summaryrefslogtreecommitdiffstats
path: root/gc.c
Commit message (Collapse)AuthorAgeFilesLines
* Synchronize license comments with LICENSE.Kaz Kylheku2016-10-011-16/+17
| | | | | | | | | | | | | | | | | | | | * Makefile, args.c, args.h, arith.c, arith.h, cadr.c, cadr.h, combi.c, combi.h, configure, debug.c, debug.h, eval.c, eval.h, filter.c, filter.h, ftw.c, ftw.h, gc.c, gc.h, glob.c, glob.h, hash.c, hash.h, jmp.S, lib.c, lib.h, lisplib.c, lisplib.h, match.c, match.h, parser.c, parser.h, parser.l, parser.y, rand.c, rand.h, regex.c, regex.h, share/txr/stdlib/awk.tl, share/txr/stdlib/build.tl, share/txr/stdlib/cadr.tl, share/txr/stdlib/conv.tl, share/txr/stdlib/except.tl, share/txr/stdlib/hash.tl, share/txr/stdlib/ifa.tl, share/txr/stdlib/path-test.tl, share/txr/stdlib/place.tl, share/txr/stdlib/socket.tl, share/txr/stdlib/struct.tl, share/txr/stdlib/termios.tl, share/txr/stdlib/txr-case.tl, share/txr/stdlib/type.tl, share/txr/stdlib/with-resources.tl, share/txr/stdlib/with-stream.tl, share/txr/stdlib/yield.tl, signal.c, signal.h, socket.c, socket.h, stream.c, stream.h, struct.c, struct.h, sysif.c, sysif.h, syslog.c, syslog.h, termios.c, termios.h, txr.1, txr.c, txr.h, unwind.c, unwind.h, utf8.c, utf8.h: Revert to verbatim 2-Clause BSD.
* Rename EXTRA_DEBUGGING to CONFIG_EXTRA_DEBUGGING.Kaz Kylheku2016-06-171-5/+5
| | | | | | | | * configure: Generate #define CONFIG_EXTRA_DEBUGGING 1 in config.h header, rather than EXTRA_DEBUGGING. * gc.c, gc.h, hash.c: Change references to EXTRA_DEBUGGING preprocessor symbol to CONFIG_EXTRA_DEBUGGING.
* Print valgrind backtraces for break_obj.Kaz Kylheku2016-06-171-2/+28
| | | | | | | | | | | | * gc.c (more): Under EXTRA_DEBUGGING, call breakpt() here also, when it is detected that break_obj is added to the free list for the firts time. Print a message and backtrace using VALGRIND_PRINTF_BACKTRACE. (make_obj): Likewise, for an object that is pulled out of the free list and returned. (mark_obj, sweep_one): Print backtrace with Valgrind for break_obj, in addition to calling breakpt().
* Useful assertion in make_obj.Kaz Kylheku2016-06-171-0/+1
| | | | | * gc.c (make_obj): Assert that the object we are pulling from the free list is marked FREE.
* gc bugfix: maintain free_tail properly.Kaz Kylheku2016-06-161-5/+6
| | | | | | | | | | | | | * gc.c (more): Only set the free_tail if the new memory is being added to an empty free_list. Otherwise free_tail is valid. (make_obj): If we empty the free_list, then reset the free_tail pointer, so it doesn't continue pointing into the object we are returning. (sweep): Remove defensive code which tries to reset the free_tail in the empty free_list case, but assumes that free_tail is correct in the non-empty case.
* Remove more vestiges of the mark_makefresh hack.Kaz Kylheku2016-06-161-16/+3
| | | | | | | | | | * gc.c (sweep_one): Don't check for gen == -1; no reachable object can have that at this stage, since mark flips -1 objects to gen 0. (sweep): Do not clear freshobj_idx here; completely revert the mark_markfresh code. (gc): Clear freshobj_idx here., as before the mark_makefresh hack.
* New --free-all option for freeing memory on exit.Kaz Kylheku2016-06-071-0/+40
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Although we are garbage-collected, being able to clean up on shutdown is nevertheless useful for uncovering leaks. Leaks can occur, for instance, due to neglect to free out-of-heap satellite data from objects that are reclaimed by gc. This feature is long overdue. * arith.c, arith.h (arith_free_all): New function. * gc.c, gc.h (gc_free_all): New function. * lib.c (init): Remove program name parameter and redundant initialization of progname globl variable. * lib.h (progname): Superfluous declaration removed. This is already declared in txr.h. (init): Declaration updated. * regex.c (char_set_destroy): Do not check the static allocation flag here; just destroy the object. Do check for a null pointer, though. (char_set_cobj_destroy): This cobj destructor now checks the static flag of the char set object and avoids freeing it. Thus our char set singletons are left alone by gc, but our global freeing function takes care of them. (wide_cs): New static variable moved out of wide_display_char_p to static scope. (regex_free_all): New function. * regex.h (regex_free_all): Declared. * txr.c (progname): const qualifier and initializer removed. (main): Ensure progname is always dynamically allocated, even in the argv[0] == 0 case. Do not pass progname to init; it doesn't take that argument any more. (free_all): New static function. (txr_main): Implement --free-all option. * txr.h (progname): Declaration updated.
