On Sat, Feb 17, 2018 at 3:18 PM, Xueming Shen <xueming.s...@oracle.com> wrote: > Hi David, > > Thanks for taking this one :-) some comments here.
Thanks for the review! > (1) I would assume you might have to do more for ByteBuffer, something like > [...] > btw, any reason go unsafe to get the byte[]? instead of > ByteBuffer.getArray()? Yes: input should always be settable from a read-only heap buffer without a performance penalty (i.e. copying the contents). This btw is why there is also an explicit check for read-only later on. > (2) It might be a concerned that you have to warp the input byte[] every time > the inflate/deflate(byte[]...) is called. (Yes, I'm constantly hearing > people > complain that you have to wrap the byte[] to ByteBuffer to use CharsetEn/ > Decoder, or even the implementation detail inside StringCoder), and it > might > be taken as a "regression". So it might be desired to wire the > "bf.hasArray()) > path inside de/encode(ByteArray) back to de/encode(byte[], int, int), to > avoid > the "unnecessary" ByteBuffer wrap. I can do some testing to see if there is an impact. I am working on the basis that the wrap may be optimized away (as it is in certain similar cases), but Inflater/Deflater is not final (nor are the inflate/deflate methods) that might not be true in practice. > (3) Same "wrap concern" argument might go to the setInput(bye[] ...) as well. > I'm not sure if it is worth keeping both byte[]/int/int and ByteBuffer > as the > "input" field of In/Deflater. Since input is stored on the object, the above theoretical optimization is much less likely. Again I can do some testing; it might be a good idea to have separate byte[]/int/int and ByteBuffer in the end, though the added expense of checking for and updating the ByteBuffer position after each call in addition to the byte[]/int/int might nullify the benefit. Testing is required... > (4) assume we keep the "wrap" approach. It appears ByteBuffer.wrap() does > check the byte[]/off/len and throw an IndexOutOfBoundsException. So it might > be better to take advantage of that check. I wanted to however it throws the wrong exception type; I could catch and rethrow I guess though if you think that is better (assuming we keep the "wrap" approach as you say). > (5) Deflater.input need to be initialized to a non-null value. > Btw ZipUtil.defaultBuf needs to be "direct"? It doesn't need to be, but the direct buffer code path is possibly somewhat "friendlier" to GC since there's no GetPrimitiveArrayCritical call. > (6) It might be desired to have some jmh measure to make sure byte[] case > does not have regression. I'll work on this when I get a chance (maybe not until Friday though). One other thought I had this weekend was that I could possibly improve things by getting the direct buffer address on the Java side, and pass it in to JNI to avoid the calls to GetDirectBufferAddress. Another thought was to eliminate the remaining SetBooleanField calls by using the remaining two bits in the doInflate/doDeflate methods' return values. I'm not sure how expensive these calls are though. I could also replace the JNI field reads with more method parameters if this is a valuable thing to do. -- - DML
