX(sv);
2647 assert(SvTYPE(sv) >= SVt_PVMG);
2648 /* This falls through to the report_uninit near the end of the
2650 } else if (SvTHINKFIRST(sv)) {
2655 if (flags & SV_SKIP_OVERLOAD)
2657 tmpstr = AMG_CALLunary(sv, numer_amg);
2658 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
2659 return SvNV(tmpstr);
2662 return PTR2NV(SvRV(sv));
2664 if (SvREADONLY(sv) && !SvOK(sv)) {
2665 if (ckWARN(WARN_UNINITIALIZED))
2670 if (SvTYPE(sv) < SVt_NV) {
2671 /* The logic to use SVt_PVNV if necessary is in sv_upgrade. */
2672 sv_upgrade(sv, SVt_NV);
2673 CLANG_DIAG_IGNORE_STMT(-Wthread-safety);
2675 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
2676 STORE_LC_NUMERIC_SET_STANDARD();
2677 PerlIO_printf(Perl_debug_log,
2678 "0x%" UVxf " num(%" NVgf ")\n",
2679 PTR2UV(sv), SvNVX(sv));
2680 RESTORE_LC_NUMERIC();
2682 CLANG_DIAG_RESTORE_STMT;
2685 else if (SvTYPE(sv) < SVt_PVNV)
2686 sv_upgrade(sv, SVt_PVNV);
2691 SvNV_set(sv, SvIsUV(sv) ? (NV)SvUVX(sv) : (NV)SvIVX(sv));
2692 #ifdef NV_PRESERVES_UV
2698 /* Only set the public NV OK flag if this NV preserves the IV */
2699 /* Check it's not 0xFFFFFFFFFFFFFFFF */
2701 SvIsUV(sv) ? ((SvUVX(sv) != UV_MAX)&&(SvUVX(sv) == U_V(SvNVX(sv))))
2702 : (SvIVX(sv) == I_V(SvNVX(sv))))
2708 else if (SvPOKp(sv)) {
2710 const int numtype = grok_number(SvPVX_const(sv), SvCUR(sv), &value);
2711 if (!SvIOKp(sv) && !numtype && ckWARN(WARN_NUMERIC))
2713 #ifdef NV_PRESERVES_UV
2714 if ((numtype & (IS_NUMBER_IN_UV | IS_NUMBER_NOT_INT))
2715 == IS_NUMBER_IN_UV) {
2716 /* It's definitely an integer */
2717 SvNV_set(sv, (numtype & IS_NUMBER_NEG) ? -(NV)value : (NV)value);
2719 S_sv_setnv(aTHX_ sv, numtype);
2726 SvNV_set(sv, Atof(SvPVX_const(sv)));
2727 /* Only set the public NV OK flag if this NV preserves the value in
2728 the PV at least as well as an IV/UV would.
2729 Not sure how to do this 100% reliably. */
2730 /* if that shift count is out of range then Configure's test is
2731 wonky. We shouldn't be in here with NV_PRESERVES_UV_BITS ==
2733 if (((UV)1 << NV_PRESERVES_UV_BITS) >
2734 U_V(SvNVX(sv) > 0 ? SvNVX(sv) : -SvNVX(sv))) {
2735 SvNOK_on(sv); /* Definitely small enough to preserve all bits */
2736 } else if (!(numtype & IS_NUMBER_IN_UV)) {
2737 /* Can't use strtol etc to convert this string, so don't try.
2738 sv_2iv and sv_2uv will use the NV to convert, not the PV. */
2741 /* value has been set. It may not be precise. */
2742 if ((numtype & IS_NUMBER_NEG) && (value >= (UV)IV_MIN)) {
2743 /* 2s complement assumption for (UV)IV_MIN */
2744 SvNOK_on(sv); /* Integer is too negative. */
2749 if (numtype & IS_NUMBER_NEG) {
2750 /* -IV_MIN is undefined, but we should never reach
2751 * this point with both IS_NUMBER_NEG and value ==
2753 assert(value != (UV)IV_MIN);
2754 SvIV_set(sv, -(IV)value);
2755 } else if (value <= (UV)IV_MAX) {
2756 SvIV_set(sv, (IV)value);
2758 SvUV_set(sv, value);
2762 if (numtype & IS_NUMBER_NOT_INT) {
2763 /* I believe that even if the original PV had decimals,
2764 they are lost beyond the limit of the FP precision.
2765 However, neither is canonical, so both only get p
2766 flags. NWC, 2000/11/25 */
2767 /* Both already have p flags, so do nothing */
2769 const NV nv = SvNVX(sv);
2770 /* XXX should this spot have NAN_COMPARE_BROKEN, too? */
2771 if (SvNVX(sv) < (NV)IV_MAX + 0.5) {
2772 if (SvIVX(sv) == I_V(nv)) {
2775 /* It had no "." so it must be integer. */
2779 /* between IV_MAX and NV(UV_MAX).
2780 Could be slightly > UV_MAX */
2782 if (numtype & IS_NUMBER_NOT_INT) {
2783 /* UV and NV both imprecise. */
2785 const UV nv_as_uv = U_V(nv);
2787 if (value == nv_as_uv && SvUVX(sv) != UV_MAX) {
2796 /* It might be more code efficient to go through the entire logic above
2797 and conditionally set with SvNOKp_on() rather than SvNOK(), but it
2798 gets complex and potentially buggy, so more programmer efficient
2799 to do it this way, by turning off the public flags: */
2801 SvFLAGS(sv) &= ~(SVf_IOK|SVf_NOK);
2802 #endif /* NV_PRESERVES_UV */
2805 if (isGV_with_GP(sv)) {
2806 glob_2number(MUTABLE_GV(sv));
2810 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
2812 assert (SvTYPE(sv) >= SVt_NV);
2813 /* Typically the caller expects that sv_any is not NULL now. */
2814 /* XXX Ilya implies that this is a bug in callers that assume this
2815 and ideally should be fixed. */
2818 CLANG_DIAG_IGNORE_STMT(-Wthread-safety);
2820 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
2821 STORE_LC_NUMERIC_SET_STANDARD();
2822 PerlIO_printf(Perl_debug_log, "0x%" UVxf " 2nv(%" NVgf ")\n",
2823 PTR2UV(sv), SvNVX(sv));
2824 RESTORE_LC_NUMERIC();
2826 CLANG_DIAG_RESTORE_STMT;
2833 Return an SV with the numeric value of the source SV, doing any necessary
2834 reference or overload conversion. The caller is expected to have handled
2841 Perl_sv_2num(pTHX_ SV *const sv)
2843 PERL_ARGS_ASSERT_SV_2NUM;
2848 SV * const tmpsv = AMG_CALLunary(sv, numer_amg);
2849 TAINT_IF(tmpsv && SvTAINTED(tmpsv));
2850 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv))))
2851 return sv_2num(tmpsv);
2853 return sv_2mortal(newSVuv(PTR2UV(SvRV(sv))));
2856 /* int2str_table: lookup table containing string representations of all
2857 * two digit numbers. For example, int2str_table.arr[0] is "00" and
2858 * int2str_table.arr[12*2] is "12".