* Fix incorrect comment.Kaz Kylheku2016-05-151-1/+1
| | | | | * gc.c (gc_set): Should say "promote obj to gen 1", not "to gen 0".
* Mechanism for recycling conses outside of GC.Kaz Kylheku2016-04-201-0/+1
| | | | | | | | | | | | * lib.c (recycled_conses): New static variable. (rcyc_pop, rcyc_cons, rcyc_list, rcyc_empty): New functions. (cons): Take a recycled cons, if available. * lib.h (rcyc_pop, rcyc_cons, rcyc_list, rcyc_empty): Declared. * gc.c (gc): Call rcyc_empty to make recycle_list unreachable.
* Replace all stray C style casts with macros.Kaz Kylheku2016-03-291-1/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | * gc.c (gc_report_copies): C style casts found in this function. * linenoise.c (strip_qual, convert, coerce): Copy and paste the macros here. (record_undo, compare_completions, lino_add_completion, history_search, ab_append, sync_data_to_buf, refresh_singleline, screen_rows, refresh_multiline, find_nearest_paren, paren_jump, yank_sel, edit_move_matching_paren, edit, lino_make, lino_copy, lino_hist_add, lino_hist_set_max_len): C style casts replaced. * mpi/mpi-types.h (MP_DIGIT_BIT, MP_DIGIT_MAX, MP_WORD_BIT, MP_WORD_MAX, RADIX): C style casts replaced. * mpi/mpi.c (convert, coerce): Copy and paste the macros here. (mp_init_size, mp_init_copy, mp_copy, mp_set_int, mp_div_d, mp_bit, mp_to_double, mp_to_signed_bin, mp_to_unsigned_bin, mp_to_unsigned_buf, mp_toradix_case, mp_grow, s_mp_set_bit, s_mp_mod_2d, s_mp_mul_2d, s_mp_div_2d, s_mp_mul_d, s_mp_mul, s_mp_sqr, s_mp_div, s_mp_2expt, s_mp_todigit): C style casts replaced. * mpi/mplogic (convert): Macro copy and pasted here. (mpl_num_set, mpl_num_clear): C style casts replaced. * parser.c (provide_completions): Likewise. * signal.c (small_sigfillset): Likewise. * stream.c (stdio_truncate, test_set_indent_mode, set_indent_mode): Likewise.
* Use struct instead of cons for lazy string fields.Kaz Kylheku2016-01-091-2/+6
| | | | | | | | | | | | | | | | * gc.c (finalize): Must free the dynamic structure attached to the LSTR type now. (mark_obj): Must mark interior of LSTR type's props structure. * lib.c (lazy_sub_str, copy_lazy_str): Copy props structure. (lazy_str): Allocate and initialize props structure. (lazy_str_force, lazy_str_put, lazy_str_force_upto, lazy_str_get_trailing_list, out_lazy_str): Follow representation change. * lib.h (struct lazy_string_props): New struct type. (strut lazy_string): Member opts replaced with props pointer to struct lazy_string_props.
* Copyright year bump.Kaz Kylheku2015-12-311-1/+1
| | | | | | | | | | | | | | | | | | | | | | | * LICENSE, METALICENSE, Makefile, args.c, args.h, arith.c, arith.h, cadr.c, cadr.h, combi.c, combi.h, configure, debug.c, debug.h, eval.c, eval.h, filter.c, filter.h, gc.c, gc.h, glob.c, glob.h, hash.c, hash.h, jmp.S, lib.c, lib.h, lisplib.c, lisplib.h, match.c, match.h, parser.c, parser.h, parser.l, parser.y, rand.c, rand.h, regex.c, regex.h, share/txr/stdlib/cadr.tl, share/txr/stdlib/except.tl, share/txr/stdlib/hash.tl, share/txr/stdlib/ifa.tl, share/txr/stdlib/path-test.tl, share/txr/stdlib/place.tl, share/txr/stdlib/struct.tl, share/txr/stdlib/txr-case.tl, share/txr/stdlib/type.tl, share/txr/stdlib/with-resources.tl, share/txr/stdlib/with-stream.tl, share/txr/stdlib/yield.tl, signal.c, signal.h, stream.c, stream.h, struct.c, struct.h, sysif.c, sysif.h, syslog.c, syslog.h, txr.1, txr.c, txr.h, unwind.c, unwind.h, utf8.c, utf8.h: Add 2016 copyright. * linenoise/LICENSE, linenoise/linenoise.c, linenoise/linenoise.h: Bump one principal author's copyright from 2014 to 2015. The code is based on a snapshot of 2015 upstream work.
* Useful debugging function for spurious retention.Kaz Kylheku2015-11-201-0/+12
| | | | | | * gc.c (gc_report_copies): New function * gc.h (gc_report_copies): Declared.
* Remove unused rel1 function.Kaz Kylheku2015-11-061-7/+0
| | | | | | * gc.c (rel1): Removed. * gc.h (rel1): Declaration removed.
* New range type, distinct from cons cell.Kaz Kylheku2015-11-011-0/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | * eval.c (eval_init): Register intrinsic functions rcons, rangep from and to. (eval_init): Register rangep intrinsic. * gc.c (mark_obj): Traverse RNG objects. (finalize): Handle RNG in switch. * hash.c (equal_hash, eql_hash): Hashing for for RNG objects. * lib.c (range_s, rcons_s): New symbol variables. (code2type): Handle RNG type. (eql, equal): Equality for ranges. (less_tab_init): Table extended to cover RNG. (less): Semantics defined for ranges. (rcons, rangep, from, to): New functions. (obj_init): range_s and rcons_s variables initialized. (obj_print_impl): Produce #R notation for ranges. (generic_funcall, dwim_set): Recognize range objects for indexing * lib.h (enum type): New enum member, RNG. MAXTYPE redefined to RNG value. (TYPE_SHIFT): Increased to 5 since there are now 16 type codes. (struct range): New struct type. (union obj): New member rn, of type struct range. (range_s, rcons_s, rcons, rangep, from, to): Declared. (range_bind): New macro. * parser.l (grammar): New rule for recognizing the #R sequence as HASH_R token. * parser.y (HASH_R): New terminal symbol. (range): New nonterminal symbol. (n_expr): Derives the new range symbol. The n_expr DOTDOT n_expr rule produces rcons expression rather than const. * match.c (format_field): Recognize rcons syntax in fields which is now what ranges translate to. Also recognize range object. * tests/013/maze.tl (neigh): Fix code which destructures range as a cons. That can't be done any more. * txr.1: Document ranges.
* Expose memory region marking function.Kaz Kylheku2015-10-251-0/+5
| | | | | | * gc.c (gc_mark_mem): New function. * gc.h (gc_mark_mem): Declared.
* Stop using C library setjmp/longjmp.Kaz Kylheku2015-10-251-3/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | TXR is moving to custom assembly-language routines. This is mainly motivated by a very dubious thing done in the GNU C Library setjmp and longjmp in the name of security. Evidently, glibc's setjmp "mangles" certain pointer values which are stored into the jmp_buf buffer. It's been that way since 2005, evidently. This means that, firstly, all along, the use of setjmp in gc.c to get registers into a buffer so they can be scanned has not actually worked properly. More importantly, this pointer mangling in setjmp and longjmp is very hostile to a stack copying implementation of delimited continuations. The reason is that continuations contain jmp_buf buffers, which get relocated in the process of capturing and reviving a continuation. Any pointers in a jmp_buf which point into the captured stack segment have to be fixed up to point into the relocated location. Mangled pointers make this difficult, requiring hacks which are specific to glibc and the machine architecture. We might as well implement a clean, well-behaved setjmp and longjmp. * Makefile (jmp.o): New object file. (dbg/%.o, opt/%.o): New rules for .S prerequisites. * args.c, arith.c, cadr.c, combi.c, cadr.c, combi.c, debug.c, eval.c, filter.c, glob.c, hash.c, lib.c, match.c, parser.c, rand.c, regex.c, signal.c, stream.c, struct.c, sysif.c, syslog.c, txr.c, unwind.c, utf8.c: Removed <setjmp.h> include. * gc.c: Switch to struct jmp and jmp_save, instead of jmp_buf and setjmp. * jmp.S: New source file. * signal.h (struct jmp): New struct type. (jmp_save, jmp_restore): New function declarations denoting assembly language routines in jmp.S. (extended_jmp_buf): Uses struct jmp instead of setjmp. (extended_setjmp): Use jmp_save instead of setjmp. (extended_longjmp): Use jmp_restore instead of longjmp.
* New function: call-finalizers.Kaz Kylheku2015-10-061-0/+33
| | | | | | | | | * gc.c (gc_call_finalizers): New function. (gc_late_init): Register call-finalizers intrinsic. * gc.h (gc_call_finalizers): Declared. * txr.1: documented call-finalizers.
* Support for reverse order in finalization.Kaz Kylheku2015-10-011-5/+15
| | | | | | | | | | | | | | | | | | | | | | A new optional argument on finalize allows it to be expressed that multiple finalizers on the same object are to be called in reverse order, which is potentially for objects with inheritance. * gc.c (gc_finalize): New optional argument, rev_order_p. Insert at the head of the list if this argument is specified and true. (gc_late_init): Register finalize as three-argument function with optional argument. * gc.h (gc_finalize): Declaration updated. * share/txr/stdlib/struct.tl (defstruct): Register :fini functions in reverse, so that derived finalizers are called before supertype finalizers. * txr.1: Documented new finalize argument, and behavior of :fini.
