8213754: PPC64: Add Intrinsics for isDigit/isLowerCase/isUpperCase/isWhitespace
Doerr, Martin
martin.doerr at sap.com
Wed Nov 21 16:33:27 UTC 2018
Hi Michihiro,
thanks. This proposal makes the java code much more intrinsics friendly.
We should wait for feedback from the core lib folks. Maybe they have some requirements or other proposals.
I think these intrinsics should be interesting for Oracle, Intel and others, too.
Btw. I think you can further simplify the code in library_call.cpp (no phi). But we can discuss details when the direction regarding the java classes is clear.
Best regards,
Martin
From: Michihiro Horie <HORIE at jp.ibm.com>
Sent: Mittwoch, 21. November 2018 17:14
To: core-libs-dev at openjdk.java.net; Doerr, Martin <martin.doerr at sap.com>
Cc: hotspot-compiler-dev at openjdk.java.net; Simonis, Volker <volker.simonis at sap.com>; Gustavo Romero <gromero at linux.vnet.ibm.com>
Subject: RE: 8213754: PPC64: Add Intrinsics for isDigit/isLowerCase/isUpperCase/isWhitespace
Hi Martin,
I send this RFR to core-libs-dev too, because it changes the class library.
Through trial and error, I separated Latin1 block as in the
http://cr.openjdk.java.net/~mhorie/8213754/webrev.01/
I followed the coding way of Character.isWhitespace() that invokes each ChracterData's isWhitespace() to refactor isDigit(), isLowerCase(), and isUpperCase().
I think this change is also useful on x86, using STTNI.
Best regards,
--
Michihiro,
IBM Research - Tokyo
----- Original message -----
From: "Michihiro Horie" <HORIE at jp.ibm.com<mailto:HORIE at jp.ibm.com>>
Sent by: "hotspot-compiler-dev" <hotspot-compiler-dev-bounces at openjdk.java.net<mailto:hotspot-compiler-dev-bounces at openjdk.java.net>>
To: "Doerr, Martin" <martin.doerr at sap.com<mailto:martin.doerr at sap.com>>
Cc: "Simonis, Volker" <volker.simonis at sap.com<mailto:volker.simonis at sap.com>>, "ppc-aix-port-dev at openjdk.java.net<mailto:ppc-aix-port-dev at openjdk.java.net>" <ppc-aix-port-dev at openjdk.java.net<mailto:ppc-aix-port-dev at openjdk.java.net>>, "hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>" <hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>>
Subject: RE: 8213754: PPC64: Add Intrinsics for isDigit/isLowerCase/isUpperCase/isWhitespace
Date: Wed, Nov 21, 2018 1:53 AM
Hi Martin,
Thank you for giving your helpful comments. I did not recognize "generate_method_call_static" prevents any optimizations, but I now checked it actually degraded the performance, thanks.
>Please note that we don't have a machine, yet. So other people will have to test.
I think Gustavo can help testing this change when its' ready.
>Would it be possible to introduce more fine-grained intrinsics such that the "slow" path is outside of them?
>
>Maybe you can factor out as in the following example?
>if (latin1) return isLatin1Digit(codePoint);
>with isLatin1Digit as HotSpotIntrinsicCandidate.
Thanks for an example, please let me try to separate the Latin block from other blocks for some time.
Best regards,
--
Michihiro,
IBM Research - Tokyo
[Inactive hide details for "Doerr, Martin" ---2018/11/20 01:55:27---Hi Michihiro, first of all, thanks for working on Power9 opt]"Doerr, Martin" ---2018/11/20 01:55:27---Hi Michihiro, first of all, thanks for working on Power9 optimizations. Please note that we don't ha
From: "Doerr, Martin" <martin.doerr at sap.com<mailto:martin.doerr at sap.com>>
To: Michihiro Horie <HORIE at jp.ibm.com<mailto:HORIE at jp.ibm.com>>, "hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>" <hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>>, "ppc-aix-port-dev at openjdk.java.net<mailto:ppc-aix-port-dev at openjdk.java.net>" <ppc-aix-port-dev at openjdk.java.net<mailto:ppc-aix-port-dev at openjdk.java.net>>
Cc: "Simonis, Volker" <volker.simonis at sap.com<mailto:volker.simonis at sap.com>>, "Lindenmaier, Goetz" <goetz.lindenmaier at sap.com<mailto:goetz.lindenmaier at sap.com>>, Gustavo Romero <gromero at linux.vnet.ibm.com<mailto:gromero at linux.vnet.ibm.com>>
Date: 2018/11/20 01:55
Subject: RE: 8213754: PPC64: Add Intrinsics for isDigit/isLowerCase/isUpperCase/isWhitespace
________________________________
Hi Michihiro,
first of all, thanks for working on Power9 optimizations. Please note that we don't have a machine, yet. So other people will have to test.
I think it may be problematic to insert a slow path by "generate_method_call_static". This may be a performance disadvantage for some users of other encodings because your intrinsics prevent inlining and further optimizations.
Would it be possible to introduce more fine-grained intrinsics such that the "slow" path is outside of them?
Maybe you can factor out as in the following example?
if (latin1) return isLatin1Digit(codePoint);
with isLatin1Digit as HotSpotIntrinsicCandidate.
I can't judge if this is needed, but I think this should be discussed first before going into the details.
Best regards,
Martin
From: Michihiro Horie <HORIE at jp.ibm.com<mailto:HORIE at jp.ibm.com>>
Sent: Freitag, 16. November 2018 12:53
To: hotspot-compiler-dev at openjdk.java.net<mailto:hotspot-compiler-dev at openjdk.java.net>; ppc-aix-port-dev at openjdk.java.net<mailto:ppc-aix-port-dev at openjdk.java.net>
Cc: Doerr, Martin <martin.doerr at sap.com<mailto:martin.doerr at sap.com>>; Simonis, Volker <volker.simonis at sap.com<mailto:volker.simonis at sap.com>>; Lindenmaier, Goetz <goetz.lindenmaier at sap.com<mailto:goetz.lindenmaier at sap.com>>; Gustavo Romero <gromero at linux.vnet.ibm.com<mailto:gromero at linux.vnet.ibm.com>>
Subject: RFR: 8213754: PPC64: Add Intrinsics for isDigit/isLowerCase/isUpperCase/isWhitespace
Dear all,
Would you please review following change?
Bug: https://bugs.openjdk.java.net/browse/JDK-8213754
Webrev: http://cr.openjdk.java.net/~mhorie/8213754/webrev.00
This change includes the intrinsics of Character isDigit, isLowerCase, isUpperCase, and isWhitespace to support the Latin1 block using POWER9's instructions cmprb and cmpeqb. The cmprb enables to compare a character with 1 or 2 ranged bytes, while the cmpeqb compares one with 1 to 8 values. Simple micro benchmark attached showed improvements by 20-40%.
(See attached file: Latin1Test.java)
Best regards,
--
Michihiro,
IBM Research - Tokyo
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20181121/7a1d5d43/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.gif
Type: image/gif
Size: 105 bytes
Desc: image001.gif
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20181121/7a1d5d43/image001-0001.gif>
More information about the hotspot-compiler-dev
mailing list