| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Paul A. Patience noted that the canonname_s variable is not
used in the C code. This indicates that the AI_CANONNAME
functionality of getaddrinfo isn't implemented. Let's do that.
* stdlib/socket.tl (sockaddr): New slot, canonname.
(addrinfo): Default canonname to nil, not 0, since it is
a string that may be absent.
* socket.c (getaddrinfo_wrap): If the first address object has
a non-null ai_canonname and it was requested via the flags,
then stick that name into every returned structure.
* txr.1: Documented.
|
|
|
|
|
|
|
|
|
|
|
|
| |
* RELNOTES: Fix various minor issues and stylistic issues.
* configure: Remove repeated word and add missing word.
* txr.1: Fix various minor, not-so-minor and stylistic issues.
In particular, struct-from-args was misspelled in the .mets line and
open-subprocess was missing from the .coNP line.
* stdlib/doc-syms.tl: Updated.
|
|
|
|
| |
* ffi.c (int8_s): Remove redundant assignment.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* lisplib.c (match_set_entries): Intern the match-error symbol.
Register autoloads for must-match and must-match-case.
* stdlib/match.tl (match-error): Register exception symbol, as subtype
of match-error.
(must-match, must-match-case): New macros.
* tests/011/patmatch.tl: Test cases.
* txr.1: Documented.
|
|
|
|
|
|
|
|
|
|
|
|
| |
* tests/common.tl (os-symbol): Look for the substring BSD in
the system name, and map to symbol :bsd. Do not produce
the :openbsd symbol.
* tests/014/socket-basic.tl: Refer to :bsd, not :openbsd.
* tests/017/glob-carray.tl: Likewise.
* tests/018/chmod.tl: Likewise.
|
|
|
|
|
|
| |
* termios.c (termios_init): Separate the #ifdef conditions for the
TAB* identifiers, instead of assuming that TABDLY covers them
all.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* lib.c (int_str): The problem here is that we are recognizing
and skipping the 0x prefix for all bases. So for instance
(int-str "0x123") produces 123. The correct requirement, and
the intent of the code, is that the C conventions are only
honored if the base is specified as the character #\c. In any
other base, including omitted base defaulting to 10, a leading
zero is just a leading zero, and 0x is a zero followed by the
junk character x. Therefore, if we have any valid base that
isn't #\c, and 0x has been seen, we must return zero. We must
not do this only in the base 16 case.
* tests/016/conv.tl: New file.
|
|
|
|
|
|
|
| |
* configure: In the test for an alignming malloc, fix a
copy-and-paste error. The $try_header expansion is passed to
printf, but there is nothing in the format string.
FreeBSD printf diagnoses this and dies.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* parser.y (elem): When the elem is a list, if it starts with mdo, then
we evaluate it immediately and substitute (do) as the semantic
value. We no longer not allow mdo to precipitate into
match_expand_elem, where expand_meta will be unleashed on it.
Doing so causes the bug that expessions like @1, denoting
the form (sys:var 1), are rewritten to (sys:expr 1), which op
syntax. So two bugs are fixed: the incorrect treatment of
meta-syntax, and the neglect to perform mdo in nested
contexts.
(check_parse_time_action): We need not handle mdo here any
more; it will never make it into this function.
Only actions done in the main clause list belong here,
not parse time actions done at any nesting level.
* tests/008/mdo.txr: New file.
* y.tab.c.shipped: Updated.
|
|
|
|
|
|
|
|
|
|
| |
* tests/007/except-3.txr: New file.
* tests/007/except-3.expected: Likewise.
* tests/007/except-4.txr: Likewise.
* tests/007/except-4.expected: Likewise.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* RELNOTES: Updated.
* configure (txr_ver): Bumped version.
* stdlib/ver.tl (lib-version): Bumped.
* txr.1: Bumped version and date.
* txr.vim, tl.vim: Regenerated.
