https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94963

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-05-06
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot 
gnu.org
     Ever confirmed|0                           |1
   Target Milestone|---                         |11.0

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  I've met the underlying issue when developing the patch and for
this reason marked the conditional store inserted by LIM with no-warning.
But for the testcase that's not enough since now PRE comes along and
optimizes the var.field load away, re-exposing the issue.

LIM transforms the testcase to (simplified a bit)

void
f (void)
{
  if (pv != 0)
    {
      bool v2_set = false;
      bool varfield_set = false;
      int v2tem, varfield_tem;
    for (const P *ph = pv; ph < &pv[ps]; ++ph)
      switch (ph->p1)
        {
        case 1:
          v2tem = ph->p2;
          v2_set = true;
          break;
        case 2:
          varfield_tem = ph->p3;
          varfield_set = true;
          break;
        }
      if (varfield_set)
        var.field = varfield_tem;
      if (v2_set)
        v2 = v2tem;
     }
  if (var.field != 0)
    foo (&var);
}

where the uninit predicate analysis doesn't grok the relation between
varfield_set and varfield_tem being initialized.

The patch changed code generation to elide the previously emitted
unconditional load of v2 and var.field.  I suspected that for
the case where there is no load the loop PHI for varfield_tem
might be eliminated, but it is not in all cases it seems.  Now
apart from marking the store no-warning we could easily initialize
the tems on loop entry, just not with their true value but for example
with zero.  That might result in less optimal out-of-SSA though
(no coalescing with constants, the constant move needs to be emitted...)
at least when the loop PHI is not eliminated.

What works is initializing with an uninitialized variable marked
TREE_NO_WARNING.  I'm going to test that (eliding the no-warning
on the conditional stores).

Reply via email to