Merged definitions of member functions distinguished by non-functionally-equivalent constraints and by deduced return type

[ecatmur]

template<int I> concept C = true;
template<typename T> struct A {
    auto f() requires C<42> { return 1; }  // #1
    auto f() requires true { return 2u; }  // #2
};
int main() {
    int (A<int>::*p)() = &A<int>::f;
    unsigned (A<int>::*q)() = &A<int>::f;
    return (A<int>().*p)() + (A<int>().*q)();
}

The above program is accepted and returns 4; the expected result would be 3. Inspection of the code (by O0 or by volatile) determines that a single A<int>::f() is emitted, corresponding to #.2.

Very strangely, replacing 2u by 2l and unsigned by long causes clang to reject: "definition with same mangled name '_ZN1AIiE1fEv' as another definition". This seems more sensible, but is apparently contra the Standard says: cplusplus/CWG#256. I have pointed out that following the Standard would require mangling constraints, which I view as a non starter since it would break ABI now and in future.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[shafik]

CC @erichkeane

I believe this is #49884 and maybe #48216

[erichkeane]

Definitely looks like it. We don't have guidance from the Itanium ABI folks yet as far as I know (and this would be something for one of the codegen code owners to take a look at: @rjmccall @efriedma-quic @asl).

[zygoloid]

Fixed in 4b163e3.