This GC bug seems to have been fixed, so the check is no longer
needed in production code. From a suggestion by Pip Cet in:
https://lists.gnu.org/r/emacs-devel/2019-08/msg00316.html
* src/alloc.c (SUSPICIOUS_OBJECT_CHECKING) [!ENABLE_CHECKING]:
Do not define.
(find_suspicious_object_in_range, detect_suspicious_free):
Expand to proper dummy expressions if !SUSPICIOUS_OBJECT_CHECKING.
* src/data.c (Ffset): Convert test to an eassert.
This uses ‘const’ to be better at catching bugs that
mistakenly attempt to modify a bignum value.
Lisp bignums are supposed to be immutable.
* src/alloc.c (make_pure_bignum):
* src/fns.c (sxhash_bignum):
Accept Lisp_Object instead of struct Lisp_Bignum *, as that’s
simpler now. Caller changed.
* src/bignum.h (bignum_val, xbignum_val): New inline functions.
Prefer them to &i->value and XBIGNUM (i)->value, since they
apply ‘const’ to the result.
* src/timefns.c (lisp_to_timespec): Use mpz_t const *
to point to a bignum value.
* src/alloc.c (free_cons):
* src/casefiddle.c (do_casify_multibyte_string):
* src/editfns.c (styled_format):
* src/image.c (png_load_body):
Remove recent workarounds for INT_ADD_WRAPV bugs since
the bugs have been fixed (Bug#37006).
This builds on the previous patch.
* src/alloc.c (consing_until_gc): Now of type intmax_t,
since gc-cons-threshold can be up to INTMAX_MAX. All uses changed.
* src/lisp.h (CONSING_CT_MAX, consing_ct): Remove.
Suggested by Eli Zaretskii (Bug#37006#46).
* src/alloc.c (consing_until_gc): Now of type consing_ct.
All uses changed, so gc-cons-threshold no longer saturates
against OBJECT_CT_MAX.
(object_ct): Move typedef here from lisp.h.
* src/lisp.h (consing_ct, CONSING_CT_MAX): New type and macro.
(OBJECT_CT_MAX): Remove. Replace all uses with CONSING_CT_MAX.
Problem reported by Eli Zaretskii (Bug#37006#25).
* src/alloc.c (garbage_collect_1): Fix typo in threshold calc.
Go back to dividing by 10 since the numerator’s a constant now.
Problem introduced in 2019-07-21T02:40:03Z!eggert@cs.ucla.edu.
* src/alloc.c (free_cons):
* src/casefiddle.c (do_casify_multibyte_string):
* src/editfns.c (styled_format):
* src/image.c (png_load_body):
Use signed arguments to INT_MULTIPLY_WRAPV etc. This doesn’t fix
any bugs, but GCC emits better code when all args are signed.
Also, this removes the need for an if in free_cons (Bug#37006).
* src/alloc.c (c_symbol_p): Move from here ...
* src/lisp.h (c_symbol_p): ... to here, and make it more portable
to hypothetical platforms where pointers are wider than ptrdiff_t.
* src/pdumper.c (dump_builtin_symbol_p): Use c_symbol_p.
Say that pdumping cannot redump unless -batch is used. Say that
the traditional unexec dumping method is by default not available,
and is deprecated. Don't call dump files "portable", as dump files
are not any more portable than the Emacs executables themselves.
Just call them "dump files". Similar, prefer "portable dumper"
(since the dumper code is portable) to "portable dumping" (since
the dump file is not). Be more systematic about calling them
"dump files" instead of "dumped images" or whatnot.
Problem reported by Pip Cet in:
https://lists.gnu.org/r/emacs-devel/2019-07/msg00520.html
* src/alloc.c (inhibit_garbage_collection): Use new function.
(allow_garbage_collection): Accept intmax_t, not pointer.
* src/eval.c (default_toplevel_binding, do_one_unbind)
(backtrace_eval_unrewind, Fbacktrace__locals, mark_specpdl):
Support SPECPDL_UNWIND_INTMAX.
