Thanks again for the paper, Corentin.

The following comments are based on the P2626R0 paper revision submitted for the August mailing. I'm sorry for the delay in responding.

I have a number of reservations about the paper as currently proposed though I am strongly in favor of what it is trying to accomplish.

My primary concern is whether these interfaces suffice to solve the issues as experienced in real world code. The before/after presentation in the Tony table section does a nice job of illustrating why the existing cast operations do not suffice, but does not illustrate the inherent danger in using the proposed interfaces. Consider the following (some suspension of disbelief is required here since std::print won't accept char8_t-based text, but ignore that for now; as the paper says, "charN_t types are poorly supported in the standard library. We should notably support them in format").

void rename(const char *from, const char *to);
const char8_t from[] = ...;
const char8_t to[] = ...;
rename(cast_utf_to<char>(from, sizeof(from)), cast_utf_to<char>(to, sizeof(to)));
std::print("Renamed {} to {}\n", from, to); // UB; the lifetime of the char8_t objects from and to held were ended.

To fix this, it is necessary to rebuild the prior objects (it is not valid to reference an object of type char via an lvalue of type char8_t; that is the aliasing problem). But attempts to do so run into the problem that from and to are not actually associated with the replacement objects other than through a common region of storage; references to the replaced objects are needed to re-bless the storage as holding objects of the previous type. However, even in that case, I'm not sure that the association between the variables and the objects can really be reestablished; I think problems similar to those for which std::launder was introduced are present here. Access to the new objects must go through a pointer/reference provided by the operation that created the replacement objects.

I think what is most needed, at least for the example above, is an operation that temporarily (e.g., via RAII and an object with full expression lifetime) mutates the type of an object temporarily and then restores it.

The paper lists the ICU character casts as an inspiration for the paper. However, the proposed interfaces don't match the semantics of the ICU casts. The ICU casts don't invalidate the source objects; they coerce the compiler into forgetting what it knows about the contents of memory. As such, I'm skeptical that the proposed interfaces could be used to replace the ICU casts. If they can be, a diff of the changes that ICU would require to adopt the new interfaces would be very helpful.

As specified, the operations do not end the lifetime of, nor change the type of, the array objects that hold the sequence of elements that are being converted. This creates the strange result that, for example, an array object of type char16_t could have elements of type wchar_t. I'm not sure if that is ok from a core language perspective or not, but it seems problematic to me.

The proposed operations are not really casts; they behave more like a destructive in-place move. Describing them as casts is, I think, misleading.

With a few exceptions, each of the existing cast operators supports symmetric conversion between types. If a cast operation supports conversion from type A to type B, then it also (usually) supports conversion from type B to type A. I'm not aware of a precedent in the language for a cast in one direction to have a different name than for a cast in the other direction. And I don't think such a distinction is needed in this case. The fundamental operation that is needed is the ability to cast/convert/swap an object of one type to a type that is, or shares, an underlying type (same kind (integer, float, etc...), size, alignment). Once that operation is available, encoding aware interfaces that limit conversion direction can be built on top if desired.

The paper includes a link to godbolt.org containing the following example use. The call to in.size() is UB since it follows a call to std::move(in). I think this illustrates how difficult these interfaces may be to use correctly.

std::string utf8_to_iso8859_7(std::u8string_view in) {
    iconv_t conv = iconv_open("ISO-8859-7", "UTF-8");
    auto as_char = cast_utf_to<char>(std::move(in));
    std::string out(10, 0);
    std::size_t out_size = out.size();
    std::size_t in_size  = in.size();
    char* inptr = const_cast<char*>(as_char.data()); // C interfaces are fantastic
    char* outptr = out.data();  
    int ret = iconv(conv, &inptr, &in_size, &outptr, &out_size);
    return out;
}

The basic_string_view and span overloads seem deeply problematic to me. Consider:

std::string s = ...;
auto u8sv = cast_as_utf_unchecked(s); // the implicitly constructed temporary string_view binds to the rvalue reference.
// The buffer managed by s is of type char, but now holds objects of type char8_t.
// Any use of s, including destruction (at least in constexpr context) is UB.

The wording for the basic_string_view and span overloads of cast_as_utf_unchecked references a From type that does not appear in the declaration.

Tom.

On 7/30/22 2:16 PM, Corentin via SG16 wrote:
Early draft, feedback welcome.

Thanks, 
Corentin 

https://isocpp.org/files/papers/D2626R0.pdf