2860 * We are going to read two bytes at a time, so we have to ensure that
2861 * the array is aligned to a 2 byte boundary. That's why it was made a
2862 * union with a dummy U16 member. */
2863 static const union {
2866 } int2str_table = {{
2867 '0', '0', '0', '1', '0', '2', '0', '3', '0', '4', '0', '5', '0', '6',
2868 '0', '7', '0', '8', '0', '9', '1', '0', '1', '1', '1', '2', '1', '3',
2869 '1', '4', '1', '5', '1', '6', '1', '7', '1', '8', '1', '9', '2', '0',
2870 '2', '1', '2', '2', '2', '3', '2', '4', '2', '5', '2', '6', '2', '7',
2871 '2', '8', '2', '9', '3', '0', '3', '1', '3', '2', '3', '3', '3', '4',
2872 '3', '5', '3', '6', '3', '7', '3', '8', '3', '9', '4', '0', '4', '1',
2873 '4', '2', '4', '3', '4', '4', '4', '5', '4', '6', '4', '7', '4', '8',
2874 '4', '9', '5', '0', '5', '1', '5', '2', '5', '3', '5', '4', '5', '5',
2875 '5', '6', '5', '7', '5', '8', '5', '9', '6', '0', '6', '1', '6', '2',
2876 '6', '3', '6', '4', '6', '5', '6', '6', '6', '7', '6', '8', '6', '9',
2877 '7', '0', '7', '1', '7', '2', '7', '3', '7', '4', '7', '5', '7', '6',
2878 '7', '7', '7', '8', '7', '9', '8', '0', '8', '1', '8', '2', '8', '3',
2879 '8', '4', '8', '5', '8', '6', '8', '7', '8', '8', '8', '9', '9', '0',
2880 '9', '1', '9', '2', '9', '3', '9', '4', '9', '5', '9', '6', '9', '7',
2884 /* uiv_2buf(): private routine for use by sv_2pv_flags(): print an IV or
2885 * UV as a string towards the end of buf, and return pointers to start and
2888 * We assume that buf is at least TYPE_CHARS(UV) long.
2891 PERL_STATIC_INLINE char *
2892 S_uiv_2buf(char *const buf, const IV iv, UV uv, const int is_uv, char **const peob)
2894 char *ptr = buf + TYPE_CHARS(UV);
2895 char * const ebuf = ptr;
2897 U16 *word_ptr, *word_table;
2899 PERL_ARGS_ASSERT_UIV_2BUF;
2901 /* ptr has to be properly aligned, because we will cast it to U16* */
2902 assert(PTR2nat(ptr) % 2 == 0);
2903 /* we are going to read/write two bytes at a time */
2904 word_ptr = (U16*)ptr;
2905 word_table = (U16*)int2str_table.arr;
2907 if (UNLIKELY(is_uv))
2913 /* Using 0- here to silence bogus warning from MS VC */
2914 uv = (UV) (0 - (UV) iv);
2919 *--word_ptr = word_table[uv % 100];
2922 ptr = (char*)word_ptr;
2925 *--ptr = (char)uv + '0';
2927 *--word_ptr = word_table[uv];
2928 ptr = (char*)word_ptr;
2938 /* Helper for sv_2pv_flags and sv_vcatpvfn_flags. If the NV is an
2939 * infinity or a not-a-number, writes the appropriate strings to the
2940 * buffer, including a zero byte. On success returns the written length,
2941 * excluding the zero byte, on failure (not an infinity, not a nan)
2942 * returns zero, assert-fails on maxlen being too short.
2944 * XXX for "Inf", "-Inf", and "NaN", we could have three read-only
2945 * shared string constants we point to, instead of generating a new
2946 * string for each instance. */
2948 S_infnan_2pv(NV nv, char* buffer, size_t maxlen, char plus) {
2950 assert(maxlen >= 4);
2951 if (Perl_isinf(nv)) {
2953 if (maxlen < 5) /* "-Inf\0" */
2963 else if (Perl_isnan(nv)) {
2967 /* XXX optionally output the payload mantissa bits as
2968 * "(unsigned)" (to match the nan("...") C99 function,
2969 * or maybe as "(0xhhh...)" would make more sense...
2970 * provide a format string so that the user can decide?
2971 * NOTE: would affect the maxlen and assert() logic.*/
2976 assert((s == buffer + 3) || (s == buffer + 4));
2982 =for apidoc sv_2pv_flags
2984 Returns a pointer to the string value of an SV, and sets C<*lp> to its length.
2985 If flags has the C<SV_GMAGIC> bit set, does an C<mg_get()> first. Coerces C<sv> to a
2986 string if necessary. Normally invoked via the C<SvPV_flags> macro.
2987 C<sv_2pv()> and C<sv_2pv_nomg> usually end up here too.
2993 Perl_sv_2pv_flags(pTHX_ SV *const sv, STRLEN *const lp, const I32 flags)
2997 PERL_ARGS_ASSERT_SV_2PV_FLAGS;
2999 assert (SvTYPE(sv) != SVt_PVAV && SvTYPE(sv) != SVt_PVHV
3000 && SvTYPE(sv) != SVt_PVFM);
3001 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
3006 if (flags & SV_SKIP_OVERLOAD)
3008 tmpstr = AMG_CALLunary(sv, string_amg);
3009 TAINT_IF(tmpstr && SvTAINTED(tmpstr));
3010 if (tmpstr && (!SvROK(tmpstr) || (SvRV(tmpstr) != SvRV(sv)))) {
3012 /* char *pv = lp ? SvPV(tmpstr, *lp) : SvPV_nolen(tmpstr);
3016 if ((SvFLAGS(tmpstr) & (SVf_POK)) == SVf_POK) {
3017 if (flags & SV_CONST_RETURN) {
3018 pv = (char *) SvPVX_const(tmpstr);
3020 pv = (flags & SV_MUTABLE_RETURN)
3021 ? SvPVX_mutable(tmpstr) : SvPVX(tmpstr);
3024 *lp = SvCUR(tmpstr);
3026 pv = sv_2pv_flags(tmpstr, lp, flags);
3039 SV *const referent = SvRV(sv);
3043 retval = buffer = savepvn("NULLREF", len);
3044 } else if (SvTYPE(referent) == SVt_REGEXP &&
3045 (!(PL_curcop->cop_hints & HINT_NO_AMAGIC) ||
3046 amagic_is_enabled(string_amg))) {
3047 REGEXP * const re = (REGEXP *)MUTABLE_PTR(referent);
3051 /* If the regex is UTF-8 we want the containing scalar to
3052 have an UTF-8 flag too */
3059 *lp = RX_WRAPLEN(re);
3061 return RX_WRAPPED(re);
3063 const char *const typestring = sv_reftype(referent, 0);
3064 const STRLEN typelen = strlen(typestring);
3065 UV addr = PTR2UV(referent);
3066 const char *stashname = NULL;
3067 STRLEN stashnamelen = 0; /* hush, gcc */
3068 const char *buffer_end;
3070 if (SvOBJECT(referent)) {
3071 const HEK *const name = HvNAME_HEK(SvSTASH(referent));
3074 stashname = HEK_KEY(name);
3075 stashnamelen = HEK_LEN(name);
3077 if (HEK_UTF8(name)) {
3083 stashname = "__ANON__";
3086 len = stashnamelen + 1 /* = */ + typelen + 3 /* (0x */
3087 + 2 * sizeof(UV) + 2 /* )\0 */;
3089 len = typelen + 3 /* (0x */
3090 + 2 * sizeof(UV) + 2 /* )\0 */;
3093 Newx(buffer, len, char);
3094 buffer_end = retval = buffer + len;
3096 /* Working backwards */
3100 *--retval = PL_hexdigit[addr & 15];
3101 } while (addr >>= 4);
3107 memcpy(retval, typestring, typelen);
3111 retval -= stashnamelen;
3112 memcpy(retval, stashname, stashnamelen);
3114 /* retval may not necessarily have reached the start of the
3116 assert (retval >= buffer);
3118 len = buffer_end - retval - 1; /* -1 for that \0 */
3130 if (flags & SV_MUTABLE_RETURN)
3131 return SvPVX_mutable(sv);
3132 if (flags & SV_CONST_RETURN)
3133 return (char *)SvPVX_const(sv);
3138 /* I'm assuming that if both IV and NV are equally valid then
3139 converting the IV is going to be more efficient */
3140 const U32 isUIOK = SvIsUV(sv);
3141 /* The purpose of this union is to ensure that arr is aligned on
3142 a 2 byte boundary, because that is what uiv_2buf() requires */
