[Bug middle-end/91606] [9/10 regression] Optimization leads to invalid code

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91606

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This is miscompiled during dse1 (or the IL is already wrong earlier).
Before dse1 we have in bb2
  <bb 2> :
  D.10827._M_elems[0].D.10649.f1.__pfn = fun1;
  D.10827._M_elems[0].D.10649.f1.__delta = 0;
  D.10827._M_elems[0].index = 0;
  D.10827._M_elems[1].D.10649.f2.__pfn = fun2;
  D.10827._M_elems[1].D.10649.f2.__delta = 0;
  D.10827._M_elems[1].index = 1;

  <bb 3> :
  # idx_6 = PHI <0(2), idx_27(9)>
  # __for_begin_7 = PHI <&MEM <struct variant[2]> [(void *)&D.10827](2),
__for_begin_28(9)>
  if (&MEM <struct variant[2]> [(void *)&D.10827 + 48B] == __for_begin_7)
    goto <bb 10>; [INV]
  else
    goto <bb 4>; [INV]

  <bb 4> :
  _18 = MEM[(const struct variant *)__for_begin_7].index;
  if (_18 != 0)
    goto <bb 9>; [17.38%]
  else
    goto <bb 5>; [82.62%]

  <bb 5> :
  _31 = &MEM[(const struct variant *)__for_begin_7].D.10649.f1;
  _1 = _31->__pfn;
  _2 = memFuncPtr.__pfn;
  if (_1 == _2)
and so in the first iteration it goes bb2, bb3, bb4 and bb5 and bb5 reads
_31->__pfn and thus D.10827._M_elems[0].D.10649.f1, but DSE1 removes the
  D.10827._M_elems[0].D.10649.f1.__pfn = fun1;
  D.10827._M_elems[0].D.10649.f1.__delta = 0;
and
  D.10827._M_elems[1].D.10649.f2.__pfn = fun2;
  D.10827._M_elems[1].D.10649.f2.__delta = 0;
stores.