refactor(bench): Allocate IDs for benches (#14757)

Cargo.lock

@@ -772,6 +772,7 @@ dependencies = [
  "fwdansi",
  "http",
  "import_map",
+ "indexmap",
  "jsonc-parser",
  "libc",
  "log",

cli/Cargo.toml

@@ -72,6 +72,7 @@ eszip = "=0.20.0"
 fancy-regex = "=0.9.0"
 http = "=0.2.6"
 import_map = "=0.9.0"
+indexmap = "1.8.1"
 jsonc-parser = { version = "=0.19.0", features = ["serde"] }
 libc = "=0.2.126"
 log = { version = "=0.4.16", features = ["serde"] }

cli/ops/bench.rs

@@ -1,4 +1,6 @@
+use crate::tools::bench::BenchDescription;
 use crate::tools::bench::BenchEvent;
+use crate::tools::test::TestFilter;
 use deno_core::error::generic_error;
 use deno_core::error::AnyError;
 use deno_core::op;
@@ -8,22 +10,32 @@ use deno_core::OpState;
 use deno_runtime::permissions::create_child_permissions;
 use deno_runtime::permissions::ChildPermissionsArg;
 use deno_runtime::permissions::Permissions;
+use serde::Deserialize;
+use serde::Serialize;
+use std::sync::atomic::AtomicUsize;
+use std::sync::atomic::Ordering;
 use std::time;
 use tokio::sync::mpsc::UnboundedSender;
 use uuid::Uuid;
 
-pub fn init(sender: UnboundedSender<BenchEvent>, unstable: bool) -> Extension {
+pub fn init(
+  sender: UnboundedSender<BenchEvent>,
+  filter: TestFilter,
+  unstable: bool,
+) -> Extension {
   Extension::builder()
     .ops(vec![
       op_pledge_test_permissions::decl(),
       op_restore_test_permissions::decl(),
       op_get_bench_origin::decl(),
+      op_register_bench::decl(),
       op_dispatch_bench_event::decl(),
       op_bench_now::decl(),
       op_bench_check_unstable::decl(),
     ])
     .state(move |state| {
       state.put(sender.clone());
+      state.put(filter.clone());
       state.put(Unstable(unstable));
       Ok(())
     })
@@ -97,6 +109,44 @@ fn op_get_bench_origin(state: &mut OpState) -> String {
   state.borrow::<ModuleSpecifier>().to_string()
 }
 
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct BenchInfo {
+  name: String,
+  origin: String,
+  baseline: bool,
+  group: Option<String>,
+}
+
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+struct BenchRegisterResult {
+  id: usize,
+  filtered_out: bool,
+}
+
+static NEXT_ID: AtomicUsize = AtomicUsize::new(0);
+
+#[op]
+fn op_register_bench(
+  state: &mut OpState,
+  info: BenchInfo,
+) -> Result<BenchRegisterResult, AnyError> {
+  let id = NEXT_ID.fetch_add(1, Ordering::SeqCst);
+  let filter = state.borrow::<TestFilter>().clone();
+  let filtered_out = !filter.includes(&info.name);
+  let description = BenchDescription {
+    id,
+    name: info.name,
+    origin: info.origin,
+    baseline: info.baseline,
+    group: info.group,
+  };
+  let sender = state.borrow::<UnboundedSender<BenchEvent>>().clone();
+  sender.send(BenchEvent::Register(description)).ok();
+  Ok(BenchRegisterResult { id, filtered_out })
+}
+
 #[op]
 fn op_dispatch_bench_event(state: &mut OpState, event: BenchEvent) {
   let sender = state.borrow::<UnboundedSender<BenchEvent>>().clone();

cli/tools/bench.rs

@@ -21,18 +21,21 @@ use crate::ops;
 use crate::proc_state::ProcState;
 use crate::resolver::ImportMapResolver;
 use crate::resolver::JsxResolver;
+use crate::tools::test::format_test_error;
+use crate::tools::test::TestFilter;
 
 use deno_core::error::generic_error;
 use deno_core::error::AnyError;
+use deno_core::error::JsError;
 use deno_core::futures::future;
 use deno_core::futures::stream;
 use deno_core::futures::FutureExt;
 use deno_core::futures::StreamExt;
-use deno_core::serde_json::json;
 use deno_core::ModuleSpecifier;
 use deno_graph::ModuleKind;
 use deno_runtime::permissions::Permissions;
 use deno_runtime::tokio_util::run_basic;
+use indexmap::IndexMap;
 use log::Level;
 use serde::Deserialize;
 use serde::Serialize;
@@ -48,12 +51,6 @@ struct BenchSpecifierOptions {
   filter: Option<String>,
 }
 
