1 #define PERL_NO_GET_CONTEXT
8 /* Not yet in ppport.h */
10 # define CvISXSUB(cv) (CvXSUB(cv) ? TRUE : FALSE)
13 # define SvRV_const(rv) SvRV(rv)
17 /* "structured exception" handling is a Microsoft extension to C and C++.
18 It's *not* C++ exception handling - C++ exception handling can't capture
19 SEGVs and suchlike, whereas this can. There's no known analagous
20 functionality on other platforms. */
22 # define TRY_TO_CATCH_SEGV __try
23 # define CAUGHT_EXCEPTION __except(EXCEPTION EXCEPTION_EXECUTE_HANDLER)
25 # define TRY_TO_CATCH_SEGV if(1)
26 # define CAUGHT_EXCEPTION else
30 # define __attribute__(x)
33 #if 0 && defined(DEBUGGING)
34 #define dbg_printf(x) printf x
39 #define TAG /* printf( "# %s(%d)\n", __FILE__, __LINE__ ) */
42 /* The idea is to have a tree structure to store 1 bit per possible pointer
43 address. The lowest 16 bits are stored in a block of 8092 bytes.
44 The blocks are in a 256-way tree, indexed by the reset of the pointer.
45 This can cope with 32 and 64 bit pointers, and any address space layout,
46 without excessive memory needs. The assumption is that your CPU cache
47 works :-) (And that we're not going to bust it) */
50 #define LEAF_BITS (16 - BYTE_BITS)
51 #define LEAF_MASK 0x1FFF
59 /* My hunch (not measured) is that for most architectures pointers will
60 start with 0 bits, hence the start of this array will be hot, and the
61 end unused. So put the flags next to the hot end. */
66 Checks to see if thing is in the bitstring.
67 Returns true or false, and
68 notes thing in the segmented bitstring.
71 check_new(struct state *st, const void *const p) {
72 unsigned int bits = 8 * sizeof(void*);
73 const size_t raw_p = PTR2nat(p);
74 /* This effectively rotates the value right by the number of low always-0
75 bits in an aligned pointer. The assmption is that most (if not all)
76 pointers are aligned, and these will be in the same chain of nodes
77 (and hence hot in the cache) but we can still deal with any unaligned
80 = (raw_p >> ALIGN_BITS) | (raw_p << (bits - ALIGN_BITS));
81 const U8 this_bit = 1 << (cooked_p & 0x7);
85 void **tv_p = (void **) (st->tracking);
87 if (NULL == p) return FALSE;
89 const char c = *(const char *)p;
93 warn( "Devel::Size: Encountered invalid pointer: %p\n", p );
99 /* bits now 24 (32 bit pointers) or 56 (64 bit pointers) */
101 /* First level is always present. */
103 i = (unsigned int)((cooked_p >> bits) & 0xFF);
105 Newxz(tv_p[i], 256, void *);
106 tv_p = (void **)(tv_p[i]);
108 } while (bits > LEAF_BITS + BYTE_BITS);
109 /* bits now 16 always */
110 #if !defined(MULTIPLICITY) || PERL_VERSION > 8 || (PERL_VERSION == 8 && PERL_SUBVERSION > 8)
111 /* 5.8.8 and early have an assert() macro that uses Perl_croak, hence needs
112 a my_perl under multiplicity */
115 leaf_p = (U8 **)tv_p;
116 i = (unsigned int)((cooked_p >> bits) & 0xFF);
118 Newxz(leaf_p[i], 1 << LEAF_BITS, U8);
123 i = (unsigned int)((cooked_p >> BYTE_BITS) & LEAF_MASK);
125 if(leaf[i] & this_bit)
133 free_tracking_at(void **tv, int level)
141 free_tracking_at(tv[i], level);
155 free_state(struct state *st)
157 const int top_level = (sizeof(void *) * 8 - LEAF_BITS - BYTE_BITS) / 8;
158 free_tracking_at((void **)st->tracking, top_level);
162 /* For now, this is somewhat a compatibility bodge until the plan comes
163 together for fine grained recursion control. total_size() would recurse into
164 hash and array members, whereas sv_size() would not. However, sv_size() is
