Improve `FrozenDictionary.ToFrozenDictionary` for Dictionary sources with different comparers #120795

prozolic · 2025-10-16T12:39:35Z

Summary

This PR improves the performance of FrozenDictionary.ToFrozenDictionary method.

Implementation Approach

When the source is already Dictionary<TKey, TValue> and satisfies the condition Dictionary.Count != 0 && !Dictionary.Comparer.Equals(comparer), the processing is changed to the following point.

Iterate directly over newDictionary instead of source to eliminate casting overhead

Benchmark

BenchmarkDotNet v0.14.1-nightly.20250107.205, Windows 11 (10.0.26100.6584)
11th Gen Intel Core i7-11700 2.50GHz, 1 CPU, 16 logical and 8 physical cores
.NET SDK 10.0.100-rc.1.25451.107
  [Host]     : .NET 10.0.0 (10.0.25.45207), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-SADIXB : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

OutlierMode=DontRemove  PowerPlanMode=00000000-0000-0000-0000-000000000000  Toolchain=CoreRun  
IterationTime=250ms  MaxIterationCount=20  MemoryRandomization=True  
MinIterationCount=15  WarmupCount=1  

| Method              | Size  | Mean          | Error         | StdDev        | Median        | Min           | Max           | Code Size | Gen0     | Gen1     | Gen2     | Allocated |
|-------------------- |------ |--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|----------:|---------:|---------:|---------:|----------:|
| Original_Dictionary | 1     |      50.77 ns |      5.017 ns |      5.778 ns |      48.29 ns |      45.60 ns |      68.21 ns |   2,243 B |   0.0409 |        - |        - |     344 B |
| PR_Dictionary       | 1     |      36.30 ns |      1.882 ns |      2.167 ns |      35.54 ns |      34.84 ns |      42.11 ns |   1,892 B |   0.0354 |        - |        - |     296 B |
| Original_Dictionary | 10    |     270.57 ns |     17.663 ns |     20.341 ns |     264.05 ns |     250.11 ns |     317.83 ns |   3,179 B |   0.1185 |        - |        - |     992 B |
| PR_Dictionary       | 10    |     173.65 ns |      3.315 ns |      3.256 ns |     173.26 ns |     169.91 ns |     182.74 ns |   2,634 B |   0.1124 |        - |        - |     944 B |
| Original_Dictionary | 100   |   2,308.68 ns |    171.190 ns |    197.143 ns |   2,239.27 ns |   2,146.19 ns |   2,915.64 ns |   3,194 B |   0.9931 |   0.0170 |        - |    8328 B |
| PR_Dictionary       | 100   |   1,371.95 ns |     31.095 ns |     35.809 ns |   1,362.03 ns |   1,348.04 ns |   1,513.92 ns |   2,646 B |   0.9881 |   0.0165 |        - |    8280 B |
| Original_Dictionary | 1000  |  19,235.59 ns |    697.651 ns |    803.416 ns |  18,891.83 ns |  18,611.36 ns |  21,727.45 ns |   2,992 B |   9.6957 |   1.8646 |        - |   81304 B |
| PR_Dictionary       | 1000  |  13,123.78 ns |    558.138 ns |    642.753 ns |  12,917.55 ns |  12,450.91 ns |  14,648.06 ns |   2,981 B |   9.6845 |   1.3478 |        - |   81256 B |
| Original_Dictionary | 10000 | 257,255.96 ns | 26,300.558 ns | 30,287.762 ns | 246,032.88 ns | 224,345.99 ns | 362,544.12 ns |   2,993 B | 124.0672 | 124.0672 | 124.0672 |  753284 B |
| PR_Dictionary       | 10000 | 179,099.95 ns |  6,462.982 ns |  7,442.779 ns | 178,413.09 ns | 168,169.41 ns | 198,089.87 ns |   1,573 B | 124.3421 | 124.3421 | 124.3421 |  753236 B |

[BenchmarkCategory(Categories.Libraries, Categories.LINQ)]
public class Perf_FrozenDictionary
{
    private Dictionary<int, int> _dictionary;

    [Params(1, 10, 100, 1000, 10000)]
    public int Size;

    [GlobalSetup]
    public void GlobalSetup()
    {
        _dictionary = Enumerable.Range(0, Size).ToDictionary(i => i, NonDefaultEqualityComparer<int>.Instance);
    }

    [Benchmark]
    [MemoryRandomization]
    public FrozenDictionary<int, int> Original_Dictionary()
    {
        return _dictionary.ToFrozenDictionary();
    }

