Why would noexcept(true) fail where noexcept works?

Question

I have the following code:

struct S {
    // conditional noexcept specification which depends on noexceptness of foo()
    void bar() noexcept(noexcept(foo()));
    // conditional noexcept specification (true)
    void foo() noexcept(true);
};

Clang rejects this code but GCC allows it (https://godbolt.org/z/nxqa5Tr1x):

<source>:2:34: error: exception specification is not available until end of class definition
    2 |     void bar() noexcept(noexcept(foo()));
      | 

Replacing noexcept(true) with noexcept is allowed by both compilers.

Questions

1. Why would noexcept(true) fail but noexcept work?
2. Which compiler conforms with the standard, if any?

---

Note: this example is obviously minimal. My actual use case is that I have a clear() member function which is noexcept only if clear_impl() is noexcept, in a custom container.

Answer 1

1. There is no reason why noexcept and noexcept(true) should have different behaviour.
2. This issue appears to be related to CWG issues 1360, 1397, 1890, and 2335. A noexcept-specifier is a complete-class context. This implies that within a noexcept-specifier, S is complete (e.g. you can see the full list of members of S), but it does not imply that within a noexcept-specifier, you have access to information that is in later complete-class contexts. In some complete-class contexts, some compilers don't even let you access information that is in earlier complete-class contexts of a different kind, and the wording is not clear enough to say for sure that this is a compiler bug. I explain in more detail here.