Support ES6 module syntax in Acorn and Closure

[kleisauke]

Avoids the need to (temporarily) replace import.meta, await import and export default usages with placeholders during the preprocess phase.

Alternatively, removing the dependency on Closure (see #23358) and updating the JSDCE pass in Acorn to support export default could also be a solution.

Marked as draft due to dependence on PR #23261.

[sbc100]

This is pretty big regression sadly.

[kleisauke]

Here's the consolidated and prettified diff between main and this PR for other.test_minimal_runtime_code_size_hello_webgl_wasm:

Details

--- a/a.js
+++ b/a.js
@@ -1,4 +1,4 @@
-var Module = async function (B = {}) {
+async function Module(B = {}) {
   var l,
     p,
     r,
@@ -7,6 +7,88 @@ var Module = async function (B = {}) {
     y,
     D,
     E,
+    h = {
+      xa: 0,
+      Ha: 0,
+      Ka: 0,
+      Qa: 0,
+      hb: 0,
+      eb: 0,
+      Ea: 0,
+      Ca: 0,
+      Sa: 0,
+      Ba: 0,
+      Ga: 0,
+      Ta: 0,
+      gb: 0,
+      cb: 0,
+      Oa(a, b, c) {
+        r.set(r.subarray(b, b + c), a);
+      },
+      Ya() {
+        for (; h.L.length; ) h.V(h.L.length - 1);
+        h.N = [];
+      },
+      X: 0,
+      N: [],
+      za(a, b, c) {
+        function d(k, p) {
+          if (k.length != p.length) return !1;
+          for (var q in k) if (k[q] != p[q]) return !1;
+          return !0;
+        }
+
+        for (var f of h.N) if (f.Y == a && d(f.$, c)) return;
+        h.N.push({Y: a, da: b, $: c});
+        h.N.sort((k, p) => k.da < p.da);
+      },
+      Za(a) {
+        h.N = h.N.filter((b) => b.Y != a);
+      },
+      ja() {
+        return navigator.userActivation ? navigator.userActivation.isActive : h.X && h.la.wa;
+      },
+      ea() {
+        if (h.ja()) {
+          var a = h.N;
+          h.N = [];
+          for (var b of a) b.Y(...b.$);
+        }
+      },
+      L: [],
+      sa: (a, b) => {
+        for (var c = 0; c < h.L.length; ++c) h.L[c].target != a || (b && b != h.L[c].T) || h.V(c--);
+      },
+      V(a) {
+        var b = h.L[a];
+        b.target.removeEventListener(b.T, b.ba, b.va);
+        h.L.splice(a, 1);
+      },
+      Xa(a) {
+        if (!a.target) return -4;
+        if (a.ya)
+          (a.ba = function (c) {
+            ++h.X;
+            h.la = a;
+            h.ea();
+            a.Ja(c);
+            h.ea();
+            --h.X;
+          }),
+            a.target.addEventListener(a.T, a.ba, a.va),
+            h.L.push(a);
+        else
+          for (var b = 0; b < h.L.length; ++b)
+            h.L[b].target == a.target && h.L[b].T == a.T && h.V(b--);
+        return 0;
+      },
+      Ia(a) {
+        return a ? (a == window ? '#window' : a == screen ? '#screen' : a?.nodeName || '') : '';
+      },
+      fullscreenEnabled() {
+        return document.fullscreenEnabled || document.webkitFullscreenEnabled;
+      },
+    },
     M = new TextDecoder(),
     w = (a, b) => {
       if (!a) return '';
@@ -283,4 +365,8 @@ var Module = async function (B = {}) {
     L();
   });
   return {};
-};
+}
+
+'object' === typeof exports && 'object' === typeof module
+  ? ((module.exports = Module), (module.exports.default = Module))
+  : 'function' === typeof define && define.amd && define([], () => Module);

One factor is that the CJS/AMD module exports are now present under MINIMAL_RUNTIME (see PR #23282). The other factor is likely(?) that JSDCE is unable to remove certain parts of libhtml5.js.

[kleisauke]

Commit 507b455 workarounds this, but I'm not sure why this is necessary. 🤷‍♂️

How could -sMINIMAL_RUNTIME=2 have removed that? It doesn't seem like it should.

[kleisauke]

... it looks like that should be moved to libhtml5_webgl.js instead, see commit f53a92d.

[sbc100]

I think we should try to land f53a92d as its own patch maybe?

[kleisauke]

Split to PR #23847.

FWIW, the *.gzsize code sizes also cause slight differences when I run tools/maint/rebaseline_tests.py on my Fedora 41 workstation, likely because zlib was replaced with zlib-ng.

[kleisauke]

PR #23850 resolves this size regression instead.

It seems that other.test_minimal_runtime_code_size_hello_webgl_wasm depends on:

• emscripten_set_element_css_size()
• emscripten_set_canvas_element_size()
• emscripten_webgl_create_context()

All of these had an unused dependency on $JSEvents, which, for some reason, JSDCE fails to remove in this changeset.

[sbc100]

Are you sure its JSDCE and not closure compile that is failing to remove these things?

That was my experience too in #23261.. I figured the optimizers were just less good at removing non-top-level code?

[kleisauke]

Closure might also be failing to remove this unused dependency - I haven't checked.

We could try replacing Closure with https://github.com/swc-project/swc, as it shows promising results in https://github.com/privatenumber/minification-benchmarks. However, I'm unsure how much effort that would require.

[sbc100]

I would love to replace closure with pretty much anything else :)

However, so far nothing as come along that can't do as good of a job at minifying. Maybe its worth looking into swc!