* Full gc should be done if driven by malloc.Kaz Kylheku2015-09-271-0/+5
| | | | | | | | | * gc.c (gc): If the number of bytes malloced since the last gc exceeds the gc delta, we must do do a full garbage collection. The mature generation is almost certainly holding on to a lot of that memory. A test case for this is (lazy-str-force (lazy-str (get-lines (open-file "file")))). The file doesn't even have to be very large.
* Small optimization in finalization.Kaz Kylheku2015-09-091-1/+6
| | | | | | | | | | | | | * gc.c (prepare_finals): In the second pass, only mark objects that were identified in the first pass as unreachable. Reachable ones don't require a redundant call to mark_obj. Also, objects identified as unreachable can be moved to gen 0. This is something like what the mark_makefresh hack was trying to do. It only does anything in a full GC pass, because in an incremental pass, anything identified as unreachable is necessarily gen 0. It can help prevent some objects that are revived by finalization from sliding back to gen 1 and hanging around.
* Remove useless mark_makefresh hack.Kaz Kylheku2015-09-091-12/+1
| | | | | | | | | This contraption doesn't actually do anything. * gc.c (mark_makefresh): Static variable removed. (mark_obj): Remove logic conditional no mark_makefresh. (prepare_finals): Remove setting of mark_makefresh. (call_finals): Remove clearing of mark_makefresh.
* Finalization fix: incorrect for generational gc.Kaz Kylheku2015-09-091-17/+19
| | | | | | | | | | | | | | | | | | | | | | | | | | Another bug found in the finalization hooks. The reachability test being applied is not correct for generational GC, relying only on the REACHABLE flag. This bug means that we call finalization hooks on mature objects, while they remain reachable. It must use the gc_is_reachable logic, just like weak hash table processing does, because under an incremental garbage collection pass, all gen 1 objects are considered reachable (though they don't have the REACHABLE flag set). * gc.c (struct fin_reg): Remving obj_type member, replacing with reachable flag. (is_reachable): New static function based on gc_is_reachable, minus the is_ptr test. (GC code which knows it is dealing with a pointer doesn't need the test). (prepare_finals): In the first pass, instead of memorizing the prior object type, simply calculate the reachable flag, correctly taking into account generational GC. (call_finals): Rely on the new reachable flag rather than the REACHABLE flag in the saved obj_type. (gc_is_reachable): Becomes a public wrapper for is_reachable which adds the pointer test. (gc_finalize): Initialize the reachable member of fin_reg. Remove initialization for obj_type.
* Fix broken finalization support.Kaz Kylheku2015-09-021-1/+1
| | | | | | | * gc.c (call_finals): Correct inverted test for reachability. The function was processing entries that are still reachable and calling their finalizers, and retaining unreachable entries in the finalization list.
* Fix incorrect finalization of struct-type.Kaz Kylheku2015-09-021-1/+1
| | | | | | | | | | | | | | | | | | | | | | The issue is that we can't call remhash in the cobj destroy handler, because we are in the middle of gc. The fix is to register a proper finalization hook. This is the first use of finalize from within TXR. * gc.c (gc_finalize): static becomes extern. * gc.h (gc_finalize): Declared. * struct.c (struct_type_finalize_f): New static variable. (struct_init): GC-protect new static variable, and initialize it to point to new struct_type_finalize function. (struct_type_finalize): New function: performs the remhash calls previously done in struct_type_free. (make_struct_type): Register a finalizer hook for every new structure type, which will call struct_type_finalize. (struct_type_free): Static function removed. (struct_type_ops): Use generic cobj_destroy_free_op since all that is left to do is to free the handle.
* Introducing structs.Kaz Kylheku2015-09-021-1/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | * args.c (args_cat_zap): New function. * args.h: (args_cat_zap): Declared. * eval.c (struct_lit_s): New symbol variable. (eval_init): Initialize struct_lit_s. * eval.h (struct_lit_s): Declared. * gc.c (finalize): If a symbol has a struct slot hash attached to it, we must free it when the symbol is reclaimed. * lib.c (make_sym): Initialize symbol's slot_cache pointer to null. (copy): Copy structure objects. (init): Call struct_init to initialize struct module. * lib.h (SLOT_CACHE_SIZE): New preprocessor symbol (slot_cache_line_t, slot_cache_t): New typedefs. (struct sym): New member, slot_cache. * lisplib.c (struct_set_entries, struct_instantiate): New static functions. (liplib_init): Register new functions in dl_table. parser.y (HASH_S): New terminal symbol. (struct): New grammar rule. (n_expr): Derive struct. (yybadtoken): Map HASH_S to #S string. parser.l (grammar): Recognize #S and return HASH_S token. share/txr/stdlib/place.tl (slot): New defplace. share/txr/stdlib/struct.tl: New file. struct.c: New file. struct.h: New file. * Makefile (OBJS): Adding struct.o.
