name: Benchmark

# Runs bench/run.sh and publishes the result table to the job summary.
#
# GitHub-hosted runners are noisier and weaker than a dedicated machine, so
# treat these numbers as regression signal against previous runs of this
# workflow, not as the headline figures for the README — see bench/README.md
# for the reference run those figures are backed by.
#
# Also runs the full test262 conformance suite (~53k tests, language/ +
# built-ins/ + intl402/) and reports 3va's real pass rate — this is why the
# job can take noticeably longer than a pure perf benchmark would; see
# docs/09-testing/06-test262.md for the methodology and known gaps.

on:
  workflow_dispatch:
  push:
    tags:
      - "v[0-9]+.[0-9]+.[0-9]*"
  schedule:
    - cron: "0 6 * * 1" # weekly, Monday 06:00 UTC

env:
  CARGO_TERM_COLOR: always

# Least privilege: jobs get a read-only token unless they ask for more.
permissions:
  contents: read

jobs:
  benchmark:
    name: Run benchmarks
    runs-on: ubuntu-latest
    permissions:
      contents: read # the README sync commit is pushed with the deploy key
    steps:
      - uses: actions/checkout@v4
        with:
          # main is protected by a ruleset; this deploy key is the only
          # bypass, used for the README benchmark sync commit below.
          ssh-key: ${{ secrets.MAIN_PUSH_DEPLOY_KEY }}

      - uses: dtolnay/rust-toolchain@stable
      # Deliberately no Swatinem/rust-cache here (unlike other workflows in
      # this repo that don't need one either — release.yml builds `v8` from
      # scratch every time and never hits this). rusty_v8's build script
      # produces a ~100+ MB static lib that rust-cache's size-based pruning
      # strips before saving, but the "build script already ran" fingerprint
      # gets cached anyway — cargo then skips re-running it and the linker
      # can't find the .a file that was never actually restored. Three
      # consecutive runs of this workflow failed on exactly that ("could not
      # find native static library `rusty_v8`") before this got removed.

      - uses: actions/setup-node@v4
        with:
          node-version: "24"

      - uses: oven-sh/setup-bun@v2
        with:
          bun-version: latest

      - name: Install benchmark tools
        # hyperfine 2.0 changed its --export-json schema (no top-level `mean`), which bench/run.sh parses.
        run: cargo install hyperfine@1.20.0 oha --locked

      - name: Build 3va (release)
        run: cargo build --release -p vvva_cli

      - name: Run benchmarks
        run: |
          set -o pipefail
          {
            echo "# Benchmark results"
            echo
            echo "Runner: $(nproc) vCPU, $(free -h | awk '/Mem:/ {print $2}') RAM, $(uname -r)"
            echo "Commit: \`${{ github.sha }}\`"
            echo
            bash bench/run.sh
          } | tee "$GITHUB_STEP_SUMMARY" /tmp/bench-output.txt

      - name: Sync README benchmark cells
        # Tag-triggered runs check out a detached tag ref, not a branch —
        # nothing sane to push to. Only commit back on main (schedule /
        # manual dispatch), same as any other push-to-main automation.
        if: github.ref == 'refs/heads/main'
        run: |
          bash bench/sync-readme.sh /tmp/bench-output.txt
          if ! git diff --quiet -- README.md; then
            git config user.name "github-actions[bot]"
            git config user.email "github-actions[bot]@users.noreply.github.com"
            git add README.md
            git commit -s -m "chore(bench): update benchmark figures [skip ci]"
            git push
          fi