Consider the following file: __attribute__((visibility("default"))) struct foo { virtual ~foo(); }; struct bar { virtual ~bar(); } __attribute__((visibility("default"))); foo::~foo() { } bar::~bar() { } My expectation would be that classes foo and bar would both get default visibility. That's not what happens: [isolde:tmp]$ /work/root/bin/g++ -fvisibility=hidden -c v5.cc [isolde:tmp]$ nm -m v5.o 00000208 (__TEXT,__text) external __ZN3barD0Ev 00000000 (absolute) external __ZN3barD0Ev.eh 000001a0 (__TEXT,__text) external __ZN3barD1Ev 00000000 (absolute) external __ZN3barD1Ev.eh 00000138 (__TEXT,__text) external __ZN3barD2Ev 00000000 (absolute) external __ZN3barD2Ev.eh 000000d0 (__TEXT,__text) private external __ZN3fooD0Ev 00000000 (absolute) private external __ZN3fooD0Ev.eh 00000068 (__TEXT,__text) private external __ZN3fooD1Ev 00000000 (absolute) private external __ZN3fooD1Ev.eh 00000000 (__TEXT,__text) private external __ZN3fooD2Ev 00000000 (absolute) private external __ZN3fooD2Ev.eh 00000270 (__DATA,__datacoal_nt) weak external __ZTI3bar 00000280 (__DATA,__datacoal_nt) weak private external __ZTI3foo 00000278 (__DATA,__datacoal_nt) weak external __ZTS3bar 00000288 (__DATA,__datacoal_nt) weak private external __ZTS3foo 000002a0 (__DATA,__const) external __ZTV3bar 00000290 (__DATA,__const) private external __ZTV3foo (undefined) external __ZTVN10__cxxabiv117__class_type_infoE (undefined [lazy bound]) external __ZdlPv (undefined) external dyld_stub_binding_helper [isolde:tmp]$ This is certainly wrong. If putting the visibility attribute at the beginning of the class definition is syntactly valid, then that visibility should be respected. If this is a syntax error then it should be diagnosed as such. Silently ignoring the attribute is wrong.
Confirmed.
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.
Subject: Re: Visibility attribute ignored when it precedes class head On Fri, 1 Oct 2004, austern at apple dot com wrote: > 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. It is at least documented to some extent ("Attribute Syntax"), although with a warning that C++ may vary from C. > 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. FWIW, I've been considering such a warning for C, to go along with the warnings for useless type qualifiers and storage class specifiers on empty declarations.
Subject: Bug 17542 CVSROOT: /cvs/gcc Module name: gcc Changes by: austern@gcc.gnu.org 2004-10-30 21:17:32 Modified files: gcc/cp : ChangeLog cp-tree.h decl.c error.c gcc/testsuite : ChangeLog Added files: gcc/testsuite/g++.dg/ext: attrib18.C Log message: PR c++/17542 * cp-tree.h (class_key_or_enum_as_string): Declare. * error.c (class_key_or_enum): Rename to class_key_or_enum_as_string and remove static qualifier. * decl.c (shadow_tag): Warn about ignored attributes in class/struct/ union/enum declaration. * g++.dg/ext/attrib18.C: New test. Patches: http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/cp/ChangeLog.diff?cvsroot=gcc&r1=1.4465&r2=1.4466 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/cp/cp-tree.h.diff?cvsroot=gcc&r1=1.1067&r2=1.1068 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/cp/decl.c.diff?cvsroot=gcc&r1=1.1322&r2=1.1323 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/cp/error.c.diff?cvsroot=gcc&r1=1.268&r2=1.269 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/ChangeLog.diff?cvsroot=gcc&r1=1.4524&r2=1.4525 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/g++.dg/ext/attrib18.C.diff?cvsroot=gcc&r1=NONE&r2=1.1
Fixed.