Could performance of wiring be improved?

[retronym]

I'm adding compilation time tracing to scalac to show how long each file/class/method takes to typecheck. The trace drills down further to measure time spent in implicit searches and in macro expansions.

I generated traces for a relatively large Scala build: https://github.com/guardian/frontend. This project uses MacWire for DI within a Play application.

The generated report suggests that 4% of total compilation time is spent in the wire macro:

I'm having trouble getting a good look at this through Flight Recorder to make concrete suggestions about what could be optimized, but I thought I'd lodge this ticket as-is in the meantime.

[adamw]

Interesting, thanks for the info! I don't think anybody ever looked at optimizing the code - if you would have any data on what specifically takes so long, that could be helpful.

[retronym]

Okay, I've gotten past my profiing problems.

• I was profiing compilation in SBT, in which it was the bottleneck in reading JSON caches of dependency resolution, details in Avoid temporary string, array in JSON reading sbt/util#186
• Macro code is dynamically classloaded by the compiler, which can hide it from Flight Recorder / Async Profiler if class unloading happens too quickly, I think, and which causes different Lambda$NNNN classnames which make profiles harder to read.

Here's what I came out with:

$ cat /tmp/profile.csv | grep wire_impl | perl -pne 's/.*wire_impl/wire_impl/g' | perl -pne 's/\$\$Lambda\$\d+\/\d+/LambdaX.X/' | /code/FlameGraph/flamegraph.pl --minwidth 2 --width 2000 - > /tmp/flame.svg

$ cat /tmp/profile.csv | grep wire_impl | perl -pne 's/.*wire_impl/wire_impl/g' | perl -pne 's/\$\$Lambda\$\d+\/\d+/LambdaX.X/g' | /code/FlameGraph/flamegraph.pl --minwidth 2 --width 2000 - > /tmp/flame.svg

https://drive.google.com/drive/folders/1T5cWwKljhDuoyqmvovdb3CllCz1Dfkrm?usp=sharing

That still shows that most of the time is spent in one-off costs of classloading and linking the macro implementation. Maybe because the implementation is quite lambda heavy, macwire pays a bigger cost than other macros?

We're slowly getting towards a situation where macros classloaders could be reused in later compilation runs.

[retronym]

https://github.com/adamw/macwire/blob/33112fe4e18bee4d41bb763e0aa4b9e03803b5e2/macros/src/main/scala/com/softwaremill/macwire/internals/ConstructorCrimper.scala#L21

This could be changed to use .declarations, constructors are not inherited and calling .members needs to do more work.

[retronym]

If i'm reading things correctly, I think the problem is that for each call to the wire macro, EligibleValuesFinder will typecheck duplicates of the DefTree-s in the body of the enclosing template. That sounds like O(N^2) complexity to me.

In guardian/frontend, a typical wiring looks like:

import akka.actor.ActorSystem
import app.{FrontendApplicationLoader, FrontendComponents}
import assets.DiscussionExternalAssetsLifecycle
import com.softwaremill.macwire._
import common.dfp.DfpAgentLifecycle
import common.{ApplicationMetrics, CloudWatchMetricsLifecycle, ContentApiMetrics, EmailSubsciptionMetrics}
import _root_.commercial.targeting.TargetingLifecycle
import common.Logback.{LogbackOperationsPool, LogstashLifecycle}
import common.commercial.OrielCacheLifecycle
import conf.CachedHealthCheckLifeCycle
import conf.switches.SwitchboardLifecycle
import contentapi.{CapiHttpClient, ContentApiClient, HttpClient, SectionsLookUp, SectionsLookUpLifecycle}
import controllers._
import dev.{DevAssetsController, DevParametersHttpRequestHandler}
import http.{CommonFilters, CorsHttpErrorHandler}
import jobs.{SiteMapJob, SiteMapLifecycle}
import model.{ApplicationContext, ApplicationIdentity}
import services.ophan.SurgingContentAgentLifecycle
import play.api.ApplicationLoader.Context
import play.api.BuiltInComponentsFromContext
import play.api.http.{HttpErrorHandler, HttpRequestHandler}
import play.api.libs.ws.WSClient
import play.api.mvc.EssentialFilter
import play.api.routing.Router
import services._
import router.Routes

import scala.concurrent.ExecutionContext

class AppLoader extends FrontendApplicationLoader {
  override def buildComponents(context: Context): FrontendComponents = new BuiltInComponentsFromContext(context) with AppComponents
}

trait ApplicationsServices {
  def wsClient: WSClient
  implicit val executionContext: ExecutionContext
  lazy val capiHttpClient: HttpClient = wire[CapiHttpClient]
  lazy val contentApiClient = wire[ContentApiClient]
  lazy val siteMapJob = wire[SiteMapJob]
  lazy val sectionsLookUp = wire[SectionsLookUp]
  lazy val ophanApi = wire[OphanApi]
  lazy val facebookGraphApiClient = wire[FacebookGraphApiClient]
  lazy val facebookGraphApi = wire[FacebookGraphApi]
}

trait AppComponents extends FrontendComponents with ApplicationsControllers with ApplicationsServices {

  lazy val devAssetsController = wire[DevAssetsController]
  lazy val healthCheck = wire[HealthCheck]
  lazy val emailSignupController = wire[EmailSignupController]
  lazy val surveyPageController = wire[SurveyPageController]
  lazy val signupPageController = wire[SignupPageController]
  lazy val logbackOperationsPool = wire[LogbackOperationsPool]

  override lazy val lifecycleComponents = List(
    wire[LogstashLifecycle],
    wire[ConfigAgentLifecycle],
    wire[CloudWatchMetricsLifecycle],
    wire[DfpAgentLifecycle],
    wire[SurgingContentAgentLifecycle],
    wire[IndexListingsLifecycle],
    wire[SectionsLookUpLifecycle],
    wire[SwitchboardLifecycle],
    wire[SiteMapLifecycle],
    wire[CachedHealthCheckLifeCycle],
    wire[TargetingLifecycle],
    wire[DiscussionExternalAssetsLifecycle],
    wire[SkimLinksCacheLifeCycle],
    wire[OrielCacheLifecycle]
  )

  lazy val router: Router = wire[Routes]

  lazy val appIdentity = ApplicationIdentity("applications")

  override lazy val appMetrics = ApplicationMetrics(
    ContentApiMetrics.HttpTimeoutCountMetric,
    ContentApiMetrics.HttpLatencyTimingMetric,
    ContentApiMetrics.ContentApiErrorMetric,
    ContentApiMetrics.ContentApiRequestsMetric,
    EmailSubsciptionMetrics.EmailSubmission,
    EmailSubsciptionMetrics.EmailFormError,
    EmailSubsciptionMetrics.NotAccepted,
    EmailSubsciptionMetrics.APIHTTPError,
    EmailSubsciptionMetrics.APINetworkError,
    EmailSubsciptionMetrics.ListIDError,
    EmailSubsciptionMetrics.AllEmailSubmission
  )


  override lazy val httpErrorHandler: HttpErrorHandler = wire[CorsHttpErrorHandler]
  override lazy val httpFilters: Seq[EssentialFilter] = wire[CommonFilters].filters
  override lazy val httpRequestHandler: HttpRequestHandler = wire[DevParametersHttpRequestHandler]

  def actorSystem: ActorSystem
}

Maybe explicit type ascriptions on the members avoid that typechecking?

[adamw]

Hm if the typechecks aren't cached (by the compiler) that might indeed be n^2. There's always the option of introducing a magnolia-style global cache which would remember previous type check results?