(record_unwind_protect_excursion): New function.
* src/lisp.h (enum specbind_tag): New constant SPECPDL_UNWIND_INTMAX.
(union specbinding): New member unwind_intmax.
* src/alloc.c (allow_garbage_collection)
(inhibit_garbage_collection): Temporarily bump
consing_until_gc, to improve performance of maybe_gc while
garbage collection is inhibited. Suggested by Stefan Monnier in:
https://lists.gnu.org/r/emacs-devel/2019-07/msg00511.html
* src/alloc.c (pure_alloc): Clear any heap-allocated storage.
This is simpler than auditing all the callers to make sure
they don’t assume pure memory is cleared memory, and the
performance implication is nonexistent except when Emacs
is misconfigured. Also, add an assertion to catch
caller misuse when pure space is exhausted.
* src/fns.c (restore_mutability)
(hash_table_user_defined_call): New functions.
(cmpfn_user_defined, hashfn_user_defined): Use them.
(make_hash_table, copy_hash_table):
Mark new hash table as mutable.
(check_mutable_hash_table): New function.
(Fclrhash, Fputhash, Fremhash): Use it instead of CHECK_IMPURE.
* src/lisp.h (struct hash_table_test): User-defined functions
now take pointers to struct Lisp_Hash_Table, not to struct
hash_table_test. All uses changed.
(struct Lisp_Hash_Table): New member ‘mutable’.
* src/pdumper.c (dump_hash_table): Copy it.
* test/src/fns-tests.el (test-hash-function-that-mutates-hash-table):
New test, which tests for the bug.
Without this patch, there are unlikely ways that garbage
collection could occur (sometimes causing undefined behavior)
even when inhibit_garbage_collection is in effect.
* src/alloc.c (garbage_collection_inhibited): New var.
(pure_alloc): Increment it if pure space is exhausted, so that
garbage_collect_1 no longer needs to inspect
pure_bytes_used_before_overflow.
(allow_garbage_collection): New function.
(inhibit_garbage_collection): Increment the new variable rather
than specbinding a user variable.
(garbage_collect_1): Do not garbage collect if the new variable
is set, rather than if pure_bytes_used_before_overflow is set.
* src/alloc.c (consing_until_gc): New variable, replacing the
combination of consing_since_gc and gc_relative_threshold.
All uses changed.
(byte_ct): Move decl here from lisp.h.
(memory_full_cons_threshold): New an enum constant.
(free_cons): Check for integer overflow in
statistics calculation.
* src/lisp.h (object_ct): Move decl here from alloc.c.
(OBJECT_CT_MAX): New macro.
(maybe_gc): Simplify accordingly.
* src/lisp.h (struct Lisp_Hash_Table): Rename ‘pure’ member to
‘purecopy’, as the old name was quite confusing (it did not
mean the hash table was pure). All uses changed.
This speeds up ‘make compile-always’ by 0.1% on my platform.
Suggested by Pip Cet in:
https://lists.gnu.org/r/emacs-devel/2019-07/msg00257.html
* src/.gdbinit (pwinx, pgx, xbuffer, xprintstr):
Output dead_object () as "DEAD".
* src/alloc.c (Vdead, DEADP): Remove.
All uses replaced by dead_object () / deadp.
(deadp): New function.
(init_alloc_once_for_pdumper): Remove no-longer-needed
initialization.
* src/lisp.h (dead_object): New function.
Without this fix, Emacs can get into a tight loop reporting
a range error when calculating timestamps.
* doc/lispref/numbers.texi (Integer Basics):
* src/alloc.c (syms_of_alloc): Document this.
* src/bignum.c (make_bignum_bits): Always allow bignums
of at least twice the width of (u)intmax_t.
The GCC + valgrind fix caused the m68k build to fail (Bug#35711).
Simplify string allocation a bit to make similar problems less
likely in the future.
* src/alloc.c (sdata, SDATA_NBYTES, SDATA_DATA) [GC_CHECK_STRING_BYTES]:
Use the same implementation as with !GC_CHECK_STRING_BYTES,
as the special case is no longer needed.
