Vladimir Ivanov

hacking HotSpot JVM @ Oracle

[Joseph Darcy @jddarcy]

JEP 306 (https://t.co/53XH6PnPG5), Restore Always-Strict Floating-Point Semantics, now integrated into JDK 17.
As it was in the beginning, is again, and (ever?) shall be, predictable floating-point arithmetic end-to-end.

[Mark Reinhold @mreinhold]

Java 16 / JDK 16: General Availability: https://t.co/C0bNCFHlkX #jdk16 #java16 #openjdk #java

[Vladimir Ivanov @iwan0www]

@tstuefe @shipilev @SonarCloud Come on! You should have been more ashamed by the bugs they regularly find. Personally, I am :-)

[OpenJDK @OpenJDK]

Targeted to JDK 16: JEP 338: Vector API (Incubator): https://t.co/16ZRe1QbK1 #jdk16 #openjdk #java

[Vladimir Ivanov @iwan0www]

@tfoldi @richardstartin ATM there are a couple of known performance issues [1] after recent API refactoring. Your example is affected by #2 (inlining). Until the fix is integrated in the repo, please, give the following patch a try: https://t.co/uyEuKq9c8y

[1] https://t.co/gCERMqP778

[Vladimir Ivanov @iwan0www]

@richardstartin I took a look and found 2 issues:
https://t.co/gCERMqP778

(Posting it as a gist since I'm not on SO. Frankly speaking, it's strange to see such questions posted on SO and not on the development mailing list [1].)

[1] https://t.co/tBQCimkzh7

[Vladimir Ivanov @iwan0www]

@richardstartin As others pointed out, (as of now) the JEP is neither Targeted nor Proposed to Target to JDK15. But the implementation has been under official review on relevant mailing lists since late March and there's still some time left before feature freeze in JDK15 (early June).

[Vladimir Ivanov @iwan0www]

@tagir_valeev @rafaelcodes @lukaseder BTW be aware of a long-standing JVM bug: https://t.co/7GAoQMKlY6

[Vladimir Ivanov @iwan0www]

@kuksenk0 @stokito @miha_x64 JVM и так умная, но почему бы просто не сделать её ещё умней? :-)

[Vladimir Ivanov @iwan0www]

@ChrisGSeaton @mwolczko But then you would still need to overwrite some fields in constructors. Unless there's a very efficient way to do bulk zeroing supported by the hardware, avoiding redundant zeroing makes sense.

[Vladimir Ivanov @iwan0www]

@cl4es @mjpt777 @nipafx @pivovarit @mreinhold I'd like stress that --illegal-access only affects visibility of _non-critical_ internal APIs. Critical internal APIs (sun.misc.Unsafe is among them) were not encapsulated and JEP 260 contains more details on the topic: https://t.co/J7GEOTxj03

[Vladimir Ivanov @iwan0www]

@jhorstmann23 Most likely, it is related to redundant barriers around off-heap accesses https://t.co/v6p20jJKtF

[Vladimir Ivanov @iwan0www]

@DrDeprecator @javawithjiva @stuartmarks @TonyPrintezis @ysr1729 @phohensee Fortunately, no work for you here :-) It was deprecated and then completely removed over a decade ago.

[Vladimir Ivanov @iwan0www]

Anybody remember when a full-blown GC was removed from HotSpot JVM last time? :-)

[Vladimir Ivanov @iwan0www]

@uncle_robot @alexanius Meanwhile, a decade ago :-)
https://t.co/OiulgKSlUl

[Vladimir Ivanov @iwan0www]

@leventov @richardstartin Yes, it should work (excerpt vector trigonometric math stubs which are available only on Linux/Windows now).

[Vladimir Ivanov @iwan0www]

@leventov @chvest Still inline classes don't look like the best tool. F14 would benefit from fine-grained layout control and ability to align on cache lines (to perfectly fit it into a single cache line and vectorize accesses). But those are deliberately left out of scope for inline classes.

[Vladimir Ivanov @iwan0www]

@leventov @chvest F14 case is definitely quite far from the sweet spot, but since a chunk fits into a cache line, inline classes may work quite well here. (Not necessarily in the first version though; number of components poses some challenges for the implementation.)