165 called with CvSTASH() of a CV, which means that if it (also) starts to
166 recurse fully, then the size of any CV now becomes the size of the entire
167 symbol table reachable from it, and potentially the entire symbol table, if
168 any subroutine makes a reference to a global (such as %SIG). The historical
169 implementation of total_size() didn't report "everything", and changing the
170 only available size to "everything" doesn't feel at all useful. */
172 #define NO_RECURSION 0
173 #define SOME_RECURSION 1
174 #define TOTAL_SIZE_RECURSION 2
176 static bool sv_size(pTHX_ struct state *, const SV *const, const int recurse);
194 cc_opclass(const OP * const o)
200 return (o->op_flags & OPf_KIDS) ? OPc_UNOP : OPc_BASEOP;
202 if (o->op_type == OP_SASSIGN)
203 return ((o->op_private & OPpASSIGN_BACKWARDS) ? OPc_UNOP : OPc_BINOP);
206 if (o->op_type == OP_GV || o->op_type == OP_GVSV || o->op_type == OP_AELEMFAST)
210 if ((o->op_type == OP_TRANS)) {
214 switch (PL_opargs[o->op_type] & OA_CLASS_MASK) {
239 case OA_PVOP_OR_SVOP: TAG;
241 * Character translations (tr///) are usually a PVOP, keeping a
242 * pointer to a table of shorts used to look up translations.
243 * Under utf8, however, a simple table isn't practical; instead,
244 * the OP is an SVOP, and the SV is a reference to a swash
245 * (i.e., an RV pointing to an HV).
247 return (o->op_private & (OPpTRANS_TO_UTF|OPpTRANS_FROM_UTF))
248 ? OPc_SVOP : OPc_PVOP;
256 case OA_BASEOP_OR_UNOP: TAG;
258 * UNI(OP_foo) in toke.c returns token UNI or FUNC1 depending on
259 * whether parens were seen. perly.y uses OPf_SPECIAL to
260 * signal whether a BASEOP had empty parens or none.
261 * Some other UNOPs are created later, though, so the best
262 * test is OPf_KIDS, which is set in newUNOP.
264 return (o->op_flags & OPf_KIDS) ? OPc_UNOP : OPc_BASEOP;
266 case OA_FILESTATOP: TAG;
268 * The file stat OPs are created via UNI(OP_foo) in toke.c but use
269 * the OPf_REF flag to distinguish between OP types instead of the
270 * usual OPf_SPECIAL flag. As usual, if OPf_KIDS is set, then we
271 * return OPc_UNOP so that walkoptree can find our children. If
272 * OPf_KIDS is not set then we check OPf_REF. Without OPf_REF set
273 * (no argument to the operator) it's an OP; with OPf_REF set it's
274 * an SVOP (and op_sv is the GV for the filehandle argument).
276 return ((o->op_flags & OPf_KIDS) ? OPc_UNOP :
278 (o->op_flags & OPf_REF) ? OPc_PADOP : OPc_BASEOP);
280 (o->op_flags & OPf_REF) ? OPc_SVOP : OPc_BASEOP);
282 case OA_LOOPEXOP: TAG;
284 * next, last, redo, dump and goto use OPf_SPECIAL to indicate that a
285 * label was omitted (in which case it's a BASEOP) or else a term was
286 * seen. In this last case, all except goto are definitely PVOP but
287 * goto is either a PVOP (with an ordinary constant label), an UNOP
288 * with OPf_STACKED (with a non-constant non-sub) or an UNOP for
289 * OP_REFGEN (with goto &sub) in which case OPf_STACKED also seems to
292 if (o->op_flags & OPf_STACKED)
294 else if (o->op_flags & OPf_SPECIAL)
299 warn("Devel::Size: Can't determine class of operator %s, assuming BASEOP\n",
300 PL_op_name[o->op_type]);
311 /* Figure out how much magic is attached to the SV and return the
314 magic_size(const SV * const thing, struct state *st) {
315 MAGIC *magic_pointer;
318 if (!SvMAGIC(thing)) {
323 /* Get the base magic pointer */
324 magic_pointer = SvMAGIC(thing);