(SDATA_ALIGN): New constant.
(SDATA_SIZE): Remove this macro, replacing with ...
(sdata_size): ... this new function. All uses changed.
Properly account for sizes and alignments even in the m68k case,
and even if GC_CHECK_STRING_BYTES is not defined.
* src/alloc.c: No need to include stdalign.h; it’s pervasive.
(GC_STRING_OVERRUN_COOKIE_SIZE): Align to sdata’s alignment,
so that the code works even if alignof (sdata) exceeds 8.
Don’t require the cookie size to be 8, as this overly fattens
32-bit platforms and one DEADBEEF should be enough.
(GC_STRING_EXTRA): Omit now-unnecessary ‘verify’.
(allocate_string_data): Omit unnecessary cast.
Issue found by Clang’s UBSan.
* src/alloc.c (GC_STRING_OVERRUN_COOKIE_SIZE): Increase to 8.
(string_overrun_cookie): Extend accordingly.
(GC_STRING_EXTRA): Ensure that it’s properly aligned for ‘sdata’.
(allocate_string_data): Verify that ‘sdata’ blocks remain aligned.
It doesn’t work anymore, and these days ‘gcc -fsanitize=address’
does a better job anyway.
* configure.ac: Remove the option.
* configure.ac (ac_xmalloc_overrun, XMALLOC_OVERRUN_CHECK):
* src/alloc.c (XMALLOC_OVERRUN_CHECK_OVERHEAD)
(XMALLOC_OVERRUN_CHECK_SIZE, XMALLOC_OVERRUN_SIZE_SIZE)
(xmalloc_overrun_check_header, xmalloc_overrun_check_trailer)
(xmalloc_put_size, xmalloc_get_size, overrun_check_malloc)
(overrun_check_realloc, overrun_check_free):
Remove. All uses removed.
* etc/NEWS: Mention this.
Because XPNTR now uses ATTRIBUTE_NO_SANITIZE_UNDEFINED,
it can be simplified.
* src/alloc.c (macro_PNTR_ADD, PNTR_ADD, macro_XPNTR): Remove.
(XPNTR): Open-code rather than using the removed macros and
functions. Also, simplify by using LISP_WORD_TAG.
* src/alloc.c (XPNTR): Add ATTRIBUTE_NO_SANITIZE_UNDEFINED and
remove ATTRIBUTE_UNUSED. Do not define as a macro, so that
ATTRIBUTE_NO_SANITIZE_UNDEFINED works.
* src/lisp.h (lisp_h_XSYMBOL): Remove. All uses removed.
With recent GCC the macro does not work with -fsanitize=undefined,
and the macro can be omitted as its only function is to optimize -O0.
Even if Emacs is compiled without module support, we don't have to
comment out every bit of user pointer support. Defining the basic
structures and functions and detecting user pointers in switch
statements is harmless, and we're already doing the same for module
functions. Removing these #ifdefs makes the code a bit easier to
read.
* src/lisp.h (PVEC_USER_PTR, struct Lisp_User_Ptr, USER_PTRP)
(XUSER_PTR): Define unconditionally.
* src/data.c (Ftype_of):
* src/alloc.c (cleanup_vector):
* src/print.c (print_vectorlike):
* src/pdumper.c (dump_vectorlike): Remove #ifdef for user pointers.
This partially reverts commit
09b2b8a5ce.
* src/alloc.c (mark_maybe_pointer): Remove special-casing of tagged
pointers. After commit 09d746dad3,
modules no longer rely on tagged pointers.
Problem reported by Keith David Bershatsky in:
https://lists.gnu.org/r/emacs-devel/2019-04/msg00259.html
Solution suggested by Stefan Monnier in:
https://lists.gnu.org/r/emacs-devel/2019-04/msg00282.html
* src/buffer.h (BUFFER_LISP_SIZE): Simplify by using PSEUDOVECSIZE.