* Provide conservative marking interface.Kaz Kylheku2015-08-191-24/+26
| | | | | | | | | | | | | | | | | | | GC now exports a function for marking something that might not be an object, instead of a function for testing. The previous way wasn't integrated with Valgrind properly, and didn't observe the FREE flag. * gc.c (gc_is_heap_obj): Function removed. (mark_obj_maybe): New static function, with body consisting of code moved from mark_mem_region. (mark_mem_region): Moved code replaced by call to mark_obj_maybe. (gc_conservative_mark): New function, wraps mark_obj_maybe. * gc.h (gc_conservative_mark): Declared. (gc_is_heap_obj): Declaration removed. * parser.c (yy_tok_mark): Use gc_conservative_mark instead of gc_is_heap_obj check and gc_mark.
* Crafting a better parser-priming hack.Kaz Kylheku2015-08-121-0/+5
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The method of inserting a character sequence which generates a SECRET_TOKEN_E token is being replaced with a purely token based method. Because we don't manipulate the input stream, the lexer is not involved. We don't have to flush its state and deal with the carry-over of the yy_hold_char. This comes about because recent changes expose a weakness in the old scheme. Now that a top-level expression can have the form expr.expr, it means that the Yacc parser reads one token ahead, to see whether there is a dot or something else. This lookahead token is discarded. We must re-create it when we call yyparse again. This re-creation is done by creating a custom yylex function, which can maintain pushback tokens. We can prime this array of pushback tokens to generate the SECRET_TOKEN_E, as well as to re-inject the lookahead symbol that was thrown away by the previous yyparse. To know which lookahead symbol to re-inject is simple: the scanner just keeps a copy of the most recent token that it returns to the parser. When the parser returns, that token must be the lookahead one. The tokens we keep now in the parser structure are subject to garbage collection, and so we must mark them. Since the YYSTYPE union has no type field, a new API is opened up into the garbage collector to help implement a conservative GC technique. * gc.c (gc_is_heap_obj): New function. * gc.h (gc_is_heap_obj): Declared. * match.c: Include y.tab.h. This is now needed by any module that needs to instantiate a parser_t structure, because members of type YYSTYPE occur in the structure. (parser.h can still be included without y.tab.h, but only an incomplete declaration for the parser strucure is then given, and a few functions are not declared.) * parser.c (yy_tok_mark): New static function. (parser_mark): Mark the recent token and the pushback tokens. (parser_common_init): Initialize the recent token, the pushback tokens, and the pushback stack index. (pushback_token): New static function. (prime_parser): hold_byte argument removed. Body considerably simplified. The catenated stream trick is no longer required. All we do here is set up two pushback tokens and prime the scanner, if necessary, so it is in the right start state for Lisp. * parser.l (YY_DECL): Take over definition of scanning function, renaming to yylex_impl, so we can implement yylex. (grammar): Rule which produces SECRET_ESCAPE_E token removed. (reset_scanner): Function removed. (yylex): New function. * parser.h (struct parser): Now only forward-declared unless y.tab.h has been included. New members, recent_tok, tok_pushback and tok_idx. (yyset_hold_char): Declared. (reset_scanner): Declaration removed. (yylex): Declared (if y.tab.h included). (prime_parser): Declaration updated. (prime_scanner): Declared. * Makefile: express new dependency on existence of y.tab.h of txr.o, match.o and parser.o.
* Remove never-used member of symbol structure.Kaz Kylheku2015-08-091-1/+0
| | | | | | | | * lib.h (struct sym): Remove value member. * lib.c (make_sym): Don't initialize removed member. * gc.c (mark_obj): Do not mark removed member.
* * gc.c (finalize): Null out cobj handle after calling theKaz Kylheku2015-07-291-0/+1
| | | | destructor. This can catch some use-after-free errors.
* * gc.c (make_obj, gc): Move resetting of prev_malloc_bytesKaz Kylheku2015-05-061-2/+1
| | | | out of make_obj into gc.
* Fix corruption triggered by extended gc disabling.Kaz Kylheku2015-05-011-39/+61
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | The issue is that when gc is disabled, the gc function does nothing. But various code depends on calls to gc() doing something, like making space available in various static arrays. When gc is disabled for long periods, there are issues, like array overruns. * gc.c (gc): Must no longer be called at all if gc_enabled is false, and asserts that it is true. Callers must check the gc_enabled flag and implement appropriate cases. (make_obj): Only call gc when gc_enabled is true. If there is no space in the freshobj array after trying gc, or gc couldn't be tried because it is disabled, then schedule a full gc. (gc_set): If the checkobj array is full, only call gc if gc is enabled, otherwise schedule a full_gc. (gc_mutated): Do not assume that the mutobj array has room for another object; only set the object's generation to -1 and put it into the array if there is room. Similarly to gc_set, do a gc if there is no room, but if gc is not enabled, then schedule a full gc. (gc_wrap): Only call gc if gc_enabled is true, and return t in that case rathe than nil. * txr.1: Document return value of sys:gc function.
* Regression: runaway allocation of heaps.Kaz Kylheku2015-04-241-1/+1
| | | | | | | | | | * gc.c (make_obj): Fix regression introduced on 2014-10-18, released in TXR 100. Because make_obj discontinues placing objects into the freshobj_array when the full_gc flag has been set, it is incorrect to allocate another heap when the free list is exhausted and the freshobj_array has room remaining. It has room remaining because it is disused.
