I had this symbol as C23 only for a reason. I think it was because things go really wrong in C++ land if this header defines char8_t otherwise. Please check this carefully (e.g. by building world with the patch applied) or just leave things as they were.

It would be great if the design patterns and symbol naming could be similar to the existing files c16rtomb.c and c32rtomb.c for consistency.

That 64 is a magic number. Where is it from? Is there a macro we can use instead?

I don't think this comment is relevant. \0 is the string terminator, which does not end up being transcoded. The function is meant to be restartable, meaning, that you can deliver input in chunks that may not themselves all start or end at boundaries of full characters (if that wasn't the case, then the mbstate_t argument would be pointless, as you'd always be in initial state when calling this function).

What is the meaning of these structure members?

What we do now is the same idea as char16_t: _types.h sets _CHAR8_T_DECLARED when the language already has the type, and uchar.h only typedefs if that flag is clear. We key it off __cpp_char8_t, not __cplusplus >= 202002L, so -fchar8_t / -fno-char8_t are correct. Prototypes are emitted in C++ only when __cpp_char8_t is set, so they always use the builtin, never unsigned char pretending to be char8_t.

C++11/17 without the builtin do get typedef unsigned char char8_t if they include <uchar.h>. That cannot collide with a keyword (there isn’t one). It can collide with a user identifier named char8_t, same class of issue as char16_t before C++11. The keyword case is protected.

So, in fact, by improving it slightly and guarding against __cpp_char8_t, I tried fixing a rare bug that had been lurking there for a while. I ran pretty much every test that sprang to my mind, and I think this is considerably safer than the previous check. Bu I would really appreciate it if you could take a look as well and let me know if you spot anything I’ve missed.

Sure, I can rename _C8rtombExtra / _Mbrtoc8Extra to _Char8State.

It’s half the size of the mbstate_t blob, and I think that’s an OK promise. Perhaps I could do this:

#define _MBSTATE_CONV_PREFIX (sizeof(mbstate_t) / 2)

I’m still not sure it’s the best we can do, though. I’d rather use sizeof(extra) + sizeof(_ConversionState) <= sizeof(mbstate_t), but that would pull in the iconv types, and our case here is MK_ICONV=no.

What would you do, Robert?

But u8'\0' is a char8_t value, not a C-string terminator. Also, restartability remains unchanged. If the comment is misleading, we can certainly drop it altogether (or perhaps just the second sentence). However, if I’m missing something here, I’d appreciate it if you could point it out and correct me.

nleft / next / buf are the leftover UTF-8 units after we already returned the first one ((size_t)-3), and incomp means mbrtoc32 last returned (size_t)-2, so a following NUL is EILSEQ instead of being swallowed by iconv.