Equal typeid for nonequal types in anonymous namespaces

[Alcaro]

#include <typeinfo>

namespace {
	struct my_x {
		static int member;
	};
	int my_x::member;
}

int* get_member() { return &my_x::member; }
const std::type_info* inf() { return &typeid(my_x); }

#include <typeinfo>

namespace {
	struct my_x {
		static int member;
	};
	int my_x::member;
}

int* get_member();
const std::type_info* inf();

int* get_member2() { return &my_x::member; }
const std::type_info* inf2() { return &typeid(my_x); }

#include <stdio.h>

int main()
{
	printf("%p %p\n", get_member(), get_member2());
	printf("%p %p\n", inf(), inf2());
	printf("%d\n", get_member() == get_member2());
	printf("%d\n", *inf() == *inf2());
}

https://godbolt.org/z/Pjnz6jbGb

This prints four pointers, then 0 1.

I believe that is an illegal output. Either the two my_x are the same, and it should print 1 1, or they're different, and it should print 0 0. I think 0 0 is the correct answer, but I'm not very familiar with this corner of the C++ specification.

Similar to #10492 - that one's long gone, but maybe a similar fix would apply here. (GCC has a similar bug, but only with -fmerge-all-constants. edit: And that one's documented as giving noncompliant behavior, so I'm not sure if it is a bug.)

[llvmbot]

@llvm/issue-subscribers-clang-frontend