On Wed, Jan 23, 2019 at 03:34:04PM -0500, Jason Merrill wrote: > On Wed, Jan 23, 2019 at 12:57 PM Marek Polacek <pola...@redhat.com> wrote: > > > > On Wed, Jan 23, 2019 at 09:00:36AM -0500, Jason Merrill wrote: > > > I was talking about digest_init, not reshape_init. digest_init calls > > > convert_for_initialization. > > > > /facepalm > > > > So yes, digest_init calls convert_for_initialization which will end up > > calling perform_implicit_conversion_flags which could call convert_like_real > > where the narrowing warnings are given, but it doesn't, we go to this case: > > > > else if (processing_template_decl && conv->kind != ck_identity) > > { > > /* In a template, we are only concerned about determining the > > type of non-dependent expressions, so we do not have to > > perform the actual conversion. But for initializers, we > > need to be able to perform it at instantiation > > (or instantiate_non_dependent_expr) time. */ > > expr = build1 (IMPLICIT_CONV_EXPR, type, expr); > > > > finish_decltype_type throws away the expression because it's not dependent, > > and > > only uses its type. So narrowing remains undetected. Not sure if I should > > mess > > with perform_implicit_conversion_flags. > > Let's try that; this is a situation where the comment is incorrect. > Perhaps just call check_narrowing here if appropriate, rather than go > through the whole conversion machinery.
I have not been successful. First, I modified perform_implicit_conversion_flags to go the convert_like route when dealing with something non-dependent. That breaks e.g. in build_value_init: 346 /* The AGGR_INIT_EXPR tweaking below breaks in templates. */ 347 gcc_assert (!processing_template_decl 348 || (SCALAR_TYPE_P (type) || TREE_CODE (type) == ARRAY_TYPE)); Even if I restrict the convert_like way for non-dependent exprs in a template for scalars, it still breaks elsewhere, e.g. constexpr-template3.C where it complains about taking the address of an rvalue. Second, I added check_narrowing to the processing_template_decl case in perform_implicit_conversion_flags. That works except it breaks constexpr-inst1.C -- we no longer get the error. That's because currently check_narrowing in finish_compound_literal calls maybe_constant_init, which calls instantiate_constexpr_fns and we get the desired diagnostic. But if I move check_narrowing to perform_implicit_conversion_flags, we no longer call it in this case -- processing_template_decl is 0 so we call convert_like but that doesn't do the trick. So, back to the patch that leaves check_narrowing in finish_compound_literal? Marek