<div dir="ltr">Hi Evgenii,<div><br></div><div>I was wondering what the status is of your work to attach "internal_linkage" to methods of standard library classes in libc++. The following piece of code doesn't link because a symbol (string::empty) is undefined and it sounds like your work might fix the linkage error I'm seeing.</div><div><br></div><div>$ cat test1.cpp</div><div><div>#include <string></div><div>#include <functional></div><div><br></div><div>int main() {</div><div> using namespace std::placeholders;</div><div><span style="white-space:pre"> </span>auto b = std::bind(&std::string::empty, _1);</div><div>}</div></div><div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Nov 9, 2015 at 1:58 PM, Evgenii Stepanov via llvm-dev <span dir="ltr"><<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex">In fact, during the code review we've decided against allowing this<br>
attribute on namespaces: <a href="http://reviews.llvm.org/D13925" rel="noreferrer" target="_blank">http://reviews.llvEm.org/D13925</a><br>
<br>
The intended use of this attribute in libc++ is on functions that<br>
programs are not allowed to take address of, like all methods of<br>
standard library classes. Visibility("hidden") attribute (which we are<br>
replacing with internal_linkage) violates ODR the same way.<br>
<br>
Also, this attribute (for us) is not about locality or performance.<br>
Its primary purpose is to reliably exclude a class method from the<br>
export table. This is the way libc++ manages ABI stability: all class<br>
methods that are not part of the library ABI are defined in the<br>
headers with a set of attributes that ensure that they become internal<br>
to the translation unit. Always_inline does that in almost all cases,<br>
but almost is not good enough. Internal_linkage is a straighforward<br>
way to achieve the same.<br>
<div><div><br>
<br>
On Mon, Nov 9, 2015 at 1:34 PM, Martin J. O'Riordan<br>
<<a href="mailto:martin.oriordan@movidius.com" target="_blank">martin.oriordan@movidius.com</a>> wrote:<br>
> With respect only to '__attribute__((internal_linkage))', not 'nodebug' and other parts of this topic; does hiding "some" members and not others not introduce a violation of the ODR because some members of the class as it appears in one translation unit are not the same actual definitions as the apparently "same" members of the class in another translation unit? In particular the ODR requirement that would ensure that the address of a method was the same regardless of where the address was taken, including 'static' methods (no 'this' pointer)? Or hidden things like the guard variables for local static objects, default argument initialisers (which are instanced), etc.?<br>
><br>
> I must admit that having spent years trying to lock down the definition of the ODR, that this seems to fly in the opposite direction of the intent of that rule.<br>
><br>
> I can (sort of) see why for particular reasons of locality you might want multiple actual definitions, even though they might exhibit the same behaviour, but maybe the answer for locality optimisation would be for the linker technology to evolve so that multiple copies of the same definition (sic) could be made; perhaps even at the Basic Block level [Block Level Linking]. But this too would lie under the auspices of the "as if" rule and respect the ODR.<br>
><br>
> There is something I am missing in this that is not making me at all comfortable about the use of this attribute for namespaces (apart from the reopening issues raised by others), and in particular the application to parts of a class but not the whole class. I just feel that this is leading to semantic trouble.<br>
><br>
> Thanks,<br>
><br>
> MartinO<br>
><br>
> -----Original Message-----<br>
> From: Evgenii Stepanov [mailto:<a href="mailto:eugeni.stepanov@gmail.com" target="_blank">eugeni.stepanov@gmail.com</a>]<br>
> Sent: 09 November 2015 18:49<br>
> To: Robinson, Paul <<a href="mailto:Paul_Robinson@playstation.sony.com" target="_blank">Paul_Robinson@playstation.sony.com</a>><br>
