https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123818
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Anyway, for the casts, I haven't succeeded making a testcase that would be
miscompiled.
This particular bug is in the CONSTRUCTOR handling,
--- gcc/tree.cc.jj 2026-01-17 14:37:22.171617643 +0100
+++ gcc/tree.cc 2026-02-03 16:16:02.637247532 +0100
@@ -6826,13 +6826,27 @@ simple_cst_equal (const_tree t1, const_t
vec<constructor_elt, va_gc> *v2 = CONSTRUCTOR_ELTS (t2);
if (vec_safe_length (v1) != vec_safe_length (v2))
- return false;
+ return 0;
for (idx = 0; idx < vec_safe_length (v1); ++idx)
- /* ??? Should we handle also fields here? */
- if (!simple_cst_equal ((*v1)[idx].value, (*v2)[idx].value))
- return false;
- return true;
+ {
+ if ((*v1)[idx].index
+ && TREE_CODE ((*v1)[idx].index) == FIELD_DECL)
+ {
+ if ((*v1)[idx].index != (*v2)[idx].index)
+ return 0;
+ }
+ else
+ {
+ cmp = simple_cst_equal ((*v1)[idx].index, (*v2)[idx].index);
+ if (cmp <= 0)
+ return cmp;
+ }
+ cmp = simple_cst_equal ((*v1)[idx].value, (*v2)[idx].value);
+ if (cmp <= 0)
+ return cmp;
+ }
+ return 1;
}
case SAVE_EXPR:
should fix that. The index stuff isn't strictly necessary for this bug, I've
just included it because the comment mentioned that.
The actual bug was that simple_cst_equal returns tristate rather than bool, 1
for known to be equal, 0 for known to be not equal and -1 for don't know,
contains stuff I don't understand (which is here AGGR_INIT_EXPR). The cmp =
simple_cst_equal (...); if (cmp <= 0) return cmp; is what is used elsewhere for
calls which aren't tail recursions.
