On Tue, Nov 21, 2023 at 04:44:01PM -0500, Jason Merrill wrote:
> > Or do you want to just use
> > error_at (location, "%<static_assert%> message must be a "
> > "unevaluated string literal or object with "
> > "%<size%> and %<data%> members");
> > wording (even when it is in C++26 term) regardless of the -std= level?
>
> No, I think "unevaluated string literal" will be confusing to users. I
> guess it's fine as it is, let's just print the type (as commented inline
> below).
Ok.
> > + error_at (location, "%<static_assert%> message must be a string "
> > + "literal or object with %<size%> and "
> > + "%<data%> members");
>
> Let's print the type of the message as well.
so add " while it has type %qT", TREE_TYPE (message) or something else?
> > + releasing_vec size_args, data_args;
> > + message_sz = finish_call_expr (message_sz, &size_args, false, false,
> > + tf_warning_or_error);
> > + message_data = finish_call_expr (message_data, &data_args, false,
> > false,
> > + tf_warning_or_error);
> > + if (message_sz == error_mark_node || message_data == error_mark_node)
> > + return;
> > + if (tree s
> > + = cp_get_callee_fndecl_nofold (extract_call_expr (message_sz)))
> > + if (!DECL_DECLARED_CONSTEXPR_P (s))
> > + warning_at (location, 0, "%qD used in %<static_assert%> message "
> > + "is not %<constexpr%>", s);
>
> I don't think we need this check, it should be covered by the later
> constant-expression checks.
If the static_assert condition is true, we won't diagnose anything then.
clang++ there incorrectly errors, but I thought a warning could be useful
to users. Perhaps it could warn only if the condition is true?
> > + error_at (location, "%<static_assert%> message %<size()%> "
> > + "must be implicitly convertible to "
> > + "%<std::size_t%>");
>
> Let's also print the type of size().
" while it has type %qT" ?
> > @@ -11485,9 +11544,96 @@ finish_static_assert (tree condition, tr
> > if (processing_template_decl)
> > goto defer;
> > - int sz = TREE_INT_CST_LOW (TYPE_SIZE_UNIT
> > - (TREE_TYPE (TREE_TYPE (message))));
> > - int len = TREE_STRING_LENGTH (message) / sz - 1;
> > + int len;
> > + const char *msg = NULL;
> > + char *buf = NULL;
> > + if (message_sz && message_data)
> > + {
> > + tree msz
> > + = fold_non_dependent_expr (message_sz, complain,
> > + /*manifestly_const_eval=*/true);
>
> We can call cxx_constant_value here instead of fold_non_dependent_expr,
> since we don't get here in a template.
Ok.
> > + t = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (t)), t);
> > + tree t2
> > + = fold_non_dependent_expr (t, complain,
> > + /*manifestly_const_eval=*/
> > + true);
>
> This can also be cxx_constant_value.
>
> > + if (!tree_fits_shwi_p (t2))
> > + {
> > + cxx_constant_value (t);
But in that case I don't have to call it again, right?
> Let's add a comment here about how you're using (data(), 0) to test
> core-constant.
Ok.
Jakub
