Re: [patch] fix libstdc++/56012 - narrowing conversion in std::atomic_flag

[Jonathan Wakely <jwakely dot gcc at gmail dot com>]

On 17 January 2013 07:53, Daniel Krügler <daniel.kruegler@gmail.com> wrote:
> 2013/1/17 Jonathan Wakely <jwakely.gcc@gmail.com>:
>> This fixes a regression since 4.6 when -Wsystem-headers is used.  The
>> initialization of the __atomic_flag_base base class has a narrowing
>> conversion from int (the macro) to either bool or unsigned char.  The
>> patch fixes it by calling a constexpr function which implicitly
>> converts the value to the return type instead of doing the conversion
>> inside a braced-init-list.  Doing that requires naming the return
>> type, so I defined a new typedef for to avoid duplicating the
>> preprocessor conditional.  The patch also adds a missing assignment
>> operator in atomic<bool>.
>>
>>         PR libstdc++/56012
>>         * include/bits/atomic_base.h (atomic_flag): Fix narrowing conversion.
>>         * testsuite/29_atomics/atomic/operators/56012.cc: New.
>>
>>         PR libstdc++/56011
>>         * include/std/atomic (atomic<bool>::operator=(bool) volatile): Add
>>         missing overload.
>>         * testsuite/29_atomics/atomic/operators/56011.cc: New.
>>
>> Tested x86_64-linux, it's a regression so I want to commit it to the
>> trunk and 4.7 branch, any objections from the atomics experts?
>
> Isn't here a typedef missing:
>
> +    /* The target's "set" value for test-and-set may not be exactly 1.  */
> +#if __GCC_ATOMIC_TEST_AND_SET_TRUEVAL == 1
> +    typedef bool __atomic_flag_data_type;
> +#else
> +    unsigned char __atomic_flag_data_type;
> +#endif

Gah! quite right, thanks.  I'll try to find a target to test it on
where the set value isn't 1, just to be sure it works.