Convert to IIncrementalGenerator about stronginject HOT 18 CLOSED

@jnm2 here's my first pass at creating an incremental generator which should reduce the amount of time iterating through syntax trees:

https://github.com/YairHalberstadt/stronginject/tree/incremental-generators

How annoying would it be to ask you to build and reference from source, and tell me if it reduces light bulb wait time?

Thanks!

from stronginject.

In the largest project where we use StrongInject, this is causing the light bulb to take 5-20 seconds to appear and 5-20 seconds to preview or apply any suggested action. (Or invoke renames, etc.)

from stronginject.

I'm investigating the work required to make this change.

from stronginject.

If it's mainly the syntax trees that are expensive, then it should be relatively straightforward to use an incremental generator for that step.

from stronginject.

Easy. I checked out that branch (24057fd), added StrongInject.csproj and StrongInject.Generator.csproj to the solution in question, and replaced the StrongInject package reference with a project reference to StrongInject.csproj and added an additional project reference with OutputItemType="Analyzer" to StrongInject.Generator.csproj to the project that needs source generation. The solution builds fine.

I then restarted VS, built again (incremental build succeeded immediately), and invoked the light bulb on random things. It seems like the light bulb takes the same amount of time to open after any source edit, and it seems like the fixes take 2-3 times longer to preview or apply than they did before. Very unscientific perceptual guess.

I'm not sure what my testing methodology should be. Sam said:

the problem here was indirect. extreme allocations from StrongInject are leading to long GC pauses. So far attributed 37GB allocs to it in your trace

So if the light bulb and fixes started applying very fast, I'd potentially still have to spend some time doing real work in the solution in order to fully evaluate the improvement?

I also checked out the commit in this project just prior to introducing StrongInject, and the light bulb and fixes became perceptually instant. That gives me something to perceptually compare to. Checking out the commit that introduced StrongInject brings back the long waits.

from stronginject.

Well that's disappointing...

Can I get an idea for how big your project is, say in number of lines of code and number of files? And do you know if you get similar slowness in CLI build times, or when working in VSCode or rider? I want to make sure that I can reproduce this without having access to Visual Studio.

from stronginject.

2510 files, 280k lines including blanks. I couldn't see a significant effect in CLI build times. (I got anywhere from 1-3 seconds additional CLI build time when I initially added StrongInject, but my recent comparisons fluctuated enough that it might be closer to 0 seconds.)

Rider's light bulb is always instant and the fixes vary from instant to less than a second to apply, for comparable operations.

from stronginject.

Ok, that's a reasonable size for a single project. As a temporary workaround have you considered separating the StrongInject code to a different project?

from stronginject.

By introducing a new entry point assembly that defines the container? (All application-level code where the light bulb needs to work is in classes injected by StrongInject.)

Every project in the solution references StrongInject, also, because Owned<T> is sprinkled throughout. Does that incur a cost, or is the cost delayed until an IContainer<T> implementation is seen?

from stronginject.

I imagine everything will incur the cost because there's no way to reference StrongInject without the generator.

from stronginject.

@sharwell I've created an incremental generator which I assume should avoid scanning all trees on every edit, but according to @jnm2 it's not improving performance. Would you be able to advise if it does what it says in the tin?

The generator is at https://github.com/YairHalberstadt/stronginject/tree/incremental-generators

from stronginject.

I wonder if ExcludeAssets="analyzers" affects generators.

from stronginject.

Busy today but will try to take a look later

from stronginject.

• The largest single allocation (6GB) is GenericDecoratorsResolver.ResolveDecorators, which would be fixed by either calling this method less frequently or changing it to not use yield return.

• There is nearly 7.5GB allocations coming from GenericRegistrationsResolver.TryResolve and methods which it calls, mostly from various use of LINQ

  

from stronginject.

@sharwell @jnm2 Is there some way you can share the full trace with me?
Also what's the easiest way to visualise the trace?

Thanks!

from stronginject.

(I shared dotMemory and dotTrace traces of devenv.exe while invoking the light bulb directly with Yair, and it matches closely with Sam's screenshot above)

from stronginject.

I submitted #182 to fix TryResolve and ResolveDecorators.

from stronginject.

Fixed by #180

from stronginject.