@jensmaurer Please examine the conclusion. It seems that in the most major implementations, they consider such an implicitly call as the constructs that appear in the lexical part that corresponds to the expression E instead of these functions that would be dynamically invoked by E. It could infer from [except.spec#6.6]. If were that, then the following cases would print false.

#include <iostream>
struct A{
    A() noexcept(false){
    }
};
void fun() noexcept(true){
    A a{};  // initialization for `a` implicitly invokes A::A()
}
int main(){
  std::cout<< noexcept(fun());
}

But it's not true, the result indicates the expression fun() is not potentially-throwing.

So, my conclusion is that

The immediate subexpressions of an expression E are

• any function call that E implicitly invokes, which is the part of the full-expression of E

This modification could restrict that these function call should be lexical appears with E rather than dynamically invoked by E

Yes, "immediate subexpression" is intended on the almost-syntax level.

Yes, "immediate subexpression" is intended on the almost-syntax level.

So, I thik the sentence "which is the part of the full-expression of E" would help to state that meaning.

It would not, because we use "subexpression" in the definition of "full-expression", thus the definition would become circular. Not good.

It would not, because we use "subexpression" in the definition of "full-expression", thus the definition would become circular. Not good.

Could we say "which is lexical". Or any other wording as long as that meaning could be read. In addition, AFAIK, it seems there's no conflict if we use full-expression in that sentence. That means, such an implicitly call is never a full-expression. That approach is merely a bit obscure as you said it becomes circular.