Cache FakeOptions constructor to improve performance

[blairconrad]

For a lark, I profiled creating 100,000 instances of a small interface (after warming up FakeItEasy by faking a different interface). Using dotTrace, I found that DefaultFakeAndDummyManager.BuildProxyOptions was a hotspot. So I started caching the constructor by type of Fake and found this difference:

BuildProxyOptions drops from 19.10% of the total time to 9.95%. Or if you prefer the diff view:

Of course we're trading away some storage space. If we end up faking many different types once, this is a bad tradeoff, but if types are faked more than once, it may be a benefit.

Also, in many cases we don't even need a FakeOptions object. I can explore dropping the construction step as well.

[blairconrad]

Actually, I may have come up with a nice way to not construct the FakeOptions object. Will explore later.

[thomaslevesque]

Well done, @blairconrad! Looks like a good improvement.

[thomaslevesque]

Actually, I may have come up with a nice way to not construct the FakeOptions object. Will explore later.

Yes, it would be even better, I guess.

[blairconrad]

Rerunning with FakeOptions creator cached as a delegate (and in release mode this time; other numbers were in debug):

27.04% to 2.93%. Not bad. Thanks for the push, @thomaslevesque. I'll work on avoiding construction next.

[blairconrad]

I was originally thinking we could avoid creating the object when users don't use creation method that specifies an option builder action, but we also have to make sure there's no IFakeOptionsBuilder, and that's more involved. It might still eventually be worthwhile, but if the new 2.93% number is to be believed, maybe it's not our next stop.

[blairconrad]

Using BenchmarkMockNet to test constructors. With FakeItEasy 4.8.1

Method	Mean	Error	StdDev	Scaled	ScaledSD
Stub	8.375 ns	6.764 ns	1.757 ns	1.00	0.00
FakeItEasy	11,059.153 ns	2,079.941 ns	540.257 ns	1,364.10	243.18
Moq	5,059.672 ns	1,041.998 ns	270.655 ns	624.09	111.94
NSubstitute	8,111.013 ns	5,264.036 ns	1,367.314 ns	1,000.46	230.82
Rocks	4,221.011 ns	4,767.949 ns	1,238.457 ns	520.65	165.25

Using BenchmarkMockNet to test constructors. With a build including this issue:

Method	Mean	Error	StdDev	Scaled	ScaledSD
Stub	7.039 ns	2.711 ns	0.7042 ns	1.00	0.00
FakeItEasy	8,046.398 ns	470.824 ns	122.2949 ns	1,151.62	96.61
Moq	6,175.403 ns	4,300.547 ns	1,117.0510 ns	883.84	161.06
NSubstitute	5,959.644 ns	129.430 ns	33.6191 ns	852.96	70.74
Rocks	3,990.786 ns	3,982.395 ns	1,034.4122 ns	571.17	141.03

Take with grain of salt. The computer was doing other things. But still, I'm pleased with the change.

[thomaslevesque]

It certainly looks much better! I think I'll do some profiling tonight, to see if there's more to be gained.

[blairconrad]

I'm sure there's some fat to trim.

[thomaslevesque]

This change has been released in FakeItEasy 5.0.0.