Single constructor reflection optimisation

[alistairjevans]

We didn't actually need to generate extra delegates using the mechanism added in #1321, I've added fast-paths through the reflection activator for single-constructor and matched-sig-constructor selection that does "early binding" of sorts on the constructors where possible.

The approach taken actually means I don't need to change the binding exception handling at all, and the existing resolve check behaviour stays.

We get noticeable perf and memory usage improvements across the board.

Some examples:

DeepGraphResolveBenchmark

# develop
|  Method |     Mean |    Error |   StdDev |  Gen 0 | Allocated |
|-------- |---------:|---------:|---------:|-------:|----------:|
| Resolve | 13.97 μs | 0.061 μs | 0.054 μs | 4.1351 |     17 KB |

# optimised
|  Method |     Mean |    Error |   StdDev |   Median |  Gen 0 | Allocated |
|-------- |---------:|---------:|---------:|---------:|-------:|----------:|
| Resolve | 12.82 μs | 0.254 μs | 0.451 μs | 12.45 μs | 3.5553 |     15 KB |

ChildScopeResolveBenchmark

# develop
|  Method |     Mean |    Error |   StdDev |  Gen 0 |  Gen 1 | Allocated |
|-------- |---------:|---------:|---------:|-------:|-------:|----------:|
| Resolve | 32.08 μs | 0.195 μs | 0.173 μs | 7.4463 | 1.4648 |     31 KB |

# optimised 
|  Method |     Mean |    Error |   StdDev |  Gen 0 |  Gen 1 | Allocated |
|-------- |---------:|---------:|---------:|-------:|-------:|----------:|
| Resolve | 31.47 μs | 0.084 μs | 0.070 μs | 7.3853 | 1.4648 |     30 KB |

EnumerableResolveBenchmark

# develop
|               Method |     Mean |     Error |    StdDev |  Gen 0 | Allocated |
|--------------------- |---------:|----------:|----------:|-------:|----------:|
|   ResolveIEnumerable | 3.066 μs | 0.0223 μs | 0.0186 μs | 0.9727 |      4 KB |
| ResolveIReadOnlyList | 3.070 μs | 0.0072 μs | 0.0064 μs | 0.9804 |      4 KB |
|         ResolveArray | 3.114 μs | 0.0106 μs | 0.0094 μs | 0.9727 |      4 KB |

# optimised
|               Method |     Mean |     Error |    StdDev |   Median |  Gen 0 | Allocated |
|--------------------- |---------:|----------:|----------:|---------:|-------:|----------:|
|   ResolveIEnumerable | 2.747 μs | 0.0122 μs | 0.0114 μs | 2.749 μs | 0.8125 |      3 KB |
| ResolveIReadOnlyList | 2.795 μs | 0.0189 μs | 0.0167 μs | 2.791 μs | 0.8202 |      3 KB |
|         ResolveArray | 2.787 μs | 0.0557 μs | 0.1175 μs | 2.704 μs | 0.8125 |      3 KB |

Multi-Constructor Benchmarks

# develop
|             Method |       Mean |   Error |  StdDev | Ratio |  Gen 0 | Allocated |
|------------------- |-----------:|--------:|--------:|------:|-------:|----------:|
|             Single |   813.9 ns | 2.08 ns | 1.84 ns |  1.00 | 0.3557 |      1 KB |
|                Two | 1,345.0 ns | 3.56 ns | 3.33 ns |  1.65 | 0.5455 |      2 KB |
| TwoValidOneInvalid | 1,684.5 ns | 6.15 ns | 5.75 ns |  2.07 | 0.6866 |      3 KB |
|              Three | 1,944.2 ns | 6.86 ns | 6.08 ns |  2.39 | 0.7782 |      3 KB |
|               Four | 2,568.4 ns | 8.19 ns | 7.66 ns |  3.16 | 1.0529 |      4 KB |

# optimised
|             Method |       Mean |    Error |  StdDev | Ratio | RatioSD |  Gen 0 | Allocated |
|------------------- |-----------:|---------:|--------:|------:|--------:|-------:|----------:|
|             Single |   735.8 ns |  2.58 ns | 2.16 ns |  1.00 |    0.00 | 0.3204 |      1 KB |
|                Two | 1,220.8 ns |  8.80 ns | 6.87 ns |  1.66 |    0.01 | 0.4921 |      2 KB |
| TwoValidOneInvalid | 1,555.0 ns |  4.23 ns | 3.96 ns |  2.11 |    0.01 | 0.6313 |      3 KB |
|              Three | 1,769.7 ns |  6.51 ns | 6.09 ns |  2.40 |    0.01 | 0.7000 |      3 KB |
|               Four | 2,359.6 ns | 10.00 ns | 8.86 ns |  3.21 |    0.02 | 0.9460 |      4 KB |

Improvements on the multi-constructor benchmarks (which use the old path) may be down to the use of the ReferenceEquals check against the NoParameters constant, rather than Any().

[codecov]

Codecov Report

Merging #1325 (efa63ac) into develop (56af5b3) will increase coverage by 0.10%.
The diff coverage is 92.77%.

@@             Coverage Diff             @@
##           develop    #1325      +/-   ##
===========================================
+ Coverage    77.69%   77.80%   +0.10%     
===========================================
  Files          190      190              
  Lines         5412     5488      +76     
  Branches      1084     1097      +13     
===========================================
+ Hits          4205     4270      +65     
- Misses         705      710       +5     
- Partials       502      508       +6     

Impacted Files	Coverage Δ
...Reflection/MatchingSignatureConstructorSelector.cs	72.50% <75.00%> (+1.66%)	⬆️
.../Core/Activators/Reflection/ReflectionActivator.cs	89.51% <96.42%> (+2.01%)	⬆️
src/Autofac/Core/Activators/InstanceActivator.cs	100.00% <100.00%> (ø)
...fac/Core/Activators/Reflection/BoundConstructor.cs	81.08% <100.00%> (+1.08%)	⬆️
...ac/Core/Activators/Reflection/ConstructorBinder.cs	91.93% <100.00%> (+0.13%)	⬆️
src/Autofac/Core/Lifetime/LifetimeScope.cs	81.94% <100.00%> (+0.12%)	⬆️
...rs/Reflection/MostParametersConstructorSelector.cs	92.30% <0.00%> (-7.70%)	⬇️
src/Autofac/Core/ResolvedParameter.cs	50.00% <0.00%> (-4.55%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 56af5b3...efa63ac. Read the comment docs.

[tillig]

Couple minor things, question about whether we should change the IConstructorSelector interface.

[alistairjevans]

Added IConstructorSelectorWithEarlyBinding. Also made it optional for the SelectConstructorBinder function in that interface to return a ConstructorBinder instance. If it returns null, it falls back to the old selection mechanism at resolve time; means implementations can make "best effort" selection attempts.

[tillig]

Looking good. Only real question of substance is about whether IConstructorSelectorWithEarlyBinding should inherit IConstructorSelector.

[tillig]

🚨

[jimmycartrette]

Improvements on the multi-constructor benchmarks (which use the old path) may be down to the use of the ReferenceEquals check against the NoParameters constant, rather than Any().

Would it be because the multi-contructors still have dependencies, and those dependencies have single constructors which are now optimized?

[alistairjevans]

Yes, that's also quite possible; I didn't really investigate the reason that closely.