* * gc.c (make_obj, gc): The check for insufficient space in freshobjKaz Kylheku2015-04-241-3/+4
| | | | array after a gc is moved into the gc function.
* * gc.c (gc): Minor off-by-one: compare gc_counter to FULL_GC_INTERVALKaz Kylheku2015-04-241-1/+1
| | | | rather than FULL_GC_INTERVAL - 1, since it is pre-incremented.
* * gc.c (sweep): Fix comment referring to nonexistent identifier.Kaz Kylheku2015-04-101-1/+1
|
* Update copyright notices from 2014 to 2015.Kaz Kylheku2015-02-011-1/+1
| | | | | | | | | | | * arith.c, arith.h, combi.c, combi.h, debug.c, debug.h, eval.c, eval.h, filter.c, filter.h, gc.c, gc.h, hash.c, hash.h, lib.c, lib.h, match.c, match.h, parser.h, rand.c, rand.h, regex.c, regex.h, signal.c, signal.h, stream.c, stream.h, sysif.c, sysif.h, syslog.c, syslog.h, txr.c, txr.h, unwind.c, unwind.h, utf8.c, utf8.h: Update. * LICENSE, METALICENSE: Likewise.
* Implementing finalization hooks.Kaz Kylheku2014-10-301-4/+110
| | | | | | | | | | | | | | | | | | | | | | | | | | | * gc.c (struct fin_reg): New struct type. (final_list, final_tail, mark_makefresh): New static variables. (mark_obj): Under generational GC, if make_makefresh is in effect, set the generation to -1 on all marked objects. (sweep_one): In an EXTRA_DEBUGGING build, call breakpt if the object being swept is the one in break_obj. Under generational GC, place reachable objects that are in generation -1 the freshobj nursery and assign them to generation 0, rather than sticking them into the mature generation 1. (sweep): Under generational gc, reset the freshobj_idx variable here, so that sweep_one has an empty nursery in which to place the generation -1 objects. (prepare_finals, call_finals): New static functions. (gc): Call prepare_finals before sweep, and call call_finals just before re-enabling GC and returning. Do not reset freshobj_idx to zero; this was done in sweep, which may have added entries into it. (gc_finalize): New function. (gc_late_init): Register gc_finalize as intrinsic function finalize. * txr.1: Documented finalize. * HACKING: Documented finalization, described the additional meaning of the -1 generation, and added a section on debugging with break_obj and breakpt.
* * Makefile: Removing trailing spaces.Kaz Kylheku2014-10-241-1/+1
| | | | | | | | | | (GREP_CHECK): New macro. (enforce): Rewritten using GREP_CHECK, with new checks. * arith.c, combi.c, debug.c, eval.c, filter.c, gc.c, hash.c, lib.c, * lib.h, match.c, parser.l, parser.y, rand.c, regex.c, signal.c, * signal.h, stream.c, syslog.c, txr.c, unwind.c, utf8.c: Remove trailing spaces.
* Ensure that a hash reorganization doesn't take placeKaz Kylheku2014-10-221-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | during a traversal, which could cause items to be visited twice or skipped. * gc.c (full_gc): Changed from static to exter (full_gc): Changed to internal linkage. * gc.h (full_gc): Declared. * hash.c (struct hash): New member, usecount. (struct hash_iter): New member, next. (reachable_iters): New static variable. (hash_mark): Reset the usecount of every reachable hash table. (hash_iter_mark): Add every reachable iterator to reachable_iters list. (make_hash, make_similar_hash, copy_hash): Initialize usecount. (gethash_c): Do not call hash_grow if usecount is nonzero. (hash_begin): Initialize next field of hash_iter to null. Increment use count on hash table. (hash_next): When traversal ends, release use count, and null out the hash table reference. (do_weak_tables): New static function. (do_iters): New static function. (hash_process_weak): Weak hash processing logic moved to do_weak_tables. Also calls do_iters. * txr.1: Describe hash table traversal, and the assurances and caveats about inserting and deleting items.
* Deal with situation when GC is disabled and the freshobj array runs outKaz Kylheku2014-10-181-3/+5
| | | | | | | | | | | | | | | | of room. This could happen when parsing a really large TXR program, since gc is disabled during parsing. Currently it asserts, which is not acceptable. * gc.c (make_obj): If after gc, the freshobj array has not been emptied (obviously because gc is disabled), do not assert. Rather, set the full_gc flag to request a full garbage collection when gc is re-enabled. Furthermore, only place newly allocated objects into freshobj if full_gc has not been set. Thus, if we exhaust the freshobj array while gc is off, the full_gc flag is set, and we discontinue use of that array. When gc is re-enabled, we will do a full gc pass. A full gc pass ignores freshobj array, so it doesn't matter that its use was discontinued.
