Use Span to fill List<T> in more ToList scenarios

[brantburnett]

Adds span-based filling of List<T> during calls to ToList using CollectionsMarshal.SetCount to the following additional scenarios:

• ToList of a Union or Distinct
• ToList of a Lookup<TKey, TElement> including cases within grouping
• ToList of a Select projection against an IPartition (i.e. sorted)
• ToList of an OrderedEnumerable returning a single element (i.e. Skip/Take applied to a sorted list for a single result)
• ToList of Append and Prepend scenarios, mostly multiple appends/prepends

Secondary improvements:

• ToArray of an empty Lookup<TKey, TElement> now uses Array.Empty
• ToList of multiple Prepend operations no longer allocates an intermediate array

Fixes #80760

[ghost]

Tagging subscribers to this area: @dotnet/area-system-collections
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Adds span-based filling of List<T> during calls to ToList using CollectionsMarshal.SetCount to the following additional scenarios:

• ToList of a Union or Distinct
• ToList of a Lookup<TKey, TElement> including cases within grouping
• ToList of a Select projection against an IPartition (i.e. sorted)
• ToList of an OrderedEnumerable returning a single element (i.e. Skip/Take applied to a sorted list for a single result)
• ToList of a DefaultIfEmpty where the source is empty and an IIListProvider
• ToList of Append and Prepend scenarios

Secondary improvements:

• ToArray of an empty Lookup<TKey, TElement> now uses Array.Empty

Fixes #80760

Author:	brantburnett
Assignees:	-
Labels:	area-System.Collections
Milestone:	-

[stephentoub]

Thanks. Can you share some benchmarks that demonstrate these improvements are meaningful, in particular in the places that are requiring a lot of code to be added, like in AppendPrepend1Iterator?

[brantburnett]

Thanks. Can you share some benchmarks that demonstrate these improvements are meaningful, in particular in the places that are requiring a lot of code to be added, like in AppendPrepend1Iterator?

Upon more careful review, the added complications in Append/Prepend were mostly not worth it. Missing the optimizations in List<T>.AddRange was degrading performance when the source was an ICollection<T>, such as an array. I've reworked this to be much simpler and it still gets some modest gains, especially on multiple Append.

I'll add benchmarks on some of the other changes shortly.

BenchmarkDotNet=v0.13.2.2052-nightly, OS=Windows 11 (10.0.22621.1776)
Intel Core i7-10850H CPU 2.70GHz, 1 CPU, 12 logical and 6 physical cores
.NET SDK=8.0.100-preview.4.23260.5
[Host] : .NET 8.0.0 (8.0.23.25905), X64 RyuJIT AVX2
Current : .NET 8.0.0 (42.42.42.42424), X64 RyuJIT AVX2
New : .NET 8.0.0 (42.42.42.42424), X64 RyuJIT AVX2

PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:EnableUnsafeBinaryFormatterSerialization=true IterationTime=250.0000 ms
MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	Job	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD	Gen0	Gen1	Allocated	Alloc Ratio
AppendSingleToEmptyArray_ToList	Current	45.65 ns	0.258 ns	0.228 ns	45.58 ns	45.38 ns	46.12 ns	1.00	0.00	0.0190	-	120 B	1.00
AppendSingleToEmptyArray_ToList	New	35.32 ns	0.243 ns	0.228 ns	35.22 ns	35.06 ns	35.79 ns	0.77	0.00	0.0190	-	120 B	1.00
AppendThreeToArray_ToList	Current	115.45 ns	0.436 ns	0.364 ns	115.59 ns	114.95 ns	116.10 ns	1.00	0.00	0.1313	-	824 B	1.00
AppendThreeToArray_ToList	New	101.24 ns	0.252 ns	0.210 ns	101.19 ns	100.93 ns	101.71 ns	0.88	0.00	0.1249	-	784 B	0.95
AppendThreeToEnumerable_ToList	Current	366.20 ns	8.505 ns	9.794 ns	360.39 ns	356.33 ns	389.72 ns	1.00	0.00	0.2489	-	1568 B	1.00
AppendThreeToEnumerable_ToList	New	340.89 ns	2.360 ns	2.092 ns	341.43 ns	336.62 ns	343.84 ns	0.92	0.03	0.2425	0.0014	1528 B	0.97
PrependSingleToEmptyArray_ToList	Current	44.10 ns	0.893 ns	0.836 ns	43.74 ns	42.91 ns	45.72 ns	1.00	0.00	0.0191	-	120 B	1.00
PrependSingleToEmptyArray_ToList	New	34.52 ns	0.080 ns	0.067 ns	34.53 ns	34.41 ns	34.64 ns	0.78	0.02	0.0191	-	120 B	1.00
PrependThreeToArray_ToList	Current	104.97 ns	0.147 ns	0.130 ns	104.96 ns	104.71 ns	105.15 ns	1.00	0.00	0.1247	-	784 B	1.00
PrependThreeToArray_ToList	New	101.45 ns	0.219 ns	0.183 ns	101.42 ns	101.22 ns	101.79 ns	0.97	0.00	0.1249	-	784 B	1.00
PrependThreeToEnumerable_ToList	Current	356.76 ns	7.159 ns	7.352 ns	355.84 ns	344.37 ns	369.00 ns	1.00	0.00	0.2424	-	1528 B	1.00
PrependThreeToEnumerable_ToList	New	343.00 ns	0.533 ns	0.473 ns	343.09 ns	342.02 ns	343.86 ns	0.96	0.02	0.2433	0.0014	1528 B	1.00

[brantburnett]

Here are some benchmarks of some other cases.