* protsym.c: Likewise.
|
|
|
|
|
| |
* tests/012/cont.tl: Exit before the test case that contains characters
ouside of the BMP, if (sizeof wchar) is less than 4.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* eval.c (eval_init): Update registrations of lazy-stream-cons
and get-lines with one more optional argument.
* lib.c (simple_lazy_stream_func_nt, lazy_stream_func_nt): New
static functions.
(lazy_stream_cons): Take a new argument, no_throw_close,
defaulting it to nil. When calling close_stream directly, pass
the inverted value of no_throw_close. Choose the new _nt
functions for the lazy list if no_throw_close is true; those
functions pass nil as the second argument of close_stream.
* lib.h (lazy_stream_cons): Declaration updated.
* match.c (v_next_impl, open_data_source, match_fun): Pass
down the nothrow value to lazy_stream_cons, or else nil in
situations when that is not applicable or there is no such
value. Thus the :nothrow feature of v_next will now not only
ensure that there is no exception when opening the stream but
also when closing it. Unusual situations encountered when
the lazy list reads from the stream still throw.
* txr.1: Documented.
|
|
|
|
| |
* txr.1: Document do being applied to do/op.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We extend the matching context structures to keep track of the
underlying stream from which lines are being taken via the lazy list.
Then the implementation of the @(eof) directive, when it hits the eof
condition, can use this stream to gain access to the termination status.
* match.c (match_line_ctx, match_files_ctx): New member, stream.
(ml_all): Take stream argument and initialize new member.
(h_call, do_match_line): Pass stream argument to h_call.
(mf_all, mf_file_data): Take stream argument and initialize new member.
(mf_from_ml): Propagate stream from line context to file context.
(freeform_prepare, v_next_impl, match_filter, match_fun, extract): Pass
stream argument where now needed.
(v_eof): Implement termination status binding via the stream stored
in the context.
(open_data_source): Store stream in match files context.
* tests/010/eof-status.txr: New file.
* tests/010/eof-status.expected: New file.
* Makefile (tst/tests/010/eof-status.ok): -B option for new test.
* txr.1: Documented eof directive, argument and all.
* stdlib/doc-syms.tl: Updated.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* stream.c (strm_base_init): Add new element to the initializer to
initialize the close_result member to nao, indicating that the close
operation has not been invoked.
(strm_base_mark): Mark the close_result value, if it isn't nao.
This is just in case it is a heap object. The structure delegate
mechanism opens the possibility that the stream is actually user code
that can return anything so we have to be careful.
(close_stream): Only call ops->close if close_result is nao, indicating
that close had never been called (or possibly that it had been called
bu threw an exception) and store the return value in close_result,
otherwise return the previously stored value.
* stream.h (struct strm_base): New member, close_result.
* txr.1: Documented.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Issue peported by Ethan Hawk. Our socket.c module is using struct
timeval without including <sys/time.h>, which breaks on musl.
* configure: in the select test, let's include <sys/time.h>, and
if the test passes, let's set have_sys_time, so that HAVE_SELECT
implies HAVE_SYS_TIME. This way code wrapped with HAVE_SELECT doesn't
separately have to test for HAVE_SYS_TIME.
* socket.c: If HAVE_SYS_TIME is true, then we include <sys/time.h>,
independently of HAVE_SELECT.
(sock_timeout, sock_load_init): Like the select-based code, code using
SO_SNDTIMEO or SO_RCVTIMO also uses timeval, so needs to be wrapped with
HAVE_SYS_TIME.
|
|
|
|
|
|
|
| |
* linenoise/linenoise.c (linenoise): Force the printing of prompts
if the input file descriptor is a tty.
* txr.1: Documentation updated.
|
|
|
|
|
|
| |
* txr.c (banner): If standard input isn't a tty, bail.
Now takes a self argument for the c_int function.
(txr_main): Pass self value down to banner.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The :prompt-on command will enable prompting in plain mode.
* linenoise/linenoise.c (struct lino_state): New member, show_prompt.
(line_enable_noninteractive_prompt): New function.
