[Clang] `-Wreturn-type` false negative with `switch` statement on enum

[bjacob]

In this compiler explorer experiment, GCC warns "warning: control reaches end of non-void function", while Clang does not. Pasting it here for self-containedness:

enum class B : int {
    B1, B2
};

int f(B b) {
    switch(b) {
        case B::B1: return 1;
        case B::B2: return 2;
    }
}

In this instance, GCC is making an important point: the function f cannot assume that its parameter of enum type has a value that is one of the enumerator values, because this enumerator has a a fixed underlying type (the : int in the enum declaration) and the standard says that in that case, any integer of that underlying type may be casted to this enum type: https://eel.is/c++draft/expr.static.cast#9

Currently, GCC always warns regardless of having a specified underlying type in the enum declaration ; Clang never warns. If my understanding of the spec is correct then the reachability of the end of the function (and therefore, the rationale for generating a warning) is when there is no specified underlying type.

[shafik]

I believe gcc has a bug and this is not intended behavior see: https://godbolt.org/z/nx881Ed6z

enum  B {
    B1=0, B2=1
};

int f(B b) {
  switch(b) {
    case B::B1: return 1;
    case B::B2: return 2;
  }
}

gcc still diagnoses this case but the valid value of B are only 0 and 1 since this enum does not have an fixed underlying type.

clang UBSan diagnoses the case when we try to the use the value 2: https://godbolt.org/z/vrcszGW9d

[bjacob]

Agree that GCC has that bug, and what I'm advocating for here is not GCC parity but something else - that's what I was trying to convey in the last paragraph of the issue description above:

Currently, GCC always warns regardless of having a specified underlying type in the enum declaration ; Clang never warns. If my understanding of the spec is correct then the reachability of the end of the function (and therefore, the rationale for generating a warning) is when there is no specified underlying type.

Great to hear that UBSan is able to make that diagnostic. That also confirms our reading of the spec.

[efriedma-quic]

The behavior of the warning is, as far as I know, intentional.

Generating a warning for every switch on an enum which doesn't have a "default" seems like it would lead to a lot of false-positive warnings. C/C++ allows enum variables with values outside the enumerated range, but most code doesn't take advantage of that (at least, not intentionally). And code which does use values outside the enumerated range usually won't use a switch statement like the given example.

[bjacob]

The warning I was talking about in the issue description above,

"warning: control reaches end of non-void function".

Is about a function, not about a switch statement. That one at least should be possible to generate with few false positives.

[bjacob]

Thanks though, the original issue title was misleading in this respect. Retitled.

[shafik]

I agree technically this could be an issue but as the previous comment says I don't believe folks taking advantage of enums to use value outside of the explicit enumerators would use switches and if they do they.

CC @AaronBallman wdyt?

[AaronBallman]

I think I consider this to be a bug in Clang. We don't explicitly document -Wreturn-type but we do want it to have reasonable parity with GCC and GCC's documentation says: "Also warn if execution may reach the end of the function body" which is possible in this case.