> Cc: Chris Lattner <<a href="mailto:clattner@apple.com" target="_blank">clattner@apple.com</a>>; <a href="mailto:sstewartgallus00@mylangara.bc.ca" target="_blank">sstewartgallus00@mylangara.bc.ca</a>; <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>; Martin J. O'Riordan <<a href="mailto:martin.oriordan@movidius.com" target="_blank">martin.oriordan@movidius.com</a>><br>
> Subject: Re: [llvm-dev] [cfe-dev] [RFC] __attribute__((internal_linkage))<br>
><br>
> Sorry, I totally missed updates to this thread.<br>
><br>
> Anonymous namespace does not work for libc++ because we need to hide some class members, and export the rest. Namespace would hide all of them. AFAIK, "static" is un-deprecated in C++11 for multiple reasons, including this one.<br>
><br>
> I agree that nodebug should not affect codegen. It would also kill debug info (ex. all debug locations in libc++ headers).<br>
><br>
><br>
> On Mon, Nov 2, 2015 at 11:57 AM, Robinson, Paul via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>
>><br>
>><br>
>>> -----Original Message-----<br>
>>> From: llvm-dev [mailto:<a href="mailto:llvm-dev-bounces@lists.llvm.org" target="_blank">llvm-dev-bounces@lists.llvm.org</a>] On Behalf Of<br>
>>> Chris Lattner via llvm-dev<br>
>>> Sent: Saturday, October 31, 2015 2:58 PM<br>
>>> To: <a href="mailto:sstewartgallus00@mylangara.bc.ca" target="_blank">sstewartgallus00@mylangara.bc.ca</a><br>
>>> Cc: LLVM Developers<br>
>>> Subject: Re: [llvm-dev] [cfe-dev] [RFC]<br>
>>> __attribute__((internal_linkage))<br>
>>><br>
>>> Hi Stewart,<br>
>>><br>
>>> I saw this get brought up at the LLVM devmtg in the context of<br>
>>> improving the stack size of libc++ frames.<br>
>>><br>
>>> Have you guys considered a different approach to solving this<br>
>>> problem? In the case of libc++, the _LIBCPP_INLINE_VISIBILITY macro<br>
>>> currently expands out to "__attribute__ ((__always_inline__))”. It<br>
>>> seems reasonable to me to have it also add the “nodebug” attribute as<br>
>>> well. This should allow the allocas generated by inlining lots of<br>
>>> frames to be promoted to registers (because there is no debug info to pessimize).<br>
>><br>
>> Are you suggesting that 'nodebug' should affect code generation?<br>
>> It most definitely should not...<br>
>> --paulr<br>
>><br>
>>><br>
>>> This would dramatically shrink the stack frames of code using libc++<br>
>>> at - O0, and would also make it go a lot faster.<br>
>>><br>
>>> -Chris<br>
>>><br>
>>><br>
>>> > On Oct 29, 2015, at 6:35 AM, Martin J. O'Riordan via llvm-dev<br>
>>> > <llvm-<br>
>>> <a href="mailto:dev@lists.llvm.org" target="_blank">dev@lists.llvm.org</a>> wrote:<br>
>>> ><br>
>>> > I haven't been able to figure out from this thread why this<br>
>>> > attribute is<br>
>>> necessary at all.<br>
>>> ><br>
>>> > Anonymous or unnamed namespaces were added to C++ for this very<br>
>>> > purpose,<br>
>>> but the ISO C++ Standard does not discuss "linkage" per-se because it<br>
>>> is outside the scope of the language specification. But it does<br>
>>> describes it in terms of having a hidden name which is "unique" to<br>
>>> the translation unit, and under the "as if" rule, internal linkage is<br>
>>> a common convention for achieving this.<br>
>>> ><br>
>>> > This is just Standardese for dealing with what compiler<br>
>>> > implementers<br>
>>> typically handle as "internal linkage" anyway; at the same time, the<br>
>>> use of 'static' for this purpose was deprecated. This is<br>
>>> specifically stated in C++98 section 7.3.1.1 and unnamed namespaces<br>
>>> are still similarly defined in C++17 (Working Paper) although the<br>
>>> specific reference to the deprecation of 'static' for this purpose is now gone.<br>
>>> ><br>
>>> > The closest the Standard gets to talking about linkage is with<br>
>>> > Linkage<br>
>>> Specifications, and even here it tries to avoid to avoid treading on<br>
>>> the definitions things like internal versus external linkage.<br>
>>> ><br>
>>> > So I am curious, what does this proposed attribute on namespaces<br>
>>> > achieve<br>
>>> that cannot already be achieved with an anonymous or unnamed namespace?<br>
>>> ><br>
>>> > Thanks,<br>
>>> ><br>
>>> > Martin O'Riordan - Movidius Ltd.<br>
>>> ><br>
