Re: [PATCH v4 2/3] stdlib: Implement mbrtoc8(), c8rtomb(), and the char8_t typedef.

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 21/07/22 17:51, Tom Honermann wrote:
> On 7/21/22 3:22 PM, Adhemerval Zanella Netto wrote:
>>
>> On 20/07/22 13:47, Tom Honermann wrote:
>>> Confirmed that this issue can be easily reproduced outside the testsuite.
>>>
>>> $ cat t.cpp
>>> #include <uchar.h>
>>>
>>> $ g++ --version
>>> g++ (GCC) 13.0.0 20220720 (experimental)
>>> ...
>>>
>>> $ g++ -c -I/path/to/glibc-char8_t/include -std=c++17 -Werror=c++20-compat t.cpp
>>> In file included from t.cpp:1:
>>> /home/tom/products/glibc-char8_t/include/uchar.h:38:23: error: identifier ‘char8_t’ is a keyword in C++20 [-Werror=c++20-compat]
>>>     38 | typedef unsigned char char8_t;
>>>        |                       ^~~~~~~
>>> cc1plus: some warnings being treated as errors
>>>
>>> The char8_t typedef is currently guarded by:
>>>
>>> /* Declare the C2x char8_t typedef in C2x modes, but only if the C++
>>>    __cpp_char8_t feature test macro is not defined.  */
>>> #if __GLIBC_USE (ISOC2X) && !defined __cpp_char8_t
>>> /* Define the 8-bit character type.  */
>>> typedef unsigned char char8_t;
>>> #endif
>>>
>>> __GLIBC_USE (ISOC2X) evaluates to true because gcc unconditionally defines _GNU_SOURCE. I believe otherwise, C++17 mode would only (or should only) imply __GLIBC_USE (ISOC11).
>>>
>>> Regardless, it seems that directives should be added to suppress the diagnostic. I tried prototyping such a fix, but it doesn't seem to work for me. I don't understand why.
>> I have tried as well and I can't get to work either.  It would expect to work
>> as we have done bits/stdlib-bsearch.h, could it be a gcc issue?
> 
> Yes, this appears to be a gcc issue. I spent some time looking at gcc source code, but didn't find anything obvious. I verified the same technique does work to suppress the similar warning issued for use of, e.g., constexpr, as an identifier when -Wc++11-compat is enabled. I found tests that exercise #pragma GCC diagnostic ignored "-Wc++-compat", but none for -Wc++20-compat (or -Wc++11-compat).
> 
> Tom.
> 

In any case I think the fix below is the correct way (in fact I don't see
another way so I am assuming a compiler issue here).  We also need to avoid 
declare the typedef for __cplusplus >= 202002L.

>>
>>> $ diff -U3 uchar.h.old uchar.h
>>> --- uchar.h.old 2022-07-20 12:37:55.544301692 -0400
>>> +++ uchar.h     2022-07-20 12:43:21.124365563 -0400
>>> @@ -34,8 +34,17 @@
>>>   /* Declare the C2x char8_t typedef in C2x modes, but only if the C++
>>>     __cpp_char8_t feature test macro is not defined.  */
>>>   #if __GLIBC_USE (ISOC2X) && !defined __cpp_char8_t
>>> +/* Suppress the C++20 compatability diagnostic regarding char8_t being a
>>> +   keyword.  */
>>> +#if defined __GNUC__ && 4 < __GNUC__ + (6 <= __GNUC_MINOR__)

Use __GNUC_PREREQ.

>>> +# pragma GCC diagnostic push
>>> +# pragma GCC diagnostic ignored "-Wc++20-compat"
>>> +#endif
>>>   /* Define the 8-bit character type.  */
>>>   typedef unsigned char char8_t;
>>> +#if defined __GNUC__ && 4 < __GNUC__ + (6 <= __GNUC_MINOR__)
>>> +# pragma GCC diagnostic pop
>>> +#endif
>>>   #endif
>>>
>>>   #ifndef __USE_ISOCXX11
>>>
>>> Tom.
>>>
>>> On 7/19/22 5:08 PM, Joseph Myers wrote:
>>>> This change appears to introduce a failure of
>>>> wcsmbs/check-installed-headers-cxx with GCC mainline, because uchar.h now
>>>> produces:
>>>>
>>>> ../wcsmbs/uchar.h:38:23: error: identifier 'char8_t' is a keyword in C++20 [-Werror=c++20-compat]
>>>>      38 | typedef unsigned char char8_t;
>>>>
>>>> (recall we want our installed headers to avoid warnings *without* relying
>>>> on the default disabling of warnings in system headers).
>>>>
>>>> Unfortunately, GCC for C++ doesn't disable -Wc++20-compat inside
>>>> __extension__ (unlike what the C front end does), so simply adding
>>>> __extension__ to that declaration wouldn't help, but we could use
>>>> diagnostic disabling pragmas (as already done in some installed headers).
>>>>