Significant performance drop with numbers in Java

[deathaxe]

Description

I noticed a significant performance drop with Java Rewrite with regards to highlighting numbers.

One of the benchmarking files used to track performance is a syntax_test_perf_numbers.java with 10k lines of following code snippet.

0x1.;
0x.1a2f;
0x1.a2f;
0x1ffp+1023 0x1ffp+_1023_;
0xd.aP-1074 0x_1_f_._a_d_P-_10_74_;
0D + 12345D + 12345D + 12_34_5_D - _12_34_5D - 12a45D;
0F + 12345F + 12345F + 12_34_5_F - _12_34_5F - 12a45F;
1. + 1_. + 1_2. - _1.;
1.D + 1_.D + 1_2.D - _1.D;
1.2 + 1_.2_ + 1_2.3_4 + 1_2_._3_4_ - _1.5;
1.2d + 1_.2_d + 1_2.3_4d + 1_2_._3_4_d - _1.5d;
12e34 + 12e+3_ + 1_2e3_4 + 1_2_e3_4_ + 1_2_e_3_4 + 12e+34 + 12e-34 + 12e+3_4 - _12e34;
12e34f + 12e+3_f + 1_2e3_4f + 1_2_e3_4_f + 1_2_e_3_4f + 12e+34f + 12e-34f + 12e+3_4f - _12e34f;
12.e34 + 12.e+3_ + 1_2.e3_4 + 1_2_.e3_4_ + 1_2_.e_3_4 + 12.e+34 + 12.e-34 + 12.e+3_4 - _12.e34;
12.e34f + 12.e+3_f + 1_2.e3_4f + 1_2_.e3_4_f + 1_2_.e_3_4f + 12.e+34f + 12.e-34f + 12.e+3_4f - _12.e34f;
12.34e56 + 12_.34_e+5_ + 1_2.3_4e5_6 + 1_2_.3_4_e5_6_ + 1_2_._3_4e_5_6 + 12.34e+56 + 12.34e-56 + 12.34e+5_6 - _12.34e+5_6;
12.34e56f + 12_.34_e+5_f + 1_2.3_4e5_6f + 1_2_.3_4_e5_6_f + 1_2_._3_4e_5_6f + 12.34e+56f + 12.34e-56f + 12.34e+5_6f - _12.34e+5_6f;
.2 + .2_ + .3_4 + ._3_4_;
.2d + .2_d + .3_4d + ._3_4_d;
.34e56 + .34_e+5_ + .3_4e5_6 + .3_4_e5_6_ + ._3_4e_5_6 + .34e+56 + .34e-56 + .34e+5_6;
23.45 + 23.45F + 23.45d;
.01 + .02e3+.02e3F;
23.45e67+23.45e+6F+23.45e-67D;
0b101101 + 0b10_11_01 + 0b10_11_01_ + 0b_101101 - 0_b10_1101 + 0b;
0b101101l + 0b10_11_01l + 0b10_11_01_l + 0b_101101l - 0_b10_1101l + 0bl;
0xABCD + 0xAB_CD + 0xAB_CD_ + 0x_AB_CD - 0_xAB_CD - 0x;
0xABCDl + 0xAB_CDl + 0xAB_CD_l + 0x_AB_CDl - 0_xAB_CDl;
07 + 0_ + 0_7 + 07_ + 079 + 079_ + 0_79_ - 0a - 0_a;
07l + 0_l + 0_7l + 07_l + 0792l + 079_2_l - 0al - 0_a_l;
0 + 0L;
12345 + 12_34_5 + 1_____5 + 12_34_5_ - _12_34_5 - 12a45;
12345l + 12345L + 12_34_5_L - _12_34_5L - 12a45L;
123_-_456;

Steps to reproduce

1. Start ST 4113 in SAFE MODE
2. Copy content from Java Rewrite PR into Packages
3. Create a syntax_test_perf_numbers.java and place it in User package
4. Run Syntax Tests - Performance
5. Follow step 1 - 4 with ST 4115/4116.

Expected behavior

Same (or better) performance in ST4115 compared to 4113.

Actual behavior

The benchmark shows:

• 130ms on ST 4113
• 6.5s on ST 4115

Environment

• Build: 4115
• Operating system and version: Windows 11 21H2

[BenjaminSchaaf]

I can't reproduce this on 4115 or 4116 under Linux. Does it still happen in 4116?

[deathaxe]

Yes, it's the same with 4116.

[deathaxe]

Recent investigation revealed the variables context to be the reason for the mensioned slowdown.

If the following line is commented out said benchmark file runs through within 70ms.

https://github.com/deathaxe/sublime-packages/blob/053a2ced694490c6c7fa386177a2f0d5e484e56d/Java/Java.sublime-syntax#L166

Slow down seems to be caused by the commented out branches. As soon as one of them is commented in again, parsing time increases to said value.

  variables:
    - match: (?={{id_first_char}})
      branch_point: variables
      branch:
        - variable-other
        - variable-namespace
        # - variable-qualifier
        - method-call-identifier
        # - lambda-parameter

May be an implementation issue of the syntax definition. Trying to find a hopefully better alternative.

[BenjaminSchaaf]

It's another branch-related performance problem. I've already got a fix that makes it faster than 4113.

[deathaxe]

Cool.

[BenjaminSchaaf]

Fixed in build 4117.