Performance of WHATWG ReadableStream.read()

[debadree25]

The performance of ReadableStream.read() seems to be lacking behind other runtimes and probably can be improved

Ref: anonrig/node-benchmarks#3 (comment)

[debadree25]

This could also probably impact fetch perf too!

[debadree25]

Removing primordials like ArrayPrototypeShift or ArrayPrototypePush from the hot path led to a slight perf boost but not sure thats a good thing to do

[KhafraDev]

Getting rid of ensureIsPromise improved perf by ~7.5% locally, but broke at least 2 tests from the WPTs. Maybe someone else will feel inspired:

diff --git a/lib/internal/webstreams/readablestream.js b/lib/internal/webstreams/readablestream.js
index 0b8b8ac1ef..9797a04c2c 100644
--- a/lib/internal/webstreams/readablestream.js
+++ b/lib/internal/webstreams/readablestream.js
@@ -46,6 +46,7 @@ const {
 const {
   isArrayBufferView,
   isDataView,
+  isPromise,
 } = require('internal/util/types');

 const {
@@ -2249,16 +2250,29 @@ function readableStreamDefaultControllerCallPullIfNeeded(controller) {
   }
   assert(!controller[kState].pullAgain);
   controller[kState].pulling = true;
-  PromisePrototypeThen(
-    ensureIsPromise(controller[kState].pullAlgorithm, controller),
-    () => {
-      controller[kState].pulling = false;
-      if (controller[kState].pullAgain) {
-        controller[kState].pullAgain = false;
-        readableStreamDefaultControllerCallPullIfNeeded(controller);
-      }
-    },
-    (error) => readableStreamDefaultControllerError(controller, error));
+
+  function onPullResolve() {
+    controller[kState].pulling = false;
+    if (controller[kState].pullAgain) {
+      controller[kState].pullAgain = false;
+      readableStreamDefaultControllerCallPullIfNeeded(controller);
+    }
+  }
+
+  try {
+    const value = FunctionPrototypeCall(controller[kState].pullAlgorithm, controller);
+
+    if (isPromise(value)) {
+      PromisePrototypeThen(
+        value,
+        onPullResolve,
+        (error) => readableStreamDefaultControllerError(controller, error));
+    } else {
+      onPullResolve();
+    }
+  } catch (error) {
+    return readableStreamDefaultControllerError(controller, error);
+  }
 }

 function readableStreamDefaultControllerClearAlgorithms(controller) {

benchmark:

                                               confidence improvement accuracy (*)   (**)  (***)
webstreams/readable-async-iterator.js n=100000        ***      7.47 %       ±1.31% ±1.74% ±2.27%

[anonrig]

Does Deno and Bun implement webstreams in their native languages or in javascript?

[anonrig]

Removing primordials like ArrayPrototypeShift or ArrayPrototypePush from the hot path led to a slight perf boost but not sure thats a good thing to do

How is the performance difference? @debadree25

[debadree25]

I remembered seeing around 6-7% but have to check again

[debadree25]

Does Deno and Bun implement webstreams in their native languages or in javascript?

Will investigate

[KhafraDev]

Deno implements it in javascript, https://github.com/denoland/deno/blob/fc6ba92024d76d44349c36dcedd13994116db45b/ext/web/06_streams.js#L5066; they do use native functions for detaching arraybuffers/checking if an arraybuffer is detached.

Bun I assume is using webkit's streams implementation, which would be native.

---

Deno isn't using any array methods in its implementation, which probably means it could be removed on node's side. Then there'd be no overhead of using ArrayPrototype primordials. Would need to look into that though 🤔.

So they do eventually push it to an array, but only when the queue is empty (see: https://github.com/denoland/deno/blob/fc6ba92024d76d44349c36dcedd13994116db45b/ext/web/06_streams.js#L5648)