codspeed.yml

name: CodSpeed

# The hosted twin of the CI `bench` job. That job DIFFS cachegrind Ir of the PR
# binary against an origin/main binary and is the merge gate. This lane runs
# the same class of measurement (valgrind CPU simulation, deterministic) on
# the targets in codspeed.yml and uploads per-commit results, so every PR gets
# a differential flamegraph per target and a trend line lives somewhere other
# than a checked-in baseline. It is a PROFILER and a record, not a second
# gate: nothing here is a required check.
#
# Docs-only diffs are skipped by path: they cannot change the instruction
# count of a C binary.

on:
  push:
    branches: [main]
    paths-ignore:
      - '**.md'
  pull_request:
    branches: [main]
    paths-ignore:
      - '**.md'

permissions:
  contents: read

concurrency:
  group: codspeed-${{ github.ref }}
  cancel-in-progress: true

jobs:
  simulation:
    name: codspeed (simulation)
    runs-on: ubuntu-latest
    timeout-minutes: 20
    permissions:
      contents: read
      id-token: write  # OIDC upload; the repo is public, so no token secret is needed
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1

      - name: Build the release interpreter
        run: ./build.sh

      # The target set is a claim; pin it before measuring anything. Red if a
      # consumer row is gone, the pin is a branch name, or an exec line reaches
      # a DMG outside the fetched dir. Its --selftest plants each of those.
      - name: Pin the lane's target set (tools/codspeed_targets_check.sh)
        run: bash tools/codspeed_targets_check.sh

      # A real consumer in the matrix. Pinned by SHA, not branch: a moving
      # consumer would put ITS changes into this repo's trend line (the
      # solver-policy confound). Bump deliberately, in a PR that says why.
      # Shallow fetch of one commit; `git clone` of a branch is not a pin.
      - name: Fetch consumer DMG at its pin
        env:
          DMG_REF: afd9bd52778cb0a8d1ab0473ec726aaae556fb90  # DMG master, 2026-10-06 (DMG#79: ready for the boolean type, #1637)
        run: |
          set -euo pipefail
          mkdir -p .codspeed-consumers/DMG && cd .codspeed-consumers/DMG
          git init -q && git remote add origin https://github.com/InauguralSystems/DMG
          git fetch -q --depth 1 origin "$DMG_REF"
          git -c advice.detachedHead=false checkout -q FETCH_HEAD
          test -f roms/cpu_instrs.gb && test -f eigs.json
          echo "DMG at $(git rev-parse HEAD)"

      # `run` is deliberately omitted: with no command the runner executes the
      # exec-harness targets declared in codspeed.yml at the repository root.
      #
      # cycle-estimation OFF: the headline metric is then INSTRUCTIONS, the
      # unit the in-repo Ir gate uses and the unit the dev box agrees with
      # (perf `instructions:u` matched cachegrind Ir within 1%). With it on,
      # the first run attributed ~60% of the DMG shape's time to simulated
      # cache misses; perf stat on the box the same night read IPC 1.61,
      # 1.6% last-level references and 76 K memory misses in a 1.19 G
      # instruction run — a busy core, not a stalled one. A model that
      # disagrees with the hardware by that much is not a headline number.
      - name: Run codspeed.yml targets under CPU simulation
        uses: CodSpeedHQ/action@373d6868929f444bc08d901fd0eb0ad52a8875ea  # v5.2.1
        with:
          mode: simulation
          cycle-estimation: false