RFR 8248043: Need to eliminate excessive i2l conversions
Boris Ulasevich
boris.ulasevich at bell-sw.com
Wed Jul 1 09:16:31 UTC 2020
Hi,
It is the third attempt to send a correct link. Sorry for that ;)
http://cr.openjdk.java.net/~bulasevich/8248043/webrev.02c
Thanks,
Boris
On Wednesday, July 1, 2020, Boris Ulasevich <boris.ulasevich at bell-sw.com>
wrote:
> Hi,
>
> I'm deeply sorry. Yes, webrev.02b is certainly wrong!
> Correct link is webrev.02c:
> http://cr.openjdk.java.net/~bulasevich/8248043/webrev.02
> <http://cr.openjdk.java.net/~bulasevich/8248043/webrev.02b/src/hotspot/share/opto/subnode.cpp.udiff.html>
> c
> - this is the change I described in my mail and wanted to review.
>
> my apologies,
> Boris
>
> On Wednesday, July 1, 2020, Igor Veresov <igor.veresov at oracle.com> wrote:
>
>> > On Jun 30, 2020, at 10:15 PM, Vladimir Kozlov <
>> vladimir.kozlov at oracle.com> wrote:
>> >
>> > I think Igor said that you can't swap arguments of compare without
>> changing condition test. For example, if it was CC_LT it should be CC_GT
>> after swap.
>>
>> Yes, that’s exactly what I had in mind. Condition must be inverted.
>> Otherwise your transformation [3] is not valid for anything else but
>> equality, so that’s not going to work. May be if [3] didn’t work, perhaps
>> there is another user of the CmpLNode in addition to BoolNode ?
>>
>> igor
>>
>> >
>> > It is not clear why you need swapping in CmpLNode::Ideal() if
>> BoolNode::Ideal() should do it already. If it does not you need to
>> investigate why.
>> >
>> > Also your list of steps 1.-3. does not reflect changes in webrev.02b:
>> > http://cr.openjdk.java.net/~bulasevich/8248043/webrev.02b/
>> src/hotspot/share/opto/subnode.cpp.udiff.html
>> >
>> > Regards,
>> > Vladimir
>> >
>> > On 6/30/20 9:33 PM, Boris Ulasevich wrote:
>> >> Hi Igor,
>> >> By BoolNode I mean the canonicalization that is already in place:
>> >> https://hg.openjdk.java.net/jdk/jdk/file/de6ad5f86276/src/ho
>> tspot/share/opto/subnode.cpp#l1391
>> >> thanks,
>> >> Boris
>> >> On Wed, Jul 1, 2020 at 5:07 AM Igor Veresov <igor.veresov at oracle.com>
>> wrote:
>> >>> I think you forgot to include changes to BoolNode in the webrev.
>> >>>
>> >>> igor
>> >>>
>> >>>
>> >>>
>> >>> On Jun 30, 2020, at 11:04 AM, Boris Ulasevich <
>> boris.ulasevich at bell-sw.com>
>> >>> wrote:
>> >>>
>> >>> Hi Claes,
>> >>>
>> >>>> Seems like the optimization is mostly effective, but not getting all
>> the
>> >>> way.
>> >>>
>> >>> Good point about LHS, thanks! CmpL turned to be not canonized on the
>> >>> moment.
>> >>> I moved the optimization to CmpLNode::Ideal and transformations now
>> works
>> >>> as follows:
>> >>> 1. CmpINode::Ideal: CmpI(CmpL3)->CmpL
>> >>> 2. BoolNode::Ideal:
>> >>> Bool(CmpL(const,val),test)->Bool(CmpL(val,const),test_invert)
>> >>> 3. CmpLNode::Ideal: CmpL(ConvI2L(val),ConL)->CmpI(val,ConI)
>> >>>
>> >>> I applied your test to the benchmark. The result is:
>> >>> Benchmark Mode Cnt Score Error Units
>> >>> SkipIntToLongCast.skipCastTestLeft avgt 5 14.288 ± 0.052 ns/op
>> >>> SkipIntToLongCast.skipCastTestRight avgt 5 14.338 ± 0.088 ns/op
>> >>>
>> >>> Updated webrev:
>> >>> http://cr.openjdk.java.net/~bulasevich/8248043/webrev.02b
>> >>>
>> >>> thanks,
>> >>> Boris
>> >>>
>> >>> On 26.06.2020 21:31, Claes Redestad wrote:
>> >>>
>> >>> Hi Boris,
>> >>>
>> >>> this looks like a nice improvement! I just have some comments about
>> the
>> >>> micro.
>> >>>
>> >>> I was curious whether the optimization works when the constant is on
>> >>> the LHS and added a variant of the micro to try that[1]. Results are
>> >>> interesting (Intel Xeon):
>> >>>
>> >>> Benchmark Mode Cnt Score Error Units
>> >>> SkipIntToLongCast.skipCastTest avgt 5 30.937 ± 0.056 ns/op
>> >>> SkipIntToLongCast.skipCastTestLeft avgt 5 30.937 ± 0.140 ns/op
>> >>>
>> >>> With your patch:
>> >>> Benchmark Mode Cnt Score Error Units
>> >>> SkipIntToLongCast.skipCastTest avgt 5 14.123 ± 0.035 ns/op
>> >>> SkipIntToLongCast.skipCastTestLeft avgt 5 17.420 ± 0.044 ns/op
>> >>>
>> >>> Seems like the optimization is mostly effective, but not getting all
>> >>> the way. I wouldn't worry about it for this RFE, but perhaps something
>> >>> to investigate in a follow-up. Feel free to include such a variant in
>> >>> your patch though (no attribution necessary).
>> >>>
>> >>> The micro also stabilizes very quickly, so you might want to provide
>> >>> some default tuning to keep runtime in check, e.g., something like:
>> >>>
>> >>> @Warmup(iterations = 10, time = 500, timeUnit = TimeUnit.MILLISECONDS)
>> >>> @Measurement(iterations = 5, time = 1000, timeUnit =
>> TimeUnit.MILLISECONDS)
>> >>> @Fork(3)
>> >>>
>> >>> Thanks!
>> >>>
>> >>> /Claes
>> >>>
>> >>> [1]
>> >>> @Benchmark
>> >>> public int skipCastTestLeft() {
>> >>> for (int i = 0; i < ARRAYSIZE_L; i++) {
>> >>> if (ARRAYSIZE_L == intValues[i]) {
>> >>> return i;
>> >>> }
>> >>> }
>> >>> return 0;
>> >>> }
>> >>>
>> >>> On 2020-06-26 17:05, Boris Ulasevich wrote:
>> >>>
>> >>> Hi all,
>> >>>
>> >>> Please review the change to eliminate the unnecessary i2l conversion
>> >>> for expressions like this: "if (intValue == 1L)".
>> >>>
>> >>> http://bugs.openjdk.java.net/browse/JDK-8248043
>> >>> http://cr.openjdk.java.net/~bulasevich/8248043/webrev.01
>> >>>
>> >>> The provided benchmark shows performance boost on all platforms:
>> >>> - Intel Xeon: 32.705 --> 14.234 ns/op
>> >>> - arm64: 42.060 --> 25.456 ns/op
>> >>> - arm32: 618.763 --> 314.040 ns/op
>> >>> - ppc8: 81.218 --> 63.026 ns/op
>> >>>
>> >>> Testing done: jtreg, jck.
>> >>>
>> >>> thanks,
>> >>> Boris
>> >>>
>> >>>
>> >>>
>> >>>
>>
>>
More information about the hotspot-compiler-dev
mailing list