3144 char arr[TYPE_CHARS(UV)];
3150 if (SvTYPE(sv) < SVt_PVIV)
3151 sv_upgrade(sv, SVt_PVIV);
3152 ptr = uiv_2buf(buf.arr, SvIVX(sv), SvUVX(sv), isUIOK, &ebuf);
3154 /* inlined from sv_setpvn */
3155 s = SvGROW_mutable(sv, len + 1);
3156 Move(ptr, s, len, char);
3161 else if (SvNOK(sv)) {
3162 if (SvTYPE(sv) < SVt_PVNV)
3163 sv_upgrade(sv, SVt_PVNV);
3164 if (SvNVX(sv) == 0.0
3165 #if defined(NAN_COMPARE_BROKEN) && defined(Perl_isnan)
3166 && !Perl_isnan(SvNVX(sv))
3169 s = SvGROW_mutable(sv, 2);
3174 STRLEN size = 5; /* "-Inf\0" */
3176 s = SvGROW_mutable(sv, size);
3177 len = S_infnan_2pv(SvNVX(sv), s, size, 0);
3183 /* some Xenix systems wipe out errno here */
3192 5 + /* exponent digits */
3196 s = SvGROW_mutable(sv, size);
3197 #ifndef USE_LOCALE_NUMERIC
3198 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3204 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
3205 STORE_LC_NUMERIC_SET_TO_NEEDED();
3207 local_radix = _NOT_IN_NUMERIC_STANDARD;
3208 if (local_radix && SvCUR(PL_numeric_radix_sv) > 1) {
3209 size += SvCUR(PL_numeric_radix_sv) - 1;
3210 s = SvGROW_mutable(sv, size);
3213 SNPRINTF_G(SvNVX(sv), s, SvLEN(sv), NV_DIG);
3215 /* If the radix character is UTF-8, and actually is in the
3216 * output, turn on the UTF-8 flag for the scalar */
3218 && SvUTF8(PL_numeric_radix_sv)
3219 && instr(s, SvPVX_const(PL_numeric_radix_sv)))
3224 RESTORE_LC_NUMERIC();
3227 /* We don't call SvPOK_on(), because it may come to
3228 * pass that the locale changes so that the
3229 * stringification we just did is no longer correct. We
3230 * will have to re-stringify every time it is needed */
3237 else if (isGV_with_GP(sv)) {
3238 GV *const gv = MUTABLE_GV(sv);
3239 SV *const buffer = sv_newmortal();
3241 gv_efullname3(buffer, gv, "*");
3243 assert(SvPOK(buffer));
3249 *lp = SvCUR(buffer);
3250 return SvPVX(buffer);
3255 if (flags & SV_UNDEF_RETURNS_NULL)
3257 if (!PL_localizing && ckWARN(WARN_UNINITIALIZED))
3259 /* Typically the caller expects that sv_any is not NULL now. */
3260 if (!SvREADONLY(sv) && SvTYPE(sv) < SVt_PV)
3261 sv_upgrade(sv, SVt_PV);
3266 const STRLEN len = s - SvPVX_const(sv);
3271 DEBUG_c(PerlIO_printf(Perl_debug_log, "0x%" UVxf " 2pv(%s)\n",
3272 PTR2UV(sv),SvPVX_const(sv)));
3273 if (flags & SV_CONST_RETURN)
3274 return (char *)SvPVX_const(sv);
3275 if (flags & SV_MUTABLE_RETURN)
3276 return SvPVX_mutable(sv);
3281 =for apidoc sv_copypv
3283 Copies a stringified representation of the source SV into the
3284 destination SV. Automatically performs any necessary C<mg_get> and
3285 coercion of numeric values into strings. Guaranteed to preserve
3286 C<UTF8> flag even from overloaded objects. Similar in nature to
3287 C<sv_2pv[_flags]> but operates directly on an SV instead of just the
3288 string. Mostly uses C<sv_2pv_flags> to do its work, except when that
3289 would lose the UTF-8'ness of the PV.
3291 =for apidoc sv_copypv_nomg
3293 Like C<sv_copypv>, but doesn't invoke get magic first.
3295 =for apidoc sv_copypv_flags
3297 Implementation of C<sv_copypv> and C<sv_copypv_nomg>. Calls get magic iff flags
3298 has the C<SV_GMAGIC> bit set.
3304 Perl_sv_copypv_flags(pTHX_ SV *const dsv, SV *const ssv, const I32 flags)
3309 PERL_ARGS_ASSERT_SV_COPYPV_FLAGS;
3311 s = SvPV_flags_const(ssv,len,(flags & SV_GMAGIC));
3312 sv_setpvn(dsv,s,len);
3320 =for apidoc sv_2pvbyte
3322 Return a pointer to the byte-encoded representation of the SV, and set C<*lp>
3323 to its length. If the SV is marked as being encoded as UTF-8, it will
3324 downgrade it to a byte string as a side-effect, if possible. If the SV cannot
3325 be downgraded, this croaks.
3327 Usually accessed via the C<SvPVbyte> macro.
3333 Perl_sv_2pvbyte_flags(pTHX_ SV *sv, STRLEN *const lp, const U32 flags)
3335 PERL_ARGS_ASSERT_SV_2PVBYTE_FLAGS;
3337 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
3339 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3340 || isGV_with_GP(sv) || SvROK(sv)) {
3341 SV *sv2 = sv_newmortal();
3342 sv_copypv_nomg(sv2,sv);
3345 sv_utf8_downgrade_nomg(sv,0);
3346 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3350 =for apidoc sv_2pvutf8
3352 Return a pointer to the UTF-8-encoded representation of the SV, and set C<*lp>
3353 to its length. May cause the SV to be upgraded to UTF-8 as a side-effect.
3355 Usually accessed via the C<SvPVutf8> macro.
3361 Perl_sv_2pvutf8_flags(pTHX_ SV *sv, STRLEN *const lp, const U32 flags)
3363 PERL_ARGS_ASSERT_SV_2PVUTF8_FLAGS;
3365 if (SvGMAGICAL(sv) && (flags & SV_GMAGIC))
3367 if (((SvREADONLY(sv) || SvFAKE(sv)) && !SvIsCOW(sv))
3368 || isGV_with_GP(sv) || SvROK(sv)) {
3369 SV *sv2 = sv_newmortal();
3370 sv_copypv_nomg(sv2,sv);
3373 sv_utf8_upgrade_nomg(sv);
3374 return lp ? SvPV_nomg(sv,*lp) : SvPV_nomg_nolen(sv);
3379 =for apidoc sv_2bool
3381 This macro is only used by C<sv_true()> or its macro equivalent, and only if
3382 the latter's argument is neither C<SvPOK>, C<SvIOK> nor C<SvNOK>.
3383 It calls C<sv_2bool_flags> with the C<SV_GMAGIC> flag.
3385 =for apidoc sv_2bool_flags
3387 This function is only used by C<sv_true()> and friends, and only if
3388 the latter's argument is neither C<SvPOK>, C<SvIOK> nor C<SvNOK>. If the flags
3389 contain C<SV_GMAGIC>, then it does an C<mg_get()> first.
3396 Perl_sv_2bool_flags(pTHX_ SV *sv, I32 flags)
3398 PERL_ARGS_ASSERT_SV_2BOOL_FLAGS;
3401 if(flags & SV_GMAGIC) SvGETMAGIC(sv);
3407 SV * const tmpsv = AMG_CALLunary(sv, bool__amg);
3408 if (tmpsv && (!SvROK(tmpsv) || (SvRV(tmpsv) != SvRV(sv)))) {
3411 if(SvGMAGICAL(sv)) {
3413 goto restart; /* call sv_2bool */
3415 /* expanded SvTRUE_common(sv, (flags = 0, goto restart)) */
3416 else if(!SvOK(sv)) {
3419 else if(SvPOK(sv)) {
3420 svb = SvPVXtrue(sv);
3422 else if((SvFLAGS(sv) & (SVf_IOK|SVf_NOK))) {
3423 svb = (SvIOK(sv) && SvIVX(sv) != 0)
3424 || (SvNOK(sv) && SvNVX(sv) != 0.0);
3428 goto restart; /* call sv_2bool_nomg */
3438 RX_WRAPLEN(sv) > 1 || (RX_WRAPLEN(sv) && *RX_WRAPPED(sv) != '0');
3440 if (SvNOK(sv) && !SvPOK(sv))
3441 return SvNVX(sv) != 0.0;
3443 return SvTRUE_common(sv, isGV_with_GP(sv) ? 1 : 0);
3447 =for apidoc sv_utf8_upgrade
3449 Converts the PV of an SV to its UTF-8-encoded form.
3450 Forces the SV to string form if it is not already.
3451 Will C<mg_get> on C<sv> if appropriate.
3452 Always sets the C<SvUTF8> flag to avoid future validity checks even
3453 if the whole string is the same in UTF-8 as not.