    [Benchmark]
    [MemoryRandomization]
    public FrozenDictionary<int, int> PR_Dictionary()
    {
        return _dictionary.ToFrozenDictionaryPR();
    }
}


public sealed class NonDefaultEqualityComparer<TKey> : IEqualityComparer<TKey>
{
    public static NonDefaultEqualityComparer<TKey> Instance { get; } = new();
    public bool Equals(TKey x, TKey y) => EqualityComparer<TKey>.Default.Equals(x, y);
    public int GetHashCode([DisallowNull] TKey obj) => EqualityComparer<TKey>.Default.GetHashCode(obj);
}

…th different comparers

stephentoub · 2025-10-16T13:15:02Z

src/libraries/System.Collections.Immutable/src/System/Collections/Frozen/FrozenDictionary.cs

            }
+            else if (newDictionary.Count != 0 && !newDictionary.Comparer.Equals(comparer))
+            {
+                var dictionary = new Dictionary<TKey, TValue>(newDictionary.Count, comparer);


This could end up significantly over-allocating, no? Let's say the original use an Ordinal comparer, the new uses an OrdinalIgnoreCase comparer, and the original had every string entry copied many times just with different casing. The resulting dictionary's count is going to be much smaller than the originals, yet this will cause it to allocate much more memory?

@stephentoub You're correct. Pre-allocating the dictionary with the original size may result in over-allocating when duplicate keys reduce the required capacity. I have removed the initial capacity specification.

Thanks. At that point then, is the only difference this new code path adds is iterating over a Dictionary<> in a strongly-typed way rather than iterating over a dictionary via IEnumerable<KeyValuePair<>>? In this state, have you tried running the before/after benchmarks on .NET 10?

I've updated the benchmark results accordingly.

BenchmarkDotNet v0.14.1-nightly.20250107.205, Windows 11 (10.0.26100.6584) 11th Gen Intel Core i7-11700 2.50GHz, 1 CPU, 16 logical and 8 physical cores .NET SDK 10.0.100-rc.1.25451.107 [Host] : .NET 10.0.0 (10.0.25.45207), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI Job-SADIXB : .NET 10.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI OutlierMode=DontRemove PowerPlanMode=00000000-0000-0000-0000-000000000000 Toolchain=CoreRun IterationTime=250ms MaxIterationCount=20 MemoryRandomization=True MinIterationCount=15 WarmupCount=1 | Method | Size | Mean | Error | StdDev | Median | Min | Max | Code Size | Gen0 | Gen1 | Gen2 | Allocated | |-------------------- |------ |--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|----------:|---------:|---------:|---------:|----------:| | Original_Dictionary | 1 | 50.77 ns | 5.017 ns | 5.778 ns | 48.29 ns | 45.60 ns | 68.21 ns | 2,243 B | 0.0409 | - | - | 344 B | | PR_Dictionary | 1 | 36.30 ns | 1.882 ns | 2.167 ns | 35.54 ns | 34.84 ns | 42.11 ns | 1,892 B | 0.0354 | - | - | 296 B | | Original_Dictionary | 10 | 270.57 ns | 17.663 ns | 20.341 ns | 264.05 ns | 250.11 ns | 317.83 ns | 3,179 B | 0.1185 | - | - | 992 B | | PR_Dictionary | 10 | 173.65 ns | 3.315 ns | 3.256 ns | 173.26 ns | 169.91 ns | 182.74 ns | 2,634 B | 0.1124 | - | - | 944 B | | Original_Dictionary | 100 | 2,308.68 ns | 171.190 ns | 197.143 ns | 2,239.27 ns | 2,146.19 ns | 2,915.64 ns | 3,194 B | 0.9931 | 0.0170 | - | 8328 B | | PR_Dictionary | 100 | 1,371.95 ns | 31.095 ns | 35.809 ns | 1,362.03 ns | 1,348.04 ns | 1,513.92 ns | 2,646 B | 0.9881 | 0.0165 | - | 8280 B | | Original_Dictionary | 1000 | 19,235.59 ns | 697.651 ns | 803.416 ns | 18,891.83 ns | 18,611.36 ns | 21,727.45 ns | 2,992 B | 9.6957 | 1.8646 | - | 81304 B | | PR_Dictionary | 1000 | 13,123.78 ns | 558.138 ns | 642.753 ns | 12,917.55 ns | 12,450.91 ns | 14,648.06 ns | 2,981 B | 9.6845 | 1.3478 | - | 81256 B | | Original_Dictionary | 10000 | 257,255.96 ns | 26,300.558 ns | 30,287.762 ns | 246,032.88 ns | 224,345.99 ns | 362,544.12 ns | 2,993 B | 124.0672 | 124.0672 | 124.0672 | 753284 B | | PR_Dictionary | 10000 | 179,099.95 ns | 6,462.982 ns | 7,442.779 ns | 178,413.09 ns | 168,169.41 ns | 198,089.87 ns | 1,573 B | 124.3421 | 124.3421 | 124.3421 | 753236 B |

