[Bug tree-optimization/109945] Escape analysis hates copy elision: different result with -O1 vs -O2

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109945

--- Comment #15 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 24 May 2023, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109945
> 
> --- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #13)
> > with the former for -m64 and the latter for -m32 only seems to be the
> > only fallout here.
> 
> It will penalize C and other languages without mandatory NRV in the FEs,
> without that I think the address can't escape (taking address then would either
> prevent tree-nrv.cc or
> even if not, would be still considered taking address of a local variable).
> Perhaps we could remember in some FUNCTION_DECL bit whether mandatory NRV was
> done and
> least for the cases where we know the callee, we know it hasn't done NRV and
> !TREE_ADDRESSABLE (TREE_TYPE (gimple_call_lhs (stmt)))) we could avoid this.
> But perhaps it is an overkill.

Well, but then the gimplifier doesn't look at the functions implementation
but decides based on the call alone.  It does have

              else if (TREE_CODE (*to_p) != SSA_NAME
                      && (!is_gimple_variable (*to_p)
                          || needs_to_live_in_memory (*to_p)))
                /* Don't use the original target if it's already 
addressable;
                   if its address escapes, and the called function uses 
the
                   NRV optimization, a conforming program could see *to_p
                   change before the called function returns; see 
c++/19317.
                   When optimizing, the return_slot pass marks more 
functions
                   as safe after we have escape info.  */
                use_target = false;

but as we've seen TREE_ADDRESSABLE is not consistently set on the LHS
even when it's eventually passed by reference to the call
(aka aggregate_value_p is true).  It also seems that the gimplifier
will apply RSO when the call is in a INIT_EXPR.

Note the C frontend shows the same non-escaping when massaging the
testcase to

typedef struct {
    int i;
    int a[4];
} Widget;
Widget *global;
Widget make2() { Widget w; w.i = 1; global = &w; return w; }
void g() { global->i = 42; }
int main() {
  Widget w = make2();
  int i = w.i;
  g();
  return (i == w.i);
    // Does this need to be reloaded and
    // compared? or is it obviously true?  
}

then we get

int main ()
{
  int w$i;
  int i;
  struct Widget w;
  int _1;
  _Bool _2;
  int _7;

  <bb 2> :
  w = make2 (); [return slot optimization]
  w$i_9 = w.i;
  i_5 = w$i_9;
  g ();
  _1 = w$i_9;
  _2 = _1 == i_5;
  _7 = (int) _2;
  w ={v} {CLOBBER(eol)};
  return _7;
}

and w not escaped (but it doesn't seem to miscompile then).  Of course
the testcase relies on RSO to be valid in the first place, C doesn't
make any guarantees here and I'm unsure whether C++ guarantees
that for any of the testcases.  As soon as there's a copy involved
the testcases invoke undefined behavior.