On 05/04/2016 10:50 AM, Peter Levart wrote: > On 05/03/2016 02:39 PM, Aleksey Shipilev wrote: >> *) There is ever-so-subtle difference in doing either: >> >> 171 byte[] tmp = new byte[path.length + 1]; >> 172 System.arraycopy(path, 0, tmp, 1, path.length); >> 173 tmp[0] = '/'; >> >> ...or: >> >> 1083 byte[] name = new byte[nlen + 1]; >> 1084 name[0] = '/'; >> 1085 System.arraycopy(cen, pos + CENHDR, name, 1, nlen); >> >> Excess op between allocation and arraycopy breaks zeroing elimination, >> see e.g.: >> http://cr.openjdk.java.net/~shade/scratch/ZeroingFirstBench.java > > ...so when allocation of array is immediately followed with arraycopy > where the newly allocated array is the target, only the part of the > array that is not overwritten by arraycopy will be zeroed? Because not > the whole array is overwritten in your example/benchmark. That's good to > know.
Yes, only not overwritten part would be zeroed. This is a little trick in C2. I think it works only for heads/tails, not for arbitrary overwrites. So it is a limited micro-optimization here. -Aleksey
