https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79937
--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The CONSTRUCTORS in TARGET_EXPR in the pr79937-{1,2,3}.C testcases have all
CONSTRUCTOR_IS_DIRECT_INIT and TREE_HAS_CONSTRUCTOR set, while nsdmi13.C
doesn't.
Does any of those matter?
In the nsdmi13.C case, there is just a single replace_placeholders call on the
whole testcase, so if we don't replace everything, we die during
gimplification,
but in pr79937-{1,2,3}.C the gimplification of the CONSTRUCTOR results in
cp_gimplify_init_expr and calls it there and handles the replacements that
aren't otherwise done.
