On 1/18/22 09:26, Jakub Jelinek wrote:
On Tue, Jan 18, 2022 at 09:08:03AM -0500, Jason Merrill wrote:
which makes call a TARGET_EXPR with the dtor in TARGET_EXPR_CLEANUP,
but then call cxx_constant_value on it. In cxx_eval_outermost_constant_expr
it triggers the:
else if (TREE_CODE (t) != CONSTRUCTOR)
{
r = get_target_expr_sfinae (r, tf_warning_or_error | tf_no_cleanup);
TREE_CONSTANT (r) = true;
}
which wraps the CONSTRUCTOR r into a new TARGET_EXPR, but one without
dtors (I think we need e.g. the TREE_CONSTANT for the callers),
and finally build_over_call uses that.
Looks like you added the tf_no_cleanup in r10-3661 (constexpr new). Maybe
that should only be added if TARGET_EXPR_CLEANUP (t) is null?
I thought about that, but I'm worried that if the cleanup action
is non-empty, then the TREE_CONSTANT (r) = true; after it isn't appropriate,
because the TARGET_EXPR has side-effects.
Hmm. If we e.g. pull out a subobject value from the TARGET_EXPR
temporary, we still need to run the cleanup, so we don't want to
optimize the temporary away, so not setting TREE_CONSTANT sounds right.
And the callers like cxx_constant_value_sfinae just return error_mark_node
if it isn't set:
if (sfinae && !TREE_CONSTANT (r))
r = error_mark_node;
Yes. Is that wrong? I guess we need to be more careful about adjusting
the arguments to cxx_constant_value_sfinae if we still want a constant
value.
Perhaps we also still want to set tf_no_cleanup if object is nonnull,
since in that case we know the prvalue is going to be an initializer
rather than a temporary.
Jason