gcc-4.7-20111022 fails to build on PA-RISC

[Daniel Krügler]

2011/10/30 Daniel Krügler <daniel.kruegler@googlemail.com>:
> 2011/10/30 Jonathan Wakely <jwakely.gcc@gmail.com>:
>> It turns out that the restriction on copying volatile objects
>> documented in C.1.9 [diff.special] breaks std::mutex for hppa-linux,
>> where pthread_mutex_t has a volatile member, so cannot be copied
>> without adding a user-defined copy constructor to the underlying libc
>> type, which is obviously difficult for us to do.
>>
>> A reduced test is:
>>
>> struct lock_t {
>>    int lock[4];
>> };
>>
>> struct pthread_mutex_t {
>>    volatile lock_t __spinlock;
>> };
>>
>> struct mutex {
>>  pthread_mutex_t m = { };
>>  mutex() = default;
>> };
>>
>> int main()
>> {
>>  mutex mx;
>> }
>>
>> One solution I can see is to undefine __GTHREAD_MUTEX_INIT etc. on
>> that platform, so we use the pthread_mutex_init function instead - can
>> anyone see a better alternative?
>
> My initial idea would have been to change the member initializer to
>
> struct mutex {
>  pthread_mutex_t m{ };
>  mutex() = default;
> };
>
> which should no longer use copy-initialization but direct-initialization.
> But the compiler produces the same error. I believe, this is a compiler
> defect in regard to handling brace-or-equal-initializers, though.

I would go even further and would argue that the current code should
also be well-formed, because the effect of writing

pthread_mutex_t m = { };

is to initialize the subobject __spinlock "from an empty initializer list",
which also is no copy-initialization.

- Daniel