[Bug tree-optimization/84824] DCE fails to remove dead code of std::function constructor

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84824

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |alias, missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2018-03-12
                 CC|                            |rguenth at gcc dot gnu.org
          Component|c++                         |tree-optimization
     Ever confirmed|0                           |1

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that printf () clobbers global memory and thus the std::function
object:

  <bb 2> [100.00%]:
  MEM[(struct __lambda0 *)&D.30227] = 1;
  D.30227._M_invoker = _M_invoke;
  D.30227.D.29944._M_manager = _M_manager;
  __builtin_printf ("%d", 3);

  <bb 3> [100.00%]:
  _5 = MEM[(struct _Function_base *)&D.30227]._M_manager;
  if (_5 != 0B)
    goto <bb 4>; [94.39%]
  else
    goto <bb 5>; [5.61%]

  <bb 4> [94.39%]:
  _5 (&MEM[(struct _Function_base *)&D.30227]._M_functor, &MEM[(struct
_Function_base *)&D.30227]._M_functor, 3);
..

we fail to CSE the load from ._M_manager across printf for this reason.
Note that D.30227 later escapes and given our "IPA mod-ref" analysis
is context insensitive this is enough to not optimize it.

I'm not sure if any std::function reorg could help with this issue.  Better
escape analysis might.