326 /* Have we seen the magic pointer? */
327 while (check_new(st, magic_pointer)) {
328 st->total_size += sizeof(MAGIC);
331 /* Have we seen the magic vtable? */
332 if (check_new(st, magic_pointer->mg_virtual)) {
333 st->total_size += sizeof(MGVTBL);
336 /* Get the next in the chain */
337 magic_pointer = magic_pointer->mg_moremagic;
340 if (st->dangle_whine)
341 warn( "Devel::Size: Encountered bad magic at: %p\n", magic_pointer );
347 check_new_and_strlen(struct state *st, const char *const p) {
349 st->total_size += 1 + strlen(p);
353 regex_size(const REGEXP * const baseregex, struct state *st) {
354 if(!check_new(st, baseregex))
356 st->total_size += sizeof(REGEXP);
357 #if (PERL_VERSION < 11)
358 /* Note the size of the paren offset thing */
359 st->total_size += sizeof(I32) * baseregex->nparens * 2;
360 st->total_size += strlen(baseregex->precomp);
362 st->total_size += sizeof(struct regexp);
363 st->total_size += sizeof(I32) * SvANY(baseregex)->nparens * 2;
364 /*st->total_size += strlen(SvANY(baseregex)->subbeg);*/
366 if (st->go_yell && !st->regex_whine) {
367 carp("Devel::Size: Calculated sizes for compiled regexes are incompatible, and probably always will be");
373 op_size(pTHX_ const OP * const baseop, struct state *st)
377 if(!check_new(st, baseop))
380 op_size(aTHX_ baseop->op_next, st);
382 switch (cc_opclass(baseop)) {
383 case OPc_BASEOP: TAG;
384 st->total_size += sizeof(struct op);
387 st->total_size += sizeof(struct unop);
388 op_size(aTHX_ cUNOPx(baseop)->op_first, st);
391 st->total_size += sizeof(struct binop);
392 op_size(aTHX_ cBINOPx(baseop)->op_first, st);
393 op_size(aTHX_ cBINOPx(baseop)->op_last, st);
396 st->total_size += sizeof(struct logop);
397 op_size(aTHX_ cBINOPx(baseop)->op_first, st);
398 op_size(aTHX_ cLOGOPx(baseop)->op_other, st);
400 case OPc_LISTOP: TAG;
401 st->total_size += sizeof(struct listop);
402 op_size(aTHX_ cLISTOPx(baseop)->op_first, st);
403 op_size(aTHX_ cLISTOPx(baseop)->op_last, st);
406 st->total_size += sizeof(struct pmop);
407 op_size(aTHX_ cPMOPx(baseop)->op_first, st);
408 op_size(aTHX_ cPMOPx(baseop)->op_last, st);
409 #if PERL_VERSION < 9 || (PERL_VERSION == 9 && PERL_SUBVERSION < 5)
410 op_size(aTHX_ cPMOPx(baseop)->op_pmreplroot, st);
411 op_size(aTHX_ cPMOPx(baseop)->op_pmreplstart, st);
412 op_size(aTHX_ (OP *)cPMOPx(baseop)->op_pmnext, st);
414 /* This is defined away in perl 5.8.x, but it is in there for
417 regex_size(PM_GETRE(cPMOPx(baseop)), st);
419 regex_size(cPMOPx(baseop)->op_pmregexp, st);
423 st->total_size += sizeof(struct pmop);
424 if (!(baseop->op_type == OP_AELEMFAST
425 && baseop->op_flags & OPf_SPECIAL)) {
426 /* not an OP_PADAV replacement */
427 sv_size(aTHX_ st, cSVOPx(baseop)->op_sv, SOME_RECURSION);
431 st->total_size += sizeof(struct padop);
434 check_new_and_strlen(st, cPVOPx(baseop)->op_pv);
437 st->total_size += sizeof(struct loop);
438 op_size(aTHX_ cLOOPx(baseop)->op_first, st);
439 op_size(aTHX_ cLOOPx(baseop)->op_last, st);
440 op_size(aTHX_ cLOOPx(baseop)->op_redoop, st);
441 op_size(aTHX_ cLOOPx(baseop)->op_nextop, st);
442 op_size(aTHX_ cLOOPx(baseop)->op_lastop, st);
447 basecop = (COP *)baseop;
448 st->total_size += sizeof(struct cop);
450 /* Change 33656 by nicholas@mouse-mill on 2008/04/07 11:29:51
451 Eliminate cop_label from struct cop by storing a label as the first
452 entry in the hints hash. Most statements don't have labels, so this
453 will save memory. Not sure how much.
