|Posted by Steven Schveighoffer|
in reply to dsimcha
Posted in reply to dsimcha
On Sun, 01 Aug 2010 12:46:58 -0400, dsimcha <firstname.lastname@example.org> wrote:
> == Quote from dsimcha (email@example.com)'s article
>> I've reread the relevant TDPL chapter and I still don't quite understand the
>> 1. What is shared? Is it simply a piece of syntactic salt to make it hard to
>> share data across threads by accident, or is there more to it?
>> 2. Is it fully or mostly implemented?
> Sorry, accidentally submitted the post before I was done.
> 3. How does casting to and from shared work? Under what circumstances can
> unshared data be cast to shared? Under what circumstances can shared data
> implicitly be cast to unshared?
Let me preface this by saying I don't actually use shared on a daily basis, but I'll try and respond how I think it works:
1. It indicates to the compiler that multiple threads have direct access to the data. But more importantly, the *lack* of shared means that exactly one thread has direct access to the data. I see shared not as great a feature as unshared is. For an example of optimizations that can be had with unshared data, see the LRU cache for lock-free array appending.
I think in terms of technical details, reading from/writing to a shared piece of data requires either a lock, or the compiler will insert a memory barrier around the write to ensure the write is atomic as long as the data written is small enough (I'm very fuzzy on these details, I use the term memory barrier in complete ignorance). Declaring a global variable shared also makes it a true global (not thread-local). I don't know what was decided on for shared classes/structs, I vaguely remember that the consensus was to require declaring the entire class shared at class definition, but I could be wrong.
2. I do not think it's fully implemented, but I think the intention is that it's fully implemented, so submit bugs against what it doesn't do if you find any.
3. *NO* implicit casting of shared<->unshared is allowed for references. To do so would violate the shared transitivity. It is ok to copy value-types to/from shared. Think of the relationship between unshared and shared in the same way as the relationship between mutable and immutable.
I think you can cast shared to unshared if you *know* that no other thread will be able to access the data pointed at by the shared value. For instance, you can never take a reference to a shared global and cast that reference to unshared, because globals are always available to all threads.
You can cast unshared to shared if you know that you have no other unshared references to the same data in your local thread. This one can be much easier to prove.
Neither cast is statically checked or verified, it's up to you as the programmer to ensure these properties.