Use alignas and unions to specify alignments of objects needing
addresses that are at least a multiple of GCALIGNMENT. Using
these standard C facilities should be safer than relying on ad hoc
and poorly-understood features like GCC’s __attribute__
((aligned (N))), the root cause for recent porting bugs like
Bug#29040. The alignas macro was standardized by C11 and Gnulib
supports alignas for pre-C11 platforms. I have tested this on Sun
Studio 12 sparc (2007) and GCC 4.4.7 x86-64 (2012) as well as on
more recent platforms like GCC 7.2.1 (2017) on Fedora 26 (both
x86-64 and x86).
* lib-src/make-docfile.c (close_emacs_globals): lispsym is now
just an array of struct Lisp_Symbol, since struct Lisp_Symbol is
now properly aligned. All uses changed.
* src/alloc.c (NEXT_FREE_LISP_STRING): Just use the new u.next
member; this is simpler and safer than casting a pointer that
might not be aligned properly.
(aligned_Lisp_Symbol): Remove. No longer needed, now that struct
Lisp_Symbol is aligned properly. All uses replaced with struct
Lisp_Symbol.
* src/lisp.h (GCALIGNED): Remove, as it does not work as expected:
it can cause the natural alignment to be ignored. All uses
replaced by unions with a ‘char alignas (GCALIGNMENT)’ member as
described below.
(struct Lisp_Symbol, struct Lisp_Cons, struct Lisp_String):
Change definition from ‘struct TAG { MEMBERS };’ to
‘struct TAG { union { struct { MEMBERS } s; char alignas
(GCALIGNMENT) gcaligned; } u; };’. This guarantees ‘struct TAG’
to have an alignment that at least max (GCALIGNMENT, N) where N is
its old alignment. All uses like ‘PTR->MEMBER’ changed to
‘PTR->u.s.MEMBER’; these uses were supposed to be mostly private
anyway. Verify that the resulting ‘struct TAG’ is properly
aligned for Emacs.
(union vectorlike_header): New member ‘gcaligned’ to guarantee
that this type, and its containing types like ‘struct Lisp_Subr’,
‘struct buffer’ and ‘struct thread_state’, are all properly
aligned for Emacs.
(struct Lisp_String): New union member ‘next’, for the benefit
of NEXT_FREE_LISP_STRING.
(union Aligned_Cons, union Aligned_String): Remove. All uses
replaced by struct Lisp_Cons and struct Lisp_String, since they
are now properly aligned.
(USE_STACK_CONS, USE_STACK_STRING): Simplify now that we can
assume struct Lisp_Cons and struct Lisp_String are properly
aligned.
Apparently GCC requires that ‘__attribute__ ((aligned (8)))’ must
immediately follow the ‘struct’ keyword when aligning a structure.
The attribute silently does not work if it follows a tag after the
‘struct’ keyword. Who knew? Anyway, this patch is designed to
fix a SIGSEGV problem reported by John Mastro (Bug#29183).
* lib-src/make-docfile.c (close_emacs_globals):
* src/buffer.c (buffer_defaults, buffer_local_symbols):
* src/lisp.h (DEFUN):
* src/thread.c (main_thread):
Put 'GCALIGNED' immediately after 'struct'.
Do not assume that the natural alignment of Lisp objects is a
multiple of GCALIGNMENT. This improves on the portability of the
recent fix for Bug#29040.
* lib-src/make-docfile.c (close_emacs_globals):
* src/buffer.c (buffer_defaults, buffer_local_symbols):
* src/lisp.h (DEFUN):
* src/thread.c (main_thread):
Use GCALIGNED, not alignas (GCALIGNMENT).
* src/alloc.c (COMMON_MULTIPLE):
Move back here from lisp.h, since it is no longer used elsewhere.
* src/lisp.h (GCALIGNMENT): No longer a macro, since we need not
worry about MSVC. Omit no-longer-needed consistency check.
* src/thread.c (THREAD_ALIGNMENT): Remove.
* src/lisp.h (COMMON_MULTIPLE): Move here from alloc.c.
* src/thread.c (THREAD_ALIGNMENT): New macro.
(main_thread): Use THREAD_ALIGNMENT to align propertly. (Bug#29040)
* msdos/sed1v2.inp (GETADDRINFO_A_LIBS, LIBLCMS2, XDBE_LIBS)
(XDBE_FLAGS, HYBRID_MALLOC, LIBSYSTEMD_CFLAGS)
(LIBSYSTEMD_LIBS): Edit to empty.
(LIBRESOLV, DEPFLAGS, MKDEPDIR, YMF_PASS_LDFLAGS)
(PRE_EDIT_LDFLAGS, POST_EDIT_LDFLAGS): Remove editing.
Remove editing of lines that are no longer present in
src/Makefile.in.
* msdos/sed2v2.inp (NEED_MKTIME_INTERNAL)
(NEED_MKTIME_WORKING): Define to 1.
(HAVE_STRUCT_DIRENT_D_TYPE): Define to 1 for
DJGPP >= 2.05.
