On 06/25/2013 10:13 PM, Peter Levart wrote:
>
> On 06/22/2013 12:54 PM, Aleksey Shipilev wrote:
>> T get(int index1, index2) {
>> T[][] lc = cache;
>> if (index1 >= lc.length) { // needs resizing
>> lc = <generate_new_T[][]_of_size>((index1 << 1) + 1);
>> cache = lc;
>> }
>> T[] lt = lc[*index2*];
>> if (index2 >= lt.length) { // needs resizing
>> lt = <generate_new_T[]_of_size>((index2 << 1) + 1);
>> lc[index1] = lt;
>> cache = lc; // publish
>> }
>> return lt[index2];
>> }
>>
>> -Aleksey.
>
> Hi Aleksey,
>
> 1st I think there's a typo in above bolded part. There should be index1
> instead of index2 there, right?
Nope, lc is indexed with index2; lt is indexed with index1.
> Then I think there's a data race in above code which can de-reference
> half-constructed array and an element within it:
Damn. Good point, let's discuss that in the appropriate RFR, see 8017540.
> I have studied the constraints of powerCache and have observed:
> So the caching could be performed with no synchronization (other than
> that provided by final fields descibed above).
Yeah, we can do that with finals. Too bad you can't have the final
semantics with BigInteger[] array.
> Here is a possible variant of such caching:
Seems overcomplicated. ;) Let's instead fix the code proposed in 8017540.
Thanks,
Aleksey.
