[Bug tree-optimization/108953] inefficient codegen for trivial equality

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108953

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c++                         |tree-optimization
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I don't think there is anything the C++ FE should do here, user can write it
like that
by hand too:
bool check(C const& lhs, C const& rhs) {
  if (lhs.a != rhs.a) return false;
  if (lhs.b != rhs.b) return false;
  if (lhs.c != rhs.c) return false;
  if (lhs.d != rhs.d) return false;
  if (lhs.e != rhs.e) return false;
  if (lhs.f != rhs.f) return false;
  if (lhs.g != rhs.g) return false;
  return true;
}
etc., so we should pattern match this somewhere during GIMPLE opts.

The memcmp issue is a RTL issue (so I think we want actually two bugs), at
GIMPLE level it is optimized into:
  _1 = __builtin_memcmp_eq (lhs_3(D), rhs_4(D), 12);
  _5 = _1 == 0;
  return _5;
which is I think what we want.

The reason for doing conditional branch after each comparison is
targetm.compare_by_pieces_branch_ratio (DImode) tells it not to batch more than
one.
Not really sure if we should override it to something else and on which
targets.

The reason for the weird
        mov     eax, 1
        test    eax, eax
        sete    al
instead of
        xor     eax, eax
and
        xor     eax, eax
        test    eax, eax
        sete    al
instead of
        mov     eax, 1
is that the epilogue is threaded only too late where the constant propagation
through it isn't possible.