Primitive Queue<any T> considerations

[Peter Levart]

On 11/18/2015 07:55 PM, Vitaly Davidovich wrote:
> As discussed upthread, how would that operation be atomic for arbitrarily
> sized value types?

Not atomic, but for SPSC queue discussed here and for the price of 
doubling the writes, this problem could be solved like:

array[offset] = Cell(value, 0);
UNSAFE.storeFence();
array[offset] = Cell(value, 1);
producerIndex++;


Maybe JIT could even detect such pattern and try to optimize it doing 
something equivalent but not expressible:

array[offset].value = value;
UNSAFE.storeFence();
array[offset].mark = 1;
producerIndex++;



Regards, Peter

> On Wed, Nov 18, 2015 at 1:53 PM, MacGregor, Duncan (GE Energy Management) <
> duncan.macgregor at ge.com> wrote:
>
>> On 18/11/2015, 16:22, "valhalla-dev on behalf of Dávid Karnok"
>> <valhalla-dev-bounces at openjdk.java.net on behalf of akarnokd at gmail.com>
>> wrote:
>>
>>> Let's assume I have a value type named cell:
>>>
>>> valuetype Cell<any T> {
>>>    public T value;
>>>    public int mark;
>>> }
>>>
>>> and a power-of-2 array of cells:
>>>
>>> Cell<any T>[] array;
>>>
>>>
>>> having Cell with final fields is useless here because then I have a
>>> constant array and not a queue.
>> Maybe I¹m not following you here but the fact fields on a Cell are final
>> does not stop you from replacing one Cell in array with a new Cell, just
>> as you can replace individual ints in an array of ints.
>>
>> So you offset method would be more like
>>
>> offset = producerIndex & (array.length - 1)
>> if (array[offset].mark != 0) {
>>      return false;
>> }
>> array[offset] = Cell(value, 1);
>> producerIndex++;
>>
>> return true;
>>
>> And that operation on the array will be guaranteed to be atomic (though
>> probably needs a compare and swap in case another thread is trying to do
>> the same thing).
>>
>> Duncan.
>>
>>