Fix longjmp across readline w/ --enable-sjlj-exceptions toolchains
authorPedro Alves <palves@redhat.com>
Tue, 20 Dec 2016 15:46:44 +0000 (15:46 +0000)
committerPedro Alves <palves@redhat.com>
Tue, 20 Dec 2016 15:46:44 +0000 (15:46 +0000)
commit2693a26216c329bd7ec2aae7743409f572de4fa5
treee9456ea347835745090a897e127677e61619ce43
parentff71884063d048e8f8c03de27d2bac343df4f77a
Fix longjmp across readline w/ --enable-sjlj-exceptions toolchains

Nowadays, GDB propagates C++ exceptions across readline using
setjmp/longjmp 89525768cd08 ("Propagate GDB/C++ exceptions across
readline using sj/lj-based TRY/CATCH") because DWARF-based unwinding
can't cross C functions compiled without -fexceptions (see details
from the commit above).

Unfortunately, toolchains that use SjLj-based C++ exceptions got
broken with that fix, because _Unwind_SjLj_Unregister, which is put at
the exit of a function, is not executed due to the longjmp added by
that commit.

 (gdb) [New Thread 2936.0xb80]
 kill

 Thread 1 received signal SIGSEGV, Segmentation fault.
 0x03ff662b in ?? ()
 top?bt 15
 #0  0x03ff662b in ?? ()
 #1  0x00526b92 in stdin_event_handler (error=0, client_data=0x172ed8)
    at ../../binutils-gdb/gdb/event-top.c:555
 #2  0x00525a94 in handle_file_event (ready_mask=<optimized out>,
    file_ptr=0x3ff5cb8) at ../../binutils-gdb/gdb/event-loop.c:733
 #3  gdb_wait_for_event (block=block@entry=1)
    at ../../binutils-gdb/gdb/event-loop.c:884
 #4  0x00525bfb in gdb_do_one_event ()
    at ../../binutils-gdb/gdb/event-loop.c:347
 #5  0x00525ce5 in start_event_loop ()
    at ../../binutils-gdb/gdb/event-loop.c:371
 #6  0x0051fada in captured_command_loop (data=0x0)
    at ../../binutils-gdb/gdb/main.c:324
 #7  0x0051cf5d in catch_errors (
    func=func@entry=0x51fab0 <captured_command_loop(void*)>,
    func_args=func_args@entry=0x0,
    errstring=errstring@entry=0x7922bf <VEC_interp_factory_p_quick_push(VEC_inte rp_factory_p*, interp_factory*, char const*, unsigned int)::__PRETTY_FUNCTION__+351> "", mask=mask@entry=RETURN_MASK_ALL)
    at ../../binutils-gdb/gdb/exceptions.c:236
 #8  0x00520f0c in captured_main (data=0x328feb4)
    at ../../binutils-gdb/gdb/main.c:1149
 #9  gdb_main (args=args@entry=0x328feb4) at ../../binutils-gdb/gdb/main.c:1159
 #10 0x0071e400 in main (argc=1, argv=0x171220)
    at ../../binutils-gdb/gdb/gdb.c:32

Fix this by making the functions involved in setjmp/longjmp as
noexcept, so that the compiler knows it doesn't need to emit the
_Unwind_SjLj_Register / _Unwind_SjLj_Unregister calls for C++
exceptions.

Tested on x86_64 Fedora 23 with:
 - GCC 5.3.1 w/ DWARF-based exceptions.
 - GCC 7 built with --enable-sjlj-exceptions.

gdb/ChangeLog:
2016-12-20  Pedro Alves  <palves@redhat.com>
    Yao Qi  <yao.qi@linaro.org>

PR gdb/20977
* event-top.c (gdb_rl_callback_read_char_wrapper_noexcept): New
noexcept function, factored out from ...
(gdb_rl_callback_read_char_wrapper): ... this.
(gdb_rl_callback_handler): Mark noexcept.
gdb/ChangeLog
gdb/event-top.c