[clang] New error in clang21, in-class initializer for static data member is not a constant expression

[lingege321]

Hi experts,

I'm updating the clang version from 20 to 21 and see the new error in my project.
The error is introduced from a third-party library and I reduce the code size to a minimum issue.
Please check the compiler-explorer, https://godbolt.org/z/x73rfhWfW

Seems that static member variable must be a constant expression.
However, in this case, type(BOUND+1) is not a constant expression in clang21 but I think it should be considered as a constant expression.

I have some workaround for it but I don't want to do that because it's a third party library.

Is it possible to support this code?

struct Test {
    enum type {
        Type1,
        BOUND
    };
    static const type binding_required = BOUND;
    static const type binding_completed = type(BOUND+1);
    static const type detached = type(binding_completed+1);
    static const type dying = type(detached+1);
};

[frederick-vs-ja]

This is CWG1766. Converting BOUND + 1 (which is equal to 2) to type triggers undefined behavior and thus results in constant evaluation failure.

b364535 (https://reviews.llvm.org/D130058) seemingly implemented the resolution of CWG1766. But it's strange that the status table wasn't updated and the commit only mentioned CWG2338.

It seems that #143034 completed the diagnostic. IMO Clang is correct to reject the example.

CC @shafik @efriedma-quic.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: lingege32 (lingege321)

Hi experts,

I'm updating the clang version from 20 to 21 and see the new error in my project.
The error is introduced from a third-party library and I reduce the code size to a minimum issue.
Please check the compiler-explorer, https://godbolt.org/z/x73rfhWfW

Seems that static member variable must be a constant expression.
However, in this case, type(BOUND+1) is not a constant expression in clang21 but I think it should be considered as a constant expression.

I have some workaround for it but I don't want to do that because it's a third party library.

Is it possible to support this code?

struct Test {
    enum type {
        Type1,
        BOUND
    };
    static const type binding_required = BOUND;
    static const type binding_completed = type(BOUND+1);
    static const type detached = type(binding_completed+1);
    static const type dying = type(detached+1);
};

[AaronBallman]

It seems that #143034 completed the diagnostic. IMO Clang is correct to reject the example.

Agreed, I believe this is by design and noted in the release notes: https://releases.llvm.org/21.1.0/tools/clang/docs/ReleaseNotes.html#c-specific-potentially-breaking-changes

Checking for integer to enum conversions in constant expressions is more strict; in particular, const E x = (E)-1; is not treated as a constant if it’s out of range. The Boost numeric_conversion library is impacted by this; it was fixed in Boost 1.81. (#143034)