@stephentoub from the updated benchmark result I see the original one that reserves the initial capacity allocated far less than the new one that doesn't do it. Instead of letting the backing storage to resizing its size to 2x each time (which would also result in over-allocating), pre-allocating the capacity that may be a bit more than the actual need seems definite more feasible to me.

It's going to depend on how much of a difference is expected in count from the original collection that's using a different comparer. The benchmark is using a comparer with identical semantics, so the ideal is to just use the exact same capacity. If, however and for example, the original collection was using Ordinal, the new collection was using OrdinalIgnoreCase, and the original collection contained 1000 words each with 10 different casing variations, choosing that original count is going to result in 10x the allocation required whereas the normal growth strategy will only be 2x.

I initially suggested pre-allocating, but after considering @stephentoub's concern about the worst-case scenario, I agree that not setting the initial capacity is the better approach.
Since the optimal capacity depends on on the comparer's behavior, the dynamic growth strategy is more robust.

dotnet-policy-service · 2025-10-16T13:40:35Z

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

prozolic added 2 commits October 16, 2025 20:58

Improve FrozenDictionary.ToFrozenDictionary for Dictionary sources wi…

5b26d19

…th different comparers

Fix file encoding to UTF-8 with BOM

9fff89d

github-actions bot added the needs-area-label An area label is needed to ensure this gets routed to the appropriate area owners label Oct 16, 2025

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Oct 16, 2025

stephentoub reviewed Oct 16, 2025

View reviewed changes

stephentoub added area-System.Collections and removed needs-area-label An area label is needed to ensure this gets routed to the appropriate area owners labels Oct 16, 2025

prozolic added 2 commits October 16, 2025 22:48

Remove initial capacity specification

089e68b

Merge branch 'main' into FrozenDictionary.ToFrozenDictionary

e2dbcf6

build-analysis bot mentioned this pull request Oct 19, 2025

Unable to pull image from mcr.microsoft.com #117164

Open

prozolic requested a review from stephentoub October 25, 2025 04:56

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Improve `FrozenDictionary.ToFrozenDictionary` for Dictionary sources with different comparers #120795

Improve `FrozenDictionary.ToFrozenDictionary` for Dictionary sources with different comparers #120795

Uh oh!

prozolic commented Oct 16, 2025 •

edited

Loading

Uh oh!

stephentoub Oct 16, 2025

Uh oh!

prozolic Oct 16, 2025

Uh oh!

stephentoub Oct 16, 2025 •

edited

Loading

Uh oh!

prozolic Oct 16, 2025 •

edited

Loading

Uh oh!

hez2010 Oct 16, 2025 •

edited

Loading

Uh oh!

stephentoub Oct 16, 2025

Uh oh!

prozolic Oct 19, 2025

Uh oh!

dotnet-policy-service bot commented Oct 16, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Improve FrozenDictionary.ToFrozenDictionary for Dictionary sources with different comparers #120795

Are you sure you want to change the base?

Improve FrozenDictionary.ToFrozenDictionary for Dictionary sources with different comparers #120795

Uh oh!

Conversation

prozolic commented Oct 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Implementation Approach

Benchmark

Uh oh!

stephentoub Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

prozolic Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

stephentoub Oct 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

prozolic Oct 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hez2010 Oct 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

stephentoub Oct 16, 2025

Choose a reason for hiding this comment

Uh oh!

prozolic Oct 19, 2025

Choose a reason for hiding this comment

Uh oh!

dotnet-policy-service bot commented Oct 16, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Improve `FrozenDictionary.ToFrozenDictionary` for Dictionary sources with different comparers #120795

Improve `FrozenDictionary.ToFrozenDictionary` for Dictionary sources with different comparers #120795

prozolic commented Oct 16, 2025 •

edited

Loading

stephentoub Oct 16, 2025 •

edited

Loading

prozolic Oct 16, 2025 •

edited

Loading

hez2010 Oct 16, 2025 •

edited

Loading