[Bug middle-end/77295] missed optimisation when copying/moving union members

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77295

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2016-08-19
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Why does the implementation not simply copy the whole union with an assignment?
It's easier for GCC to optimize that.

Because with the current one you'll have

  <bb 2>:
  MEM[(struct  &)this_7(D)] ={v} {CLOBBER};
  _1 = x_9(D)->which;
  this_7(D)->which = _1;
  _2 = x_9(D)->which;
  switch (_2) <default: <L2>, case 1: <L0>, case 2: <L1>>

<L0>:
  _3 = x_9(D)->D.2254.v1;
  this_7(D)->D.2254.v1 = _3;
  goto <bb 5> (<L2>);

<L1>:
  _4 = x_9(D)->D.2254.v2;
  this_7(D)->D.2254.v2 = _4;

<L2>:
  return;

where it's hard to argue the desired transform is profitable in general
(x->which might get known via inlining and if you'd optimize the above
earlier then you'd at least lose TBAA info on the stores / loads).

Also note sth else above - we read from x->which twice (this may point
to x).  There's a related PR somewhere about this missed optimization,
with a patch (that has cost even if that special situation doesn't apply,
thus I didn't pursue it).