* Converting cast expressions to macros that are retargettedKaz Kylheku2014-10-171-8/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | to C++ style casts when compiling as C++. * lib.h (strip_qual, convert, coerce): New casting macros. (TAG_MASK, tag, type, wli_noex, auto_str, static_str, litptr, num_fast, chr, lit_noex, nil, nao): Use cast macros. * arith.c (mul, isqrt_fixnum, bit): Use cast macros. * configure (INT_PTR_MAX): Define using cast macro. * debug.c (debug_init): Use cast macro. * eval.c (do_eval, expand_macro, reg_op, reg_mac, eval_init): Use cast macros. * filter.c (filter_init): Use cast macro. * gc.c (more, mark_obj, in_heap, mark, sweep_one, unmark): Use cast macros. * hash.c (hash_double, equal_hash, eql_hash, hash_equal_op, hash_hash_op, hash_print_op, hash_mark, make_hash, make_similar_hash, copy_hash, gethash_c, gethash, gethash_f, gethash_n, remhash, hash_count, get_hash_userdata, set_hash_userdata, hash_iter_destroy, hash_iter_mark, hash_begin, hash_uni, hash_diff, hash_isec): Use cast macros. * lib.c (code2type, chk_malloc, chk_malloc_gc_more, chk_calloc, chk_realloc, chk_strdup, num, c_num, string, mkstring, mkustring, upcase_str, downcase_str, string_extend, sub_str, cat_str, trim_str, c_chr, vector, vec_set_length, copy_vec, sub_vec, cat_vec, cobj_print_op, obj_init): Likewise. * match.c (do_match_line, hv_trampoline, match_files, dir_tables_init): Likewise. * parser.l (grammar): Likewise. * parser.y (parse): Likewise. * rand.c (make_state, make_random_state, random_fixnum, random): Likewise. * regex.c (CHAR_SET_L2_LO, CHAR_SET_L2_HI, CHAR_SET_L1_LO, CHAR_SET_L1_HI, CHAR_SET_L0_LO, CHAR_SET_L0_HI, L0_full, L0_fill_range, L1_full, L1_fill_range, L1_contains, L1_free, L2_full, L2_fill_range, L2_contains, L2_free, L3_fill_range, L3_contains, L3_free, char_set_create, char_set_cobj_destroy, nfa_state_accept, nfa_state_empty, nfa_state_single, nfa_state_wild, nfa_state_set,
* * gc.c (MALLOC_DELTA_THRESH): Macro remaned to DFL_MALLOC_DELTA_THRESH.Kaz Kylheku2014-09-081-3/+23
| | | | | | | | | | | | | | | | | | | | | (opt_gc_delta): New global variable. (make_obj): Use opt_gc_delta rather than MALLOC_DELTA_THRESH. (gc_set_delta, gc_wrap): New static functions. (gc_late_init): New function. * gc.h (gc_late_init): Declared. * genvim.txr: scan gc.c also. * lib.c (init): call gc_late_init. * txr.1: Document new --gc-delta option and the functions gc and gc-set-delta. * txr.c (help): Help text for --gc-delta. (txr_main): Parse --gc-delta option. * txr.h (opt_gc_delta): Declared.
* Make the garbage collector aware of malloced bytes, toKaz Kylheku2014-09-061-3/+10
| | | | | | | | | | | prevent skyrocketing memory use when the program manipulates heap objects that have large amounts of malloc memory attached, like very large bignums. * gc.c (MALLOC_DELTA_THRESH): New preprocessor symbol. (prev_malloc_bytes): New static variable. (make_obj): when the value of malloc_bytes jumps by MALLOC_DELTA_THRESH or more since the last gc, trigger gc.
* * gc.c (top): Renamed to gc_prot_top, turned extern.Kaz Kylheku2014-08-251-5/+5
| | | | | | | | | (prot1, rel1, mark): Follow rename. * gc.h (gc_prot_top): Declared. * signal.h (extended_jmp_buf): New member, gc_pt. (extended_setjmp): Save and restore gc_prot_top.
* * Makefile, arith.c, arith.h, combi.c, combi.h, configure, debug.c,Kaz Kylheku2014-07-231-16/+16
| | | | | | | | debug.h, eval.c, eval.h, filter.c, filter.h, gc.c, gc.h, hash.c, hash.h, lib.c, lib.h, match.c, match.h, parser.h, parser.l, parser.y, rand.c, rand.h, regex.c, regex.h, signal.c, signal.h, stream.c, stream.h, syslog.c, syslog.h, txr.c, txr.h, unwind.c, unwind.h, utf8.c, utf8.h: Synchronize license header with LICENSE.
* * eval.c, gc.c, rand.c, regex.c, signal.c: Remove inclusion of unneededKaz Kylheku2014-04-131-1/+0
| | | | headers.
* * gc.c (CHECKOBJ_VEC_SIZE, FRESHOBJ_VEC_SIZE): IncreaseKaz Kylheku2014-03-291-2/+2
| | | | | to 8 * HEAP_SIZE from 2 * HEAP_SIZE to reduce frequency of garbage collections.
