Reflection Constructor Binding Refactor

[alistairjevans]

I've refactored/replaced the existing constructor binding logic to do more work up front and generally be faster.

Among key behaviour changes:

• The binding process has been split into a ConstructorBinder, which at resolve time will 'bind' and output a BoundConstructor, which can then be used to instantiate the instance.
• We now build the method used to resolve the constructor at container build time, rather than when it is first resolved. This means we don't have to hit the constructor delegate dictionary every time we resolve, only when we create the activator.
• Reduced LINQ usage in the ReflectionActivator (removed Concat).

Some Benchmarks:

TLDR; This change improves across the board over v6, particularly with deep reflection graphs (DeepGraph is a full 1μs faster than develop now).

v6 is now generally faster than develop, apart from a few cases, particularly on concurrency where it still lags behind develop.

ChildScopeResolve

develop

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	50.65 us	0.395 us	0.350 us	11.9629	-	-	49.07 KB

v6

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	35.98 μs	0.279 μs	0.261 μs	7.8125	-	-	31.97 KB

constructor-binder

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	36.02 μs	0.358 μs	0.317 μs	7.7515	-	-	31.85 KB

Concurrency

v6 is still generally slower on concurrency. That's one for another day.

develop

Method	ResolveTaskCount	ResolvesPerTask	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
MultipleResolvesOnMultipleTasks	100	100	2.529 ms	0.0140 ms	0.0131 ms	1234.3750	-	-	4.9 MB

v6

Method	ResolveTaskCount	ResolvesPerTask	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
MultipleResolvesOnMultipleTasks	100	100	2.861 ms	0.0355 ms	0.0332 ms	1960.9375	-	-	7.64 MB

constructor-binder

Method	ResolveTaskCount	ResolvesPerTask	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
MultipleResolvesOnMultipleTasks	100	100	2.828 ms	0.0476 ms	0.0422 ms	1960.9375	-	-	7.64 MB

Decorators.KeylessSimple

develop

Method	repetitions	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Baseline	?	592.2 ns	4.18 ns	3.49 ns	1.00	0.00	0.5083	-	-	2.08 KB
ResolveEnumerableT	1	4,105.6 ns	18.88 ns	16.74 ns	?	?	1.5717	-	-	6.44 KB
ResolveT	1	1,973.8 ns	10.16 ns	9.01 ns	?	?	0.9918	-	-	4.05 KB
ResolveEnumerableT	2	7,380.3 ns	50.42 ns	44.70 ns	?	?	2.6398	-	-	10.8 KB
ResolveT	2	3,296.3 ns	25.38 ns	23.74 ns	?	?	1.4763	-	-	6.03 KB
ResolveEnumerableT	3	10,883.2 ns	66.92 ns	62.60 ns	?	?	3.6926	-	-	15.16 KB
ResolveT	3	4,650.1 ns	24.56 ns	21.77 ns	?	?	1.9608	-	-	8.01 KB

v6

Method	repetitions	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Baseline	?	593.0 ns	3.99 ns	3.11 ns	1.00	0.00	0.5102	-	-	2.09 KB
ResolveEnumerableT	1	4,506.0 ns	37.79 ns	35.35 ns	?	?	1.6785	-	-	6.88 KB
ResolveT	1	2,076.4 ns	11.90 ns	10.55 ns	?	?	1.0414	-	-	4.27 KB
ResolveEnumerableT	2	8,140.0 ns	69.90 ns	65.38 ns	?	?	2.8534	-	-	11.68 KB
ResolveT	2	3,514.9 ns	10.28 ns	8.58 ns	?	?	1.5755	-	-	6.45 KB
ResolveEnumerableT	3	11,660.5 ns	33.15 ns	27.68 ns	?	?	4.0283	-	-	16.48 KB
ResolveT	3	4,899.3 ns	40.67 ns	36.06 ns	?	?	2.1057	-	-	8.63 KB

constructor-binder

Method	repetitions	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Baseline	?	591.6 ns	3.72 ns	3.48 ns	1.00	0.00	0.5102	-	-	2.09 KB
ResolveEnumerableT	1	4,165.5 ns	21.89 ns	19.40 ns	?	?	1.6708	-	-	6.85 KB
ResolveT	1	1,980.0 ns	13.46 ns	11.93 ns	?	?	1.0376	-	-	4.25 KB
ResolveEnumerableT	2	7,437.9 ns	18.74 ns	14.63 ns	?	?	2.8381	-	-	11.62 KB
ResolveT	2	3,267.1 ns	23.52 ns	22.00 ns	?	?	1.5678	-	-	6.41 KB
ResolveEnumerableT	3	11,083.5 ns	82.16 ns	76.85 ns	?	?	3.9978	-	-	16.38 KB
ResolveT	3	4,581.7 ns	54.62 ns	51.09 ns	?	?	2.0981	-	-	8.58 KB

DeepGraphResolve

develop

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	16.23 us	0.066 us	0.058 us	4.4861	-	-	18.45 KB

v6

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	16.39 μs	0.071 μs	0.063 μs	4.4861	-	-	18.37 KB

constructor-binder

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
Resolve	15.20 us	0.123 us	0.115 us	4.3182	-	-	17.65 KB

