Announcing .NET 6 Preview 3
Today, we are delighted to release .NET 6 Preview 3. This release is dedicated almost entirely to low-level performance features. These are the types of improvements that many folks don’t necessarily always fully appreciate, but they help a lot for many apps. Most of these improvements apply to the CLR type system directly, either making it function faster or better interplay with modern CPUs (think “hardware accelerate the type system”). In the last few years, there have been a few key performance trends with .NET, including: using structs more liberally in libraries, and moving runtime code to C#. Both of trends are visible (directly or indirectly) in these changes. It also demonstrates continued efforts on a focused set of performance strategies.
You will notice that two of these performance changes came from Ben Adams. He’s well known for word play on his first name. In the spirit of April 1st (recently past), perhaps we can virtually rename this release to .NET 6 Preview B3n. Thanks B3n!
You’ll also see performance contributions from Nathan Moore and SingleAccretion. Thanks!
You can download .NET 6 Preview 3, for Windows, macOS, and Linux.
- Installers and binaries
- Container images
- Linux packages
- Release notes
- Known issues
- GitHub issue tracker
See the ASP.NET Core post for more detail on what’s new for web technology. For EF Core, Preview 3 includes several bug fixes and ongoing improvements to infrastructure to support upcoming features including compiled models and temporal tables
.NET 6 has been tested with Visual Studio 16.10 Preview 1 and Visual Studio for Mac 8.9. We recommend you use those builds if you want to try .NET 6.
For Linux users: .NET SDK 6 Preview 3 resolves an issue where NuGet restore fails on Linux due to expired NuGet certificates and unfortunate interactions with changes made to root certificates stores on Linux, carried by ca-certificates and nss packages. Please update your .NET 5 and .NET 6 SDK deployments on Linux, ASAP.
Support
.NET 6 will be released in November 2021, and will be supported for three years, as a Long Term Support (LTS) release. The platform matrix has been significantly expanded.
The additions are:
- Android.
- iOS.
- Mac and Mac Catalyst, for x64 and Apple Silicon (AKA “M1”).
- Windows Arm64 (specifically Windows Desktop).
.NET 6 Debian container images are based on Debian 11 (“bullseye”), which is currently in testing.
Libraries
The following APIs and improvements have been added to the .NET libraries.
Faster handling of structs as Dictionary values
A new unsafe api — CollectionsMarshal.GetValueRefOrNullRef — has been added that makes updating struct values in Dictionaries faster. The new API is intended for high performance scenarios, not for general purpose use. It returns a ref
to the struct value which can then be updated in place with typical techniques.
Prior to this change, updating struct
dictionary values can be expensive for high-performance scenarios, requiring a dictionary lookup and a copy to stack of the struct
; then after changing the struct
, it needs to be assigned to the dictionary key again resulting in another look up and copy operation. This improvement reduces the key hashing to 1 (from 2) and removes all the struct copy operations.
Example:
ref MyStruct value = ref CollectionsMarshal.GetValueRefOrNullRef(dictionary, key);
// Returns Unsafe.NullRef<TValue>() if it doesn't exist; check using Unsafe.IsNullRef(ref value)
if (!Unsafe.IsNullRef(ref value))
{
// Mutate in-place
value.MyInt++;
}
Credit to Ben Adams.
Faster interface checking and casting
Interface casting performance has been boosted by 16% – 38%. This improvement is particularly useful for C#’s pattern matching to and between interfaces.
Method | Implementation | Mean | Version | Ratio | Improvement |
---|---|---|---|---|---|
IsInterface1 | Interfaces1 | 2.166 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface1 | Interfaces1 | 1.629 ns | .NET 6 Preview 3 | 0.752 | x1.3 |
IsNotImplemented | Interfaces1 | 2.166 ns | .NET 6 Preview 2 | 1.000 | |
IsNotImplemented | Interfaces1 | 1.950 ns | .NET 6 Preview 3 | 0.900 | x1.1 |
IsInterface1 | Interfaces6 | 2.155 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface1 | Interfaces6 | 1.936 ns | .NET 6 Preview 3 | 0.898 | x1.1 |
IsInterface2 | Interfaces6 | 2.161 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface2 | Interfaces6 | 1.908 ns | .NET 6 Preview 3 | 0.883 | x1.1 |
IsInterface3 | Interfaces6 | 2.188 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface3 | Interfaces6 | 1.672 ns | .NET 6 Preview 3 | 0.764 | x1.3 |
IsInterface4 | Interfaces6 | 2.346 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface4 | Interfaces6 | 1.965 ns | .NET 6 Preview 3 | 0.838 | x1.2 |
IsInterface5 | Interfaces6 | 3.122 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface5 | Interfaces6 | 2.173 ns | .NET 6 Preview 3 | 0.696 | x1.4 |
IsInterface6 | Interfaces6 | 3.472 ns | .NET 6 Preview 2 | 1.000 | |
IsInterface6 | Interfaces6 | 2.687 ns | .NET 6 Preview 3 | 0.774 | x1.3 |
IsNotImplemented | Interfaces6 | 3.644 ns | .NET 6 Preview 2 | 1.000 | |
IsNotImplemented | Interfaces6 | 3.071 ns | .NET 6 Preview 3 | 0.843 | x1.2 |
One of the biggest advantages of moving parts of the .NET runtime from C++ to managed C# is that it lowers the barrier to contribution. This includes interface casting, which was moved to C# as an early .NET 6 change. Many more people in the .NET ecosystem are literate in C# than C++ (and the runtime isn’t even regular C++). Just being able to read some of the code that composes the runtime is a major step to developing confidence in contributing, in its various forms.