* Change to how locations are passed around, for the sake of generationalKaz Kylheku2014-03-291-8/+11
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | GC. The issue being solved here is the accuracy of the gc_set function. The existing impelmentation is too conservative. It has no generation information about the memory location being stored, and so it assumes the worst: that it is a location in the middle of a gen 1 object. This is sub-optimal, creating unacceptable pressure against the checkobj array and, worse, as a consequence causing unreachable gen 0 objects to be tenured into gen 1. To solve this problem, we replace "val *" pointers with a structure of type "loc" which keeps track of the object too, which lets us discover the generation. I tried another approach: using just a pointer with a bitfield indicating the generation. This turned out to have a serious issue: such a bitfield goes stale when the object is moved to a different generation. The object holding the memory location is in gen 1, but the annotated pointer still indicates gen 0. The gc_set function then makes the wrong decision, and premature reclamation takes place. * combi.c (perm_init_common, comb_gen_fun_common, rcomb_gen_fun_common, rcomb_list_gen_fun): Update to new interfaces for managing mutation. * debug.c (debug): Update to new interfaces for managing mutation. Avoid loc variable name. * eval.c (env_fbind, env_fbind): Update to new interfaces for managing mutation. (lookup_var_l, dwim_loc): Return loc type and update to new interfaces. (apply_frob_args, op_modplace, op_dohash, transform_op, mapcarv, mappendv, repeat_infinite_func, repeat_times_func): Update to new interfaces for managing mutation. * eval.h (lookup_var_l): Declaration updated. * filter.c (trie_add, trie_compress, trie_compress_intrinsic, * build_filter, built_filter_from_list, filter_init): Update to new * interfaces. * gc.c (gc_set): Rewritten to use loc type which provides the exact generation. We do not need the in_malloc_range hack any more, since we have the backpointer to the object. (gc_push): Take loc rather than raw pointer. * gc.h (gc_set, gc_push): Declarations updated. * hash.c (struct hash): The acons* functions use loc instead of val * now. (hash_equal_op, copy_hash, gethash_c, inhash, gethash_n, pushhash, Change to how locations are passed around, for the sake of generational GC. The issue being solved here is the accuracy of the gc_set function. The existing impelmentation is too conservative. It has no generation information about the memory location being stored, and so it assumes the worst: that it is a location in the middle of a gen 1 object. This is sub-optimal, creating unacceptable pressure against the checkobj array and, worse, as a consequence causing unreachable gen 0 objects to be tenured into gen 1. To solve this problem, we replace "val *" pointers with a structure of type "loc" which keeps track of the object too, which lets us discover the generation. I tried another approach: using just a pointer with a bitfield indicating the generation. This turned out to have a serious issue: such a bitfield goes stale when the object is moved to a different generation. The object holding the memory location is in gen 1, but the annotated pointer still indicates gen 0. The gc_set function then makes the wrong decision, and premature reclamation takes place. * combi.c (perm_init_common, comb_gen_fun_common, rcomb_gen_fun_common, rcomb_list_gen_fun): Update to new interfaces for managing mutation. * debug.c (debug): Update to new interfaces for managing mutation. Avoid loc variable name. * eval.c (env_fbind, env_fbind): Update to new interfaces for managing mutation. (lookup_var_l, dwim_loc): Return loc type and update to new interfaces. (apply_frob_args, op_modplace, op_dohash, transform_op, mapcarv, mappendv, repeat_infinite_func, repeat_times_func): Update to new interfaces for managing mutation. * eval.h (lookup_var_l): Declaration updated. * filter.c (trie_add, trie_compress, trie_compress_intrinsic, * build_filter, built_filter_from_list, filter_init): Update to new * interfaces. * gc.c (gc_set): Rewritten to use loc type which provides the exact generation. We do not need the in_malloc_range hack any more, since we have the backpointer to the object. (gc_push): Take loc rather than raw pointer. * gc.h (gc_set, gc_push): Declarations updated. * hash.c (struct hash): The acons* functions use loc instead of val * now. (hash_equal_op, copy_hash, gethash_c, inhash, gethash_n, pushhash,
* Generational GC tweaks.Kaz Kylheku2014-03-291-19/+25
| | | | | | | | | | | | | | | | | | * gc.c (make_obj): If we have room in the freshobj array, but are out of objects, then call more. Without this, we don't take proper advantage of this nursing area. (gc): Set the full_gc flag after doing gc, in preparation for next time. If we know full GC is coming, we can adjust some behaviors. Increase the threshold for calling the more() function to be the same as in the non-generational case: less than 3/4 of the size of a heap scavenged. (gc_set): Now does nothing if it is known that a full gc is coming. Also, in the checkobj array overflow case when we invoke gc, there is no point in adding obj to the array, since it must have been promoted to the mature generation. (gc_mutated): Don't bother storing the object in the array if a full GC is coming.
generated by cgit v1.2.3 (git 2.25.1) at 2025-03-10 15:52:45 +0000