[Bug tree-optimization/54906] New: write introduction incorrect wrt the C++11 memory model (case with atomic accesses)

[francesco.zappa.nardelli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54906

             Bug #: 54906
           Summary: write introduction incorrect wrt the C++11 memory
                    model (case with atomic accesses)
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: francesco.zappa.nardelli@gmail.com


The program below is miscompiled by g++ --param allow-store-data-races=0 -O2
(or -O3).

$ g++ -v
gcc version 4.8.0 20121011 (experimental) (GCC) 

#include <atomic>
using namespace std;
uint8_t g_5;
atomic_ushort a_9;
atomic_schar a_24;

void func_1 () {
  if (a_9.load ())
    for (g_5 = 0; 0; g_5++) {
    }
  a_24.store (0);
}

int main () {
  func_1 ();
  return 0;
}

The assembly code generated by -O3 for func_1 is:

_Z6func_1v:
        movzwl  a_9(%rip), %edx
        xorl    %eax, %eax
        testw   %dx, %dx
        movzbl  g_5(%rip), %edx
        cmove   %edx, %eax
        movb    %al, g_5(%rip)
        movb    $0, a_24(%rip)
        mfence
        ret

This code loads and restores the global variable g_5, while the reference
semantics for the source program does not perform a write to g_5.  It is easy
to write a non-racy context that observes this unexpected behaviour.

[Bug 54900 might be related or not; in the example here the write is introduced
only if a_24 is an atomic variable].