Performance regressions in Quaternion.Conjugate and Quaternion.Negate

[adamsitnik]

It looks like Conjugate and Negate methods of the Quaternion type have regressed compared to 3.1.

@tannergooding could you please take a look and triage it? I assume that this might be acceptable similarly to #39035 but I don't have enough knowledge to make any decisions here.

Repro

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f netcoreapp3.1 netcoreapp5.0 --filter 'System.Numerics.Tests.Perf_Quaternion.ConjugateBenchmark' 'System.Numerics.Tests.Perf_Quaternion.Negat*'

To see all the numbers please click "details" below:

System.Numerics.Tests.Perf_Quaternion.ConjugateBenchmark

Result	Base	Diff	Ratio	Alloc Delta	Modality	Operating System	Bit	Processor Name	Base V	Diff V
Slower	1.30	7.91	0.16	+0		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Slower	6.43	18.76	0.34	+0		manjaro	X64	Intel Core i7-4771 CPU 3.50GHz (Haswell)	3.1.6	5.0.20.41714
Slower	12.90	21.39	0.60	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	12.90	27.30	0.47	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.7	5.0.20.41714
Slower	14.44	24.09	0.60	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	1.00	7.67	0.13	+0		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Faster	9.44	5.20	1.81	+0		Windows 10.0.19041.450	Arm	Microsoft SQ1 3.0 GHz	3.1.6	5.0.20.41714
Slower	8.01	23.80	0.34	+0		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	3.1.6	5.0.20.41714

System.Numerics.Tests.Perf_Quaternion.NegateBenchmark

Result	Base	Diff	Ratio	Alloc Delta	Modality	Operating System	Bit	Processor Name	Base V	Diff V
Slower	1.54	7.44	0.21	+0		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Slower	6.49	17.96	0.36	+0		manjaro	X64	Intel Core i7-4771 CPU 3.50GHz (Haswell)	3.1.6	5.0.20.41714
Slower	13.67	22.12	0.62	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	12.52	20.97	0.60	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.7	5.0.20.41714
Slower	14.06	18.74	0.75	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	1.26	7.64	0.16	+0	bimodal	Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Faster	9.52	5.05	1.89	+0		Windows 10.0.19041.450	Arm	Microsoft SQ1 3.0 GHz	3.1.6	5.0.20.41714
Slower	7.99	23.96	0.33	+0		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	3.1.6	5.0.20.41714

System.Numerics.Tests.Perf_Quaternion.NegationOperatorBenchmark

Result	Base	Diff	Ratio	Alloc Delta	Modality	Operating System	Bit	Processor Name	Base V	Diff V
Slower	1.56	7.47	0.21	+0		Windows 10.0.18363.959	X64	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Slower	6.48	18.24	0.36	+0		manjaro	X64	Intel Core i7-4771 CPU 3.50GHz (Haswell)	3.1.6	5.0.20.41714
Slower	12.90	22.14	0.58	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	12.90	22.12	0.58	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.7	5.0.20.41714
Slower	13.67	24.44	0.56	+0		ubuntu 16.04	Arm64	Unknown processor	3.1.6	5.0.20.41714
Slower	1.26	7.16	0.18	+0		Windows 10.0.18363.959	X86	Intel Xeon CPU E5-1650 v4 3.60GHz	3.1.6	5.0.20.41714
Faster	9.60	5.15	1.86	+0		Windows 10.0.19041.450	Arm	Microsoft SQ1 3.0 GHz	3.1.6	5.0.20.41714
Slower	7.98	21.66	0.37	+0		macOS Mojave 10.14.5	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)	3.1.6	5.0.20.41714

@DrewScoggins this regression did not get detected by the bot as it was added very recently 👍

Tagging subscribers to this area: @tannergooding, @pgovind
See info in area-owners.md if you want to be subscribed.

[tannergooding]

This one is also a bit strange as there were no codechanges to these Quaternion methods between 3.1 and 5.0.

I'm guessing there is some codegen change by the JIT instead.

[jeffhandley]

@jeffschwMSFT Is this being reviewed on your team or is there routing still needed here? Thanks!

[jeffschwMSFT]

@JulieLeeMSFT

[tannergooding]

Looks like the Conjugate and Negate methods aren't being inlined in .NET 5, which is leading to the perf slowdown.

However, as indicated the managed implementation hasn't changed and is actually rather simple: https://github.com/dotnet/corefx/blob/release/3.1/src/System.Numerics.Vectors/src/System/Numerics/Quaternion.cs#L124, so this is likely to do with one of the JIT changes.
Here is the netcoreapp3.1 disassembly: System.Numerics.Tests.Perf_Quaternion-asm.md.txt
Here is the net5.0 disassembly: System.Numerics.Tests.Perf_Quaternion-asm.md.txt

[JulieLeeMSFT]

CC @dotnet/jit-contrib

[AndyAyersMS]

I'll take a look.