(HAVE_STRUCT_ATTRIBUTE_ALIGNED): Define to 1.
Define PACKAGE_VERSION, not VERSION.
(FLEXIBLE_ARRAY_MEMBER): Define to empty.
(HAVE_DECL_*_UNLOCKED): Define to 0.
(HAVE___BUILTIN_FRAME_ADDRESS): Define to 1.
(PENDING_*): Don't define, as Gnulib no longer supports that.
Instead, define _IOERR as it is in libc/file.h.
* msdos/sed3v2.inp: Use $(CURDIR) instead of $(shell cd) to
determine the current directory.
(UPDATE_MANIFEST, UTILITIES): Don't edit.
* msdos/sedlisp.inp (FIND_DELETE): Edit to "-delete".
* msdos/sedlibmk.inp (AUTO_DEPEND): Define to yes.
(HYBRID_MALLOC): Edit to empty.
(am__cd): Don't edit.
(../config.status): Replaces $(top_builddir)/config.status.
Define OMIT_GNULIB_MODULE_foo = true for modules not built for
MS-DOS. Convert GL_GENERATE_xxx_H_TRUE and
GL_GENERATE_xxx_H_FALSE into values of GL_GENERATE_xxx_H.
* msdos/mainmake.v2 (src): Use 'compile-one-process', and make
the command line shorter to fit into 126-char limit of
command.com.
* config.bat: Generate src/deps/*.d files. Rename more files
like djtar on plain DOS would.
Don't rename src/dir.h: it is long gone. Edit
lib/gnulib.mk.in using the same scripts as for
lib/Makefile.in.
* msdos/depfiles.bat: Create *.d files, not *.Po.
* src/thread.c (Fmake_thread) [!THREADS_ENABLED]: Improve the
error message.
* src/thread.h [MSDOS]: Include <signal.h>.
* src/sysselect.h (select) [MSDOS]: Undefine, to avoid
compilation errors.
* src/sysdep.c (block_interrupt_signal, restore_signal_mask):
Expose to MSDOS build.
* src/process.c (update_processes_for_thread_death)
[!subprocess]: No-op implementation.
[HAVE_SETRLIMIT]: Move inclusion of sys/resource.h and
declaration of nofile_limit outside "#ifdef subprocesses", as
it's needed for MSDOS.
* src/msdos.c (faccessat): Declare fullname[].
* src/msdos.h (ENOTSUP): Define to be identical to ENOSYS.
Include termhooks.h.
* src/conf_post.h [WINDOWSNT]: Include ms-w32.h only on
WINDOWSNT, not DOS_NT.
* admin/admin.el (set-version): Set version on PACKAGE_VERSION.
* src/thread.h (m_getcjmp): New member of 'struct thread_state'.
(getcjmp): Define to current thread's 'm_getcjmp'.
* src/thread.c (maybe_reacquire_global_lock): Switch to main
thread, since this is called from a SIGINT handler, which always
runs in the context of the main thread.
* src/lisp.h (sys_jmp_buf, sys_setjmp, sys_longjmp): Move the
definitions before thread.h is included, as thread.h now uses
sys_jmp_buf.
* src/keyboard.c (getcjmp): Remove declaration.
(read_char): Don't call maybe_reacquire_global_lock here.
(handle_interrupt): Call maybe_reacquire_global_lock here, if
invoked from the SIGINT handler, to make sure
quit_throw_to_read_char runs with main thread's Lisp bindings and
uses the main thread's jmp_buf buffer. (Bug#28630)
Most of this change is to boilerplate commentary such as license URLs.
This change was prompted by ftp://ftp.gnu.org's going-away party,
planned for November. Change these FTP URLs to https://ftp.gnu.org
instead. Make similar changes for URLs to other organizations moving
away from FTP. Also, change HTTP to HTTPS for URLs to gnu.org and
fsf.org when this works, as this will further help defend against
man-in-the-middle attacks (for this part I omitted the MS-DOS and
MS-Windows sources and the test tarballs to keep the workload down).
HTTPS is not fully working to lists.gnu.org so I left those URLs alone
for now.
Problem reported by Steve Kemp (Bug#27585).
* src/eval.c (near_C_stack_top): Remove. All uses replaced
by current_thread->stack_top.
(record_in_backtrace): Set current_thread->stack_top.
This is for when the Lisp interpreter calls itself.
* src/lread.c (read1): Set current_thread->stack_top.
This is for recursive s-expression reads.
* src/print.c (print_object): Set current_thread->stack_top.
This is for recursive s-expression printing.
* src/thread.c (mark_one_thread): Get stack top first.
* src/thread.h (struct thread_state.stack_top): Now void *, not char *.
* src/thread.c (lisp_mutex_lock_for_thread): New function,
with all the guts of lisp_mutex_lock.
(lisp_mutex_lock): Call lisp_mutex_lock_for_thread.
(condition_wait_callback): Don't call post_acquire_global_lock
before locking the mutex, as that could cause a signaled thread to
exit prematurely, because the condvar's mutex is recorded to be
not owned by any thread, and with-mutex wants to unlock it as part
of unwinding the stack in response to the signal.
