RFR: 8347009: Speed ​​up parseInt and parseLong [v5]

[Shaojin Wen]

On Sun, 5 Jan 2025 11:01:18 GMT, Shaojin Wen <swen at openjdk.org> wrote:

>> This is an optimization for decimal Integer.parseInt and Long.parseLong, which improves performance by about 10%. The optimization includes:
>> 1. Improve performance by parsing 2 numbers at a time, which has performance improvements for numbers with length >= 3.
>> 2. It uses charAt(0) for the first number. Assuming that the optimization can eliminate boundary checks, this will be more friendly to parsing numbers with length 1.
>> 3. It removes the reliance on the Character.digit method and eliminates the reliance on the CharacterDataLatin1#DIGITS cache array, which avoids performance degradation caused by cache misses.
>
> Shaojin Wen has updated the pull request incrementally with one additional commit since the last revision:
> 
>   vector digit2

The performance is improved again by calculating two characters at a time in the DecimalDigit::digit2 method. The performance numbers are as follows:

## 1. Script

git remote add wenshao git at github.com:wenshao/jdk.git
git fetch wenshao

#master
git checkout 5bb9644202a8b961faad02439c95a81ca6862168
make test TEST="micro:java.lang.Longs.parseLong"
make test TEST="micro:java.lang.Integers.parseInt" 

 # current
git checkout f97093d324fbb7de322c86d52db11cafcff2de87
make test TEST="micro:java.lang.Longs.parseLong"
make test TEST="micro:java.lang.Integers.parseInt" 


## 2. aliyun_ecs_c8a_x64 (CPU AMD EPYC™ Genoa)

-Benchmark          (size)  Mode  Cnt  Score   Error  Units (master 5bb9644)
-Longs.parseLong       500  avgt   15  3.539 ± 0.031  us/op
-Integers.parseInt     500  avgt   15  3.270 ± 0.028  us/op

+Benchmark          (size)  Mode  Cnt  Score   Error  Units (current f97093d)
+Longs.parseLong       500  avgt   15  3.029 ± 0.033  us/op +16.83%
+Integers.parseInt     500  avgt   15  3.120 ± 0.004  us/op  +4.80%



## 3. aliyun_ecs_c8i_x64 (CPU Intel®Xeon®Emerald Rapids)

-Benchmark          (size)  Mode  Cnt  Score   Error  Units (master 5bb9644)
-Longs.parseLong       500  avgt   15  3.469 ± 0.011  us/op
-Integers.parseInt     500  avgt   15  3.136 ± 0.005  us/op


+Benchmark          (size)  Mode  Cnt  Score   Error  Units (current f97093d)
+Longs.parseLong       500  avgt   15  2.999 ± 0.024  us/op +15.67%
+Integers.parseInt     500  avgt   15  2.819 ± 0.004  us/op +11.24%




## 4. orange_pi5_aarch64 (CPU RK3588S)

-Benchmark          (size)  Mode  Cnt  Score   Error  Units (master 5bb9644)
-Longs.parseLong       500  avgt   15  9.903 ± 0.064  us/op
-Integers.parseInt     500  avgt   15  9.527 ± 0.678  us/op

+Benchmark          (size)  Mode  Cnt  Score   Error  Units (current f97093d)
+Longs.parseLong       500  avgt   15  8.702 ± 1.287  us/op +13.80%
+Integers.parseInt     500  avgt   15  6.117 ± 0.046  us/op +55.74%




# 5. MacBook M1 Pro (aarch64)

-Benchmark          (size)  Mode  Cnt  Score   Error  Units (master 5bb9644)
-Longs.parseLong       500  avgt   15  2.617 ? 0.040  us/op
-Integers.parseInt     500  avgt   15  2.344 ? 0.027  us/op

+Benchmark          (size)  Mode  Cnt  Score   Error  Units (current f97093d)
+Longs.parseLong       500  avgt   15  2.262 ? 0.014  us/op +15.69%
+Integers.parseInt     500  avgt   15  2.033 ? 0.007  us/op +15.29%

-------------

PR Comment: https://git.openjdk.org/jdk/pull/22919#issuecomment-2571595624