On Sun, 6 Oct 2024 15:14:05 GMT, Claes Redestad <redes...@openjdk.org> wrote:
>> #14632 showed that coalescing loads in the `ZipUtils` utility methods could
>> improve performance in zip-related microbenchmarks, but doing so would
>> increase startup overheads. Progress was stalled as we backed out some
>> related early use of `ByteArray(LittleEndian)` and started exploring merge
>> store optimizations in C2.
>>
>> In this PR I instead suggest using `Unsafe` directly to coalesce `short`,
>> `int`, and `long` reads from zip data. Even with explicit bounds checking to
>> ensure these utilities are _always_ safe there are significant improvements
>> both to lookup speed and speed of opening zip files (most if not all bounds
>> checks are optimized away):
>>
>>
>> make test TEST=micro:java.util.zip.ZipFile
>>
>> Name (size) Cnt Base Error Test
>> Error Unit Change
>> GetEntry.getEntryHit 512 15 37.999 ± 0.841 34.641 ±
>> 0.389 ns/op 1.10x (p = 0.000*)
>> GetEntry.getEntryHit 1024 15 39.557 ± 0.523 36.959 ±
>> 1.488 ns/op 1.07x (p = 0.000*)
>> GetEntry.getEntryHitUncached 512 15 69.250 ± 0.931 64.851 ±
>> 0.987 ns/op 1.07x (p = 0.000*)
>> GetEntry.getEntryHitUncached 1024 15 71.628 ± 0.307 67.927 ±
>> 0.714 ns/op 1.05x (p = 0.000*)
>> GetEntry.getEntryMiss 512 15 22.961 ± 0.336 22.825 ±
>> 0.188 ns/op 1.01x (p = 0.158 )
>> GetEntry.getEntryMiss 1024 15 22.940 ± 0.115 23.502 ±
>> 0.273 ns/op 0.98x (p = 0.000*)
>> GetEntry.getEntryMissUncached 512 15 35.886 ± 0.429 35.598 ±
>> 1.296 ns/op 1.01x (p = 0.395 )
>> GetEntry.getEntryMissUncached 1024 15 38.168 ± 0.911 36.141 ±
>> 0.356 ns/op 1.06x (p = 0.000*)
>> Open.openCloseZipFile 512 15 62425.563 ± 997.455 56263.401 ±
>> 896.892 ns/op 1.11x (p = 0.000*)
>> Open.openCloseZipFile 1024 15 117491.250 ± 962.928 108055.491 ±
>> 1595.577 ns/op 1.09x (p = 0.000*)
>> Open.openCloseZipFilex2 512 15 62974.575 ± 911.095 57996.388 ±
>> 910.929 ns/op 1.09x (p = 0.000*)
>> Open.openCloseZipFilex2 1024 15 119164.769 ± 1756.065 108803.468 ±
>> 929.483 ns/op 1.10x (p = 0.000*)
>> * = significant
>>
>>
>> This PR also address some code duplication in `ZipUtils`.
>>
>> An appealing alternative would be to implement a merge load analogue to the
>> merge store optimizations in C2. Such optimizations would be very welcome
>> since it would improve similar code outside of java.base (jdk.zi...
>
> Claes Redestad has updated the pull request incrementally with one additional
> commit since the last revision:
>
> copyright
src/java.base/share/classes/java/util/zip/ZipUtils.java line 195:
> 193: * The bytes are assumed to be in Intel (little-endian) byte order.
> 194: */
> 195: public static final long get64(byte[] b, int off) {
This method returns a signed 64 bit value, which I think is not what some of
its call sites expect. It should in any case be renamed to `get64S` to align
with `get32S`. A new method `get64` should be introduced and any call site
expecting unsigned numbers (most?) should use that instead.
If you don't want to deal with this in this PR, I could file an issue and
suggest a PR for this. Let me know.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/21377#discussion_r1789144124