(BUFFER_REST_SIZE): Simplify by using VECSIZE and BUFFER_LISP_SIZE.
* src/lisp.h (PSEUDOVECSIZE): Base it on the last Lisp field,
not the first non-Lisp field. All callers changed. Callers
without Lisp fields changed to use ALLOCATE_PLAIN_PSEUDOVECTOR.
(ALLOCATE_PLAIN_PSEUDOVECTOR): New macro.
* src/.gdbinit: Simplify by removing dependency on globals
implementation. This is useful for a future performance
improvement that I have in mind.
* src/alloc.c (enum defined_HAVE_X_WINDOWS, defined_HAVE_X_WINDOWS):
New enum and constant.
(gdb_make_enums_visible) [__GNUC__]: Use it, to make
defined_HAVE_X_WINDOWS visible to GDB.
Problem found on 32-bit sparc, which has stricter alignment
checking than x86-64.
* src/alloc.c (pure_alloc): When TYPE is negative it now specifies
the negation of the required alignment of the result.
(make_pure_bignum): Specify bignum limb alignment.
Add ‘const’ to make the GC marking code a bit clearer.
This can also help the compiler in some cases, I think because
GCC can now determine more often that the value of a static C
variable can be cached when its address is now converted to
‘Lisp Object const *’ before escaping.
* src/alloc.c (staticvec, mark_maybe_objects, mark_memory)
(mark_stack, staticpro, mark_object_root_visitor)
(garbage_collect_1):
* src/pdumper.c (dump_ptr_referrer, dump_emacs_reloc_to_lv)
(dump_emacs_reloc_to_emacs_ptr_raw, dump_root_visitor):
* src/lisp.h (vcopy, struct gc_root_visitor):
* src/sysdep.c (stack_overflow):
* src/thread.c (mark_one_thread):
* src/thread.h (struct thread_state):
Use pointer-to-const instead of plain pointer in some
GC-related places where either will do.
The main new thing here is that C code can now say
‘list (a, b, c, d, e, f)’ instead of
‘listn (CONSTYPE_HEAP, 6, a, b, c, d, e, f)’,
thus relieving callers of the responsibility of counting
arguments (plus, the code feels more like Lisp). The old
list1 ... list5 functions remain, as they’re probably a bit
faster for small lists.
* src/alloc.c (cons_listn, pure_listn): New functions.
(listn): Omit enum argument.
All callers changed to use either new ‘list’ or ‘pure_list’ macros.
* src/charset.c (Fdefine_charset_internal):
* src/coding.c (detect_coding_system)
(Fset_terminal_coding_system_internal):
* src/frame.c (frame_size_history_add, adjust_frame_size):
* src/gtkutil.c (xg_frame_set_char_size):
* src/keyboard.c (command_loop_1):
* src/nsfns.m (frame_geometry):
* src/widget.c (set_frame_size):
* src/xfaces.c (Fcolor_distance):
* src/xfns.c (frame_geometry):
* src/xterm.c (x_set_window_size_1):
* src/xwidget.c (Fxwidget_size_request):
Prefer list1i, list2i, etc. to open-coding them.
* src/charset.c (Fset_charset_priority):
* src/nsterm.m (append2):
* src/window.c (window_list):
* src/xfaces.c (Fx_list_fonts):
Use nconc2 instead of open-coding it.
* src/eval.c (eval_sub, backtrace_frame_apply):
* src/kqueue.c (kqueue_generate_event):
* src/nsterm.m (performDragOperation:):
* src/pdumper.c (Fpdumper_stats):
* src/w32.c (init_environment):
Prefer list1, list2, etc. to open-coding them.
* src/font.c (font_list_entities):
Parenthesize to avoid expanding new ‘list’ macro.
* src/gtkutil.c (GETSETUP): Rename from MAKE_FLOAT_PAGE_SETUP
to get lines to fit. Move outside the ‘list’ call, since it’s
now a macro.
* src/keymap.c (Fmake_keymap): Simplify.
* src/lisp.h (list, pure_list): New macros.
(list1i): New function.