3454 Returns the number of bytes in the converted string
3456 This is not a general purpose byte encoding to Unicode interface:
3457 use the Encode extension for that.
3459 =for apidoc sv_utf8_upgrade_nomg
3461 Like C<sv_utf8_upgrade>, but doesn't do magic on C<sv>.
3463 =for apidoc sv_utf8_upgrade_flags
3465 Converts the PV of an SV to its UTF-8-encoded form.
3466 Forces the SV to string form if it is not already.
3467 Always sets the SvUTF8 flag to avoid future validity checks even
3468 if all the bytes are invariant in UTF-8.
3469 If C<flags> has C<SV_GMAGIC> bit set,
3470 will C<mg_get> on C<sv> if appropriate, else not.
3472 The C<SV_FORCE_UTF8_UPGRADE> flag is now ignored.
3474 Returns the number of bytes in the converted string.
3476 This is not a general purpose byte encoding to Unicode interface:
3477 use the Encode extension for that.
3479 =for apidoc sv_utf8_upgrade_flags_grow
3481 Like C<sv_utf8_upgrade_flags>, but has an additional parameter C<extra>, which is
3482 the number of unused bytes the string of C<sv> is guaranteed to have free after
3483 it upon return. This allows the caller to reserve extra space that it intends
3484 to fill, to avoid extra grows.
3486 C<sv_utf8_upgrade>, C<sv_utf8_upgrade_nomg>, and C<sv_utf8_upgrade_flags>
3487 are implemented in terms of this function.
3489 Returns the number of bytes in the converted string (not including the spares).
3493 If the routine itself changes the string, it adds a trailing C<NUL>. Such a
3494 C<NUL> isn't guaranteed due to having other routines do the work in some input
3495 cases, or if the input is already flagged as being in utf8.
3500 Perl_sv_utf8_upgrade_flags_grow(pTHX_ SV *const sv, const I32 flags, STRLEN extra)
3502 PERL_ARGS_ASSERT_SV_UTF8_UPGRADE_FLAGS_GROW;
3504 if (sv == &PL_sv_undef)
3506 if (!SvPOK_nog(sv)) {
3508 if (SvREADONLY(sv) && (SvPOKp(sv) || SvIOKp(sv) || SvNOKp(sv))) {
3509 (void) sv_2pv_flags(sv,&len, flags);
3511 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3515 (void) SvPV_force_flags(sv,len,flags & SV_GMAGIC);
3519 /* SVt_REGEXP's shouldn't be upgraded to UTF8 - they're already
3520 * compiled and individual nodes will remain non-utf8 even if the
3521 * stringified version of the pattern gets upgraded. Whether the
3522 * PVX of a REGEXP should be grown or we should just croak, I don't
3524 if (SvUTF8(sv) || isREGEXP(sv)) {
3525 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3530 S_sv_uncow(aTHX_ sv, 0);
3533 if (SvCUR(sv) == 0) {
3534 if (extra) SvGROW(sv, extra + 1); /* Make sure is room for a trailing
3536 } else { /* Assume Latin-1/EBCDIC */
3537 /* This function could be much more efficient if we
3538 * had a FLAG in SVs to signal if there are any variant
3539 * chars in the PV. Given that there isn't such a flag
3540 * make the loop as fast as possible. */
3541 U8 * s = (U8 *) SvPVX_const(sv);
3544 if (is_utf8_invariant_string_loc(s, SvCUR(sv), (const U8 **) &t)) {
3546 /* utf8 conversion not needed because all are invariants. Mark
3547 * as UTF-8 even if no variant - saves scanning loop */
3549 if (extra) SvGROW(sv, SvCUR(sv) + extra);
3553 /* Here, there is at least one variant (t points to the first one), so
3554 * the string should be converted to utf8. Everything from 's' to
3555 * 't - 1' will occupy only 1 byte each on output.
3557 * Note that the incoming SV may not have a trailing '\0', as certain
3558 * code in pp_formline can send us partially built SVs.
3560 * There are two main ways to convert. One is to create a new string
3561 * and go through the input starting from the beginning, appending each
3562 * converted value onto the new string as we go along. Going this
3563 * route, it's probably best to initially allocate enough space in the
3564 * string rather than possibly running out of space and having to
3565 * reallocate and then copy what we've done so far. Since everything
3566 * from 's' to 't - 1' is invariant, the destination can be initialized
3567 * with these using a fast memory copy. To be sure to allocate enough
3568 * space, one could use the worst case scenario, where every remaining
3569 * byte expands to two under UTF-8, or one could parse it and count
3570 * exactly how many do expand.
3572 * The other way is to unconditionally parse the remainder of the
3573 * string to figure out exactly how big the expanded string will be,
3574 * growing if needed. Then start at the end of the string and place
3575 * the character there at the end of the unfilled space in the expanded
3576 * one, working backwards until reaching 't'.
3578 * The problem with assuming the worst case scenario is that for very
3579 * long strings, we could allocate much more memory than actually
3580 * needed, which can create performance problems. If we have to parse
3581 * anyway, the second method is the winner as it may avoid an extra
3582 * copy. The code used to use the first method under some
3583 * circumstances, but now that there is faster variant counting on
3584 * ASCII platforms, the second method is used exclusively, eliminating
3585 * some code that no longer has to be maintained. */
3588 /* Count the total number of variants there are. We can start
3589 * just beyond the first one, which is known to be at 't' */
3590 const Size_t invariant_length = t - s;
3591 U8 * e = (U8 *) SvEND(sv);
3593 /* The length of the left overs, plus 1. */
3594 const Size_t remaining_length_p1 = e - t;
3596 /* We expand by 1 for the variant at 't' and one for each remaining
3597 * variant (we start looking at 't+1') */
3598 Size_t expansion = 1 + variant_under_utf8_count(t + 1, e);
3600 /* +1 = trailing NUL */
3601 Size_t need = SvCUR(sv) + expansion + extra + 1;
3604 /* Grow if needed */
3605 if (SvLEN(sv) < need) {
3606 t = invariant_length + (U8*) SvGROW(sv, need);
3607 e = t + remaining_length_p1;
3609 SvCUR_set(sv, invariant_length + remaining_length_p1 + expansion);
3611 /* Set the NUL at the end */
3612 d = (U8 *) SvEND(sv);
3615 /* Having decremented d, it points to the position to put the
3616 * very last byte of the expanded string. Go backwards through
3617 * the string, copying and expanding as we go, stopping when we
3618 * get to the part that is invariant the rest of the way down */
3622 if (NATIVE_BYTE_IS_INVARIANT(*e)) {
3625 *d-- = UTF8_EIGHT_BIT_LO(*e);
3626 *d-- = UTF8_EIGHT_BIT_HI(*e);
3631 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3632 /* Update pos. We do it at the end rather than during
3633 * the upgrade, to avoid slowing down the common case
3634 * (upgrade without pos).
3635 * pos can be stored as either bytes or characters. Since
3636 * this was previously a byte string we can just turn off
3637 * the bytes flag. */
3638 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3640 mg->mg_flags &= ~MGf_BYTES;
3642 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3643 magic_setutf8(sv,mg); /* clear UTF8 cache */
3653 =for apidoc sv_utf8_downgrade
3655 Attempts to convert the PV of an SV from characters to bytes.
3656 If the PV contains a character that cannot fit
3657 in a byte, this conversion will fail;
3658 in this case, either returns false or, if C<fail_ok> is not
3661 This is not a general purpose Unicode to byte encoding interface:
3662 use the C<Encode> extension for that.