>>> > -----Original Message-----<br>
>>> > From: cfe-dev [mailto:<a href="mailto:cfe-dev-bounces@lists.llvm.org" target="_blank">cfe-dev-bounces@lists.llvm.org</a>] On Behalf Of<br>
>>> Evgenii Stepanov via cfe-dev<br>
>>> > Sent: 23 October 2015 22:41<br>
>>> > To: Steven Stewart-Gallus <<a href="mailto:sstewartgallus00@mylangara.bc.ca" target="_blank">sstewartgallus00@mylangara.bc.ca</a>><br>
>>> > Cc: Clang Dev <<a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a>><br>
>>> > Subject: Re: [cfe-dev] [RFC] __attribute__((internal_linkage))<br>
>>> ><br>
>>> > Sounds right. The proposed attribute is a rough equivalent of<br>
>>> ><br>
>>> > static void foo::do_method(void)<br>
>>> > {<br>
>>> > // implementation here<br>
>>> > }<br>
>>> ><br>
>>> > where "static" does not mean a static class member, but makes the<br>
>>> declaration local to the translation unit. C-style "static".<br>
>>> ><br>
>>> > The patch in <a href="http://reviews.llvm.org/D13925" rel="noreferrer" target="_blank">http://reviews.llvm.org/D13925</a> (waiting for review,<br>
>>> > btw!)<br>
>>> supports this attribute on classes and namespaces, with this syntax:<br>
>>> ><br>
>>> > namespace __attribute__((internal_linkage)) { }<br>
>>> ><br>
>>> ><br>
>>> > On Fri, Oct 23, 2015 at 9:03 AM, Steven Stewart-Gallus via cfe-dev<br>
>>> > <cfe-<br>
>>> <a href="mailto:dev@lists.llvm.org" target="_blank">dev@lists.llvm.org</a>> wrote:<br>
>>> >> Hello,<br>
>>> >><br>
>>> >> Can I ask for some clarification?<br>
>>> >><br>
>>> >> Apparently, C++ doesn't allow to static class methods?<br>
>>> >><br>
>>> >> While in C one might write inside a header file:<br>
>>> >><br>
>>> >> static inline void mylib_foo_do_method(struct mylib_foo *foo) {<br>
>>> >> // implementation here<br>
>>> >> }<br>
>>> >><br>
>>> >> In C++ the typical style would be to write something like:<br>
>>> >><br>
>>> >> namespace mylib {<br>
>>> >><br>
>>> >> void foo::do_method(void)<br>
>>> >> {<br>
>>> >> // implementation here<br>
>>> >> }<br>
>>> >> }<br>
>>> >><br>
>>> >> Unfortunately, the C++ case then implies some linkage behaviour<br>
>>> >> that some might not want.<br>
>>> >><br>
>>> >> Apparently, one can't do things like:<br>
>>> >><br>
>>> >> namespace mylib {<br>
>>> >><br>
>>> >> static void foo::do_method(void)<br>
>>> >> {<br>
>>> >> // implementation here<br>
>>> >> }<br>
>>> >> }<br>
>>> >><br>
>>> >> Or,<br>
>>> >><br>
>>> >> namespace {<br>
>>> >> namespace mylib {<br>
>>> >><br>
>>> >> static void foo::do_method(void)<br>
>>> >> {<br>
>>> >> // implementation here<br>
>>> >> }<br>
>>> >> }<br>
>>> >> }<br>
>>> >><br>
>>> >> Also, if you wanted to an attribute to a whole namespace you<br>
>>> >> should do something like the following I think:<br>
>>> >><br>
>>> >> namespace mylib {<br>
>>> >> [[clang::internal_linkage]];<br>
>>> >><br>
>>> >> static void foo::do_method(void)<br>
>>> >> {<br>
>>> >> // implementation here<br>
>>> >> }<br>
>>> >> }<br>
>>> >><br>
>>> >><br>
>>> >> Thank you,<br>
>>> >> Steven Stewart-Gallus<br>
>>> >> _______________________________________________<br>
>>> >> cfe-dev mailing list<br>
>>> >> <a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a><br>
>>> >> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev</a><br>
>>> > _______________________________________________<br>
>>> > cfe-dev mailing list<br>
>>> > <a href="mailto:cfe-dev@lists.llvm.org" target="_blank">cfe-dev@lists.llvm.org</a><br>
>>> > <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev</a><br>
>>> ><br>
>>> > _______________________________________________<br>
>>> > LLVM Developers mailing list<br>
>>> > <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>> > <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
>>><br>
>>> _______________________________________________<br>
>>> LLVM Developers mailing list<br>
>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
>> _______________________________________________<br>
>> LLVM Developers mailing list<br>
>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
><br>
_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
</div></div></blockquote></div><br></div></div></div>