Revert changes to w32proc.c done since 2012-11-30T09:23:15Z!eliz@gnu.org.
Do NOT merge to trunk!
src/w32proc.c (new_child): Remove the loop that attempted to salvage
slots of dead processes.
(new_child, delete_child, find_child_pid): Don't insist on
procinfo.hProcess be NULL for a process slot to be considered
inactive, or be non-NULL to be considered active.
(reader_thread): Don't set the FILE_AT_EOF flag of a file
descriptor if the corresponding child was deleted and its
char_avail handle is NULL.
(reap_subprocess): Don't reset the FILE_AT_EOF flag.
(sys_select): Don't pass a NULL process handle to
WaitForMultipleObjects.
Fixes: debbugs:13735 debbugs:13546
src/xdisp.c (x_draw_vertical_border): For a window that is neither
the leftmost nor the rightmost, redraw both the left and the right
vertical borders.
src/w32proc.c (new_child): Free up to 2 slots of dead processes at a
time. Improve diagnostics in DebPrint.
(reader_thread): If cp->char_avail is NULL, set the FILE_AT_EOF
flag, so that sys_select could have a chance of noticing that this
process is dead, and call a SIGCHLD handler for it. Improve
diagnostics in DebPrint.
(reap_subprocess): Reset the FILE_AT_EOF flag set by
reader_thread.
(sys_select): Watch a process whose procinfo.hProcess is non-NULL
even if its char_avail is NULL. Allows to reap subprocesses that
were forcibly deleted by delete-process. (Bug#13546)
src/w32.c (sys_socket, sys_bind, sys_connect, sys_gethostname)
(sys_gethostbyname, sys_getservbyname, sys_getpeername)
(sys_shutdown, sys_setsockopt, sys_listen, sys_getsockname)
(sys_accept, sys_recvfrom, sys_sendto, fcntl): In case of failure,
make sure errno is set to an appropriate value. (Bug#13546)
(socket_to_fd): Add assertion against indexing fd_info[] with a
value that is out of bounds.
(sys_accept): If fd is negative, do not set up the child_process
structure for reading.
The bug was reported for AIX before today's changes.
I reproduced the problem on Fedora 17 x86-64 when setting
GC_MARK_STACK by hand, and I presume it occurs with default
configurations on HP-UX and Unixware.
Trunk fix on 2013-01-14 by Dmitry Antipov <dmantipov@yandex.ru>:
Fix compilation with GC_MARK_STACK == GC_USE_GCPROS_AS_BEFORE.
* eval.c (eval_sub): Protect `form' from being GCed before its
car and cdr becomes protected with the backtrace entry.
Fixes: debbugs:13650
* configure.ac (DATA_START, DATA_SEG_BITS): Set to 0x20000000 on AIX.
(GC_MARK_STACK): Do not set to GC_USE_GCPROS_AS_BEFORE, as that
runs afoul of some other bug in Emacs, and the default value
GC_MAKE_GCPROS_NOOPS has been tested and works.
* src/lisp.h (XPNTR) [!USE_LSB_TAG && DATA_SEG_BITS]:
Fix bug introduced in 2012-07-27 change. DATA_SEG_BITS, if set,
was #undeffed earlier, so it cannot be used as a macro here.
Use the constant and not the macro. Tested on AIX.
* src/unexaix.c: Revert 2013-02-11 and 2013-02-12 changes to this
file. They're almost surely OK but we're just before a release so
we should avoid changes unless they're clearly needed. Instead,
make the following minor change:
(ADDR_CORRECT): New macro.
Fixes: debbugs:13650
forward-sexp movement in header documentation.
(python-nav--forward-sexp): Behave like emacs-lisp-mode in
comments and strings (GH bug 114).
Fixes: debbugs:13642
src/w32.c (sys_pipe): When failing due to file descriptors above
MAXDESC, set errno to EMFILE.
(_sys_read_ahead): Update cp->status when failing to read serial
communications input, so that the status doesn't stay at
STATUS_READ_IN_PROGRESS.
With this, it should be as good as it was in 23.3, though it's
still pretty bad: the dumped emacs does not run. See Mark Fleishman in
http://lists.gnu.org/archive/html/help-gnu-emacs/2011-04/msg00287.html
* unexaix.c (start_of_text): Remove.
(_data, _text): Declare as char[], not int, as AIX manual suggests.
(bias, lnnoptr, text_scnptr, data_scnptr, load_scnptr)
(orig_load_scnptr, orig_data_scnptr):
Now off_t, not long, since they are file offsets.
(make_hdr): Use _data, not start_of_data ().
This is the key part of the fix.
(make_hdr, unrelocate_symbols): Use off_t for file offsets.
(unrelocate_symbols): Cast pointers to intptr_t, not to ulong.
Fixes: debbugs:13650
This doesn't fix the bug, but it makes progress: Emacs builds now.
* unexaix.c: Include inttypes.h, stdarg.h.
(report_error, report_error_1): Mark as _Noreturn.
(report_error): Don't report the wrong errno.
(report_error_1): Now varargs. All callers changed.
(make_hdr): Use uintptr_t, not unsigned, when converting pointers
to unsigned. Don't use ADDR_CORRECT, as it no longer exists.
(write_ptr): Use %p to print address rather than %lx and a cast
to unsigned long. Grow buffer a bit, to be safer.
Fixes: debbugs:13650
src/bidi.c (bidi_resolve_neutral): After finding the next
non-neutral character, accept NEUTRAL_ON type as well, because
directional control characters, such as LRE and RLE, have their
type converted to that by bidi_resolve_weak. This avoids aborts
when LRE/RLE follows a run of neutrals.
(bidi_move_to_visually_next): Assert that return value of
bidi_peek_at_next_level is non-negative. Negative values will
cause an infloop.
