Extended Description

Consider:

extern "C" int __declspec(dllimport) x;
extern "C" int *y = &x;

Clang creates this LLVM IR:

@x = external dllimport global i32
@y = global i32* @x, align 4

This is lowered to:

        .data
        .globl  _y
        .align  4
_y:
        .long   _x

Which is incorrect, because x is dllimported, and its address isn't a link time constant. It must be loaded from __imp_x. Clang should emit something like the following IR instead:

@y = global i32* null, align 4
@x = external dllimport global i32
@llvm.global_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @init, i8* null }]
define internal void @init() {
entry:
  store i32* @x, i32** @y, align 4
  ret void
}

This is consistent with what MSVC does, and it would be important to match them if x was selectany or a static data member of a class template.

However, this is really a bug in LLVM, because LLVM will optimize that IR back to Clang's original output:

$ opt t.ll -O2 -S -o -
@y = global i32* @x, align 4
@x = external dllimport global i32
@llvm.global_ctors = appending global [0 x { i32, void ()*, i8* }] zeroinitializer

If y were a function, this would actually link correctly, because import libraries provide thunks for functions. The only downside is that the address of a thunk isn't really the address of the function, but few programs rely on function identity. There's no way to create a forwarding thunk for data, so we get a link failure with this test case.