[Bug tree-optimization/60320] New: Redundant static initialization check

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60320

            Bug ID: 60320
           Summary: Redundant static initialization check
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org

Hello,

I think it would be nice if g++ realized that static variables that have been
initialized don't later on become uninitialized again. A simple example:

extern int e;
static int b(){
  static const int a = e;
  return a;
}
int g(){
  return b()-b();
}

gets "optimized" to:

<bb 2>:
  _7 = MEM[(char *)&_ZGVZL1bvE1a];
  if (_7 == 0)
    goto <bb 3>;
  else
    goto <bb 9>;

  <bb 3>:
  _8 = __cxa_guard_acquire (&_ZGVZL1bvE1a);
  if (_8 != 0)
    goto <bb 5>;
  else
    goto <bb 4>;

  <bb 4>:
  pretmp_26 = a;
  goto <bb 6>;

  <bb 5>:
  e.2_9 = e;
  a = e.2_9;
  __cxa_guard_release (&_ZGVZL1bvE1a);

  <bb 6>:
  # prephitmp_27 = PHI <e.2_9(5), pretmp_26(4)>
  _11 = MEM[(char *)&_ZGVZL1bvE1a];
  if (_11 == 0)
    goto <bb 7>;
  else
    goto <bb 9>;

  <bb 7>:
  _12 = __cxa_guard_acquire (&_ZGVZL1bvE1a);
  if (_12 != 0)
    goto <bb 8>;
  else
    goto <bb 9>;

  <bb 8>:
  e.2_13 = e;
  a = e.2_13;
  __cxa_guard_release (&_ZGVZL1bvE1a);
  pretmp_5 = prephitmp_27 - e.2_13;

  <bb 9>:
  # prephitmp_3 = PHI <0(6), _12(7), pretmp_5(8), 0(2)>
  return prephitmp_3;


There may be a dup but I couldn't find it. It doesn't seem that easy to teach
gcc about it. Maybe LTO-inlining of __cxa_guard_* functions would help (or
not). We could emit _ZGVZL1bvE1a=1 after the release call, but that seems ugly.
We could have special code that, for a MEM_REF[&var42], looks for a dominating
call to __cxa_guard_release(&var42), and in that case asserts it is non-zero
(is that, or a slight variant, valid?).