Re: general protection faults with "git grep" version 1.7.7.1

From: Thomas Rast <trast@student.ethz.ch>
Date: Oct 25, 2011, 16:00 UTC
Message-ID: <201110251800.28054.trast@student.ethz.ch>
In-Reply-To: <87y5w9ayoa.fsf@rho.meyering.net>

Jim Meyering wrote:

Show 7 quoted lines

> Thomas Rast wrote:
> > [GCC moves access to a file-static variable across pthread_mutex_lock()]
> 
> Thanks for the investigation.
> Actually, isn't gcc -O2's code-motion justified?
> While we *know* that those globals may be modified asynchronously,
> builtin/grep.c forgot to tell gcc about that.

I'm somewhat unwilling to believe that:

* "volatile" enforces three unrelated things, see e.g. [1].

* Removing "static" would do the same as it prevents the compiler from
  proving at compile-time that pthread_mutex_lock() cannot affect the
  variable in question.

  If this is correct, it also means that all code in all pthreads
  tutorials I can find works merely by the accident of not declaring
  their variables "static".

  Furthermore, a future smarter compiler with better link-time
  optimization might again prove the same and eliminate the
  "superfluous" load.

However, as a result of the discussion I now have a shorter testcase:

  #include <pthread.h>

  int y;
  static int x;

  pthread_mutex_t m = PTHREAD_MUTEX_INITIALIZER;

  void test ()
  {
          y = x;
          pthread_mutex_lock(&m);
          x = x + 1;
          pthread_mutex_unlock(&m);
  }

GCC 4.6.1 on F16 again assumes 'x' was not modified across the lock. I also tested GCC 4.5.1 and 4.4.5, which instead issue a direct add-to-memory instruction

        addl    $1, x(%rip)

in the locked part.

In the event that you and GCC 4.6.1 are right, I still vote for removing 'static' instead of adding 'volatile' so as to allow basic optimizations.

-- 
Thomas Rast
trast@{inf,student}.ethz.ch