RootContainerResolve

develop

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
OperatorNew	3.407 ns	0.0293 ns	0.0259 ns	1.00	0.00	0.0057	-	-	24 B
NonSharedReflectionResolve	466.116 ns	2.5108 ns	2.2258 ns	136.81	1.14	0.1645	-	-	688 B
NonSharedDelegateResolve	375.629 ns	2.7164 ns	2.2683 ns	110.25	1.10	0.1278	-	-	536 B
SharedResolve	334.323 ns	1.6897 ns	1.5806 ns	98.14	0.93	0.1221	-	-	512 B

v6

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
OperatorNew	3.443 ns	0.0254 ns	0.0225 ns	1.00	0.00	0.0057	-	-	24 B
NonSharedReflectionResolve	453.595 ns	1.2737 ns	1.0636 ns	131.70	0.86	0.2332	-	-	976 B
NonSharedDelegateResolve	377.809 ns	5.9404 ns	11.4452 ns	111.49	5.19	0.1969	-	-	824 B
SharedResolve	323.010 ns	0.5395 ns	0.4212 ns	93.76	0.62	0.1912	-	-	800 B

constructor-binder

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
OperatorNew	3.403 ns	0.0167 ns	0.0139 ns	1.00	0.00	0.0057	-	-	24 B
NonSharedReflectionResolve	421.828 ns	1.7988 ns	1.5945 ns	123.91	0.76	0.2294	-	-	960 B
NonSharedDelegateResolve	359.356 ns	0.9369 ns	0.7314 ns	105.59	0.41	0.1969	-	-	824 B
SharedResolve	323.966 ns	1.1969 ns	0.9995 ns	95.20	0.60	0.1912	-	-	800 B

OpenGeneric

develop

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
ResolveOpenGeneric	491.7 ns	1.76 ns	1.56 ns	0.1640	-	-	688 B

v6

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
ResolveOpenGeneric	464.4 ns	3.02 ns	2.36 ns	0.2332	-	-	976 B

constructor-binder

Method	Mean	Error	StdDev	Gen 0	Gen 1	Gen 2	Allocated
ResolveOpenGeneric	436.6 ns	3.66 ns	3.42 ns	0.2294	-	-	960 B

[tillig]

I like what I'm seeing here but, as usual, have a couple of questions. Let me know what you think.

[alistairjevans]

Oof, yeah, that MostParametersConstructorSelector was overdue for some love. Refactor is in, results based on the benchmark pending for develop:

develop

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Single	994.6 ns	4.53 ns	4.01 ns	1.00	0.00	0.3071	-	-	1.26 KB
Two	2,112.8 ns	8.60 ns	7.18 ns	2.12	0.01	0.6638	-	-	2.72 KB
Three	3,022.0 ns	9.73 ns	8.13 ns	3.04	0.01	0.9499	-	-	3.88 KB
Four	3,893.8 ns	14.52 ns	12.12 ns	3.91	0.02	1.2970	-	-	5.32 KB

constructor-binder

Method	Mean	Error	StdDev	Ratio	Gen 0	Gen 1	Gen 2	Allocated
Single	893.8 ns	2.50 ns	1.95 ns	1.00	0.3576	-	-	1.46 KB
Two	1,492.8 ns	6.43 ns	5.70 ns	1.67	0.5474	-	-	2.24 KB
Three	2,182.4 ns	9.87 ns	8.75 ns	2.44	0.7820	-	-	3.2 KB
Four	2,999.6 ns	12.39 ns	10.35 ns	3.36	1.0567	-	-	4.32 KB

[alistairjevans]

I've revisited how we allocate bindings, and re-allocate those invalid bindings, as you mentioned.

Where I've ended up is that the IConstructorSelector receives the original array of bindings rather than a filtered set. This means that the selectors are now responsible for ignoring invalid bindings.

This is obviously a breaking behaviour change, but the new requirement is a relatively easy one to integrate for that small set of users who have implemented a custom IConstructorSelector.

Memory and perf gains for the 'one of my constructors isn't valid' path:

develop

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Single	1.010 us	0.0069 us	0.0057 us	1.00	0.00	0.3071	-	-	1.26 KB
Two	2.153 us	0.0161 us	0.0151 us	2.13	0.02	0.6638	-	-	2.72 KB
TwoValidOneInvalid	2.738 us	0.0108 us	0.0101 us	2.71	0.02	0.8392	-	-	3.44 KB
Three	3.041 us	0.0178 us	0.0158 us	3.01	0.03	0.9499	-	-	3.88 KB
Four	3.917 us	0.0192 us	0.0170 us	3.88	0.03	1.2970	-	-	5.32 KB

constructor-binder with filtering before-hand

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Single	893.1 ns	4.59 ns	4.07 ns	1.00	0.00	0.3576	-	-	1.46 KB
Two	1,489.1 ns	9.34 ns	8.74 ns	1.67	0.01	0.5474	-	-	2.24 KB
TwoValidOneInvalid	2,188.6 ns	17.92 ns	16.77 ns	2.45	0.02	0.7324	-	-	2.99 KB
Three	2,171.2 ns	14.34 ns	13.41 ns	2.43	0.02	0.7820	-	-	3.2 KB
Four	2,975.3 ns	18.54 ns	17.35 ns	3.33	0.02	1.0567	-	-	4.32 KB

constructor-binder with selectors getting all bind results

Method	Mean	Error	StdDev	Ratio	Gen 0	Gen 1	Gen 2	Allocated
Single	896.5 ns	4.61 ns	3.85 ns	1.00	0.3576	-	-	1.46 KB
Two	1,511.2 ns	7.57 ns	7.08 ns	1.69	0.5474	-	-	2.24 KB
TwoValidOneInvalid	1,955.7 ns	8.92 ns	8.34 ns	2.18	0.6866	-	-	2.81 KB
Three	2,187.0 ns	9.01 ns	7.99 ns	2.44	0.7820	-	-	3.2 KB
Four	2,964.0 ns	11.71 ns	10.96 ns	3.31	1.0567	-	-	4.32 KB

[tillig]

One minor typo and we're good. This is shaping up to be a really great v6!

[tillig]

🦙