Credit to Ben Adams.
Runtime: Codegen
The following changes improve code generation in RyuJIT, either making the process more efficient or the resulting code faster to run. A recent post on Loop Alignment does a great job of demonstrating the level of consideration and detail that is required for optimal performance.
Optimizations:
- Remove bounds checks with tests against length Credit: Nathan Moore
- Span bounds check elision with top-level range check nodes removal Credit SingleAccretion
- Support loop cloning of byte array accesses
- JIT: Non-void ThrowHelpers
- CSE for floating-point constants
- Invariant static readonly fields to enable CSE/Hoist-from-loops optimizations
- Fold more with null checks for const string
- Zero init elimination: Don’t init tracked temps without GC fields
Dynamic PGO
- JIT: profile updates for return merges and tail calls
- Class profile: Use unknown placeholder for collectible class typehandle
Keep structs in register
Completed .NET 6 EH Write Thru
Tools: Initial .NET Hot Reload support now available for web apps
Early support for .NET Hot Reload is now available for ASP.NET Core & Blazor projects using dotnet watch
. .NET Hot Reload applies code changes to your running app without restarting it and without losing any app state. Code changes that cannot be applied to the running app can still be applied by rebuilding and restarting the app.
This is just the first step in our more comprehensive plan to bring this technology to all .NET developers including desktop (WPF, WinUI, WinForms), cross-platform client scenarios in .NET MAUI, and more. .NET Hot Reload will be supported with these additional platforms in future preview releases of .NET 6. In addition, we will also make .NET Hot Reload available as an integrated experience in future Visual Studio releases.
For more details on trying out hot reload with web app projects, see ASP.NET Core updates in .NET 6 Preview 3
Closing
We’re now about halfway through the .NET 6 release, at least in terms of feature development. You can read about other features that have been added, in the Preview 2 and Preview 1 releases. We expect some larger features to land in the next few previews before we start into focusing on quality in the release candidate builds.
I started the post raising visibility on some great community contributions. Some community folks are using the GitHub Sponsors program to support spending more of their time on contributing to .NET. That’s great. Ben is an example. I encourage you to consider supporting community developers improving .NET, perhaps focused on a part of the product you depend on significantly. It’s a good opportunity to invest in your supply chain.
25 comments
I checked and it helps, finally my site wotking better locação de computadores
Hi Richard,
We’re just about to begin a major green field implementation and trying to determine which .NET version to go with. Our first PROD release will be in November, which makes us a little tentative to commit to .NET 6. Is there recommended reading for this scenario, or perhaps someone from the .NET team we could speak with to get their thoughts/recommendations?
Thanks,
Peter
Will MAUI support such control like WebView2?
“hardware accelerate the type system”? You mean “hardware accelerated type system”??
Hiya Richard, thanks for the announcement. I always used to cringe in C# whenever I would write “if (d.TryGet(key, out value)) { value.field = x; d[key] = value; }”. That’s twice the number of dictionary searches as is needed.
Do you think we’ll get something similar for efficient get-or-add logic? C++ ordered maps ruled for this sort of use case, as the return value from a failed lookup could be passed as an argument to the insertion to skip the tree search.
Also, I think the sample might be an old version of the API, as I don’t see any GetValueRef in preview 3. Maybe try “ref MyStruct value = ref CollectionsMarshal.GetValueRefOrNullRef(dictionary, key);”
I appreciate it. Thank you.
3 years of LONG term support is a joke. The long Term Support should be at least 6 years. My softwares works without support for decades after I finish. Why a huge enterprise like MS cannot?
Standard is 2 years… Angular is even less, 1 year.
3 years is good enough, and it helps focusing on improving and bringing new features instead of supporting legacy-to-be stuff.
3 year a joke so what about:
Node 30 month : https://nodejs.org/en/about/releases/
Angular 12 month : https://angular.io/guide/releases
HAHAHAHAHAHAHAH
Support is WAYYYYY more than “Functioning”/”Running”. Support means also fixing discovered bugs and security issues, even conducting re-architecture if it so needs to be.
Lets talk more after you open source your code to the world and point out clients that are using it…
Lets see how much “support” you will be providing then
Nice work team, great progress!
Hi team,
any plan to support for RDLC reporting services in dotnet 6? It is a essential part for any web, windows or any other plateform applications.
Thanks
SSRS support for .NET 5/6 is a major blocker for my team as well.
Unfortunately, the SQL Server team own the components, and they don’t seem to be able (or lack the resources) to upgrade their own software. This is despite the fact that it’s the 4th highest voted issue on their feedback site across their entire product! :
https://feedback.azure.com/forums/908035-sql-server/suggestions/33241936-develop-a-ssrs-reportviewer-for-asp-net-core
The issue has been “under review” for over 3 years and no human response at all 🙁
This release is dedicated almost entirely to low-level performance features.
My favourite type of release!
Great work team.
Performance improvements even if they are minor does contribute to the performance of the entire application and is always welcome.
Totally agree with both of you, Martin and Praveen. Will share more of this information in upcoming posts.