On Mon, Dec 9, 2019 at 1:23 PM Eric Botcazou <ebotca...@adacore.com> wrote:
>
> Hi,
>
> this is a regression present on the mainline and 9 branch: the compiler gives
> an ICE for the attached Ada testcase on the following assertion:
>
> if (DECL_P (ref))
> {
> /* We shouldn't have true variables here. */
> gcc_assert (TREE_READONLY (ref));
> subst = ref;
> }
>
> in self_referential_size because the size function machinery is invoked again
> by the *free_lang_data pass, more precisely from fld_process_array_type:
>
> if (!existed)
> {
> array
> = build_array_type_1 (t2, TYPE_DOMAIN (t), TYPE_TYPELESS_STORAGE (t),
> false, false);
> TYPE_CANONICAL (array) = TYPE_CANONICAL (t);
> if (!fld->pset.add (array))
> add_tree_to_fld_list (array, fld);
> }
>
> through the call to build_array_type_1, and more precisely through the
> recursive call made for computing TYPE_CANONICAL:
>
> if (TYPE_CANONICAL (t) == t)
> {
> if (TYPE_STRUCTURAL_EQUALITY_P (elt_type)
> || (index_type && TYPE_STRUCTURAL_EQUALITY_P (index_type))
> || in_lto_p)
> SET_TYPE_STRUCTURAL_EQUALITY (t);
> else if (TYPE_CANONICAL (elt_type) != elt_type
> || (index_type && TYPE_CANONICAL (index_type) != index_type))
> TYPE_CANONICAL (t)
> = build_array_type_1 (TYPE_CANONICAL (elt_type),
> index_type
> ? TYPE_CANONICAL (index_type) : NULL_TREE,
> typeless_storage, shared, set_canonical);
> }
>
>
> That's a bit surprising because t2 is an incomplete type and we have these
> lines in build_array_type_1 just before:
>
> /* If the element type is incomplete at this point we get marked for
> structural equality. Do not record these types in the canonical
> type hashtable. */
> if (TYPE_STRUCTURAL_EQUALITY_P (t))
> return t;
>
> so the computation of TYPE_CANONICAL should be skipped. But it turns out that
> these lines from 2009 are obsolete because layout_type no longer forces the
> TYPE_STRUCTURAL_EQUALITY_P on the array when the element type is incomplete.
>
>
> Since fld_process_array_type overwrites TYPE_CANONICAL just after the call to
> build_array_type_1, there is no point for the latter in computing it so the
> proposed fix is to add a new SET_CANONICAL parameter to build_array_type_1.
>
> Tested on x86_64-suse-linux, OK for mainline and 9 branch?
OK.
>
> 2019-12-09 Eric Botcazou <ebotca...@adacore.com>
>
> * tree.c (build_array_type_1): Add SET_CANONICAL parameter and compute
> TYPE_CANONICAL from the element type only if it is set. Remove
> obsolete
> lines and adjust recursive call.
> (fld_process_array_type): Adjust call to build_array_type_1.
> (build_array_type): Likewise.
> (build_nonshared_array_type): Likewise.
>
>
> 2019-12-09 Eric Botcazou <ebotca...@adacore.com>
>
> * gnat.dg/lto23.adb: New test.
>
> --
> Eric Botcazou
