RFR: 8286279: [vectorapi] Only check index of masked lanes if offset is out of array boundary for masked store [v2]
Quan Anh Mai
duke at openjdk.java.net
Fri May 13 02:37:45 UTC 2022
On Fri, 13 May 2022 01:35:40 GMT, Xiaohong Gong <xgong at openjdk.org> wrote:
>> Checking whether the indexes of masked lanes are inside of the valid memory boundary is necessary for masked vector memory access. However, this could be saved if the given offset is inside of the vector range that could make sure no IOOBE (IndexOutOfBoundaryException) happens. The masked load APIs have saved this kind of check for common cases. And this patch did the similar optimization for the masked vector store.
>>
>> The performance for the new added store masked benchmarks improves about `1.83x ~ 2.62x` on a x86 system:
>>
>> Benchmark Before After Gain Units
>> StoreMaskedBenchmark.byteStoreArrayMask 12757.936 23291.118 1.826 ops/ms
>> StoreMaskedBenchmark.doubleStoreArrayMask 1520.932 3921.616 2.578 ops/ms
>> StoreMaskedBenchmark.floatStoreArrayMask 2713.031 7122.535 2.625 ops/ms
>> StoreMaskedBenchmark.intStoreArrayMask 4113.772 8220.206 1.998 ops/ms
>> StoreMaskedBenchmark.longStoreArrayMask 1993.986 4874.148 2.444 ops/ms
>> StoreMaskedBenchmark.shortStoreArrayMask 8543.593 17821.086 2.086 ops/ms
>>
>> Similar performane gain can also be observed on ARM hardware.
>
> Xiaohong Gong has updated the pull request incrementally with one additional commit since the last revision:
>
> Wrap the offset check into a static method
However, we seem to lack the ability to do an unsigned comparison reliably. C2 can transform `x + MIN_VALUE <=> y + MIN_VALUE` into `x u<=> y` but it will fail if `x` or `y` is an addition with constant in such cases the constants will be merged together. As a result, I think we need an intrinsic for this. `Integer.compareUnsigned` may fit but it manifests the result into an integer register which may lead to suboptimal materialisation of flags, another approach would be to have a separate method `Integer.lessThanUnsigned` which only returns `boolean` and C2 can have better time splitting the boolean comparison through `IfNode`, which will prevent the materialisation of `boolean` values. What do you two think?
I.e, after splitting if through merge point, the shape of `if (Integer.lessThanUnsigned(a, b))` would be transformed from
a b
\ /
CmpU
|
Bool
|
If
/ \
IfTrue IfFalse
\ /
Region 1 0
\ | /
Phi 0
\ /
CmpI
into
a b
\ /
CmpU
Thanks.
-------------
PR: https://git.openjdk.java.net/jdk/pull/8620
More information about the core-libs-dev
mailing list