https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81112
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> Ah, no. This is ipa-prop.c special code (bah).
>
> Index: gcc/ipa-prop.c
> ===================================================================
> --- gcc/ipa-prop.c (revision 249245)
> +++ gcc/ipa-prop.c (working copy)
> @@ -3030,7 +3030,10 @@ find_constructor_constant_at_offset (tre
>
> if (index)
> {
> - off = wi::to_offset (index);
> + if (TREE_CODE (index) == RANGE_EXPR)
> + off = wi::to_offset (TREE_OPERAND (index, 0));
> + else
> + off = wi::to_offset (index);
> if (TYPE_DOMAIN (type) && TYPE_MIN_VALUE (TYPE_DOMAIN (type)))
> {
> tree low_bound = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
> @@ -3039,6 +3042,9 @@ find_constructor_constant_at_offset (tre
> TYPE_PRECISION (TREE_TYPE (index)));
> }
> off *= wi::to_offset (unit_size);
> + if (TREE_CODE (index) == RANGE_EXPR)
> + off *= wi::sub (wi::to_offset (TREE_OPERAND (index, 1)),
> + wi::to_offset (TREE_OPERAND (index, 0)));
> }
> else
> off = wi::to_offset (unit_size) * ix;
>
> Fixes this.
Actually the 2nd hunk is bogus, and w/o it we'll simply handle RANGE_EXPR
conservatively (only matching with the very first index in it). I'm
testing that.
Enhancements can be done on trunk.