454 The check below will be incorrect fail on bleadperls
455 before 5.11 @33656, but later than 5.10, producing slightly too
456 small memory sizes on these Perls. */
457 #if (PERL_VERSION < 11)
458 check_new_and_strlen(st, basecop->cop_label);
461 check_new_and_strlen(st, basecop->cop_file);
462 check_new_and_strlen(st, basecop->cop_stashpv);
464 sv_size(aTHX_ st, (SV *)basecop->cop_stash, SOME_RECURSION);
465 sv_size(aTHX_ st, (SV *)basecop->cop_filegv, SOME_RECURSION);
475 if (st->dangle_whine)
476 warn( "Devel::Size: Encountered dangling pointer in opcode at: %p\n", baseop );
480 #if PERL_VERSION > 9 || (PERL_VERSION == 9 && PERL_SUBVERSION > 2)
481 # define NEW_HEAD_LAYOUT
485 sv_size(pTHX_ struct state *const st, const SV * const orig_thing,
487 const SV *thing = orig_thing;
489 if(!check_new(st, thing))
492 st->total_size += sizeof(SV);
494 switch (SvTYPE(thing)) {
498 /* Just a plain integer. This will be differently sized depending
499 on whether purify's been compiled in */
501 #ifndef NEW_HEAD_LAYOUT
503 st->total_size += sizeof(sizeof(XPVIV));
505 st->total_size += sizeof(IV);
508 if(recurse && SvROK(thing))
509 sv_size(aTHX_ st, SvRV_const(thing), recurse);
511 /* Is it a float? Like the int, it depends on purify */
514 st->total_size += sizeof(sizeof(XPVNV));
516 st->total_size += sizeof(NV);
519 #if (PERL_VERSION < 11)
520 /* Is it a reference? */
522 #ifndef NEW_HEAD_LAYOUT
523 st->total_size += sizeof(XRV);
525 if(recurse && SvROK(thing))
526 sv_size(aTHX_ st, SvRV_const(thing), recurse);
529 /* How about a plain string? In which case we need to add in how
530 much has been allocated */
532 st->total_size += sizeof(XPV);
533 if(recurse && SvROK(thing))
534 sv_size(aTHX_ st, SvRV_const(thing), recurse);
536 st->total_size += SvLEN(thing);
538 /* A string with an integer part? */
540 st->total_size += sizeof(XPVIV);
541 if(recurse && SvROK(thing))
542 sv_size(aTHX_ st, SvRV_const(thing), recurse);
544 st->total_size += SvLEN(thing);
546 st->total_size += SvIVX(thing);
549 /* A scalar/string/reference with a float part? */
551 st->total_size += sizeof(XPVNV);
552 if(recurse && SvROK(thing))
553 sv_size(aTHX_ st, SvRV_const(thing), recurse);
555 st->total_size += SvLEN(thing);
558 st->total_size += sizeof(XPVMG);
559 if(recurse && SvROK(thing))
560 sv_size(aTHX_ st, SvRV_const(thing), recurse);
562 st->total_size += SvLEN(thing);
563 magic_size(thing, st);
565 #if PERL_VERSION <= 8
567 st->total_size += sizeof(XPVBM);
568 if(recurse && SvROK(thing))
569 sv_size(aTHX_ st, SvRV_const(thing), recurse);
571 st->total_size += SvLEN(thing);
572 magic_size(thing, st);
576 st->total_size += sizeof(XPVLV);
577 if(recurse && SvROK(thing))
578 sv_size(aTHX_ st, SvRV_const(thing), recurse);
580 st->total_size += SvLEN(thing);
581 magic_size(thing, st);
583 /* How much space is dedicated to the array? Not counting the
584 elements in the array, mind, just the array itself */
586 st->total_size += sizeof(XPVAV);
587 /* Is there anything in the array? */
588 if (AvMAX(thing) != -1) {
