------- Additional Comments From austern at apple dot com 2004-10-01 23:22 -------
Actually, this is almost straightforward. It has nothing to do with the visibility
attribute: it has to do
with attributes and C++ classes in general. Looking at cp_parser_class, and
especially at
cp_parser_class_head, attributes can appear in one of two places. The parser will
recognize either
struct __attribute__((visibility("hidden"))) foo { virtual ~foo(); };
or
struct foo { virtual ~foo(); } __attribute__((visibility("hidden")));
But, as the code and the comments both make quite clear, the syntax we're recognizing
does not
include an attribute list before the class-key.
So then how come the __atrtribute__ is being swallowed and ignored? Answer: what
we've got here is a
simple-declaration with two decl-specifiers, an attribute list and a class definition,
and no declarators.
The attribute list applies to a declarator, which in this case is missing. We could
instead have written:
__attribute__((visibility("hidden"))) struct foo { virtual ~foo() { } } x;
In this case we can see that the attribute isn't being ignored; it just applies to x,
not to foo.
I hesitate to call this "behaves correctly", since this behavior is unexpected, hard
to understand, and
leads to the user silently not getting what they expected. I'm afraid that with
visibility, in particular, it'll
lead to problems because users will want to hide this attribute list behind macros
that expand to
different things on different platforms. But I'm also not completely sure what the
best thing to do is.
Here are my two two choices:
1. Special-case this construct. If a simple-declaration consists of a class
definition with no declarator,
then any attributes preceding the class head get applied to the class.
2. If cp_parser_simple_declaration collects attributes in
cp_parser_decl_specifier_seq and it's throwing
them away because there's no declarator to apply them to, then warn the user and
suggest a better
place to put the attribute list.
Option 1 is admittedly a hack, but that doesn't necessarily mean it's a bad idea.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17542