3664 This function process get magic on C<sv>.
3666 =for apidoc sv_utf8_downgrade_nomg
3668 Like C<sv_utf8_downgrade>, but does not process get magic on C<sv>.
3670 =for apidoc sv_utf8_downgrade_flags
3672 Like C<sv_utf8_downgrade>, but with additional C<flags>.
3673 If C<flags> has C<SV_GMAGIC> bit set, processes get magic on C<sv>.
3679 Perl_sv_utf8_downgrade_flags(pTHX_ SV *const sv, const bool fail_ok, const U32 flags)
3681 PERL_ARGS_ASSERT_SV_UTF8_DOWNGRADE_FLAGS;
3683 if (SvPOKp(sv) && SvUTF8(sv)) {
3687 U32 mg_flags = flags & SV_GMAGIC;
3690 S_sv_uncow(aTHX_ sv, 0);
3692 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3694 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3695 if (mg && mg->mg_len > 0 && mg->mg_flags & MGf_BYTES) {
3696 mg->mg_len = sv_pos_b2u_flags(sv, mg->mg_len,
3697 mg_flags|SV_CONST_RETURN);
3698 mg_flags = 0; /* sv_pos_b2u does get magic */
3700 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3701 magic_setutf8(sv,mg); /* clear UTF8 cache */
3704 s = (U8 *) SvPV_flags(sv, len, mg_flags);
3706 if (!utf8_to_bytes(s, &len)) {
3711 Perl_croak(aTHX_ "Wide character in %s",
3714 Perl_croak(aTHX_ "Wide character");
3725 =for apidoc sv_utf8_encode
3727 Converts the PV of an SV to UTF-8, but then turns the C<SvUTF8>
3728 flag off so that it looks like octets again.
3734 Perl_sv_utf8_encode(pTHX_ SV *const sv)
3736 PERL_ARGS_ASSERT_SV_UTF8_ENCODE;
3738 if (SvREADONLY(sv)) {
3739 sv_force_normal_flags(sv, 0);
3741 (void) sv_utf8_upgrade(sv);
3746 =for apidoc sv_utf8_decode
3748 If the PV of the SV is an octet sequence in Perl's extended UTF-8
3749 and contains a multiple-byte character, the C<SvUTF8> flag is turned on
3750 so that it looks like a character. If the PV contains only single-byte
3751 characters, the C<SvUTF8> flag stays off.
3752 Scans PV for validity and returns FALSE if the PV is invalid UTF-8.
3758 Perl_sv_utf8_decode(pTHX_ SV *const sv)
3760 PERL_ARGS_ASSERT_SV_UTF8_DECODE;
3763 const U8 *start, *c, *first_variant;
3765 /* The octets may have got themselves encoded - get them back as
3768 if (!sv_utf8_downgrade(sv, TRUE))
3771 /* it is actually just a matter of turning the utf8 flag on, but
3772 * we want to make sure everything inside is valid utf8 first.
3774 c = start = (const U8 *) SvPVX_const(sv);
3775 if (! is_utf8_invariant_string_loc(c, SvCUR(sv), &first_variant)) {
3776 if (!is_utf8_string(first_variant, SvCUR(sv) - (first_variant -c)))
3780 if (SvTYPE(sv) >= SVt_PVMG && SvMAGIC(sv)) {
3781 /* XXX Is this dead code? XS_utf8_decode calls SvSETMAGIC
3782 after this, clearing pos. Does anything on CPAN
3784 /* adjust pos to the start of a UTF8 char sequence */
3785 MAGIC * mg = mg_find(sv, PERL_MAGIC_regex_global);
3787 I32 pos = mg->mg_len;
3789 for (c = start + pos; c > start; c--) {
3790 if (UTF8_IS_START(*c))
3793 mg->mg_len = c - start;
3796 if ((mg = mg_find(sv, PERL_MAGIC_utf8)))
3797 magic_setutf8(sv,mg); /* clear UTF8 cache */
3804 =for apidoc sv_setsv
3806 Copies the contents of the source SV C<ssv> into the destination SV
3807 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3808 function if the source SV needs to be reused. Does not handle 'set' magic on
3809 destination SV. Calls 'get' magic on source SV. Loosely speaking, it
3810 performs a copy-by-value, obliterating any previous content of the
3813 You probably want to use one of the assortment of wrappers, such as
3814 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3815 C<SvSetMagicSV_nosteal>.
3817 =for apidoc sv_setsv_flags
3819 Copies the contents of the source SV C<ssv> into the destination SV
3820 C<dsv>. The source SV may be destroyed if it is mortal, so don't use this
3821 function if the source SV needs to be reused. Does not handle 'set' magic.
3822 Loosely speaking, it performs a copy-by-value, obliterating any previous
3823 content of the destination.
3824 If the C<flags> parameter has the C<SV_GMAGIC> bit set, will C<mg_get> on
3825 C<ssv> if appropriate, else not. If the C<flags>
3826 parameter has the C<SV_NOSTEAL> bit set then the
3827 buffers of temps will not be stolen. C<sv_setsv>
3828 and C<sv_setsv_nomg> are implemented in terms of this function.
3830 You probably want to use one of the assortment of wrappers, such as
3831 C<SvSetSV>, C<SvSetSV_nosteal>, C<SvSetMagicSV> and
3832 C<SvSetMagicSV_nosteal>.
3834 This is the primary function for copying scalars, and most other
3835 copy-ish functions and macros use this underneath.
3837 =for apidoc Amnh||SV_NOSTEAL
3843 S_glob_assign_glob(pTHX_ SV *const dstr, SV *const sstr, const int dtype)
3845 I32 mro_changes = 0; /* 1 = method, 2 = isa, 3 = recursive isa */
3846 HV *old_stash = NULL;
3848 PERL_ARGS_ASSERT_GLOB_ASSIGN_GLOB;
3850 if (dtype != SVt_PVGV && !isGV_with_GP(dstr)) {
3851 const char * const name = GvNAME(sstr);
3852 const STRLEN len = GvNAMELEN(sstr);
3854 if (dtype >= SVt_PV) {
3860 SvUPGRADE(dstr, SVt_PVGV);
3861 (void)SvOK_off(dstr);
3862 isGV_with_GP_on(dstr);
3864 GvSTASH(dstr) = GvSTASH(sstr);
3866 Perl_sv_add_backref(aTHX_ MUTABLE_SV(GvSTASH(dstr)), dstr);
3867 gv_name_set(MUTABLE_GV(dstr), name, len,
3868 GV_ADD | (GvNAMEUTF8(sstr) ? SVf_UTF8 : 0 ));
3869 SvFAKE_on(dstr); /* can coerce to non-glob */
3872 if(GvGP(MUTABLE_GV(sstr))) {
3873 /* If source has method cache entry, clear it */
3875 SvREFCNT_dec(GvCV(sstr));
3876 GvCV_set(sstr, NULL);
3879 /* If source has a real method, then a method is
3882 GvCV((const GV *)sstr) && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3888 /* If dest already had a real method, that's a change as well */
3890 !mro_changes && GvGP(MUTABLE_GV(dstr)) && GvCVu((const GV *)dstr)
3891 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3896 /* We don't need to check the name of the destination if it was not a
3897 glob to begin with. */
3898 if(dtype == SVt_PVGV) {
3899 const char * const name = GvNAME((const GV *)dstr);
3900 const STRLEN len = GvNAMELEN(dstr);
3901 if(memEQs(name, len, "ISA")
3902 /* The stash may have been detached from the symbol table, so
3904 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
3908 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
3909 || (len == 1 && name[0] == ':')) {
3912 /* Set aside the old stash, so we can reset isa caches on
3914 if((old_stash = GvHV(dstr)))
3915 /* Make sure we do not lose it early. */
3916 SvREFCNT_inc_simple_void_NN(
3917 sv_2mortal((SV *)old_stash)
3922 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
3925 /* freeing dstr's GP might free sstr (e.g. *x = $x),
3926 * so temporarily protect it */
3928 SAVEFREESV(SvREFCNT_inc_simple_NN(sstr));
3929 gp_free(MUTABLE_GV(dstr));
3930 GvINTRO_off(dstr); /* one-shot flag */
3931 GvGP_set(dstr, gp_ref(GvGP(sstr)));
3934 if (SvTAINTED(sstr))
3936 if (GvIMPORTED(dstr) != GVf_IMPORTED
3937 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
3939 GvIMPORTED_on(dstr);
3942 if(mro_changes == 2) {
3943 if (GvAV((const GV *)sstr)) {
3945 SV * const sref = (SV *)GvAV((const GV *)dstr);
3946 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
3947 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
3948 AV * const ary = newAV();
3949 av_push(ary, mg->mg_obj); /* takes the refcount */
3950 mg->mg_obj = (SV *)ary;
3952 av_push((AV *)mg->mg_obj, SvREFCNT_inc_simple_NN(dstr));
3954 else sv_magic(sref, dstr, PERL_MAGIC_isa, NULL, 0);
3956 mro_isa_changed_in(GvSTASH(dstr));
3958 else if(mro_changes == 3) {
3959 HV * const stash = GvHV(dstr);
3960 if(old_stash ? (HV *)HvENAME_get(old_stash) : stash)
3966 else if(mro_changes) mro_method_changed_in(GvSTASH(dstr));
3967 if (GvIO(dstr) && dtype == SVt_PVGV) {
3968 DEBUG_o(Perl_deb(aTHX_
3969 "glob_assign_glob clearing PL_stashcache\n"));
3970 /* It's a cache. It will rebuild itself quite happily.
