Reduce allocations while loading deferred resources

[MrJul]

What does the pull request do?

This PR aims to remove most unnecessary allocations related to deferred content inside resources dictionaries.
While loading the Fluent theme, there's a bit less than 1450 deferred items created, and that's normal.
However, they come with several allocations, both very short lived ones, and some longer ones, stored.

This PR removes more than 15000 allocations while loading the FluentTheme, half of which were long lived.

Resuls to load the FluentTheme:

Method	Mean	Error	StdDev	Gen0	Gen1	Allocated
FluentTheme (Before)	1.670 ms	0.0192 ms	0.0160 ms	66.4063	64.4531	1.07 MB
FluentTheme (After)	633.4 us	12.60 us	6.59 us	13.6719	12.6953	228.17 KB

Note that it's still the JIT that takes most of the time when starting up the application (use NativeAOT when you can!), but I'll still take those gains :)

As part of this PR, StaticResourceExtension.ProvideValue() was also improved (not that it was slow to begin with):

Method	Mean	Error	StdDev	Gen0	Allocated
StaticResource (Before)	238.5 ns	1.45 ns	1.35 ns	0.0153	256 B
StaticResource (After)	160.6 ns	1.50 ns	1.33 ns	-	-

How was the solution implemented (if it's not obvious)?

All numbers below are for loading the Fluent theme.

Buffer resource nodes

In XamlIlRuntimeHelpers, a reusable buffer is now used to store the parent resources nodes and get an array out of that, without using Linq. The array is stored instead of the list.

Removes ≈4350 allocations:

• 1450 List<IResourceNode> (long lived)
• 1450 OfType<IResourceNode> enumerator (short lived)
• 1450 IResourceNode[] backing array from List having to grow (short lived)

Extra enumerator allocations happening when the deferred resource will be built have also been removed.

Cache default parents

DefaultAvaloniaXamlIlParentStackProvider now caches its parent as an array, since it's only the Application.Current instance, which is very unlikely to change between two calls.

Removes ≈1450 allocations:

• 1450 Parents enumerator (short lived)

List parents without allocating

A specialized version of IAvaloniaXamlIlParentStackProvider is introduced IAvaloniaXamlIlEagerParentStackProvider: it can iterate the parents without any allocation. This interface is implemented by the runtime XAML context.

Removes ≈3500 allocations:

• 3500 ParentStackEnumerable.Enumerator (short lived)

As a bonus, using this interface makes StaticResource.ProvideValue() allocation-free.

Avoid delegate allocations

XamlIlRuntimeHelpers.DeferredTransformationFactoryV3 now returns an IDeferredContent instead of a Func<IServiceProvider, object>.
The returned closure directly implements this interface, so we can get rid of the delegate.
ResourceDictionary is now aware of IDeferredContent and can store it directly.

The XAML compiler now passes a function pointer to DeferredTransformationFactoryV3 to build its content instead of creating a delegate.

Removes ≈4350 allocations:

• 2900 Func<IServiceProvider, object> (long lived)
• 1450 DeferredItem inside the resource dictionary (long lived)

Reuse parent nodes

DeferredTransformationFactoryV3 reuses the parent resource nodes if they're the same as the last time it was called. This method is often called with the same parents (e.g. all items in the same resource dictionary), so this change provides nice gains. Out of 1450 items in the fluent theme, there are only 20 different lists of parents.

Removes ≈1430 allocations:

• 1430 IResourceNode[] (long lived)

Set resource dictionary capacity

ResourceDictionary.EnsureCapacity() is called.
This is #15069 which is the only one I could really make separate (all other modifications are intertwined), but that I measured as part of this PR.

Removes ≈30 allocations, but larger ones:

• 30 Dictionary+Entry[] (short lived)

Notes

ResourceDictionary.AddDeferred(IDeferredContent) and Add(IDeferredContent) are equivalent. I didn't bother with disallowing one or the other, but this might be worth mentioning.

Dependencies

This PR is complete, but is marked as a draft because it contains merges from several other PR:

• Expose XamlILContextDefinition fields kekekeks/XamlX#107
• Advanced deferred customization kekekeks/XamlX#108
• Set capacity of XAML resource dictionaries #15069

When these get merged, I will rebase this PR to have a clean diff.

[avaloniaui-bot]

You can test this PR using the following package version. 11.2.999-cibuild0046430-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[avaloniaui-bot]

You can test this PR using the following package version. 11.2.999-cibuild0046841-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[avaloniaui-bot]

You can test this PR using the following package version. 11.2.999-cibuild0047665-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[maxkatz6]

TODO: we should really avoid XamlIL in public API when possible. Not a job for this PR, as it would end up in breaking changes anyway.

[kekekeks]

@MrJul @maxkatz6 It seems that this PR has introduced a breaking change. IAvaloniaXamlIlEagerParentStackProvider isn't supported by XAML compiled by Avalonia 11.0.x, so attempting to cast to it will trigger exceptions on XAML load.

[kekekeks]

You are only doing a type check for the current stack provider, but EagerParentStackEnumerator assumes that all providers in the chain support the new interface