InlineArray in generic class

Steps to Reproduce:

This example code does nothing with the parametric type value except to read and return as a result.
Therefore, it was surprising to find a strong difference in the code size and performance for value and reference types.
Is this expected?

See sharplab

Expected Behavior:

Code for value types and reference types is almost the same, performance is almost the same.

Actual Behavior:

Code for reference types is much larger and works 2-3 times slower.

Moving to runtime as this is a question about the ASM code generation.

@JulieLeeMSFT

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

The generic type is used for InlineArrayAsSpan calls generated under-the-hood by Roslyn. The JIT does not manage to remove all those runtime lookups as some of them feed control flow (in the end, there is no difference in taking the true and false path for these tests, but the JIT's DCE is not able to reason that far).
We could clean this up if we taught the early liveness passes how to DCE other things than stores, since then we can do it before we expand the runtime lookups into control flow here.

For current 9.0 milestone I'd ask to add a warning in docs "Don't use InlineArray with reference types for performance reason". This will save someone a lot of time searching for the reasons why their code is slow.

Can you share a benchmark that shows a perf problem warranting such a statement? Your code here always throws NullReferenceException.

Yes, sharplab example is minimal asm demo.

I attached working benchmark project for you.
genericlist.zip

@jkotas

I get the following results:

Method	Mean	Error	StdDev	Ratio	RatioSD	Code Size
Array	3.570 ns	0.0861 ns	0.0763 ns	1.00	0.00	71 B
Array2	5.888 ns	0.0634 ns	0.0593 ns	1.65	0.05	43 B
IPAddress1	6.196 ns	0.1363 ns	0.1208 ns	1.74	0.06	145 B
IPAddress21	3.590 ns	0.0596 ns	0.0558 ns	1.01	0.03	53 B

In general we expect shared generic code to be slightly slower than unshared code. For example, you will also see a slowdown for List<string>.set_Item compared to List<int>.set_Item. We don't try to document this anywhere, so I'm not sure it's warranted to do it for this case either, especially considering that I expect we can improve on this (either by improving our DCE, or by implementing #105586).

Thoughts @dotnet/jit-contrib @jkotas?

JIT: Teach middle-end liveness to DCE dead calls without sideeffects

In general we expect shared generic code to be slightly slower than unshared code. For example, you will also see a slowdown for List.set_Item compared to List.set_Item.

While I was researching this problem, I measured this example too and the differences there are scanty, there is an understanding of this and it is acceptable.

If we compare 1 and 3 lines (InlineArray code - value vs referense) - diff is 2.47!

| Method      | Mean      | Error     | StdDev    | Ratio | RatioSD | Code Size |
|------------ |----------:|----------:|----------:|------:|--------:|----------:|
| Array       | 13.817 ns | 0.1581 ns | 0.1479 ns |  1.00 |    0.00 |      81 B |
| Array2      |  9.376 ns | 0.2055 ns | 0.1922 ns |  0.68 |    0.02 |      47 B |
| IPAddress1  | 34.252 ns | 0.4945 ns | 0.4129 ns |  2.47 |    0.04 |     184 B |
| IPAddress21 | 11.089 ns | 0.2526 ns | 0.2363 ns |  0.80 |    0.02 |      54 B |

I don't insist on documenting. But the problem is that there are intuitive expectations that the new InlineArray is designed for high-performance code, although it works worse than the usual language constructs.

JIT: Fix `gtNodeHasSideEffects` checking call arguments

I think we can get a fix for this into .NET 9. #106183 and #106185 both fix the performance disparity here.

Navigation Menu

Repository navigation

InlineArray in generic class #106129

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

InlineArray in generic class #106129

Description

Activity

jaredpar commented on Aug 8, 2024

dotnet-policy-service commented on Aug 8, 2024

jakobbotsch commented on Aug 8, 2024

iSazonov commented on Aug 9, 2024

jakobbotsch commented on Aug 9, 2024

iSazonov commented on Aug 9, 2024

jakobbotsch commented on Aug 9, 2024

iSazonov commented on Aug 9, 2024

jakobbotsch commented on Aug 9, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions