Is it correct that compiled lambda expressions on UWP aren't really compiled and are instead interpreted? If so does that mean that on UWP and other AOT platforms like Xamarin/Unity with no support for IL emit it is faster to use MemberInfo and ConstructorInfo?
Related: https://github.com/JamesNK/Newtonsoft.Json/issues/968
If normal reflection is faster on AOT platforms I would like to know is if there is a way at runtime to discover that IL emit isn't supported?
A way to discover this capability at runtime will allow libraries to use compiled lambda expressions to boost on platforms it is available, and to fallback to traditional reflection on platforms where it isn't, while still using a single netstandard assembly.
cc @VSadov
That is correct for UWP the expressions are interpreted and not compiled at runtime.
@JamesNK I know you'll hate this answer, but cross-compiling can ensure that each platform gets the most optimized version for it. With project.json and whatever's coming with the 1.1 tooling transitive dependencies will resolve correctly and ensure that the best version of your library wind up in the runtime dir.
This needs to be a runtime light up API that is netstandard. The project N toolchain and the JIT can treat it as an intrinsic and eliminate it at runtime.
+1
This is a HUGE performance issue for us!
I know you'll hate this answer, but cross-compiling can ensure that each platform gets the most optimized version for it.
That would require knowing every platform that doesn't support compiling expressions and having a different version in the package for each of those platforms, UWP, iOS, Android, Unity, etc
And as new platforms come out they would get the slow netstandard binary until someone explicitly adds a different version in the package.
Or we could have a simple property that library authors can use to adapt the netstandard assembly to the platform at runtime. One assembly that works for now and forever.
Agree - hardcoding the list of platforms is fragile. It does not work well for new platforms. Also, the answer can differ for given platform over time; or both interpreter-only and compiled System.Linq.Expressions can be available for some platforms.
Having an API that returns whether expressions are compiled sounds pretty reasonable.
Since now any expression can be asked to be compiled for interpretation, the API would need to take a compiled expression.
We generally cannot tell if a delegate is a compiled expression (could be not related to expressions at all), but we can recognize interpreted expression delegates.
Would something like the following work for you?
/// returns true if d is a delegate that represents an interpreted expression
bool IsInterpretedExpression(Delegate d)
I read this as more an environment query; code being able to know if calling Compile(false) is going to result in interpretation or not.
Yes, I see it as an environment query.
IsInterpretedExpression might be useful for some people but I'm interested in discovering whether an environment supports compiled expressions before I choose to use expressions (or not, depending on the environment query result).
I read this as more an environment query; code being able to know if calling Compile(false) is going to result in interpretation or not.
Yes. It needs to be before not after. Framework code would likely switch logic altogether and use reflection invoke vs expression tree compilation.
Man, this really burned me - I simply ported some really well working code from an "old" style codebase to UWP and this absolutely hammered me... slowed down by roughly 20x.
+1 to fixing this... please.
Another person being hurt by this - https://github.com/JamesNK/Newtonsoft.Json/issues/968#issuecomment-270204118
Should such a property live in S.L.Expressions or perhaps elsewhere?
Am I right in thinking that a complementary CanInterpret would also be true now, and hence have no value?
Should such a property live in S.L.Expressions
I think so.
It would be simple enough to add:
C#
namespace System.Linq.Expressions
{
public abstract class LambdaExpression : Expression
{
public static CanCompileToIL { get; }
}
}
(Assuming I'm correct above that there couldn't be a case where CanInterpret would also be wanted).
But could a similar question, which boils down to "is Reflection.Emit available?" going to have similar implications elsewhere so that other assemblies would want such a property, in which case maybe it should live in S.Reflection.
is
Reflection.Emitavailable?
This is not the same question.
Consider some of the following situations:
Reflection.Emit is available, but it is implemented using IL interpreterReflection.Emit is available, but the version of System.Linq.Expression used is the interpreter only versionReflection.Emit is not available, but System.Linq.Expression uses some other fast underlying execution systemCool. That's enough variance to argue well that the property should live in S.L.E. Shall I open the above as a separate API proposal?
Yes, this needs to live in S.L.E.
