Re: [PATCH] [python] Avoid saving signal masks for Python API.

[Pedro Alves <palves at redhat dot com>]

On 02/05/2015 01:40 AM, Yichun Zhang (agentzh) wrote:
> Saving the siginal mask in sigsetjmp involes the expensive
> sigprocmask syscall while in our Python API, it is not likely
> for the C primitives to mess up with signals. Also, value_rtti_type
> is heavily called from within Python API, which does not mess up with
> signals either (feel free to prove me wrong).
> 
> This change leads to 11%+ speedup in our "lgcstat" [1] GDB python script
> and 14%+ speedup in our "lgcpath" [2] script while Python 3.2 is used.
> 
> Alas, even the sigsetjmp calls without signal mask saving still
> introduce significant performance overhead (also for the surrounding
> code). Nevertheless, this patch can serve as a good start anyway and
> we can try further optimizations along this way in the future.
> 
> If you have any better suggestions, please don't hesitate to drop
> me a note.

Thanks.  Impressive the speed we can get out of this.

It seems to me this tackles things from the wrong end though.

I think we should enumerate the cases where we throw
from a signal handler, and treat _those_ specially.

Off the top of my head, that will be when immediate_quit it set,
the SIGFPE handler, though that seems to be broken anyway,
and when calling into the demangler and catch_demangler_crashes
is set (though that uses explicit sigsetjmp instead of TRY_CATCH),
and maybe the various SIGINT handlers.

We'll need to contain these for C++ too.

I'd be curious if you already see the speedup with this series

 https://sourceware.org/ml/gdb-patches/2015-02/msg00202.html

given that that switches TRY/CATCH to use C++ exceptions / Dwarf
unwinding instead of sigsetjmp/longjmp.

Thanks,
Pedro Alves