-#[derive(Debug, Clone, PartialEq, Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub enum BenchOutput {
-  Console(String),
-}
-
 #[derive(Debug, Clone, PartialEq, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct BenchPlan {
@@ -67,50 +64,36 @@ pub struct BenchPlan {
 #[serde(rename_all = "camelCase")]
 pub enum BenchEvent {
   Plan(BenchPlan),
-  Output(BenchOutput),
+  Output(String),
-  Wait(BenchMetadata),
+  Register(BenchDescription),
-  Result(String, BenchResult),
+  Wait(usize),
+  Result(usize, BenchResult),
 }
 
 #[derive(Debug, Clone, PartialEq, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub enum BenchResult {
-  Ok(BenchMeasurement),
+  Ok(BenchStats),
-  Failed(BenchFailure),
+  Failed(Box<JsError>),
 }
 
-#[derive(Debug, Clone, Serialize)]
+#[derive(Debug, Clone)]
 pub struct BenchReport {
   pub total: usize,
   pub failed: usize,
-  pub failures: Vec<BenchFailure>,
+  pub failures: Vec<(BenchDescription, Box<JsError>)>,
-  pub measurements: Vec<BenchMeasurement>,
+  pub measurements: Vec<(BenchDescription, BenchStats)>,
 }
 
 #[derive(Debug, Clone, PartialEq, Deserialize, Eq, Hash)]
-pub struct BenchMetadata {
+pub struct BenchDescription {
+  pub id: usize,
   pub name: String,
   pub origin: String,
   pub baseline: bool,
   pub group: Option<String>,
 }
 
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
-pub struct BenchMeasurement {
-  pub name: String,
-  pub baseline: bool,
-  pub stats: BenchStats,
-  pub group: Option<String>,
-}
-
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
-pub struct BenchFailure {
-  pub name: String,
-  pub error: String,
-  pub baseline: bool,
-  pub group: Option<String>,
-}
-
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct BenchStats {
   pub n: u64,
@@ -142,9 +125,10 @@ pub trait BenchReporter {
   fn report_group_summary(&mut self);
   fn report_plan(&mut self, plan: &BenchPlan);
   fn report_end(&mut self, report: &BenchReport);
-  fn report_wait(&mut self, wait: &BenchMetadata);
+  fn report_register(&mut self, desc: &BenchDescription);
-  fn report_output(&mut self, output: &BenchOutput);
+  fn report_wait(&mut self, desc: &BenchDescription);
-  fn report_result(&mut self, result: &BenchResult);
+  fn report_output(&mut self, output: &str);
+  fn report_result(&mut self, desc: &BenchDescription, result: &BenchResult);
 }
 
 struct ConsoleReporter {
@@ -152,8 +136,8 @@ struct ConsoleReporter {
   show_output: bool,
   has_ungrouped: bool,
   group: Option<String>,
-  baseline: Option<BenchMeasurement>,
+  baseline: bool,
-  group_measurements: Vec<BenchMeasurement>,
+  group_measurements: Vec<(BenchDescription, BenchStats)>,
   options: Option<mitata::reporter::Options>,
 }
 
@@ -163,7 +147,7 @@ impl ConsoleReporter {
       show_output,
       group: None,
       options: None,
-      baseline: None,
+      baseline: false,
       name: String::new(),
       has_ungrouped: false,
       group_measurements: Vec::new(),
@@ -181,7 +165,7 @@ impl BenchReporter for ConsoleReporter {
     self.report_group_summary();
 
     self.group = None;
-    self.baseline = None;
+    self.baseline = false;
     self.name = String::new();
     self.group_measurements.clear();
     self.options = Some(mitata::reporter::Options::new(
@@ -218,10 +202,12 @@ impl BenchReporter for ConsoleReporter {
     );
   }
 
