In a function-call expression, if the value F invoked as a function is of type reference-to-function or pointer-to-function, a value of reference-to-array type passed as an argument to F appears to be immediately converted to the corresponding pointer type to which it would decay. This prevents implicit converting constructors for the nominal argument types of F which would convert from reference-to-array from being found. This does not occur if F is of function type, nor if it is of some object type with an `operator()`; in these cases the array type is passed without decay. Example code (also at https://godbolt.org/z/jjKE31G7r): struct S { S(const int (&)[2]) {} // S(const int *) {} // footnote [0] }; static const int arr[] = {17, 42}; void func(S s) {} void test() { void (&f)(S) = func; f(arr); // fails on GCC; succeeds on clang & MSVC // func(arr); // succeeds everywhere // (std::function{f})(arr); // succeeds everywhere } [0] If this alternate constructor is uncommented, the above code will compile successfully under GCC... but will fail to compile on clang and MSVC, which reject `f(arr);` as ambiguous. Per testing on godbolt, this is not a recent regression, but appears to happen on every version of GCC back to at least 4.1.2.
Maybe a dup of one of the bugs linked from PR 24666.
(In reply to Andrew Pinski from comment #1) > Maybe a dup of one of the bugs linked from PR 24666. It's none of the open bugs, as far as I can tell. Since it still happens on trunk, I assume it's none of the closed bugs, either... although bug 41426 is suspiciously similar.