https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79232
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |jsm28 at gcc dot gnu.org
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
On the gimplifier side, this is caused by the PR17855 changes which look weird.
We have COND_EXPR <cond, INDIRECT_REF <something>, INDIRECT_REF
<somethingelse>> = whatever;
(of course the question is if the FE should emit that and not INDIRECT_REF
<COND_EXPR <cond, something, somethingelse>> = whatever), and there are 2
problems. Before the
/* C99 code may assign to an array in a structure value of a
conditional expression, and this has undefined behavior
only on execution, so create a temporary if an lvalue is
required. */
if (fallback == fb_lvalue)
{
*expr_p = get_initialized_tmp_var (*expr_p, pre_p, post_p,
false);
mark_addressable (*expr_p);
ret = GS_OK;
}
block for COND_EXPR we have *expr_p a MEM_REF with iftmp.0 as address, exactly
what we need. Without the mark_addressable (not sure why that was ever needed)
we would actually end up writing to a temporary variable (first store there the
previous value of the MEM_REF, then again the value we want into the MEM_REF).
But the mark_addressable, if *expr_p has gimple ref type, actually causes also
invalid IL in addition to wrong-code.
I've tried the struct-non-lval-2.c testcase and this code is not reached in
that case (because fallback is just 0, not fb_lvalue).
That said, even if we fix the gimplifier, we probably want to change the FE so
that it actually makes it clear what the store does. For COND_EXPR without the
COMPOUND_EXPR, the FE actually emits if (cond) { (&c)[d] = e; } else { a[d] =
e; }, so we probably want to emit that too.