-  fn report_wait(&mut self, wait: &BenchMetadata) {
+  fn report_register(&mut self, _desc: &BenchDescription) {}
-    self.name = wait.name.clone();
 
-    match &wait.group {
+  fn report_wait(&mut self, desc: &BenchDescription) {
+    self.name = desc.name.clone();
+
+    match &desc.group {
       None => {
         self.has_ungrouped = true;
       }
@@ -249,56 +235,52 @@ impl BenchReporter for ConsoleReporter {
     }
   }
 
-  fn report_output(&mut self, output: &BenchOutput) {
+  fn report_output(&mut self, output: &str) {
     if self.show_output {
-      match output {
+      print!("{} {}", colors::gray(format!("{}:", self.name)), output)
-        BenchOutput::Console(line) => {
-          print!("{} {}", colors::gray(format!("{}:", self.name)), line)
-        }
-      }
     }
   }
 
-  fn report_result(&mut self, result: &BenchResult) {
+  fn report_result(&mut self, desc: &BenchDescription, result: &BenchResult) {
     let options = self.options.as_ref().unwrap();
 
     match result {
-      BenchResult::Ok(bench) => {
+      BenchResult::Ok(stats) => {
-        let mut bench = bench.to_owned();
+        let mut desc = desc.clone();
 
-        if bench.baseline && self.baseline.is_none() {
+        if desc.baseline && !self.baseline {
-          self.baseline = Some(bench.clone());
+          self.baseline = true;
         } else {
-          bench.baseline = false;
+          desc.baseline = false;
         }
 
-        self.group_measurements.push(bench.clone());
-
         println!(
           "{}",
           mitata::reporter::benchmark(
-            &bench.name,
+            &desc.name,
             &mitata::reporter::BenchmarkStats {
-              avg: bench.stats.avg,
+              avg: stats.avg,
-              min: bench.stats.min,
+              min: stats.min,
-              max: bench.stats.max,
+              max: stats.max,
-              p75: bench.stats.p75,
+              p75: stats.p75,
-              p99: bench.stats.p99,
+              p99: stats.p99,
-              p995: bench.stats.p995,
+              p995: stats.p995,
             },
             options
           )
         );
+
+        self.group_measurements.push((desc, stats.clone()));
       }
 
-      BenchResult::Failed(failure) => {
+      BenchResult::Failed(js_error) => {
         println!(
           "{}",
           mitata::reporter::benchmark_error(
-            &failure.name,
+            &desc.name,
             &mitata::reporter::Error {
               stack: None,
-              message: failure.error.clone(),
+              message: format_test_error(js_error),
             },
             options
           )
@@ -314,8 +296,7 @@ impl BenchReporter for ConsoleReporter {
     };
 
     if 2 <= self.group_measurements.len()
-      && (self.group.is_some()
+      && (self.group.is_some() || (self.group.is_none() && self.baseline))
-        || (self.group.is_none() && self.baseline.is_some()))
     {
       println!(
         "\n{}",
@@ -323,18 +304,18 @@ impl BenchReporter for ConsoleReporter {
           &self
             .group_measurements
             .iter()
-            .map(|b| mitata::reporter::GroupBenchmark {
+            .map(|(d, s)| mitata::reporter::GroupBenchmark {
-              name: b.name.clone(),
+              name: d.name.clone(),
-              baseline: b.baseline,
+              baseline: d.baseline,
-              group: b.group.as_deref().unwrap_or("").to_owned(),
+              group: d.group.as_deref().unwrap_or("").to_owned(),
 
               stats: mitata::reporter::BenchmarkStats {
-                avg: b.stats.avg,
+                avg: s.avg,
-                min: b.stats.min,
+                min: s.min,
-                max: b.stats.max,
+                max: s.max,
-                p75: b.stats.p75,
+                p75: s.p75,
-                p99: b.stats.p99,
+                p99: s.p99,
-                p995: b.stats.p995,
+                p995: s.p995,
               },
             })
             .collect::<Vec<mitata::reporter::GroupBenchmark>>(),
@@ -343,7 +324,7 @@ impl BenchReporter for ConsoleReporter {
       );
     }
 
-    self.baseline = None;
+    self.baseline = false;
     self.group_measurements.clear();
   }
 
@@ -380,11 +361,12 @@ async fn bench_specifier(
   channel: UnboundedSender<BenchEvent>,
   options: BenchSpecifierOptions,
 ) -> Result<(), AnyError> {
+  let filter = TestFilter::from_flag(&options.filter);
   let mut worker = create_main_worker(
     &ps,
     specifier.clone(),
     permissions,
-    vec![ops::bench::init(channel.clone(), ps.flags.unstable)],
+    vec![ops::bench::init(channel.clone(), filter, ps.flags.unstable)],
     Default::default(),
   );
 
@@ -413,12 +395,7 @@ async fn bench_specifier(
 
   let bench_result = worker.js_runtime.execute_script(
     &located_script_name!(),
-    &format!(
+    r#"Deno[Deno.internal].runBenchmarks()"#,
-      r#"Deno[Deno.internal].runBenchmarks({})"#,
-      json!({
-        "filter": options.filter,
-      }),
-    ),
   )?;
 
   worker.js_runtime.resolve_value(bench_result).await?;
@@ -462,6 +439,7 @@ async fn bench_specifiers(
       let mut used_only = false;
       let mut report = BenchReport::new();
       let mut reporter = create_reporter(log_level != Some(Level::Error));
+      let mut benches = IndexMap::new();
 
       while let Some(event) = receiver.recv().await {
         match event {
@@ -474,27 +452,32 @@ async fn bench_specifiers(
             reporter.report_plan(&plan);
           }
 
-          BenchEvent::Wait(metadata) => {
+          BenchEvent::Register(desc) => {
-            reporter.report_wait(&metadata);
+            reporter.report_register(&desc);
+            benches.insert(desc.id, desc);
+          }
+
+          BenchEvent::Wait(id) => {
+            reporter.report_wait(benches.get(&id).unwrap());
           }
 
           BenchEvent::Output(output) => {
             reporter.report_output(&output);
           }
 
-          BenchEvent::Result(_origin, result) => {
+          BenchEvent::Result(id, result) => {
-            match &result {
+            let desc = benches.get(&id).unwrap();
-              BenchResult::Ok(bench) => {
+            reporter.report_result(desc, &result);
-                report.measurements.push(bench.clone());
+            match result {
+              BenchResult::Ok(stats) => {
+                report.measurements.push((desc.clone(), stats));
               }
 
               BenchResult::Failed(failure) => {
                 report.failed += 1;
-                report.failures.push(failure.clone());
+                report.failures.push((desc.clone(), failure));
               }
             };
-
-            reporter.report_result(&result);
           }
         }
       }

cli/tools/test.rs

@@ -77,6 +77,58 @@ pub enum TestMode {
   Both,
 }
 
+// TODO(nayeemrmn): This is only used for benches right now.
+#[derive(Clone, Debug, Default)]
+pub struct TestFilter {
+  pub substring: Option<String>,
+  pub regex: Option<Regex>,
+  pub include: Option<Vec<String>>,
+  pub exclude: Vec<String>,
+}
+
+impl TestFilter {
+  pub fn includes(&self, name: &String) -> bool {
+    if let Some(substring) = &self.substring {
+      if !name.contains(substring) {
+        return false;
+      }
+    }
+    if let Some(regex) = &self.regex {
+      if !regex.is_match(name) {
+        return false;
+      }
+    }
+    if let Some(include) = &self.include {
+      if !include.contains(name) {
+        return false;
+      }
+    }
+    if self.exclude.contains(name) {
+      return false;
+    }
+    true
+  }
+
+  pub fn from_flag(flag: &Option<String>) -> Self {
+    let mut substring = None;
+    let mut regex = None;
+    if let Some(flag) = flag {
+      if flag.starts_with('/') && flag.ends_with('/') {
+        let rs = flag.trim_start_matches('/').trim_end_matches('/');
+        regex =
+          Some(Regex::new(rs).unwrap_or_else(|_| Regex::new("$^").unwrap()));
+      } else {
+        substring = Some(flag.clone());
+      }
+    }
+    Self {
+      substring,
+      regex,
+      ..Default::default()
+    }
+  }
+}
+
 #[derive(Debug, Clone, PartialEq, Deserialize, Eq, Hash)]
 #[serde(rename_all = "camelCase")]
 pub struct TestLocation {