* src/thread.c (last_thread_error): New static variable.
(syms_of_threads): Staticpro it.
(record_thread_error, Fthread_last_error): New functions.
(syms_of_threads): Defsubr Fthread_last_error.
* doc/lispref/threads.texi (Basic Thread Functions): Document
thread-last-error.
* test/src/thread-tests.el (thread-errors, thread-signal-early)
(threads-condvar-wait): Test the values returned by
thread-last-error.
* src/thread.c (lisp_mutex_lock, lisp_mutex_unlock)
(lisp_mutex_unlock_for_wait, condition_wait_callback)
(condition_notify_callback): Improve commentary.
(condition_wait_callback): Call post_acquire_global_lock before
attempting to lock the mutex, to make sure the lock's owner is
recorded correctly.
* test/src/thread-tests.el (threads-condvar-wait): New test.
This avoids the confusion of using two different phrases "main thread"
and "primary thread" internally to mean the same thing. See:
http://lists.gnu.org/archive/html/emacs-devel/2016-12/msg01142.html
* src/thread.c (main_thread): Rename from primary_thread,
since the new name no longer clashes with main_thread_id
and Emacs internals normally call this the "main thread".
(init_main_thread): Rename from init_primary_thread.
(main_thread_p): Rename from primary_thread_p.
All uses changed.
This improves performance overall on my benchmark on x86-64,
since the interpreted program-counter resides in a machine
register rather than in RAM.
* etc/DEBUG, src/.gdbinit: Remove xbytecode GDB command, as there
is no longer a byte stack to decode.
* src/bytecode.c (struct byte_stack, byte_stack_list)
(relocate_byte_stack): Remove. All uses removed.
(FETCH): Simplify now that pc is now local (typically, in a
register) and no longer needs to be relocated.
(CHECK_RANGE): Remove. All uses now done inline, in a different way.
(BYTE_CODE_QUIT): Remove; now done by op_relative_branch.
(exec_byte_code): Allocate a copy of the function’s bytecode,
so that there is no problem if GC moves it.
* src/lisp.h (struct handler): Remove byte_stack member.
All uses removed.
* src/thread.c (unmark_threads): Remove. All uses removed.
* src/thread.h (struct thread_state): Remove m_byte_stack_list member.
All uses removed. m_stack_bottom is now the first non-Lisp field.
* src/thread.h (struct thread_state): New member not_holding_lock.
(maybe_reacquire_global_lock): Add prototype.
* src/thread.c: Include syssignal.h.
(maybe_reacquire_global_lock): New function.
(really_call_select): Set the not_holding_lock member of the
thread state before releasing the lock, and rest it after
re-acquiring the lock when the select function returns. Block
SIGINT while doing this to make sure we are not interrupted on TTY
frames.
* src/sysdep.c (block_interrupt_signal, restore_signal_mask): New
functions.
* src/syssignal.h (block_interrupt_signal, restore_signal_mask):
Add prototypes.
* src/keyboard.c (read_char) [THREADS_ENABLED]: Call
maybe_reacquire_global_lock. (Bug#25178)
* src/thread.c (post_acquire_global_lock): Don't raise the pending
signal if the thread's handlers were not yet set up, as that will
cause Emacs to exit with a fatal error. This can happen if a
thread is signaled as soon as make-thread returns, before the new
thread had an opportunity to acquire the global lock, set up the
handlers, and call the thread function.
* test/src/thread-tests.el (thread-signal-early): New test.
* src/thread.c (post_acquire_global_lock): Call
set_buffer_internal_2 instead of tricking set_buffer_internal_1
into resetting the current buffer even if it didn't change. This
avoids bug#25165, caused by failing to record the modified values
of point and mark, because current_buffer was set to NULL. Also,
don't bother re-setting the buffer if there was no thread switch,
as that just wastes cycles.
* src/buffer.c (set_buffer_internal_2): New function, with most of
the body of set_buffer_internal_1, but without the test for B
being identical to the current buffer.
(set_buffer_internal_1): Call set_buffer_internal_2 if B is not
identical to the current buffer.
* src/buffer.h (set_buffer_internal_2): Add prototype.
* test/src/thread-tests.el (thread-sticky-point): New test.
* src/thread.c (mark_one_thread): Use NILP to compare with
m_saved_last_thing_searched, which is a Lisp object. Reported by
Andreas Politz <politza@hochschule-trier.de>.
src/eval.c (unbind_for_thread_switch): Accept a 'struct
thread_state *' argument and use specpdl_ptr and specpdl of that
thread. Fixes crashes if find_symbol_value signals an error.
src/thread.c (post_acquire_global_lock): Update current_thread
before calling unbind_for_thread_switch. Pass the previous thread
to unbind_for_thread_switch.
This introduces the thread_alive_p macro and changes
thread-alive-p to use it. This is a minor cleanup.
It also changes all-threads to ignore dead threads.
