[Bug tree-optimization/41186] New: VN doesn't look through non-aliasing by offset memcpy

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

The C++ FE now creates memcpy for class copies to avoid copying objects in
tail-padding.

struct Foo {
  Foo() {};
  int i;
  short f;
};
struct Bar : public Foo {
  Bar() {};
  short b;
};

extern "C" void abort(void);

int main()
{
  Bar b1, b2;
  b2.i = 0;
  b1.f = 0;
  b1.b = 1;
  b2.f = 1;
  b2.b = 2;
  static_cast<Foo&>(b1) = static_cast<Foo&>(b2);
  if (b1.i != 0 || b1.b != 1)
    abort ();
  if (b1.f != 1)
    abort ();
  return 0;
}

The alias-oracle does consider the memcpy to clobber b1.b and the
look-through doesn't find b1.f or b1.i either.

Note that with combining the tests like

  if (b1.i != 0 || b1.b != 1 || b1.f != 1)
    abort ();

fold creates a BIT_FIELD_REF covering both b and f which confuses us
even more.  But I have considered this a bad habit of fold earlier.


-- 
           Summary: VN doesn't look through non-aliasing by offset memcpy
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Keywords: missed-optimization, alias
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: rguenth at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41186