runtime/js/40_testing.js

@@ -4,11 +4,10 @@
 ((window) => {
   const core = window.Deno.core;
   const { setExitHandler } = window.__bootstrap.os;
-  const { Console, inspectArgs } = window.__bootstrap.console;
+  const { Console } = window.__bootstrap.console;
   const { serializePermissions } = window.__bootstrap.permissions;
   const { assert } = window.__bootstrap.infra;
   const {
-    AggregateErrorPrototype,
     ArrayFrom,
     ArrayPrototypeFilter,
     ArrayPrototypeJoin,
@@ -537,8 +536,23 @@
     };
   }
 
+  /**
+   * @typedef {{
+   *   id: number,
+   *   name: string,
+   *   fn: BenchFunction
+   *   origin: string,
+   *   filteredOut: boolean,
+   *   ignore: boolean,
+   *   only: boolean.
+   *   sanitizeExit: boolean,
+   *   permissions: PermissionOptions,
+   * }} BenchDescription
+   */
+
   const tests = [];
-  const benches = [];
+  /** @type {BenchDescription[]} */
+  const benchDescs = [];
 
   // Main test function provided by Deno.
   function test(
@@ -655,12 +669,11 @@
     maybeFn,
   ) {
     core.opSync("op_bench_check_unstable");
-    let benchDef;
+    let benchDesc;
     const defaults = {
       ignore: false,
+      baseline: false,
       only: false,
-      sanitizeOps: true,
-      sanitizeResources: true,
       sanitizeExit: true,
       permissions: null,
     };
@@ -670,7 +683,7 @@
         throw new TypeError("The bench name can't be empty");
       }
       if (typeof optionsOrFn === "function") {
-        benchDef = { fn: optionsOrFn, name: nameOrFnOrOptions, ...defaults };
+        benchDesc = { fn: optionsOrFn, name: nameOrFnOrOptions, ...defaults };
       } else {
         if (!maybeFn || typeof maybeFn !== "function") {
           throw new TypeError("Missing bench function");
@@ -685,7 +698,7 @@
             "Unexpected 'name' field in options, bench name is already provided as the first argument.",
           );
         }
-        benchDef = {
+        benchDesc = {
           ...defaults,
           ...optionsOrFn,
           fn: maybeFn,
@@ -702,7 +715,7 @@
       if (maybeFn != undefined) {
         throw new TypeError("Unexpected third argument to Deno.bench()");
       }
-      benchDef = {
+      benchDesc = {
         ...defaults,
         fn: nameOrFnOrOptions,
         name: nameOrFnOrOptions.name,
@@ -732,28 +745,18 @@
       if (!name) {
         throw new TypeError("The bench name can't be empty");
       }
-      benchDef = { ...defaults, ...nameOrFnOrOptions, fn, name };
+      benchDesc = { ...defaults, ...nameOrFnOrOptions, fn, name };
     }
 
+    benchDesc.origin = getBenchOrigin();
     const AsyncFunction = (async () => {}).constructor;
-    benchDef.async = AsyncFunction === benchDef.fn.constructor;
+    benchDesc.async = AsyncFunction === benchDesc.fn.constructor;
 
-    ArrayPrototypePush(benches, benchDef);
+    const { id, filteredOut } = core.opSync("op_register_bench", benchDesc);
-  }
+    benchDesc.id = id;
+    benchDesc.filteredOut = filteredOut;
 
