Volatile arrays in Java
A slight complication of Java volatile fields, and one sometimes overlooked,
is that declaring an array volatile does not give volatile access to its fields!.
At least, it doesn't when elements of the array are accessed with "normal" Java syntax. In other words:
- it is unsafe to call arr[x] = y on an array (even if declared volatile) in one
thread and then expect arr[x] to return y from another thread;
- on the other hand, it is safe to call arr = new int (or whatever) and expect another
thread to then read the new array as that referenced by arr: this is what is meant by declaring
the array reference volatile.
So, what do we do if we want a truly volatile array in Java, where the array's individual fields
have volatile semantics?
Solution 1: use AtomicIntegerArray or AtomicLongArray
The AtomicIntegerArray class implements an int array whose
individual fields can be accessed with volatile semantics, via the class's
get() and set() methods. Calling arr.set(x, y) from one thread
will then guarantee that another thread calling arr.get(x) will read the value y
(until another value is read to position x).
Solution 2: re-write the array reference after each field write
This is slightly kludgy and slightly inefficient (since what would be one write now
involves two writes) but I believe it is theoretically correct. After setting an element of
the array, we re-set the array reference to be itself:
volatile int arr = new int[...];
arr = 100;
arr = arr;
The marginal benefit of this technique could be:
- it may allow you to fix some broken code with minimal changes if you have
code already (incorrectly) using a volatile array and expecting thread-safe element access;
- it saves us creating the wrapper object around the array (which is what happens with
IntegerArray and LongArray);
- it's more or less the only option for array types with no available atomic wrapper (e.g. a float array).