Remove resolved perf-check TODOs from UInt256 shift helpers

[parthdagia05]

What

Remove the two // TODO: check perf - wiring [shiftRight|shiftLeft] callers with this one comments above UInt256.sar0 and UInt256.shl0, added in #10216.

Why

The TODOs asked whether routing the public shift methods (sar, shr, shl) through the new internal helpers (sar0, shl0) preserves performance versus the pre-refactor implementation in StackArithmetic. I ran the existing V2 shift JMH benchmarks before and after #10216 to find out.

Method

• PRE state: 8e4ef80c40^ (commit 67dbd9a4c5, parent of Move SAR, SHR and SHL to UInt256 #10216)
• POST state: upstream/main at 80b381ee9f
• Benchmarks: SarOperationBenchmarkV2, ShlOperationBenchmarkV2, ShrOperationBenchmarkV2
• JMH config: 3 forks × (3 warmup + 5 measurement iterations × 1s) repo default
• Hardware: local x86_64 laptop, JDK 21
• Significance: marked when 99.9% confidence intervals do not overlap

Results

Benchmark              Scenario              PRE (ns/op)     POST (ns/op)      Δ%      Significant?
----------------------------------------------------------------------------------------------------
SarOperationBenchmarkV2  SHIFT_0               4.953 ± 0.140   5.486 ± 0.461   +10.8%  noise
SarOperationBenchmarkV2  NEGATIVE_SHIFT_1      7.522 ± 0.387   7.445 ± 0.348    -1.0%  noise
SarOperationBenchmarkV2  ALL_BITS_SHIFT_1      7.363 ± 0.251   7.238 ± 0.435    -1.7%  noise
SarOperationBenchmarkV2  POSITIVE_SHIFT_1      7.083 ± 0.279   6.891 ± 0.335    -2.7%  noise
SarOperationBenchmarkV2  NEGATIVE_SHIFT_128    5.698 ± 0.107   5.830 ± 0.145    +2.3%  noise
SarOperationBenchmarkV2  NEGATIVE_SHIFT_255    5.942 ± 0.163   6.116 ± 0.232    +2.9%  noise
SarOperationBenchmarkV2  POSITIVE_SHIFT_128    5.649 ± 0.107   6.099 ± 0.143    +8.0%  REGRESSION
SarOperationBenchmarkV2  POSITIVE_SHIFT_255    5.655 ± 0.259   5.868 ± 0.204    +3.8%  noise
SarOperationBenchmarkV2  OVERFLOW_SHIFT_256    5.608 ± 0.233   6.045 ± 0.544    +7.8%  noise
SarOperationBenchmarkV2  OVERFLOW_LARGE_SHIFT  5.785 ± 0.196   5.939 ± 0.215    +2.7%  noise
SarOperationBenchmarkV2  FULL_RANDOM          16.628 ± 1.596  17.642 ± 1.630    +6.1%  noise
ShlOperationBenchmarkV2  SHIFT_0               4.903 ± 0.235   5.352 ± 0.260    +9.2%  noise
ShlOperationBenchmarkV2  SHIFT_1               7.465 ± 0.262   7.391 ± 0.401    -1.0%  noise
ShlOperationBenchmarkV2  SHIFT_128             5.869 ± 0.139   5.835 ± 0.152    -0.6%  noise
ShlOperationBenchmarkV2  SHIFT_255             5.944 ± 0.230   5.982 ± 0.200    +0.6%  noise
ShlOperationBenchmarkV2  OVERFLOW_SHIFT_256    5.058 ± 0.140   5.779 ± 0.192   +14.3%  REGRESSION
ShlOperationBenchmarkV2  OVERFLOW_LARGE_SHIFT  5.220 ± 0.143   5.913 ± 0.217   +13.3%  REGRESSION
ShlOperationBenchmarkV2  FULL_RANDOM          12.975 ± 1.183  11.084 ± 0.608   -14.6%  IMPROVEMENT
ShrOperationBenchmarkV2  SHIFT_0               4.993 ± 0.202   5.240 ± 0.264    +4.9%  noise
ShrOperationBenchmarkV2  SHIFT_1               7.341 ± 0.375   6.990 ± 0.365    -4.8%  noise
ShrOperationBenchmarkV2  SHIFT_128             5.514 ± 0.164   5.903 ± 0.150    +7.1%  REGRESSION
ShrOperationBenchmarkV2  SHIFT_255             5.663 ± 0.176   5.892 ± 0.168    +4.0%  noise
ShrOperationBenchmarkV2  OVERFLOW_SHIFT_256    5.079 ± 0.128   5.710 ± 0.197   +12.4%  REGRESSION
ShrOperationBenchmarkV2  OVERFLOW_LARGE_SHIFT  5.227 ± 0.136   5.984 ± 0.142   +14.5%  REGRESSION
ShrOperationBenchmarkV2  FULL_RANDOM          12.658 ± 1.378  10.826 ± 0.742   -14.5%  noise

Summary: 6 statistically significant regressions (+7%–+14%, all sub-1ns absolute) on overflow / specific fixed-shift cases, 1 statistically significant improvement (SHL FULL_RANDOM, −14.6%), 18 within noise.

Conclusion

The refactor is performance-neutral on realistic workloads (random-shift is unchanged or faster). The small overflow-path regressions are sub-nanosecond and not worth a follow-up at this absolute scale. The TODOs have been answered closing them.

A separate question worth a follow-up issue (not this PR): whether re-introducing public-method early returns for shift >= 256 would recover the OVERFLOW regression without bloating the API.

Test plan

• ./gradlew :evm:compileJava passes
• :evm:test - UInt256Test (91), UInt256PropertyBasedTest (118), Shift256OperationsTest (7), ShiftOperationsPropertyBasedTest (10), ShiftOperationsV2PropertyBasedTest (25) - 251 tests, all pass

[parthdagia05]

@lu-pinto @ahamlat