[consteval] clang++ sometimes reports false-positive errors for static consteval member functions

[ldalessa]

When I have a CRTP that uses it's client's function, if that client's function happens to be static consteval, clang++ reports an error.

template <class T>
struct base
{
    constexpr auto foobar() const -> int {
        return static_cast<T const*>(this)->bar(); // error: call to consteval function 'foo::bar' is not a constant expression
    }
};

struct foo : base<foo> {
    static consteval auto bar() -> int {
        return 0;
    }
};

foo x;
int i = x.foobar();

live: https://godbolt.org/z/KG5oT6GTj

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[llvmbot]

@llvm/issue-subscribers-c-2b

[llvmbot]

@llvm/issue-subscribers-c-20

[usx95]

This does not work even for simpler cases of this and consteval interactions.

struct S {
    static consteval int hello() { return 1;}

    int foo() const { 
        return this->hello(); 
        // ^ error: call to consteval function 'S::hello' is not a constant expression
        // ^ note: use of 'this' pointer is only allowed within the evaluation of a call to a 'constexpr' member function
    }
    constexpr int bar() const {
        return this->hello(); // same error and note as above.
    }
    consteval int baz() const {
        // works fine.
        return this->hello();
    }
} s;
static_assert(s.baz() == 1);

Clang errors and diverges from all the rest of the compilers. https://godbolt.org/z/93WTdbK1v

[usx95]

9fcca8b#diff-f10defc3f20fb095bf22b3a79bead200d494bde9d503e283067a57aff483936c

I think this patch fixed a crash but clang still produces an error.

struct A {
  consteval int a() const { return 1; }
  void c() {
    a() + d(); // expected-error {{call to consteval function 'PR48235::A::a' is not a constant expression}} \
               // expected-note {{use of 'this' pointer is only allowed within the evaluation of a call to a 'constexpr' member function}}
  }

My reading of https://eel.is/c++draft/dcl.constexpr is that consteval members are valid and using them in any context (constant or unevaluated) show also be legal.

Would be interested in knowing thoughts of @AaronBallman

[AaronBallman]

9fcca8b#diff-f10defc3f20fb095bf22b3a79bead200d494bde9d503e283067a57aff483936c

I think this patch fixed a crash but clang still produces an error.

struct A {
  consteval int a() const { return 1; }
  void c() {
    a() + d(); // expected-error {{call to consteval function 'PR48235::A::a' is not a constant expression}} \
               // expected-note {{use of 'this' pointer is only allowed within the evaluation of a call to a 'constexpr' member function}}
  }

My reading of https://eel.is/c++draft/dcl.constexpr is that consteval members are valid and using them in any context (constant or unevaluated) show also be legal.

Would be interested in knowing thoughts of @AaronBallman

I believe http://eel.is/c++draft/expr.prim.id.general#4 is the cause of that diagnostic. https://godbolt.org/z/PP51G3WWW sure suggests everyone implements this the same way, too.

[usx95]

Interesting. The above example looks right now.

Coming back to

struct S {
    static consteval int hello() { return 1;}

    int foo() const { 
        return this->hello(); 
        // ^ error: call to consteval function 'S::hello' is not a constant expression
        // ^ note: use of 'this' pointer is only allowed within the evaluation of a call to a 'constexpr' member function
    }
    constexpr int bar() const {
        return this->hello(); // same error and note as above.
    }
    consteval int baz() const {
        // works fine.
        return this->hello();
    }
} s;
static_assert(s.baz() == 1);

I guess clang is almost right but it should consider this->hello() in both foo() and bar() to be a constant expression according to http://eel.is/c++draft/dcl.constexpr#5.1 and http://eel.is/c++draft/expr.const#5.1

[AaronBallman]

I think you're correct because of http://eel.is/c++draft/expr.const#5.1.2 specifically. While the this isn't necessary for the member access expression (because the expression denotes a static function), that this is still evaluated per https://eel.is/c++draft/expr.ref#1.sentence-2 (see the associated footnote as well), so it is being used as the postfix-expression and so it should be fine as a core constant expression.

So I think the diagnostic in foo() is correct, but the one in bar() is incorrect, and accepting the code in baz() is correct. Do you come to the same conclusion?

[usx95]

I come to the conclusion that
this->static_consteval_member() is a core constant expression and thus the invocation inside foo() is also valid and clang diagnostic is incorrect.
(http://eel.is/c++draft/dcl.constexpr#5.1 : an invocation of a constexpr function can appear in a constant expression [expr.const])

I agree with the rest of the conclusion (bar() and baz()).

[usx95]

Although I suspect I might be wrong again due to http://eel.is/c++draft/expr.prim.id.general#4. I am mostly confused about the definition of id-expression. Basically:

consteval int func() {return 0;}
struct S {
    static consteval int hello() { return 1;}
    int foo() const { 
        return this->hello(); // Why this should be invalid
        // return func() ; // but this is valid. Both are core constant expressions!
    }
}

Is func() not an id-expression and http://eel.is/c++draft/expr.prim.id.general#4 does not apply to it ?

[AaronBallman]

func is an id-expression (it's the primary-expression component of the postfix-expression and the () denotes a function call expression http://eel.is/c++draft/expr.compound#nt:postfix-expression). But id-expression is also used in a postfix-expression denoting a member access in the same paragraph:

postfix-expression . template_opt id-expression
postfix-expression -> template_opt id-expression

http://eel.is/c++draft/expr.post#expr.ref-6 determines what is denoted by the member access expression, and that id-expression will denote an immediate function whether called through this->func() or func().

[cor3ntin]

Should be fixed by #63139

[shafik]

This currently working in trunk: https://godbolt.org/z/4P8GdPhWM