3971 It's a lot of effort to work out exactly which key (or keys)
3972 might be invalidated by the creation of the this file handle.
3974 hv_clear(PL_stashcache);
3980 Perl_gv_setref(pTHX_ SV *const dstr, SV *const sstr)
3982 SV * const sref = SvRV(sstr);
3984 const int intro = GvINTRO(dstr);
3987 const U32 stype = SvTYPE(sref);
3989 PERL_ARGS_ASSERT_GV_SETREF;
3992 GvINTRO_off(dstr); /* one-shot flag */
3993 GvLINE(dstr) = CopLINE(PL_curcop);
3994 GvEGV(dstr) = MUTABLE_GV(dstr);
3999 location = (SV **) &(GvGP(dstr)->gp_cv); /* XXX bypassing GvCV_set */
4000 import_flag = GVf_IMPORTED_CV;
4003 location = (SV **) &GvHV(dstr);
4004 import_flag = GVf_IMPORTED_HV;
4007 location = (SV **) &GvAV(dstr);
4008 import_flag = GVf_IMPORTED_AV;
4011 location = (SV **) &GvIOp(dstr);
4014 location = (SV **) &GvFORM(dstr);
4017 location = &GvSV(dstr);
4018 import_flag = GVf_IMPORTED_SV;
4021 if (stype == SVt_PVCV) {
4022 /*if (GvCVGEN(dstr) && (GvCV(dstr) != (const CV *)sref || GvCVGEN(dstr))) {*/
4023 if (GvCVGEN(dstr)) {
4024 SvREFCNT_dec(GvCV(dstr));
4025 GvCV_set(dstr, NULL);
4026 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4029 /* SAVEt_GVSLOT takes more room on the savestack and has more
4030 overhead in leave_scope than SAVEt_GENERIC_SV. But for CVs
4031 leave_scope needs access to the GV so it can reset method
4032 caches. We must use SAVEt_GVSLOT whenever the type is
4033 SVt_PVCV, even if the stash is anonymous, as the stash may
4034 gain a name somehow before leave_scope. */
4035 if (stype == SVt_PVCV) {
4036 /* There is no save_pushptrptrptr. Creating it for this
4037 one call site would be overkill. So inline the ss add
4041 SS_ADD_PTR(location);
4042 SS_ADD_PTR(SvREFCNT_inc(*location));
4043 SS_ADD_UV(SAVEt_GVSLOT);
4046 else SAVEGENERICSV(*location);
4049 if (stype == SVt_PVCV && (*location != sref || GvCVGEN(dstr))) {
4050 CV* const cv = MUTABLE_CV(*location);
4052 if (!GvCVGEN((const GV *)dstr) &&
4053 (CvROOT(cv) || CvXSUB(cv)) &&
4054 /* redundant check that avoids creating the extra SV
4055 most of the time: */
4056 (CvCONST(cv) || ckWARN(WARN_REDEFINE)))
4058 SV * const new_const_sv =
4059 CvCONST((const CV *)sref)
4060 ? cv_const_sv((const CV *)sref)
4062 HV * const stash = GvSTASH((const GV *)dstr);
4063 report_redefined_cv(
4066 ? Perl_newSVpvf(aTHX_
4067 "%" HEKf "::%" HEKf,
4068 HEKfARG(HvNAME_HEK(stash)),
4069 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr))))
4070 : Perl_newSVpvf(aTHX_
4072 HEKfARG(GvENAME_HEK(MUTABLE_GV(dstr))))
4075 CvCONST((const CV *)sref) ? &new_const_sv : NULL
4079 cv_ckproto_len_flags(cv, (const GV *)dstr,
4080 SvPOK(sref) ? CvPROTO(sref) : NULL,
4081 SvPOK(sref) ? CvPROTOLEN(sref) : 0,
4082 SvPOK(sref) ? SvUTF8(sref) : 0);
4084 GvCVGEN(dstr) = 0; /* Switch off cacheness. */
4085 GvASSUMECV_on(dstr);
4086 if(GvSTASH(dstr)) { /* sub foo { 1 } sub bar { 2 } *bar = \&foo */
4087 if (intro && GvREFCNT(dstr) > 1) {
4088 /* temporary remove extra savestack's ref */
4090 gv_method_changed(dstr);
4093 else gv_method_changed(dstr);
4096 *location = SvREFCNT_inc_simple_NN(sref);
4097 if (import_flag && !(GvFLAGS(dstr) & import_flag)
4098 && CopSTASH_ne(PL_curcop, GvSTASH(dstr))) {
4099 GvFLAGS(dstr) |= import_flag;
4102 if (stype == SVt_PVHV) {
4103 const char * const name = GvNAME((GV*)dstr);
4104 const STRLEN len = GvNAMELEN(dstr);
4107 (len > 1 && name[len-2] == ':' && name[len-1] == ':')
4108 || (len == 1 && name[0] == ':')
4110 && (!dref || HvENAME_get(dref))
4113 (HV *)sref, (HV *)dref,
4119 stype == SVt_PVAV && sref != dref
4120 && memEQs(GvNAME((GV*)dstr), GvNAMELEN((GV*)dstr), "ISA")
4121 /* The stash may have been detached from the symbol table, so
4122 check its name before doing anything. */
4123 && GvSTASH(dstr) && HvENAME(GvSTASH(dstr))
4126 MAGIC * const omg = dref && SvSMAGICAL(dref)
4127 ? mg_find(dref, PERL_MAGIC_isa)
4129 if (SvSMAGICAL(sref) && (mg = mg_find(sref, PERL_MAGIC_isa))) {
4130 if (SvTYPE(mg->mg_obj) != SVt_PVAV) {
4131 AV * const ary = newAV();
4132 av_push(ary, mg->mg_obj); /* takes the refcount */
4133 mg->mg_obj = (SV *)ary;
4136 if (SvTYPE(omg->mg_obj) == SVt_PVAV) {
4137 SV **svp = AvARRAY((AV *)omg->mg_obj);
4138 I32 items = AvFILLp((AV *)omg->mg_obj) + 1;
4142 SvREFCNT_inc_simple_NN(*svp++)
4148 SvREFCNT_inc_simple_NN(omg->mg_obj)
4152 av_push((AV *)mg->mg_obj,SvREFCNT_inc_simple_NN(dstr));
4158 sref, omg ? omg->mg_obj : dstr, PERL_MAGIC_isa, NULL, 0
4160 for (i = 0; i <= AvFILL(sref); ++i) {
4161 SV **elem = av_fetch ((AV*)sref, i, 0);
4164 *elem, sref, PERL_MAGIC_isaelem, NULL, i
4168 mg = mg_find(sref, PERL_MAGIC_isa);
4170 /* Since the *ISA assignment could have affected more than
4171 one stash, don't call mro_isa_changed_in directly, but let
4172 magic_clearisa do it for us, as it already has the logic for
4173 dealing with globs vs arrays of globs. */
4175 Perl_magic_clearisa(aTHX_ NULL, mg);
4177 else if (stype == SVt_PVIO) {
4178 DEBUG_o(Perl_deb(aTHX_ "gv_setref clearing PL_stashcache\n"));
4179 /* It's a cache. It will rebuild itself quite happily.
