Well instead of the usual "measure it" stuff an actual answer - because that stuff is actually real fun math. Although the compiler could and probably does this as well (at least modern optimizing c++ compilers, javac certainly won't and I've got no idea if the JVM does this) - so better check if it isn't already doing the work for you.
But still fun to know the theory behind the optimization: I'll use assembly because we need the higher 32bit word of a multiplication. The following is from Warren's book on bit twiddling:
n is the input integer we want the modulo from:
li M, 0x55555556 ; load magical number (2^32 + 2) / 3
mulhs q, M, n ; q = higher word of M * n; i.e. q = floor(M*n / 2^32)
shri t, n, 31 ; add 1 to q if it is negative
add q, q, t
Here q contains the divisor of n / 3 so we just compute the remainder as usual: r = n - q*3
The math is the interesting part - latex would be rather cool here:
q = Floor( (2^32+2)/ 3 * (n / 2^32) ) = Floor( n/3 + 2*n/(3*2^32) )
Now for n = 2^31-1 (largest n possible for signed 32bit integers) the error term is less than 1/3 (and non negative) which makes it quite easy to show that the result is indeed correct. For n = -2^31 we have the correction by 1 above and if you simplify that you'll see that the error term is always larger than -1/3 which means it holds for negative numbers as well.
I leave the proof with the error term bounds for the interested - it's not that hard.