(linenoise): In the plain mode loop, the show_prompt flag is on, show
the prompt. For continuation lines, show a condensed prompt, which
consists of the suffix of the full prompt, starting on the last
non-whitespace character.
* linenoise/linenoise.h (lino_enable_noninteractive): Declared.
* parser.c (repl): Implement :prompt-on command which enables the
above mode.
* txr.1: Documented.
|
|
|
|
|
| |
* txr.1: Document that if close-stream is applied to a closed stream,
then nil is returned without throwing an exception.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* stream.c (stdio_close, pipe_close): Fix throw_on_error
argument not being defaulted correctly, so that errors are
thrown even when the argument is omitted.
* strudel.c (strudel_close): Here, we also must default the
argument. The corresponding close method does not have an
optional argument; it is mandatory. The documentation is
bungled for it, though.
* txr.1: Fix documentation of structure delegate streams with
regard to the close method. It does not take offs and whence
parametrs, but throw-on-error-p, which is mandatory.
|
|
|
|
|
| |
* tests/012/oop.tl: Adjust one recently added test case to
eliminate undefined variable warning.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* linenoise/linenoise.c (linenoise): If we are in
noninteractive mode, then do not just read one line and return
it. If an enter_callback is defined then keep accumulating
lines while the callback indicates incomplete syntax, until
EOF occurs or the syntax appears complete. Return the lines
glued together, with \n characters replaced by \r, so the line
is correctly entered into the history, and the trailing LF
obliterated, as usual.
* txr.1: Documented new multi-line behavior of plain mode.
|
|
|
|
|
|
|
|
|
|
| |
* stdlib/struct.tl (sys:new-expander): If the argument of
new* or lnew* is dwim, then treat that as an expression,
rather than as a boa-style construction.
* tests/012/oop.tl: Tests for new* focusing on this issue.
* txr.1: Documented.
|
|
|
|
|
|
| |
* txr.1: Fix code -> codn.
* stdlib/doc-syms.tl: Updated.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This issue doesn't affect the tests. This is for the benefit
of someone who happens to be copy-and-pasting the amb
implementation from here.
* tests/012/cont.tl (amb): This function has an issue in that
it calls the continuation (future calculation) and then if
that succeeds, it normally returns the value. This means that
the future is executed again. In the case of N amb
expressions, the successful future is executed 2**N times.
What amb must do is this: call the continuation and capture
the value. If the value is successful, then that is the master
return value; just return that from amb-scope, bypassing the
second re-execution of the future.
|
|
|
|
|
| |
* tests/012/cont.tl: New test case. This aborts prior to
recent gc fixes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* parser.c (read_bad_json_s): New symbol variable.
(parser_common_init): Propagate value of *read-bad-json* into
read_bad_json flag in parser structure.
(parser_init): Initialize read_bad_json_s and register the
*read-bad-json* dynamic variable.
* parser.h (struct parser): New member, read_bad_json.
(read_bad_json_s): Declared.
* parser.y (json_val): Support an opt_comma symbol just before
the closing bracket or brace.
(opt_comma): New nonterminal symbol. Recognizes ',' or nothing.
Error is flagged if ',' is recognized, and *read-bad-json*
is nil.
* y.tab.c.shipped: Updated.
* tests/010/json.tl: New tests.
* txr.1: Documented.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* eval.c (copy_env, deep_copy_env): These functions are not following
a protocol for object construction that is correct under generational
GC. They are allocating a new object with make_obj first, and then
calling functions to copy the constituent elements to populate into
the object with a direct assignment. This direct assignment is wrong;
the set macro is required. A better fix, rather than using the set
macro, is to copy the constituent parts first, holding them in local
variables, then allocate the new object, and finally, without doing any
other memory allocating operations, assign the constituent parts into
the new object.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The functions copy-cons, copy-tree, copy-fun and copy-tnode have a
problem. They copy the original object bitwise with a structure
assignment, and then make some adjustments. The problem is that this
inappropriately copies the object's metadata related to gc, such as its
generation number or finalization count. To fix this, we introduce a
copy_obj function, which is a companion to make_obj. This performs a
shallow copy of an object without incorrectly propagating inappropriate
metadata.
