optimization bug in gfortran 6.3?

[James Spencer]

Hi,

Sorry in advance if this isn't the appropriate forum but I am confused
with behaviour I see with gcc 6.2 and 6.3 compared to 5.2 and 7.1. I
haven't been able to cut this down to a minimal test case but the
critical lines are lines 794-797 in
https://github.com/hande-qmc/hande/blob/master/src/energy_evaluation.F90#L794.

With gcc 6.3-O3 (similar behaviour with gcc 6.2 and other optimization
levels), -fdump-tree-optimized gives for this region:

  <bb 10>:
  _64 = &excitation_8(D)->from_orb[1];
  _65 = &excitation_8(D)->from_orb[0];
  ij_sym_66 = cross_product_basis_read_in (sys_15(D), _65, _64);
  _67 = &excitation_8(D)->to_orb[1];
  _68 = &excitation_8(D)->to_orb[0];
  ab_sym_69 = cross_product_basis_read_in (sys_15(D), _68, _67);
  if (ij_sym_66 == ab_sym_69)
    goto <bb 11>;
  else
    goto <bb 12> (<L23>);

  <bb 11>:
  _70 = &excitation_8(D)->perm;
  *hmatel_6(D) = slater_condon2_mol_excit (sys_15(D), _65, _64, _68,
_67, _70);
  _72 = *estimators_10(D).proj_energy;
  _73 = *pop.0_5[0];
  _74 = _73 * 0.0;
  _75 = _72 + _74;
  *estimators_10(D).proj_energy = _75;

I am not familiar with the syntax given here but it seems that rather
than accumulating hmatel%r*pop(1), 0.0*pop(1) is instead accumulated.
Given slater_condon2_mol_excit is not a compile-time constant, this
looks like an optimization bug to me. (A debug build behaves as
expected.)

With the same optimization settings on gcc 7.1, -fdump-tree-optimized gives

  <bb 10> [9.42%]:
  _47 = &excitation_64(D)->from_orb[1];
  _48 = &excitation_64(D)->from_orb[0];
  ij_sym_66 = cross_product_basis_read_in (sys_65(D), _48, _47);
  _49 = &excitation_64(D)->to_orb[1];
  _50 = &excitation_64(D)->to_orb[0];
  ab_sym_67 = cross_product_basis_read_in (sys_65(D), _50, _49);
  if (ij_sym_66 == ab_sym_67)
    goto <bb 11>; [37.68%]
  else
    goto <bb 12> (<L25>); [62.32%]

  <bb 11> [3.55%]:
  _51 = &excitation_64(D)->perm;
  *hmatel_62(D) = slater_condon2_mol_excit (sys_65(D), _48, _47, _50,
_49, _51);
  _52 = *estimators_69(D).proj_energy;
  _53 = *hmatel_62(D).r;
  _54 = *pop.0_61[0];
  _55 = _53 * _54;
  _56 = _52 + _55;
  *estimators_69(D).proj_energy = _56;

(The same output for gfortran 5.2 looks similar.)

Here the accumulation is being performed as I would expect .

I tried to find a bug report for this in the bug tracker but failed.
Some questions:

1. Is this indeed a (known) bug?
2. If so, where was it fixed between gfortran 6.3 and 7.1? When was it
introduced?
3. If so, can the fix be backported to the 6 release series?

Thanks,

    James Spencer