PerfLabTests.LowLevelPerf.StaticDelegate has regressed on Linux x64

[adamsitnik]

PerfLabTests.LowLevelPerf.StaticDelegate has regressed on Linux x64.

Reporting System:

Looking at the commit diff, #65738 seems to be to blame (cc @janvorli).

Repro:

git clone https://github.com/dotnet/performance.git
python3 ./performance/scripts/benchmarks_monthly.py net7.0-preview2 net7.0-preview3 --filter PerfLabTests.LowLevelPerf.StaticDelegate

PerfLabTests.LowLevelPerf.StaticDelegate

Result	Base	Diff	Ratio	Operating System	Bit	Processor Name
Same	723785.36	723696.04	1.00	alpine 3.12	Arm64	Unknown processor
Same	406036.94	406404.81	1.00	debian 11	Arm64	Unknown processor
Same	723245.26	723231.85	1.00	ubuntu 18.04	Arm64	Unknown processor
Slower	234832.09	267772.62	0.88	Windows 10	Arm64	Microsoft SQ1 3.0 GHz
Slower	235887.13	269869.45	0.87	Windows 11	Arm64	Microsoft SQ1 3.0 GHz
Same	205188.98	205032.81	1.00	macOS Monterey 12.3	Arm64	Apple M1 Max
Same	257062.70	246622.22	1.04	Windows 10	X64	Intel Xeon Platinum 8272CL CPU 2.60GHz
Slower	238081.16	264541.67	0.90	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Same	210286.75	246663.03	0.85	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Same	143409.46	142170.65	1.01	Windows 11	X64	AMD Ryzen 9 5900X
Same	332630.65	350222.08	0.95	Windows 11	X64	Intel Core i5-4300U CPU 1.90GHz (Haswell)
Same	304547.76	319492.63	0.95	Windows 11	X64	Unknown processor
Same	199884.10	199594.98	1.00	Windows 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Same	230565.99	233570.36	0.99	Windows 11	X64	Intel Core i9-9900T CPU 2.10GHz
Same	200459.38	217517.14	0.92	alpine 3.13	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Same	380576.28	376321.13	1.01	centos 7	X64	Intel Xeon CPU E5530 2.40GHz
Slower	194035.49	219647.31	0.88	debian 11	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	195775.15	220128.56	0.89	opensuse-leap 15.3	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	241212.25	275027.68	0.88	pop 20.04	X64	Intel Core i7-6600U CPU 2.60GHz (Skylake)
Slower	202248.52	228340.18	0.89	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Faster	358265.98	307771.13	1.16	ubuntu 18.04	X64	Intel Core i7-2720QM CPU 2.20GHz (Sandy Bridge)
Faster	435897.22	368043.75	1.18	Windows 10	Arm	Microsoft SQ1 3.0 GHz
Same	867438.71	867460.14	1.00	ubuntu 18.04	Arm	ARMv7 Processor rev 3 (v7l)
Slower	238061.08	264550.78	0.90	Windows 10	X86	Intel Xeon CPU E5-1650 v4 3.60GHz
Same	184812.57	184752.06	1.00	Windows 11	X86	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Same	454396.72	516626.95	0.88	Windows 11	X86	Intel Core i7-10510U CPU 1.80GHz
Same	242192.88	252374.60	0.96	Windows 7 SP1	X86	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	235410.75	264838.38	0.89	macOS Monterey 12.2.1	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)
Slower	218997.66	246448.38	0.89	macOS Monterey 12.3.1	X64	Intel Core i7-4870HQ CPU 2.50GHz (Haswell)

[janvorli]

I know of this regression. The reason is that the delegate always goes to the target method through a FixupPrecode. With my change, the precode in stable state changed from a direct jump to an indirect jump to the target. I have discussed possible ways to make delegates jump directly to the target method without going through the fixup with David Wrighton and it would require a non-trivial change doing some kind of back patching of the delegates. I am planning to look into it, but the timeframe is not clear yet.

[janvorli]

I did an experiment to asses the possibility of making delegates call the target method directly. While it improved this benchmark, I haven't seen any perf gain on real world ASPNet benchmarks that uses delegates a lot (the Orchard / Fortunes). Given this and the complexity of the necessary changes it seems it is not worth the effort.
Please note that this benchmark calls the delegate pointing to an empty function, so the effect of my change is very amplified compared to real world scenarios.
So I am closing this as won't fix, @adamsitnik please let me know if you disagree.