-  function formatError(error) {
+    ArrayPrototypePush(benchDescs, benchDesc);
-    if (ObjectPrototypeIsPrototypeOf(AggregateErrorPrototype, error)) {
-      const message = error
-        .errors
-        .map((error) =>
-          inspectArgs([error]).replace(/^(?!\s*$)/gm, " ".repeat(4))
-        )
-        .join("\n");
-
-      return error.name + "\n" + message + error.stack;
-    }
-
-    return inspectArgs([error]);
   }
 
   /**
@@ -848,7 +851,8 @@
     };
   }
 
-  async function benchMeasure(timeBudget, fn, step, sync) {
+  async function benchMeasure(timeBudget, desc) {
+    const fn = desc.fn;
     let n = 0;
     let avg = 0;
     let wavg = 0;
@@ -859,11 +863,10 @@
 
     // warmup step
     let c = 0;
-    step.warmup = true;
     let iterations = 20;
     let budget = 10 * 1e6;
 
-    if (sync) {
+    if (!desc.async) {
       while (budget > 0 || iterations-- > 0) {
         const t1 = benchNow();
 
@@ -890,13 +893,11 @@
     wavg /= c;
 
     // measure step
-    step.warmup = false;
-
     if (wavg > lowPrecisionThresholdInNs) {
       let iterations = 10;
       let budget = timeBudget * 1e6;
 
-      if (sync) {
+      if (!desc.async) {
         while (budget > 0 || iterations-- > 0) {
           const t1 = benchNow();
 
@@ -906,7 +907,7 @@
           n++;
           avg += iterationTime;
           budget -= iterationTime;
-          all.push(iterationTime);
+          ArrayPrototypePush(all, iterationTime);
           if (iterationTime < min) min = iterationTime;
           if (iterationTime > max) max = iterationTime;
         }
@@ -920,7 +921,7 @@
           n++;
           avg += iterationTime;
           budget -= iterationTime;
-          all.push(iterationTime);
+          ArrayPrototypePush(all, iterationTime);
           if (iterationTime < min) min = iterationTime;
           if (iterationTime > max) max = iterationTime;
         }
@@ -929,7 +930,7 @@
       let iterations = 10;
       let budget = timeBudget * 1e6;
 
-      if (sync) {
+      if (!desc.async) {
         while (budget > 0 || iterations-- > 0) {
           const t1 = benchNow();
           for (let c = 0; c < lowPrecisionThresholdInNs; c++) fn();
@@ -937,7 +938,7 @@
 
           n++;
           avg += iterationTime;
-          all.push(iterationTime);
+          ArrayPrototypePush(all, iterationTime);
           if (iterationTime < min) min = iterationTime;
           if (iterationTime > max) max = iterationTime;
           budget -= iterationTime * lowPrecisionThresholdInNs;
@@ -962,21 +963,15 @@
     return benchStats(n, wavg > lowPrecisionThresholdInNs, avg, min, max, all);
   }
 
-  async function runBench(bench) {
+  async function runBench(desc) {
-    const step = new BenchStep({
-      name: bench.name,
-      sanitizeExit: bench.sanitizeExit,
-      warmup: false,
-    });
-
     let token = null;
 
     try {
-      if (bench.permissions) {
+      if (desc.permissions) {
-        token = pledgePermissions(bench.permissions);
+        token = pledgePermissions(desc.permissions);
       }
 
-      if (bench.sanitizeExit) {
+      if (desc.sanitizeExit) {
         setExitHandler((exitCode) => {
           assert(
             false,
@@ -986,24 +981,31 @@
       }
 
       const benchTimeInMs = 500;
-      const fn = bench.fn.bind(null, step);
+      const stats = await benchMeasure(benchTimeInMs, desc);
-      const stats = await benchMeasure(benchTimeInMs, fn, step, !bench.async);
 
-      return { ok: { stats, ...bench } };
+      return { ok: stats };
     } catch (error) {
-      return { failed: { ...bench, error: formatError(error) } };
+      return { failed: core.destructureError(error) };
     } finally {
       if (bench.sanitizeExit) setExitHandler(null);
       if (token !== null) restorePermissions(token);
     }
   }
 
