Re: [boost] [thread] Compiler eoorrs while running the tests

Subject: Re: [boost] [thread] Compiler eoorrs while running the tests
From: vicente.botet (vicente.botet_at_[hidden])
Date: 2010-05-21 04:32:47

----- Original Message ----- 
From: "Anthony Williams" <anthony.ajw_at_[hidden]>
To: <boost_at_[hidden]>
Sent: Friday, May 21, 2010 10:01 AM
Subject: Re: [boost] [thread] Compiler eoorrs while running the tests

> 
> "vicente.botet" <vicente.botet_at_[hidden]> writes:
> 
>> ----- Original Message ----- 
>> From: "Anthony Williams" <anthony.ajw_at_[hidden]>
>>
>> Have 10.1/Darwin, vacpp/AIX and gcc 4.3.4/FreeBSD rvalue references when running on C++098 mode?
> 
> I don't know about Intel 10.1 or vacpp, but gcc 4.3.4 only has rvalue
> references in C++0x mode.
> 
>> I don't undesrtand how this test can work with your move semantics emulation?
>>
>> boost::thread make_thread_return_local(boost::thread::id* the_id)
>> {
>>     boost::thread t(do_nothing,the_id);
>>     return t;
>> }
> 
> It's the NRVO optimization --- the named object is constructed directly
> in the return value rather than copied. Many compilers do the
> optimization, but they are supposed to check for an accessible copy
> constructor anyway.
> 
> The compilers it works on must skip the required check for an accessible
> copy constructor, so they are not strictly conforming to C++98, but in
> this case it's a bonus. In C++0x this is also required to work if the
> object being returned has an accessible move constructor.

Well, as the behaviour is not conforming to C++98, I would say that the test must be part of compile-fails for C++98 compilers, and state  that in these compiler the test fails as expected. Of course the test must succeed on C++0x as boost::thread defines a constructor from a rvalue reference.

Any other opinions?

Vicente