NVP's Quest Issue for Investigating Performance

[NullVoxPopuli]

Original Issue: #1590

As a part of this issue, it was reported that, in consuming ember.js apps, the vendor.js (where the VM lives), grew by ~ 50KB.
See: emberjs/ember.js#20628

Suspicions to investigate

• we need embroider and auto-import to use the real @glimmer/* dependencies and not the ember-source/dist/dependencies stuff
  • embroider already supports handling export conditions
  • it would likely be too accidentally disruptive to switch to using real glimmer packages in a patch release
  • ember-source's deps should instead use the correct glimmer (and we wshould confirm this)
• debug code is ending up in production (even in 0.84.3)
  • if any debug things that weren't present in 0.84.3 should be changed to LOCAL_DEBUG
  • are there any DEBUG things that are not for ember-debug?
• dead code elimination of the GlimmerVM
  • all "checks" code should not make it to production builds
    • All the @glimmer/debug stack-checks are in the prod build
      This is most of what was added post-0.84.3 -- making debugging experience better / making working in the VM easier about what's going on.
      • WIP: Fix stripping local debug code so can that dead-code-elimination produces smaller bundles with fewer function calls (part 1) #1606
• are there possibly duplicate modules between the packages?
• destruction, while async, is autotracked -- it doesn't need to be autotracked -- can we remove that?
• iteration changes in 64eb186

Things that don't work

• explicit / strict terser settings from the end-consumer
  • https://github.com/NullVoxPopuli/ember-performance/pull/17/files

No change in perf

• Removing the dynamic checker
  • https://github.com/glimmerjs/glimmer-vm/pull/1603/files
  • This is surprising to me, but does demonstrate that the JS engines are doing good work

Things that have a small improvement

• about 6% improvement
  • Explore performance 3 #1598
  • However, this begs the question if our CI's performance benchmark is using development or production
    • 2024-09-06: the benchmark in CI does use the prod build
• about a 1% improvement
  • Explore Performance 2 #1597
  • This implies that single-use function inlining is not happening in consuming projects

Needs further exploration

• Use terser in the development -- shrinks the build in dev by 3kb

  • Use terser in dev #1602

• More aggressive build / rollup changes
  Dev shrinks by ~ 17KB, Prod shrinks by ~ 7KB

  • Rollup/Build tweaks #1601

• Using the production assets as the development assets
  Explore performance 4 - don't use development assets #1599

Notes

Which packages are in ember-source?

ember-source @ v5.4	ember-source @ 6 alpha
@glimmer/destroyable @glimmer/encoder @glimmer/env @glimmer/global-context @glimmer/low-level @glimmer/manager @glimmer/node @glimmer/opcode-compiler @glimmer/owner @glimmer/program @glimmer/reference @glimmer/runtime @glimmer/tracking/index @glimmer/tracking/primitives/cache @glimmer/util @glimmer/validator @glimmer/vm @glimmer/wire-format	@glimmer/debug @glimmer/destroyable @glimmer/encoder @glimmer/env @glimmer/global-context @glimmer/manager @glimmer/node @glimmer/opcode-compiler @glimmer/owner @glimmer/program @glimmer/reference @glimmer/runtime @glimmer/tracking/index @glimmer/tracking/primitives/cache @glimmer/util @glimmer/validator @glimmer/vm @glimmer/wire-format