+  let origin = null;
+
   function getTestOrigin() {
-    return core.opSync("op_get_test_origin");
+    if (origin == null) {
+      origin = core.opSync("op_get_test_origin");
+    }
+    return origin;
   }
 
   function getBenchOrigin() {
-    return core.opSync("op_get_bench_origin");
+    if (origin == null) {
+      origin = core.opSync("op_get_bench_origin");
+    }
+    return origin;
   }
 
   function reportTestPlan(plan) {
@@ -1036,30 +1038,6 @@
     });
   }
 
-  function reportBenchPlan(plan) {
-    core.opSync("op_dispatch_bench_event", {
-      plan,
-    });
-  }
-
-  function reportBenchConsoleOutput(console) {
-    core.opSync("op_dispatch_bench_event", {
-      output: { console },
-    });
-  }
-
-  function reportBenchWait(description) {
-    core.opSync("op_dispatch_bench_event", {
-      wait: description,
-    });
-  }
-
-  function reportBenchResult(origin, result) {
-    core.opSync("op_dispatch_bench_event", {
-      result: [origin, result],
-    });
-  }
-
   function benchNow() {
     return core.opSync("op_bench_now");
   }
@@ -1120,50 +1098,52 @@
     }
   }
 
-  async function runBenchmarks({
+  async function runBenchmarks() {
-    filter = null,
-  } = {}) {
     core.setMacrotaskCallback(handleOpSanitizerDelayMacrotask);
 
     const origin = getBenchOrigin();
     const originalConsole = globalThis.console;
 
-    globalThis.console = new Console(reportBenchConsoleOutput);
+    globalThis.console = new Console((s) => {
+      core.opSync("op_dispatch_bench_event", { output: s });
+    });
 
-    const only = ArrayPrototypeFilter(benches, (bench) => bench.only);
+    const only = ArrayPrototypeFilter(benchDescs, (bench) => bench.only);
     const filtered = ArrayPrototypeFilter(
-      only.length > 0 ? only : benches,
+      only.length > 0 ? only : benchDescs,
-      createTestFilter(filter),
+      (desc) => !desc.filteredOut && !desc.ignore,
     );
 
     let groups = new Set();
-    const benchmarks = ArrayPrototypeFilter(filtered, (bench) => !bench.ignore);
-
     // make sure ungrouped benchmarks are placed above grouped
     groups.add(undefined);
 
-    for (const bench of benchmarks) {
+    for (const desc of filtered) {
-      bench.group ||= undefined;
+      desc.group ||= undefined;
-      groups.add(bench.group);
+      groups.add(desc.group);
     }
 
     groups = ArrayFrom(groups);
     ArrayPrototypeSort(
-      benchmarks,
+      filtered,
       (a, b) => groups.indexOf(a.group) - groups.indexOf(b.group),
     );
 
-    reportBenchPlan({
+    core.opSync("op_dispatch_bench_event", {
+      plan: {
         origin,
-      total: benchmarks.length,
+        total: filtered.length,
         usedOnly: only.length > 0,
-      names: ArrayPrototypeMap(benchmarks, (bench) => bench.name),
+        names: ArrayPrototypeMap(filtered, (desc) => desc.name),
+      },
     });
 
-    for (const bench of benchmarks) {
+    for (const desc of filtered) {
-      bench.baseline = !!bench.baseline;
+      desc.baseline = !!desc.baseline;
-      reportBenchWait({ origin, ...bench });
+      core.opSync("op_dispatch_bench_event", { wait: desc.id });
-      reportBenchResult(origin, await runBench(bench));
+      core.opSync("op_dispatch_bench_event", {
+        result: [desc.id, await runBench(desc)],
+      });
     }
 
     globalThis.console = originalConsole;
@@ -1364,27 +1344,6 @@
     }
   }
 
-  /**
-   * @typedef {{
-   *   name: string;
-   *   sanitizeExit: boolean,
-   *   warmup: boolean,
-   * }} BenchStepParams
-   */
-  class BenchStep {
-    /** @type {BenchStepParams} */
-    #params;
-
-    /** @param params {BenchStepParams} */
-    constructor(params) {
-      this.#params = params;
-    }
-
-    get name() {
-      return this.#params.name;
-    }
-  }
-
   /** @param parentStep {TestStep} */
   function createTestContext(parentStep) {
     return {