4180 It's a lot of effort to work out exactly which key (or keys)
4181 might be invalidated by the creation of the this file handle.
4183 hv_clear(PL_stashcache);
4187 if (!intro) SvREFCNT_dec(dref);
4188 if (SvTAINTED(sstr))
4196 #ifdef PERL_DEBUG_READONLY_COW
4197 # include <sys/mman.h>
4199 # ifndef PERL_MEMORY_DEBUG_HEADER_SIZE
4200 # define PERL_MEMORY_DEBUG_HEADER_SIZE 0
4204 Perl_sv_buf_to_ro(pTHX_ SV *sv)
4206 struct perl_memory_debug_header * const header =
4207 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4208 const MEM_SIZE len = header->size;
4209 PERL_ARGS_ASSERT_SV_BUF_TO_RO;
4210 # ifdef PERL_TRACK_MEMPOOL
4211 if (!header->readonly) header->readonly = 1;
4213 if (mprotect(header, len, PROT_READ))
4214 Perl_warn(aTHX_ "mprotect RW for COW string %p %lu failed with %d",
4215 header, len, errno);
4219 S_sv_buf_to_rw(pTHX_ SV *sv)
4221 struct perl_memory_debug_header * const header =
4222 (struct perl_memory_debug_header *)(SvPVX(sv)-PERL_MEMORY_DEBUG_HEADER_SIZE);
4223 const MEM_SIZE len = header->size;
4224 PERL_ARGS_ASSERT_SV_BUF_TO_RW;
4225 if (mprotect(header, len, PROT_READ|PROT_WRITE))
4226 Perl_warn(aTHX_ "mprotect for COW string %p %lu failed with %d",
4227 header, len, errno);
4228 # ifdef PERL_TRACK_MEMPOOL
4229 header->readonly = 0;
4234 # define sv_buf_to_ro(sv) NOOP
4235 # define sv_buf_to_rw(sv) NOOP
4239 Perl_sv_setsv_flags(pTHX_ SV *dstr, SV* sstr, const I32 flags)
4244 unsigned int both_type;
4246 PERL_ARGS_ASSERT_SV_SETSV_FLAGS;
4248 if (UNLIKELY( sstr == dstr ))
4251 if (UNLIKELY( !sstr ))
4252 sstr = &PL_sv_undef;
4254 stype = SvTYPE(sstr);
4255 dtype = SvTYPE(dstr);
4256 both_type = (stype | dtype);
4258 /* with these values, we can check that both SVs are NULL/IV (and not
4259 * freed) just by testing the or'ed types */
4260 STATIC_ASSERT_STMT(SVt_NULL == 0);
4261 STATIC_ASSERT_STMT(SVt_IV == 1);
4262 if (both_type <= 1) {
4263 /* both src and dst are UNDEF/IV/RV, so we can do a lot of
4269 /* minimal subset of SV_CHECK_THINKFIRST_COW_DROP(dstr) */
4270 if (SvREADONLY(dstr))
4271 Perl_croak_no_modify();
4273 if (SvWEAKREF(dstr))
4274 sv_unref_flags(dstr, 0);
4276 old_rv = SvRV(dstr);
4279 assert(!SvGMAGICAL(sstr));
4280 assert(!SvGMAGICAL(dstr));
4282 sflags = SvFLAGS(sstr);
4283 if (sflags & (SVf_IOK|SVf_ROK)) {
4284 SET_SVANY_FOR_BODYLESS_IV(dstr);
4285 new_dflags = SVt_IV;
4287 if (sflags & SVf_ROK) {
4288 dstr->sv_u.svu_rv = SvREFCNT_inc(SvRV(sstr));
4289 new_dflags |= SVf_ROK;
4292 /* both src and dst are <= SVt_IV, so sv_any points to the
4293 * head; so access the head directly
4295 assert( &(sstr->sv_u.svu_iv)
4296 == &(((XPVIV*) SvANY(sstr))->xiv_iv));
4297 assert( &(dstr->sv_u.svu_iv)
4298 == &(((XPVIV*) SvANY(dstr))->xiv_iv));
4299 dstr->sv_u.svu_iv = sstr->sv_u.svu_iv;
4300 new_dflags |= (SVf_IOK|SVp_IOK|(sflags & SVf_IVisUV));
4304 new_dflags = dtype; /* turn off everything except the type */
4306 SvFLAGS(dstr) = new_dflags;
4307 SvREFCNT_dec(old_rv);
4312 if (UNLIKELY(both_type == SVTYPEMASK)) {
4313 if (SvIS_FREED(dstr)) {
4314 Perl_croak(aTHX_ "panic: attempt to copy value %" SVf
4315 " to a freed scalar %p", SVfARG(sstr), (void *)dstr);
4317 if (SvIS_FREED(sstr)) {
4318 Perl_croak(aTHX_ "panic: attempt to copy freed scalar %p to %p",
4319 (void*)sstr, (void*)dstr);
4325 SV_CHECK_THINKFIRST_COW_DROP(dstr);
4326 dtype = SvTYPE(dstr); /* THINKFIRST may have changed type */
4328 /* There's a lot of redundancy below but we're going for speed here */
4333 if (LIKELY( dtype != SVt_PVGV && dtype != SVt_PVLV )) {
4334 (void)SvOK_off(dstr);
4342 /* For performance, we inline promoting to type SVt_IV. */
4343 /* We're starting from SVt_NULL, so provided that define is
4344 * actual 0, we don't have to unset any SV type flags
4345 * to promote to SVt_IV. */
4346 STATIC_ASSERT_STMT(SVt_NULL == 0);
4347 SET_SVANY_FOR_BODYLESS_IV(dstr);
4348 SvFLAGS(dstr) |= SVt_IV;
4352 sv_upgrade(dstr, SVt_PVIV);
4356 goto end_of_first_switch;
4358 (void)SvIOK_only(dstr);
4359 SvIV_set(dstr, SvIVX(sstr));
4362 /* SvTAINTED can only be true if the SV has taint magic, which in
4363 turn means that the SV type is PVMG (or greater). This is the
4364 case statement for SVt_IV, so this cannot be true (whatever gcov
4366 assert(!SvTAINTED(sstr));
4371 if (dtype < SVt_PV && dtype != SVt_IV)
4372 sv_upgrade(dstr, SVt_IV);
4376 if (LIKELY( SvNOK(sstr) )) {
4380 sv_upgrade(dstr, SVt_NV);
4384 sv_upgrade(dstr, SVt_PVNV);
4388 goto end_of_first_switch;
4390 SvNV_set(dstr, SvNVX(sstr));
4391 (void)SvNOK_only(dstr);
4392 /* SvTAINTED can only be true if the SV has taint magic, which in
4393 turn means that the SV type is PVMG (or greater). This is the
4394 case statement for SVt_NV, so this cannot be true (whatever gcov
4396 assert(!SvTAINTED(sstr));
4403 sv_upgrade(dstr, SVt_PV);
4406 if (dtype < SVt_PVIV)
4407 sv_upgrade(dstr, SVt_PVIV);
4410 if (dtype < SVt_PVNV)
4411 sv_upgrade(dstr, SVt_PVNV);
4415 invlist_clone(sstr, dstr);
4419 const char * const type = sv_reftype(sstr,0);
4421 /* diag_listed_as: Bizarre copy of %s */
4422 Perl_croak(aTHX_ "Bizarre copy of %s in %s", type, OP_DESC(PL_op));
4424 Perl_croak(aTHX_ "Bizarre copy of %s", type);
4426 NOT_REACHED; /* NOTREACHED */
4430 if (dtype < SVt_REGEXP)
4431 sv_upgrade(dstr, SVt_REGEXP);
4437 if (SvGMAGICAL(sstr) && (flags & SV_GMAGIC)) {
4439 if (SvTYPE(sstr) != stype)
4440 stype = SvTYPE(sstr);
4442 if (isGV_with_GP(sstr) && dtype <= SVt_PVLV) {
4443 glob_assign_glob(dstr, sstr, dtype);
