.NET Native Performance
blogs.msdn.com
blogs.msdn.com
Although saying that, it's quite amazing how much of the core libraries are written in C#, you what have thought it would be wrappers over C++ libraries.
.NET Reflector came in handy once or twice to copy out a few library functions to avoid adding another reference dll.
How will .Net Native deal with this sort of stuff if the framework parts are directly baked into the executable?
The core framework libraries (CoreFX) - https://github.com/dotnet/corefx - are used for all .NET Core scenarios, including .NET Native (UWP). This means that your code does the same thing in all of these different environments, since it's using the same underlying framework libraries. Separately, the Mono project is taking a lot of the same code, which means that the base framework for Xamarin apps are becoming more compatible with CoreFX, too. Yeahh! We hope to make this more formal in the future. We talk to @migueldeicaza about this frequently.
Today, .NET Native and .NET Core use two different runtimes, MRT and CoreCLR, respectively. MRT expands to the extremely creative "managed runtime". Colloquially, we call it "Mr. T". Earlier in the project, everyone working on .NET Native had posters of Mr. T (yes, that one) on our doors. Mohawks were entirely optional on the parts of team members.
MRT was built for static ahead-of-time (AOT) compilation. It is the child of the Redhawk project (http://www.zdnet.com/article/microsoft-codename-redhawk-live...). Redhawk was built to be a .NET-esque (using C# + extensions) systems programming environment. MRT was built to have many of the same benefits, but be compatible with .NET (not just "esque") and support nothing less, nothing more than C#.
CoreCLR is a child of the .NET Framework CLR. CLR (and by extention, CoreCLR) was built for dynamic execution, with a JIT. It supports AOT compilation in the pre-jit/NGEN sense.
The GC is the primary shared component between the two (MRT, CLR/CoreCLR) runtimes.
We plan to bring static compilation to more scenarios, however, we strongly believe that both JIT and (real) AOT are legitimate and compelling technologies and experiences and both are on our long-term roadmap. We'd like to leave the compilation choice/appraoch up to our users/customers. Novel idea, eh?
We have a lot of fun runtime and compiler tech on the team. It's a fun place to work. Much of it is now open source, meaning that we get to work on this fun tech in the open. This is our latest compiler announcement: http://blogs.msdn.com/b/dotnet/archive/2015/07/20/announcing....
Why isn't this just rolled into ngen? Or run as a second stage JIT, caching the results to disk? I cannot think of any scenarios where anyone wants to re-JIT every time they start a program. At least not on desktop and server scenarios (maybe some embedded system with no storage space).
Or, I guess the real question is: apart from reflection which other features don't work the same on .NET Native?
Reflection is supported on .NET Native Native. See: http://blogs.msdn.com/b/dotnet/archive/2014/05/20/net-native.... Reflection.Emit is not.
There are effectively three codegen strategies: JIT, pre-jit and static compilation/AOT. They all have their own characteristics and reasons why they are a good choice. I can see that you have depth here, but I'm going to provide an answer with more context, for a broader audience.
- JIT has the most flexibility since the code is generated (you guessed it) just in time. This means that the code generated can be specialized in terms of hardware or use a layered approach to codegen (prefer throughput and then CQ for hotter methods). It is also the best for versioning, since the code is always generated in terms of the actual dependencies, as opposed to the ones you happened to compile with.
- Pre-JIT can be very high performance since you don't have to pay for JIT and other related costs up-front, as is the case with NGEN. Depending on the code generated, it may be subject to the fragile base class problem: https://en.wikipedia.org/wiki/Fragile_base_class, as is the case with NGEN. MDIL is also subject to it, but we made the cost to regenerate images very low. There can be a problem of when to generate the images and where to store them. These two issues have been a constant design challenge for us (mostly the former). At least on the surface, pre-jit is much cheaper than AOT, particularly since you can rely on a JIT for the code that you can generate deterministically (at least into a single bit of code) ahead of time (think generics). ReadyToRun (https://github.com/dotnet/coreclr/issues/227) is a project we're working that makes NGEN/CrossGen images more resilient to the fragile base class problem.
- AOT can be very high performance, for the same reasons. We've found that it is even higher performance since you can skip a number of architecural requirements for a dynamic runtime environment. If you couple this with app-local framework and runtime distribution, as is the case with .NET Native, then you also skip the fragile base class problem. Building a high-performance industrial quality AOT system with a fabulous developer experience is a fantastic computer science challenge. We've certainly found it to be that. We're several years into AOT investments and we're finding that it is paying off.
"...we only support C# and VB in a technical sense. You've picked up on one reason why[1]; we do IL scanning and rewriting and look for various patterns."
They do use the normal, unmodified csc compiler to generate the IL[2]. But it sounds like they are partially decompiling the IL to get C#-level semantics? There's also this comment[3] regarding expression trees:
"the JIT runtime does emit those as dynamic code, but .NET Native can interpret those"
So I'm guessing they have to look through your source and figure out when you exec one and then do it head of time? I'm not sure how that can be a general solution, as you can construct arbitrary expression trees at runtime, even using external input. Maybe that part doesn't work, but normal programs don't do that?
I do wish they were a bit more open and clear on all this. (And wow, the public messaging is really confusing other developers. I've had a customer come super excited because .NET Native would obviously cut his server requirements by a nice chunk... had to explain it wouldn't work at all in his scenario.)
1: I wrote asking "Are dependencies being taken on specific quirks of the C# compiler? Like, I dunno, how certain members are named, or the way C# creates Expression Trees"
2: http://blogs.msdn.com/b/dotnet/archive/2014/05/09/the-net-na...
3: http://blogs.msdn.com/b/dotnet/archive/2014/05/20/net-native...
The open source .NET Core provides a subset of the libraries in the .NET Framework. Namely, .NET Core does not include WinForms, etc. The open source effort of .NET Core is divided into CoreFX (C# libraries) and CoreCLR "Runtime" (JIT, GC, C++ libraries, for the most part).
.NET Native is built on .NET Core, but does not depend on the runtime (unclear to me what that means vis-a-vis GC, and C++ libraries, which I assume are still there... but certainly no JIT).
If you think MS ever would "pull out" of mobile and be satisfied with "being a guest in someone else house", your insane :)
Fewer models, to be exact. In contrast to their old strategy of saturating the low end with multitudes of confusingly numbered Lumias they're consolidating into three distinct lines. Presumably they want to have volume!
Native and allowing reflection? I am now much less interested in M#.