BenchmarkDotNet=v0.13.2.2052-nightly, OS=Windows 11 (10.0.22621.1776)
Intel Core i7-10850H CPU 2.70GHz, 1 CPU, 12 logical and 6 physical cores
.NET SDK=8.0.100-preview.4.23260.5
[Host] : .NET 8.0.0 (8.0.23.25905), X64 RyuJIT AVX2
Current : .NET 8.0.0 (42.42.42.42424), X64 RyuJIT AVX2
New : .NET 8.0.0 (42.42.42.42424), X64 RyuJIT AVX2

PowerPlanMode=00000000-0000-0000-0000-000000000000 Arguments=/p:EnableUnsafeBinaryFormatterSerialization=true IterationTime=250.0000 ms
MaxIterationCount=20 MinIterationCount=15 WarmupCount=1

Method	Job	input	Mean	Error	StdDev	Median	Min	Max	Ratio	RatioSD
Distinct_ToList	Current	Array	1,030.0 ns	16.17 ns	15.12 ns	1,026.4 ns	1,007.4 ns	1,059.2 ns	1.00	0.00
Distinct_ToList	New	Array	904.0 ns	8.63 ns	7.21 ns	902.9 ns	890.5 ns	918.3 ns	0.88	0.01
Distinct_ToList	Current	IEnumerable	1,500.5 ns	49.13 ns	56.58 ns	1,486.0 ns	1,437.1 ns	1,622.8 ns	1.00	0.00
Distinct_ToList	New	IEnumerable	1,402.4 ns	10.73 ns	8.96 ns	1,403.7 ns	1,387.5 ns	1,420.4 ns	0.92	0.03
Distinct_ToList	Current	IList	919.4 ns	6.54 ns	5.11 ns	919.9 ns	909.1 ns	927.4 ns	1.00	0.00
Distinct_ToList	New	IList	867.8 ns	10.52 ns	9.32 ns	864.0 ns	857.5 ns	887.2 ns	0.94	0.01
Distinct_ToList	Current	List	950.9 ns	17.87 ns	19.12 ns	943.2 ns	922.2 ns	991.8 ns	1.00	0.00
Distinct_ToList	New	List	886.4 ns	5.57 ns	4.94 ns	885.8 ns	880.3 ns	897.9 ns	0.93	0.02
GroupBy_ToList	Current	Array	1,942.0 ns	50.05 ns	57.64 ns	1,914.4 ns	1,862.5 ns	2,038.9 ns	1.00	0.00
GroupBy_ToList	New	Array	1,840.7 ns	34.90 ns	32.64 ns	1,831.2 ns	1,798.7 ns	1,898.6 ns	0.94	0.03
GroupBy_ToList	Current	IEnumerable	1,863.5 ns	15.90 ns	12.42 ns	1,858.5 ns	1,850.1 ns	1,889.9 ns	1.00	0.00
GroupBy_ToList	New	IEnumerable	1,854.5 ns	51.87 ns	59.73 ns	1,844.6 ns	1,782.0 ns	1,973.5 ns	0.99	0.03
GroupBy_ToList	Current	IList	1,841.0 ns	7.91 ns	7.40 ns	1,841.0 ns	1,829.5 ns	1,853.3 ns	1.00	0.00
GroupBy_ToList	New	IList	1,796.9 ns	5.67 ns	5.31 ns	1,796.0 ns	1,788.1 ns	1,805.2 ns	0.98	0.01
GroupBy_ToList	Current	List	1,887.6 ns	46.44 ns	53.48 ns	1,916.0 ns	1,790.1 ns	1,938.7 ns	1.00	0.00
GroupBy_ToList	New	List	1,767.3 ns	19.99 ns	17.72 ns	1,761.3 ns	1,741.7 ns	1,805.6 ns	0.92	0.02
OrderSelect_ToList	Current	Array	552.8 ns	9.48 ns	8.40 ns	550.3 ns	540.3 ns	575.5 ns	1.00	0.00
OrderSelect_ToList	New	Array	521.4 ns	15.42 ns	17.76 ns	513.1 ns	502.0 ns	544.8 ns	0.95	0.04
OrderSelect_ToList	Current	IEnumerable	848.0 ns	21.32 ns	23.70 ns	835.6 ns	820.8 ns	894.6 ns	1.00	0.00
OrderSelect_ToList	New	IEnumerable	822.6 ns	7.27 ns	6.07 ns	820.5 ns	815.5 ns	831.9 ns	0.97	0.03
OrderSelect_ToList	Current	IList	520.7 ns	2.68 ns	2.51 ns	520.2 ns	517.5 ns	525.3 ns	1.00	0.00
OrderSelect_ToList	New	IList	478.6 ns	2.94 ns	2.45 ns	478.6 ns	474.6 ns	483.3 ns	0.92	0.01
OrderSelect_ToList	Current	List	520.8 ns	8.18 ns	7.25 ns	517.9 ns	514.3 ns	541.3 ns	1.00	0.00
OrderSelect_ToList	New	List	478.7 ns	9.21 ns	9.04 ns	477.1 ns	466.7 ns	492.3 ns	0.92	0.02

[brantburnett]

Are there any further changes required on this?

[eiriktsarpalis]

LGTM @brantburnett, apologies for the delay in review. Before we get this merged, any chance you could remove the singleton optimization since it doesn't appear to be contributing substantially?

[brantburnett]

LGTM @brantburnett, apologies for the delay in review. Before we get this merged, any chance you could remove the singleton optimization since it doesn't appear to be contributing substantially?

Sure, I'll try to get it done today.

[eiriktsarpalis]

Thanks!