[Bug c++/105260] Union with user-defined empty destructor leads to worse code-gen

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105260

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |jamborm at gcc dot gnu.org
           Keywords|ABI, wrong-code             |

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I don't see any ABI differences, nor wrong-code.
If you look at *.optimized dump differences, there are pretty much none, except
for one extra CLOBBER.
But there are 2 differences not visible in the dump.
One is that while neither un variable is TREE_ADDRESSABLE, the no WORKAROUND
one has TREE_ADDRESSABLE type (as it has non-trivial destructor).
The other difference that matters for this testcase is that TYPE_MODE of the
NoDestroy union type is BLKmode, while for WORKAROUND=1 case it is DImode.

This is both done in:
  /* If this type has a copy constructor or a destructor, force its
     mode to be BLKmode, and force its TREE_ADDRESSABLE bit to be
     nonzero.  This will cause it to be passed by invisible reference
     and prevent it from being returned in a register.  */
  if (type_has_nontrivial_copy_init (t)
      || TYPE_HAS_NONTRIVIAL_DESTRUCTOR (t))
    {
      tree variants;
      SET_DECL_MODE (TYPE_MAIN_DECL (t), BLKmode);
      for (variants = t; variants; variants = TYPE_NEXT_VARIANT (variants))
        {
          SET_TYPE_MODE (variants, BLKmode);
          TREE_ADDRESSABLE (variants) = 1;
        }
    }
Now, for the ABI passing the above is really required, such types are indeed
passed differently from ones that have trivialy copy ctor and dtor, but for
this testcase it is just an optimization decision (use_register_for_decl).
We do there:
  /* Only register-like things go in registers.  */
  if (DECL_MODE (decl) == BLKmode)
    return false;
Guess we'd need to recompute mode for non-TREE_ADDRESSABLE vars before
expansion.
Or find out why SRA doesn't optimize this (remove the useless union, replace
all the un.value occurrences with a var with Foo type.