fno-semantic-interposition breaks code relying on address uniqueness of a function

[aelovikov-intel]

// a.cpp
#include <functional>
#include <cassert>
int foo(int x) { return x + 42; }

bool bar(std::function<int(int)> F) {
  auto f = F.target<int(*)(int)>();
  return f && *f == foo;
}

// b.cpp
#include <functional>
int foo(int x);

bool bar(std::function<int(int)> F);

int main() {
  return bar(foo);
}

$ /usr/bin/clang++ -std=c++17 -fPIC -fno-semantic-interposition a.cpp -shared -o liba.so
$ /usr/bin/clang++ -std=c++17 b.cpp -L. -la
$ LD_LIBRARY_PATH=. ./a.out ; echo $?
0
$ /usr/bin/clang++ -std=c++17 -fPIC a.cpp -shared -o liba.so
$ LD_LIBRARY_PATH=. ./a.out ; echo $?
1

GCC is consistent with/without the option.

I'm not sure if that is a bug or desired behavior, but https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100483#c2 suggests that the case above doesn't use semantic interposition and so should work with the option. I'm not unlikely to misread/misunderstand it though.

[aelovikov-intel]

Tagging @MaskRay .

[llvmbot]

@llvm/issue-subscribers-clang-codegen

```cpp // a.cpp #include <functional> #include <cassert> int foo(int x) { return x + 42; }

bool bar(std::function<int(int)> F) {
auto f = F.target<int(*)(int)>();
return f && *f == foo;
}

```cpp
// b.cpp
#include <functional>
int foo(int x);

bool bar(std::function<int(int)> F);

int main() {
  return bar(foo);
}

$ /usr/bin/clang++ -std=c++17 -fPIC -fno-semantic-interposition a.cpp -shared -o liba.so
$ /usr/bin/clang++ -std=c++17 b.cpp -L. -la
$ LD_LIBRARY_PATH=. ./a.out ; echo $?
0
$ /usr/bin/clang++ -std=c++17 -fPIC a.cpp -shared -o liba.so
$ LD_LIBRARY_PATH=. ./a.out ; echo $?
1

GCC is consistent with/without the option.

I'm not sure if that is a bug or desired behavior, but https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100483#c2 suggests that the case above doesn't use semantic interposition and so should work with the option. I'm not unlikely to misread/misunderstand it though.