4446 if (stype == SVt_PVLV)
4448 if (isREGEXP(sstr)) goto upgregexp;
4449 SvUPGRADE(dstr, SVt_PVNV);
4452 SvUPGRADE(dstr, (svtype)stype);
4454 end_of_first_switch:
4456 /* dstr may have been upgraded. */
4457 dtype = SvTYPE(dstr);
4458 sflags = SvFLAGS(sstr);
4460 if (UNLIKELY( dtype == SVt_PVCV )) {
4461 /* Assigning to a subroutine sets the prototype. */
4464 const char *const ptr = SvPV_const(sstr, len);
4466 SvGROW(dstr, len + 1);
4467 Copy(ptr, SvPVX(dstr), len + 1, char);
4468 SvCUR_set(dstr, len);
4470 SvFLAGS(dstr) |= sflags & SVf_UTF8;
4471 CvAUTOLOAD_off(dstr);
4476 else if (UNLIKELY(dtype == SVt_PVAV || dtype == SVt_PVHV
4477 || dtype == SVt_PVFM))
4479 const char * const type = sv_reftype(dstr,0);
4481 /* diag_listed_as: Cannot copy to %s */
4482 Perl_croak(aTHX_ "Cannot copy to %s in %s", type, OP_DESC(PL_op));
4484 Perl_croak(aTHX_ "Cannot copy to %s", type);
4485 } else if (sflags & SVf_ROK) {
4486 if (isGV_with_GP(dstr)
4487 && SvTYPE(SvRV(sstr)) == SVt_PVGV && isGV_with_GP(SvRV(sstr))) {
4490 if (GvIMPORTED(dstr) != GVf_IMPORTED
4491 && CopSTASH_ne(PL_curcop, GvSTASH(dstr)))
4493 GvIMPORTED_on(dstr);
4498 glob_assign_glob(dstr, sstr, dtype);
4502 if (dtype >= SVt_PV) {
4503 if (isGV_with_GP(dstr)) {
4504 gv_setref(dstr, sstr);
4507 if (SvPVX_const(dstr)) {
4513 (void)SvOK_off(dstr);
4514 SvRV_set(dstr, SvREFCNT_inc(SvRV(sstr)));
4515 SvFLAGS(dstr) |= sflags & SVf_ROK;
4516 assert(!(sflags & SVp_NOK));
4517 assert(!(sflags & SVp_IOK));
4518 assert(!(sflags & SVf_NOK));
4519 assert(!(sflags & SVf_IOK));
4521 else if (isGV_with_GP(dstr)) {
4522 if (!(sflags & SVf_OK)) {
4523 Perl_ck_warner(aTHX_ packWARN(WARN_MISC),
4524 "Undefined value assigned to typeglob");
4527 GV *gv = gv_fetchsv_nomg(sstr, GV_ADD, SVt_PVGV);
4528 if (dstr != (const SV *)gv) {
4529 const char * const name = GvNAME((const GV *)dstr);
4530 const STRLEN len = GvNAMELEN(dstr);
4531 HV *old_stash = NULL;
4532 bool reset_isa = FALSE;
4533 if ((len > 1 && name[len-2] == ':' && name[len-1] == ':')
4534 || (len == 1 && name[0] == ':')) {
4535 /* Set aside the old stash, so we can reset isa caches
4536 on its subclasses. */
4537 if((old_stash = GvHV(dstr))) {
4538 /* Make sure we do not lose it early. */
4539 SvREFCNT_inc_simple_void_NN(
4540 sv_2mortal((SV *)old_stash)
4547 SvREFCNT_inc_simple_void_NN(sv_2mortal(dstr));
4548 gp_free(MUTABLE_GV(dstr));
4550 GvGP_set(dstr, gp_ref(GvGP(gv)));
4553 HV * const stash = GvHV(dstr);
4555 old_stash ? (HV *)HvENAME_get(old_stash) : stash
4565 else if ((dtype == SVt_REGEXP || dtype == SVt_PVLV)
4566 && (stype == SVt_REGEXP || isREGEXP(sstr))) {
4567 reg_temp_copy((REGEXP*)dstr, (REGEXP*)sstr);
4569 else if (sflags & SVp_POK) {
4570 const STRLEN cur = SvCUR(sstr);
4571 const STRLEN len = SvLEN(sstr);
4574 * We have three basic ways to copy the string:
4580 * Which we choose is based on various factors. The following
4581 * things are listed in order of speed, fastest to slowest:
4583 * - Copying a short string
4584 * - Copy-on-write bookkeeping
4586 * - Copying a long string
4588 * We swipe the string (steal the string buffer) if the SV on the
4589 * rhs is about to be freed anyway (TEMP and refcnt==1). This is a
4590 * big win on long strings. It should be a win on short strings if
4591 * SvPVX_const(dstr) has to be allocated. If not, it should not
4592 * slow things down, as SvPVX_const(sstr) would have been freed
4595 * We also steal the buffer from a PADTMP (operator target) if it
4596 * is ‘long enough’. For short strings, a swipe does not help
4597 * here, as it causes more malloc calls the next time the target
4598 * is used. Benchmarks show that even if SvPVX_const(dstr) has to
4599 * be allocated it is still not worth swiping PADTMPs for short
4600 * strings, as the savings here are small.
4602 * If swiping is not an option, then we see whether it is
4603 * worth using copy-on-write. If the lhs already has a buf-
4604 * fer big enough and the string is short, we skip it and fall back
4605 * to method 3, since memcpy is faster for short strings than the
4606 * later bookkeeping overhead that copy-on-write entails.
4608 * If the rhs is not a copy-on-write string yet, then we also
4609 * consider whether the buffer is too large relative to the string
4610 * it holds. Some operations such as readline allocate a large
4611 * buffer in the expectation of reusing it. But turning such into
4612 * a COW buffer is counter-productive because it increases memory
4613 * usage by making readline allocate a new large buffer the sec-
4614 * ond time round. So, if the buffer is too large, again, we use
4617 * Finally, if there is no buffer on the left, or the buffer is too
4618 * small, then we use copy-on-write and make both SVs share the
4623 /* Whichever path we take through the next code, we want this true,
4624 and doing it now facilitates the COW check. */
4625 (void)SvPOK_only(dstr);
4629 /* slated for free anyway (and not COW)? */
4630 (sflags & (SVs_TEMP|SVf_IsCOW)) == SVs_TEMP
4631 /* or a swipable TARG */
4633 (SVs_PADTMP|SVf_READONLY|SVf_PROTECT|SVf_IsCOW))
4635 /* whose buffer is worth stealing */
4636 && CHECK_COWBUF_THRESHOLD(cur,len)
4639 !(sflags & SVf_OOK) && /* and not involved in OOK hack? */
4640 (!(flags & SV_NOSTEAL)) &&
4641 /* and we're allowed to steal temps */
4642 SvREFCNT(sstr) == 1 && /* and no other references to it? */
4643 len) /* and really is a string */
4644 { /* Passes the swipe test. */
4645 if (SvPVX_const(dstr)) /* we