[erlang-patches] Per-thread FPE pointer not being initialized

[Raimo Niskanen]

On Wed, Jul 01, 2009 at 12:03:04PM +0200, Mikael Pettersson wrote:
> David Reiss writes:
>  > The per-thread last floating point exception pointer is not initialized to
>  > NULL when it is allocated.  As a result, spurious floating point
>  > exceptions are reported.
>  > 
>  > This is 100% reproducible for me on CentOS 5.1 for x86-64, with Erlang/OTP
>  > R13B built from source with gcc 2.5 and gcc 4.1.2.  It doesn't happen with
>  > R13B01, but, based on the nature of the bug, I think that is most likely
>  > due to chance.
>  > 
>  > Steps to reproduce:
>  > 
>  > 1/ Start up a node with
>  >   dreiss@REDACTED:~$ erl -noshell -name bob@`hostname`
>  > 
>  > 2/ Build and compile test program
>  >   dreiss@REDACTED:~$ cat test.erl
>  >   -module(test).
>  >   -export([start/1]).
>  >   start([Remote]) ->
>  >     rpc:call(Remote, erlang, is_float, [750.0]).
>  >   dreiss@REDACTED:~$ erlc test.erl
>  > 
>  > 3/ Run test program
>  >   dreiss@REDACTED:~$ erl -noshell -name alice@`hostname` -pz . -s test
>  > start bob@`hostname` -s init stop
>  > 
>  > The "alice" node exits with no output.  After running the "alice" command,
>  > the "bob" node prints the following
>  > 
>  >   ERTS_FP_CHECK_INIT at 0x4aaded: detected unhandled FPE at 0x2
>  > 
>  > Patch:
>  > 
>  > diff otp_src_R13B/erts/emulator/sys/unix/sys_float.c
>  > otp_src_R13B/sys/unix/sys_float.c/sys/unix/sys_float.c
>  > --- otp_src_R13B/erts/emulator/sys/unix/sys_float.c	2009-06-30
>  > 17:53:31.000000000 -0700
>  > +++ otp_src_R13B/erts/emulator/sys/unix/sys_float.c	2009-06-30
>  > 17:53:31.000000000 -0700
>  > @@ -57,7 +57,8 @@
>  >  void erts_thread_init_fp_exception(void)
>  >  {
>  >      unsigned long *fpe = erts_alloc(ERTS_ALC_T_FP_EXCEPTION, sizeof(*fpe));
>  > +    *fpe = 0;
>  >      erts_tsd_set(fpe_key, fpe);
>  >  }
> 
> Thanks, that bug was mine :-( Your fix is correct.

Would your patch in http://www.erlang.org/cgi-bin/ezmlm-cgi/2/1380
also solve this problem? Or are both needed?

--- otp-0622/erts/emulator/sys/unix/sys_float.c.~1~	2009-02-26 11:45:48.000000000 +0100
+++ otp-0622/erts/emulator/sys/unix/sys_float.c	2009-06-26 11:14:15.000000000 +0200
@@ -792,6 +792,8 @@ sys_chars_to_double(char* buf, double* f
 int
 matherr(struct exception *exc)
 {
+#if !defined(NO_FPE_SIGNALS)
     set_current_fp_exception((unsigned long)__builtin_return_address(0));
+#endif
     return 1;
 }



> 
> I believe the initialisation here wasn't needed originally,
> but became necessary as a side effect of some FP exn machinery
> redesign I did late last year.
> 
> Where did the spurious FP exception get detected, i.e.,
> what runtime system procedure does 0x4aaded correspond to?
> 
> /Mikael
> 
> ________________________________________________________________
> erlang-patches mailing list. See http://www.erlang.org/faq.html
> erlang-patches (at) erlang.org

-- 

/ Raimo Niskanen, Erlang/OTP, Ericsson AB