I still believe there is a value in
/// returns true if d is a delegate that represents an interpreted expression
bool IsInterpretedExpression(Delegate d)
We already need to detect interpreted lambdas inside the interpreter itself. Would not be surprised if someone else needs this.
Interestingly, CanCompileToIL could actually be implemented in terms of IsInterpretedExpression - compile and cache trivial delegate and then just ask if it is interpreted or not.
Of course there are more efficient solutions :-)
We already need to detect interpreted lambdas inside the interpreter itself.
We have IsInterpretedFrame but don't seem to use it. It's one of those bits of dead code that I can be wary of deleting because it certainly looks useful! Am I missing a use or another mechanism for same, or is this just dead? In any case it could be the basis of that.
I'm inclined to think of LambdaExpression as the most natural declaring class, as the place where expressions meet delegates. What think you?
Are we going to get this API anytime soon?
@weshaggard , This is along the same lines as the "capabilities" api we've talked about in the past. Can you share your thoughts?
@weshaggard , This is along the same lines as the "capabilities" api we've talked about in the past. Can you share your thoughts?
Yes this is a specific instance of a capability API which I think is best solved with the suggestions provided by @VSadov and @JonHanna in this issue by exposing an API directly in System.Linq.Expressions.
Are we going to get this API anytime soon?
It needs a formal API proposal and the owner @VSadov to drive it through the process. Given this will be a new API it will only initially be added to netcoreapp and uap, as it cannot be added to netstandard yet, so its usage will be limited for a while. For that reason to solve this in the short term the best answer is to cross-compile as @onovotny suggested.
You can't cross compile for CoreRT.
You can't cross compile for CoreRT.
Why not? Lets not derail this issue but it is my understanding corert would be represented with its own RID. See https://github.com/dotnet/corefx/pull/14142.
We don't want upstack libraries that wouldn't have to cross compile to have to cross compile just for this do we?
We don't want upstack libraries that wouldn't have to cross compile to have to cross compile just for this do we?
Ideally no, but if they need to optimize for a the corert platform specifically then they would have too. For this case assuming they target netcoreapp and we expose this new API there they wouldn't need to cross-compile for corert, they could just have a netcoreapp asset and do the runtime detection based on this new API.
We don't want upstack libraries that wouldn't have to cross compile to have to cross compile just for this do we?
The standard way to avoid cross compilation in these cases is to do light-up using reflection. Look for the (capability) API and use it if it exists. Otherwise, use a fallback logic that is not as good.
The standard way to avoid cross compilation in these cases is to do light-up using reflection. Look for the (capability) API and use it if it exists. Otherwise, use a fallback logic that is not as good.
Perfect! What do we look for? This new API that's going to be added?
@weshaggard do we have an issue tracking an API to detect whether the capability to emit IL is available or not?
@weshaggard do we have an issue tracking an API to detect whether the capability to emit IL is available or not?
No and there aren't currently any plans to do that as @jkotas called out there isn't a great way to answer that question. The API to look for would be whatever API we add for this issue.
@VSadov can we take this to the next level?
Hi,
I've looked at the linked issues above and as far as I can tell we're still missing an appropriate API for Json.Net to check? Or do we have it now, and if so, has Json.Net been updated?
We are still being burned heavily by this but I fear any fix is now too late for us. Our UWP code is running on (enterprise) W10M devices and so any new API likely to require an OS release we will never get (for example we cannot get FCU because it was never released for mobile). Please correct me if I'm wrong and this might be shippable in a new framework version that works on older OS'?
I suspect our only option now is to make a private fork of Json.Net and modify it to force an alternate strategy for the reflection/expression based code. We'll have to decide internally whether that pain would be worth it or not, given we are looking at moving to Xamarin.Android since WM10 is basically dead. Unfortunately we have a large number of deployed devices we need to continue supporting for several years so it's not simple either way. I understand limited resources and other priorities, but for us personally it is a shame this didn't get resolved quicker.
Despite the probability a fix would do us no good, I think it would still be a good idea to resolve this in both .Net/UWP and Json.Net for other platforms.
Thanks to everyone who has participated in trying to get this sorted.
Most helpful comment
Yes. It needs to be before not after. Framework code would likely switch logic altogether and use reflection invoke vs expression tree compilation.