[llvm-dev] Is it really valid to discard externally instantiated functions from a TU when marked inline?

[Louis Dionne via llvm-dev]

Hi,

While investigating the situation of visibility annotations and linkage in libc++ with the goal of removing uses of `__always_inline__`, Eric Fiselier and I stumbled upon the attached test case, which I don't think Clang compiles properly. Here's the gist of the test case, reduced to the important parts (see the attachment if you want to repro):
    
    // RUN: %cxx -shared -o %T/libtest.so %flags %compile_flags -fPIC %s
    // RUN: %cxx -c -o %T/main.o %flags %compile_flags %s -O2 -DBUILDING_MAIN
    // RUN: %cxx -o %T/test.exe -L%T/ -Wl,-rpath,%T/ -ltest %T/main.o
    // RUN: %T/test.exe

    template <class T>
    struct Foo {
        inline __attribute__((visibility("hidden")))
        int __init(int x) { /* LOTS OF CODE */ }

        inline __attribute__((visibility("default")))
        int bar(int x) { return __init(x); }
    };

    extern template struct Foo<int>;

    #ifdef BUILDING_MAIN
        int main() {
            Foo<int> f;
            f.bar(101);
        }
    #else
        template struct Foo<int>;
    #endif

When running the attached file in `lit`, we get:

    Undefined symbols for architecture x86_64:
      "Foo<int>::__init(int)", referenced from:
          _main in main.o
    ld: symbol(s) not found for architecture x86_64

The idea here is that `__init` is a pretty big function, and we're promising that an external definition of it is available through the use of the extern template declaration. With the appropriate optimization level (O2 and above), LLVM decides not to include the definition of `__init` in the executable and to use the one available externally. Unfortunately, `__init` has hidden visibility, and so the definition in the .so is not visible to the executable, and the link fails.

Where I think Clang/LLVM goes wrong is when it decides to remove the instantiation in the executable in favor of the one that is hypothetically provided externally. Indeed, the Standard says in [temp.explicit]/12 (http://eel.is/c++draft/temp.explicit#12):

    Except for inline functions and variables, declarations with types deduced from their initializer or return value ([dcl.spec.auto]), const variables of literal types, variables of reference types, and class template specializations, explicit instantiation declarations have the effect of suppressing the implicit instantiation of the definition of the entity to which they refer. [ Note: The intent is that an inline function that is the subject of an explicit instantiation declaration will still be implicitly instantiated when odr-used so that the body can be considered for inlining, but that no out-of-line copy of the inline function would be generated in the translation unit. — end note ]

Only reading the normative wording, it seems like LLVM should leave the instantiation there because it can't actually assume that there will be a definition provided elsewhere (yes, despite the extern template declaration, because the function is inline). Then, the non-normative note seems to be approving of what LLVM is doing, but I'm wondering whether that's really the intended behavior.

Questions:
1. Is what LLVM's doing there legal?
2. If it is legal, then libc++ needs a way to express that a function should either be inlined in the caller, or emitted in the TU and de-duplicated later (I think that’s linkonce_odr), despite there being an extern template declaration promising a definition elsewhere. I think the current situation is that the function gets available_externally linkage instead. Is there a way to express this at the C++ source code level?

Thank you,
Louis Dionne

-------------- next part --------------
A non-text attachment was scrubbed...
Name: linkage.sh.cpp
Type: application/octet-stream
Size: 1051 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180710/a9b08edc/attachment.obj>