* gc.c, gc.h (copy_obj): New function.
* lib.c (copy_fun, copy_cons, copy_tree): Use copy_obj, instead of
make_obj plus structure assignment.
* tree.c (copy_tnode): Likewise.
|
|
|
|
| |
* stdlib/quips.tl (sys:%quips%): New entry.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The subtypep function has poor requirements, handling only type
symbols. Let's extend it to handle structure type objects.
* lib.c (subtypep): In all cases when an argument is considered to be a
possible structure symbol, and thus subject to find_struct_type,
consider whether it already is a struct type, and just take it as-is.
* tests/012/type.tl: New tests.
* txr.1: Updated.
|
|
|
|
|
|
|
|
|
|
|
|
| |
The zap_s variable is a vestige of op_modplace. It must have been missed
when op_modplace was removed in commit
209e731429a0fd890ec6d922c1efc6f02d81a032.
* eval.c (zap_s): Delete variable.
(eval_init): Remove initialization of zap_s.
* protsym.c (zap_s): Remove extern variable declaration.
(protected_sym): Remove reference to zap_s.
|
|
|
|
|
| |
* lib.c (remql): Correct the rem_impl call's name argument from "remq"
to "remql".
|
|
|
|
|
|
|
|
| |
* lib.h (slot_cache_t): Typedef removed.
(struct sym): Use slot_cache_set_t *.
* struct.c (struct_type_finalize, lookup_slot,
lookup_static_slot, lookup_static_slot_desc): Likewise.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* RELNOTES: Updated.
* configure (txr_ver): Bumped version.
* stdlib/ver.tl (lib-version): Bumped.
* txr.1: Bumped version and date.
* txr.vim, tl.vim: Regenerated.
* protsym.c: Likewise.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The first field of each etag definition is referred to in the spec as
the "pattern", but it is supposed to contain literal text, and therefore
no characters within it need be escaped.
* tags.tl (escape): Move above tag definition.
(tag)[pattern]: Rename to...
[line]: ...this.
[text]: Update renamed slot. Escape the line here rather than on
creation.
[etext]: Update renamed slot.
(slot-tag)[text]: Update renamed slot. Escape the line here rather than
on creation.
(orig-tag)[line]: Rename to...
[orig-line]: ...this.
[text]: Update renamed slot.
(get-pos-pat): Rename to...
(get-pos-line): ...this. Don't escape the line when returning it.
(with-tag-shorthand-macro, toplevel): Rename variables and references to
functions in accordance with the above.
|
|
|
|
|
|
|
| |
* tags.tl (tags-opts): New q/qual option.
(slot-tag make-qual-tag): New method.
(toplevel): If qualified tags requested, then augment the list
of tags with qualified tags.
|
|
|
|
|
|
|
|
| |
* stream.c (volume_name_p): New static function.
(plp_regex): Static variable removed.
(pure_rel_path_p): Rewrite using lower-level string manipulation, and
using volume_name_p instead of a cached regex.
(stream_init): Remove reference to plp_regex.
|
|
|
|
| |
* tests/018/path.tl: New tests.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The make_hash function now takes the hash_weak_opt_t
enumeration instead of a pair of flags.
* hash.c (do_make_hash): Take enum argument instead of pair of
flags. Just store the option; nothing to calculate.
(weak_opt_from_flags): New static function.
(tweak_hash): Function removed.
(make_seeded_hash): Adjust to new do_make_hash interface with
help from weak_opt_from_flags.
(make_hash, make_eq_hash): Take enum argument instead of pair
of flags.
(hashv): Calculate hash_weak_opt_t enum from the extracted
flags, pass down to make_eq_hash or make_hash.
* hash.h (tweak_hash): Declration removed.
(make_hash, make_eq_hash): Declarations updated.
* eval.c (me_case, expand_switch): Update make_hash
calls to new style.
