[Bug tree-optimization/87901] partial DSE of memset doesn't work for other kind of stores

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87901

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The problem is simple.

```
      /* We want &lhs for the MEM_REF expression.  */
      tree lhs_addr = build_fold_addr_expr (gimple_assign_lhs (stmt));

      if (! is_gimple_min_invariant (lhs_addr))
        return;

```
We just need a STRIP_USELESS_TYPE_CONVERSION there; At least is what is needed
for the testcase in comment #3 which I think is enough for PR 36602 really.

We get this now:
```
  MEM <char[8]> [(char * {ref-all})&z] = {};
  MEM <unsigned char[24]> [(char * {ref-all})&z + 8B] = {};
```

For the testcase in comment 2, it is harder and not currently handled and I am
not sure it should be here; store merging handles it otherwise.

Note for s/17/16/ we still fail; I will handle that afterwards.