On 01/07/2016 02:21 AM, Martin Buchholz wrote: > A stupid benchmark (sorry Aleksey - maybe you already have a jmh > version of this?) http://cr.openjdk.java.net/~martin/Utf8Bench.java
Okay, I'll bite: http://cr.openjdk.java.net/~shade/scratch/UTF8Scan.java As Andrew already pointed out, the trick is to get the ByteBuffer byte order to native, and use JDK 9 with getLongUnaligned bells and whistles. As Paul suggested, VarHandles views could be used to always avoid ByteBuffer allocation, even though compilers are moderately successful in removing them. > It's hard to resist Unsafe when it gives 3x improvement in a > performance hotspot. Maybe C2 can optimize the simple loop even more > than it is today? It would seem that with JDK 9, you don't need Unsafe in this scenario. ByteBuffers provide a comparable performance, and VarHandles would push the envelope a bit further, since we are implicitly optimizing the internal APIs to boost them up. Thanks, -Aleksey