(eval_init): Update make_hash calls and get rid of tweak_hash
calls. This renders the tweak_hash function unused.
* ffi.c (make_ffi_type_enum, ffi_init): Update make_hash calls
to new style.
* filter.c (make_trie, trie_add, filter_init): Likewise.
* lib.c (make_package_common, obj_init, obj_print): Likewise.
* lisplib.c (lisplib_init): Likewise.
* match.c (dir_tables_init): Likewise.
* parser.c (parser_circ_def, repl, parse_init): Likewise.
* parser.l (parser_l_init): Likewise.
* struct.c (struct_init, get_slot_syms): Likewise.
* sysif.c (get_env_hash): Likewise.
* lex.yy.c.shipped, y.tab.c.shipped: Updated.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The flags field of hashes isn't really functioning as flags;
it's an enumeration whose numeric properties are exploited in
one place in the code.
* hash.h (enum hash_flags): Rename to enum hash_weak_opt.
(hash_flags_t): Renum to hash_weak_opt_t.
(tweak_hash): Declaration updated.
* hash.c (struct hash): Rename flags member to wkopt.
(hash_print_op, hash_mark, do_make_hash, tweak_hash,
make_similar_hash, copy_hash, do_weak_tables): Follow renames.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A weak table with and-semantics expires entries only when both
their key and value is unreachable. When this condition is not
met, therefore, the hash table generates a reference to both
the key and value. This gives rise to a subtlety that must
be be correctly handles in the marking phase.
* hash.c (hash_mark): When marking an and-semantics table,
whenever we find a reachable key or value, we know that the
entry is staying. Therefore we mark it: if the key is
unreachable, we mark the value and vice versa. This is
important because these unreachable objects may be the only
references for reaching reach some other objects via one or
more weak hash tables. Those secondary objects may
spontaneously disappear due to those other hash tables
removing their entries. E.g suppose H0 has and-semantics, and
some K-V entry in H1 has a reachable K, but unreachable V.
Therefore the entry is not eligible for removal, and thus
maintains references to K and V. Suppose V happens to be a key
in a weak-key hash table H1. If, while marking H0, we do not
mark V, then there is a risk that H1 will be processed first
during the later weak procesing stage, and H1 will wrongly
expire its V entry due to the key V being unreachable. Then
when H0 is processed, it will mark V, making it reachable, but
too late: the V entry in H1 is already spuriously gone.
The main principle at play is that an entry in an
and-semantics table strongly holds on to a key if the value is
reachable and vice versa. Only if both are simultaneously
unreachable does it relinquish its references.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Hash tables with weak keys and values now support a choice of both
possible semantics: under and-semantics, an entry lapses when both the
key and value are unreachable. Under or-semantics, an entry lapses if
either the key or value is unreachable.
The and-semantics is new. Until TXR 266, only or-semantics was
supported. This will be the default: when a hash table is specified
as :weak-keys and :weak-vals, it will have or-semantics.
The keywords :weak-or and :weak-and specify weak keys and values,
with the specific semantics. They are utually exclusive, but tolerate
the presence of :weak-keys and :weak-vals.
The make-hash function is being extended such that if its leftmost
argument, <weak-keys>, is specified as one of the keywords :weak-and
or :weak-or, then the hash table will have weak keys and values with the
specified semantics, and the <weak-vals> argument is ignored
(values are weak even if that argument is false).
* eval.c (eval_init): Initially register the top_vb,
top_mb, top_smb, special and builtin hashes as ordinary hashes: no weak
keys or values. Then use tweak_hash to switch to weak keys+vals
with and-semantics. We do it this way because the keywords are not yet
initialized; we cannot use them.
* hash.h (enum hash_flags, hash_flags_t): Moved to header. Member
hash_weak_both renamed to hash_weak_or. New member hash_weak_and.
(weak_and_k, weak_or_k): New keyword variables.
(hash_print_op): Handle hash_weak_and by printing :weak-and.
(hash_mark): Handle hash_weak_and by marking nothing, like hash_weak_or.
