Minor regression in System.Tests.Perf_Char.GetUnicodeCategory for non-ascii characters

[adamsitnik]

I am now running an independent 3.1 vs 5.0 comparison and I've confirmed that it's actual regression introduced in 5.0.

The issue has been originally reported by a bot owned by @DrewScoggins in DrewScoggins/performance-2#574

Repro

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp3.1 netcoreapp5.0 --filter 'System.Tests.Perf_Char.GetUnicodeCategory'

System.Tests.Perf_Char.GetUnicodeCategory(c: '?')

Conclusion	Base	Diff	Base/Diff	Modality	Operating System	Bit	Processor Name	Base Runtime	Diff Runtime
Slower	5.82	7.89	0.74		Windows 10.0.18363.1016	Arm	Microsoft SQ1 3.0 GHz	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Slower	3.15	4.08	0.77		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20404.3
Slower	3.03	4.10	0.74		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	3.40	4.80	0.71		Windows 10.0.19041.450	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Slower	2.63	3.43	0.77		Windows 10.0.19042	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	1.86	4.54	0.41		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20420.2
Slower	2.27	5.28	0.43		Windows 10.0.19041.450	X86	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20419.5
Slower	3.12	3.77	0.83		ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20403.23
Slower	3.52	5.28	0.67		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20404.2
Slower	2.01	3.01	0.67		ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3

No regressions for ascii chars:

System.Tests.Perf_Char.GetUnicodeCategory(c: '.')

Conclusion	Base	Diff	Base/Diff	Modality	Operating System	Bit	Processor Name	Base Runtime	Diff Runtime
Slower	0.66	2.31	0.29		Windows 10.0.18363.1016	Arm	Microsoft SQ1 3.0 GHz	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Same	0.51	0.54	0.94		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20404.3
Same	0.52	0.55	0.95		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Same	0.61	0.61	1.00		Windows 10.0.19041.450	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Same	0.53	0.27	1.92		Windows 10.0.19042	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	0.43	0.76	0.56		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20420.2
Slower	0.50	0.87	0.57		Windows 10.0.19041.450	X86	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20419.5
Same	0.47	0.52	0.92		ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20403.23
Same	0.58	0.59	0.98		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20404.2
Same	0.23	0.47	0.48	several?	ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3

System.Tests.Perf_Char.GetUnicodeCategory(c: '.')

Conclusion	Base	Diff	Base/Diff	Modality	Operating System	Bit	Processor Name	Base Runtime	Diff Runtime
Slower	0.66	2.31	0.29		Windows 10.0.18363.1016	Arm	Microsoft SQ1 3.0 GHz	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Same	0.51	0.54	0.94		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20404.3
Same	0.52	0.55	0.95		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Same	0.61	0.61	1.00		Windows 10.0.19041.450	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Same	0.53	0.27	1.92		Windows 10.0.19042	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	0.43	0.76	0.56		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20420.2
Slower	0.50	0.87	0.57		Windows 10.0.19041.450	X86	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20419.5
Same	0.47	0.52	0.92		ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20403.23
Same	0.58	0.59	0.98		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20404.2
Same	0.23	0.47	0.48	several?	ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3

@GrabYourPitchforks I am adding 6.0 milestone since it's most probably rare and not used on hot-paths

When you take a look at disassembly, it's clear why it has regressed. I am not pasting it here because I plan to run a workshop next week and would love to get the answer from attendees

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[adamsitnik]

FWIW I just got results from @GrabYourPitchforks and his machine is the only one which is not showing the regression. It's also the only Skylake machine I had access to:

System.Tests.Perf_Char.GetUnicodeCategory(c: '?')

Conclusion	Base	Diff	Base/Diff	Modality	Operating System	Bit	Processor Name	Base Runtime	Diff Runtime
Slower	5.82	7.89	0.74		Windows 10.0.18363.1016	Arm	Microsoft SQ1 3.0 GHz	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Slower	3.15	4.08	0.77		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20404.3
Slower	3.03	4.10	0.74		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	3.40	4.80	0.71		Windows 10.0.19041.450	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20413.9
Same	3.72	3.65	1.02		Windows 10.0.19041.450	X64	Intel Core i7-6700 CPU 3.40GHz (Skylake)	.NET Core 3.1.6	5.0.100-rc.1.20419.9
Slower	2.63	3.43	0.77		Windows 10.0.19042	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3
Slower	1.86	4.54	0.41		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20420.2
Slower	2.27	5.28	0.43		Windows 10.0.19041.450	X86	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20419.5
Slower	3.12	3.77	0.83		ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	.NET Core 3.1.6	5.0.100-rc.1.20403.23
Slower	3.52	5.28	0.67		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	.NET Core 3.1.6	5.0.100-rc.1.20404.2
Slower	2.01	3.01	0.67		ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)	.NET Core 3.1.6	5.0.100-rc.1.20418.3

[GrabYourPitchforks]

Direct link to benchmark code (see the [Arguments('\x05D0')] case):

https://github.com/dotnet/performance/blob/1c160f6fd074d104e6de39b2fae48d657673a5ab/src/benchmarks/micro/libraries/System.Runtime/Perf.Char.cs#L85-L89

[danmoseley]

@GrabYourPitchforks @adamsitnik do you believe this should block 6.0 release at this point? BTW, presumably we haven't fixed it in the meantime?

[GrabYourPitchforks]

@danmoseley Looks like average regression is 1ns, with 3ns being the worst case? It's worth investigating why this regressed, but I wouldn't block release for it.

Some really premature speculation: when the original change was checked in, we also did a little bit of size-on-disk work to try to shrink the size of the data tables we carry. Perhaps this microbenchmark shows the result of the extra bit twiddling needed as a result of this size optimization?

[danmoseley]

OK moved.

[GrabYourPitchforks]

Possible low hanging fruit optimization: consider having char.GetUnicodeCategory(char) wrap CharUnicodeInfo.GetUnicodeCategory(char) instead of CharUnicodeInfo.GetUnicodeCategory(int). The int overload performs some extra argument checks that the char overload does not. (Though I don't remember offhand if we call the int overload for codegen optimization purposes.)