589 /* an array with 10 slots has AvMax() set to 9 - te 2007-04-22 */
590 st->total_size += sizeof(SV *) * (AvMAX(thing) + 1);
591 dbg_printf(("total_size: %li AvMAX: %li av_len: $i\n", st->total_size, AvMAX(thing), av_len((AV*)thing)));
593 if (recurse >= TOTAL_SIZE_RECURSION) {
594 SSize_t i = AvFILLp(thing) + 1;
597 sv_size(aTHX_ st, AvARRAY(thing)[i], recurse);
600 /* Add in the bits on the other side of the beginning */
602 dbg_printf(("total_size %li, sizeof(SV *) %li, AvARRAY(thing) %li, AvALLOC(thing)%li , sizeof(ptr) %li \n",
603 st->total_size, sizeof(SV*), AvARRAY(thing), AvALLOC(thing), sizeof( thing )));
605 /* under Perl 5.8.8 64bit threading, AvARRAY(thing) was a pointer while AvALLOC was 0,
606 resulting in grossly overstated sized for arrays. Technically, this shouldn't happen... */
607 if (AvALLOC(thing) != 0) {
608 st->total_size += (sizeof(SV *) * (AvARRAY(thing) - AvALLOC(thing)));
610 #if (PERL_VERSION < 9)
611 /* Is there something hanging off the arylen element?
612 Post 5.9.something this is stored in magic, so will be found there,
613 and Perl_av_arylen_p() takes a non-const AV*, hence compilers rightly
614 complain about AvARYLEN() passing thing to it. */
615 sv_size(aTHX_ st, AvARYLEN(thing), recurse);
617 magic_size(thing, st);
620 /* First the base struct */
621 st->total_size += sizeof(XPVHV);
622 /* Now the array of buckets */
623 st->total_size += (sizeof(HE *) * (HvMAX(thing) + 1));
624 /* Now walk the bucket chain */
625 if (HvARRAY(thing)) {
628 for (cur_bucket = 0; cur_bucket <= HvMAX(thing); cur_bucket++) {
629 cur_entry = *(HvARRAY(thing) + cur_bucket);
631 st->total_size += sizeof(HE);
632 if (cur_entry->hent_hek) {
633 /* Hash keys can be shared. Have we seen this before? */
634 if (check_new(st, cur_entry->hent_hek)) {
635 st->total_size += HEK_BASESIZE + cur_entry->hent_hek->hek_len + 2;
638 if (recurse >= TOTAL_SIZE_RECURSION)
639 sv_size(aTHX_ st, HeVAL(cur_entry), recurse);
640 cur_entry = cur_entry->hent_next;
644 magic_size(thing, st);
647 st->total_size += sizeof(XPVCV);
648 magic_size(thing, st);
650 st->total_size += ((XPVIO *) SvANY(thing))->xpv_len;
651 sv_size(aTHX_ st, (SV *)CvSTASH(thing), SOME_RECURSION);
652 sv_size(aTHX_ st, (SV *)SvSTASH(thing), SOME_RECURSION);
653 sv_size(aTHX_ st, (SV *)CvGV(thing), SOME_RECURSION);
654 sv_size(aTHX_ st, (SV *)CvPADLIST(thing), SOME_RECURSION);
655 sv_size(aTHX_ st, (SV *)CvOUTSIDE(thing), recurse);
656 if (CvISXSUB(thing)) {
657 sv_size(aTHX_ st, cv_const_sv((CV *)thing), recurse);
659 op_size(aTHX_ CvSTART(thing), st);
660 op_size(aTHX_ CvROOT(thing), st);
665 magic_size(thing, st);
666 st->total_size += sizeof(XPVGV);
667 if(isGV_with_GP(thing)) {
668 st->total_size += GvNAMELEN(thing);
670 # if !defined(USE_ITHREADS) || (PERL_VERSION > 8 || (PERL_VERSION == 8 && PERL_SUBVERSION > 8))
671 /* With itreads, before 5.8.9, this can end up pointing to freed memory
672 if the GV was created in an eval, as GvFILE() points to CopFILE(),
673 and the relevant COP has been freed on scope cleanup after the eval.