(do_make_hash): Check first argument against the two new keywords and
set flags accordingly. This function is called from eval_init before
the keywords have been initialized, in which case weak_keys ==
weak_and_k is true when both are nil; we watch for that.
(tweak_hash): Now returns void and takes a hash_flags_t argument which
is simply planted.
(do_wak_tables): Implement hash_weak_and case. Remove the compat 266
stuff from hash_weak_or. Compatibility is no longer required since we
are not changing the default semantics of hash tables. Phew; that's a
load of worry off the plate.
(hashv): Parse the two new keywords, validate and provide semantics.
(hash_init): Initialize weak_and_k and weak_or_k kewyords.
* hash.h (enum hash_flags, hash_flags_t): Moved here now.
(weak_and_k, weak_or_k): Declared.
* lib.c (compat_fixup): Remove call to parse_compat_fixup.
* parser.c (parse_init): Create stream_parser_hash with and-semantics.
(parse_compat_fixup): Function removed.
* parser.h (parse_compat_fixup): Declaration removed.
* txr.1: Hash documentation updated.
|
|
|
|
|
|
| |
* ffi.c (carray_uint, carray_int): We must use
ffi_type_struct_checked here, otherwise we are blindly
assuming that the element type is a FFI type.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We are doing numerous compat_ver checks in various init
functions, to enact alternative symbol registrations. Only
problem is, compat_ver is always zero during initialization;
it is not set until the -C option is processed in txr_main.
Registrations must be fixed up after initialization;
that's what the compat_fixup mechanism is for.
This is an long-standing problem which affects compatibility
operation going back over 150 versions.
* arith.c (arith_init): Move compat logic to
arith_compat_fixup.
(arith_compat_fixup): New function.
* arith.h (arith_compat_fixup): Declared.
* eval.c (eval_init): Move compat logic to eval_compat_fixup.
* ffi.c (ffi_init): Move compat logic to ffi_compat_fixup.
(ffi_compat_fixup): New function.
* ffi.h (ffi_compat_fixup): Declared.
* regex.c (regex_init): Move compat logic to
regex_compat_fixup.
(regex_compat_fixup): New function.
* regex.h (regex_compat_fixup): Declared.
* stream.c (stream_init): Move compat logic to
stream_compat_fixup.
(stream_compat_fixup): New function.
* stream.h (stream_compat_fixup): Declared.
* struct.c (struct_init): Move compat logic to
struct_compat_fixup.
(struct_compat_fixup): New function.
* struct.h (stream_compat_fixup): Declared.
* lib.c (compat_fixup): Call arith_compat_fixup,
ffi_compat_fixup, regex_compat_fixup, stream_compat_fixup and
struct_compat_fixup.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This addresses the problem that
a4c376979d15323ad729e92e41ba43768e8dc163
tried to fix.
* eval.c (eval_init): Make all the top-level binding tables,
top_fb, top_vb, top_mb, top_smb, special and builtin,
weak-both tables: keys and values are weak.
This way, the entries disappear if both key and value
are unreachable, even if they refer to each other.
(eval_compat_fixup): In 266 or earlier compat mode, weak-both
tables don't have the right semantics, so we tweak
the tables to weak-key tables.
* parser.c (parse_init): Same treatment for
stream_parser_hash. We want an entry to disappear from the
hash if neither the parser nor the stream are reachable.
(parse_compat_fixup): New function.
* parser.h (parse_compat_function): Declared.
* hash.c, hash.h (tweak_hash): New function.
* lib.c (compat_fixup): Call parse_compat_fixup.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
From now on, hash tables with both weak keys and values have
dijunctive retention semantics. If either the key or value of
an entry is reachable, then the entry stays. This is subject
to compatibility.
* hash.c (do_weak_tables): Expire an entry if neither the key
nor the value is reachable. In 266 or lower compatibility
mode, expire an entry if either the key or value is
unreachable, like before.
* txr.1: Document the change, with compat notes. Add a
cautionary note about the referencing issue which defeats weak
key or weak value tables.
|