674 5.8.9 adds a binary compatible fudge that catches the vast majority
675 of cases. 5.9.something added a proper fix, by converting the GP to
676 use a shared hash key (porperly reference counted), instead of a
677 char * (owned by who knows? possibly no-one now) */
678 check_new_and_strlen(st, GvFILE(thing));
681 /* Is there something hanging off the glob? */
682 if (check_new(st, GvGP(thing))) {
683 st->total_size += sizeof(GP);
684 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_sv), recurse);
685 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_form), recurse);
686 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_av), recurse);
687 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_hv), recurse);
688 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_egv), recurse);
689 sv_size(aTHX_ st, (SV *)(GvGP(thing)->gp_cv), recurse);
694 st->total_size += sizeof(XPVFM);
695 magic_size(thing, st);
696 st->total_size += ((XPVIO *) SvANY(thing))->xpv_len;
697 sv_size(aTHX_ st, (SV *)CvPADLIST(thing), SOME_RECURSION);
698 sv_size(aTHX_ st, (SV *)CvOUTSIDE(thing), recurse);
700 if (st->go_yell && !st->fm_whine) {
701 carp("Devel::Size: Calculated sizes for FMs are incomplete");
706 st->total_size += sizeof(XPVIO);
707 magic_size(thing, st);
708 if (check_new(st, (SvPVX_const(thing)))) {
709 st->total_size += ((XPVIO *) SvANY(thing))->xpv_cur;
711 /* Some embedded char pointers */
712 check_new_and_strlen(st, ((XPVIO *) SvANY(thing))->xio_top_name);
713 check_new_and_strlen(st, ((XPVIO *) SvANY(thing))->xio_fmt_name);
714 check_new_and_strlen(st, ((XPVIO *) SvANY(thing))->xio_bottom_name);
715 /* Throw the GVs on the list to be walked if they're not-null */
716 sv_size(aTHX_ st, (SV *)((XPVIO *) SvANY(thing))->xio_top_gv, recurse);
717 sv_size(aTHX_ st, (SV *)((XPVIO *) SvANY(thing))->xio_bottom_gv, recurse);
718 sv_size(aTHX_ st, (SV *)((XPVIO *) SvANY(thing))->xio_fmt_gv, recurse);
720 /* Only go trotting through the IO structures if they're really
721 trottable. If USE_PERLIO is defined we can do this. If
722 not... we can't, so we don't even try */
724 /* Dig into xio_ifp and xio_ofp here */
725 warn("Devel::Size: Can't size up perlio layers yet\n");
729 warn("Devel::Size: Unknown variable type: %d encountered\n", SvTYPE(thing) );
735 #include "vtables.inc"
739 static struct state *
744 void **vt_p = vtables;
746 Newxz(st, 1, struct state);
748 if (NULL != (warn_flag = perl_get_sv("Devel::Size::warn", FALSE))) {
749 st->dangle_whine = st->go_yell = SvIV(warn_flag) ? TRUE : FALSE;
751 if (NULL != (warn_flag = perl_get_sv("Devel::Size::dangle", FALSE))) {
752 st->dangle_whine = SvIV(warn_flag) ? TRUE : FALSE;
754 check_new(st, &PL_sv_undef);
755 check_new(st, &PL_sv_no);
756 check_new(st, &PL_sv_yes);
758 check_new(st, *vt_p++);
762 MODULE = Devel::Size PACKAGE = Devel::Size
770 total_size = TOTAL_SIZE_RECURSION
773 SV *thing = orig_thing;
774 struct state *st = new_state(aTHX);
776 /* If they passed us a reference then dereference it. This is the
777 only way we can check the sizes of arrays and hashes */
782 sv_size(aTHX_ st, thing, ix);
783 RETVAL = st->total_size;