config-disassembler 0.11.7

name: Profile

# Profiles the crate under sf-decomposer's real workload (see
# examples/replay_workload.rs) instead of micro benches: the CPU-only
# `xml_micro` benches got 25-36% faster in #154 while sf-decomposer's
# end-to-end CPU time didn't move, so the hot path is somewhere they don't
# measure. This job answers "where does the time / memory actually go".
#
# Fixtures come from sf-decomposer's own deterministic generator
# (scripts/gen-perf-fixtures.ts), so the shapes match its perf suite.
#
# Outputs (job summary + artifact):
#   - unprofiled wall time and max RSS per phase (/usr/bin/time -v), to check
#     the replay is representative of sf-decomposer's perf comment
#   - perf: top functions by self time, by inclusive time, and a flamegraph SVG
#
# Same-runner A/B (set `compare_ref`): builds the replay from `compare_ref` and
# from the dispatched ref on ONE runner and runs them in alternating order for
# `rounds` rounds (fresh fixture copy + dropped page cache before every run).
# Cross-runner comparisons swing by tens of percent on hosted runners; this is
# what made #158's -35% decompose result conclusive. Reports per-phase medians
# and how many rounds the dispatched ref won.
#
# Heap profiling (heaptrack) was tried and removed: recording every allocation
# got the hosted runner shut down even on the `medium` fixture. Max RSS per
# phase is still reported by the unprofiled baseline.

on:
  workflow_dispatch:
    inputs:
      profile:
        description: sf-decomposer fixture profile
        type: choice
        default: large
        options: [small, medium, large, xlarge, manyfiles]
      format:
        description: Shard format passed to disassemble
        type: choice
        default: xml
        options: [xml, json, json5, yaml]
      sf_decomposer_ref:
        description: sf-decomposer ref to take the fixture generator from
        type: string
        default: main
      compare_ref:
        description: Same-runner A/B against this ref (e.g. main). Empty = profile only.
        type: string
        default: ''
      rounds:
        description: A/B rounds (only with compare_ref)
        type: string
        default: '6'
  # Run on PRs that change the profiling setup itself, so it can be iterated
  # on before it exists on main (workflow_dispatch only sees main's copy).
  pull_request:
    paths:
      - examples/replay_workload.rs
      - .github/workflows/profile.yml

permissions:
  contents: read

# PR runs supersede each other; manual dispatches never cancel one another (two
# dispatches on the same ref, e.g. repeated A/B runs, must both complete).
concurrency:
  group: profile-${{ github.event.pull_request.number || github.run_id }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

env:
  PROFILE: ${{ inputs.profile || 'large' }}
  FORMAT: ${{ inputs.format || 'xml' }}
  CARGO_TERM_COLOR: never
  # Frame pointers so perf can unwind with --call-graph fp (cheap, small
  # perf.data); line tables so symbols resolve to source lines.
  RUSTFLAGS: -C force-frame-pointers=yes
  CARGO_PROFILE_RELEASE_DEBUG: line-tables-only

jobs:
  profile:
    if: ${{ !inputs.compare_ref }}
    name: Profile (${{ inputs.profile || 'large' }}, ${{ inputs.format || 'xml' }})
    runs-on: ubuntu-latest
    timeout-minutes: 45
    steps:
      - name: Checkout
        uses: actions/checkout@v7
        with:
          persist-credentials: false

      - name: Checkout sf-decomposer (fixture generator)
        uses: actions/checkout@v7
        with:
          repository: mcarvin8/sf-decomposer
          ref: ${{ inputs.sf_decomposer_ref || 'main' }}
          path: sf-decomposer
          persist-credentials: false

      - name: Setup Node
        uses: actions/setup-node@v7
        with:
          node-version: lts/*

      - name: Generate fixtures
        # The generator only uses node built-ins, so no `npm ci` is needed.
        run: |
          npx --yes tsx scripts/gen-perf-fixtures.ts --profile "$PROFILE" --out "$RUNNER_TEMP/fixtures"
        working-directory: sf-decomposer

      - name: Install Rust toolchain
        uses: actions-rust-lang/setup-rust-toolchain@v2
        with:
          toolchain: stable
          cache: false

      - uses: Swatinem/rust-cache@v2
        with:
          cache-bin: false

      - name: Install inferno (flamegraphs)
        uses: taiki-e/install-action@v2
        with:
          tool: inferno

      - name: Install perf
        run: |
          sudo apt-get update -qq
          # Package layout differs across runner images: newer Ubuntu ships a
          # `linux-perf` package; older ones need the kernel-matched
          # linux-tools package. Try both; whichever yields a working binary wins.
          sudo apt-get install -y -qq linux-tools-common
          sudo apt-get install -y -qq linux-perf \
            || sudo apt-get install -y -qq "linux-tools-$(uname -r)" \
            || sudo apt-get install -y -qq linux-tools-azure || true
          # Prefer a versioned binary (the /usr/bin/perf wrapper on older images
          # refuses to run when the tools package doesn't match the kernel), and
          # keep the first candidate that actually executes.
          PERF=""
          for candidate in /usr/lib/linux-tools/*/perf /usr/lib/linux-tools-*/perf \
                           /usr/lib/linux-perf*/perf "$(command -v perf || true)"; do
            if [ -x "$candidate" ] && "$candidate" --version >/dev/null 2>&1; then
              PERF="$candidate"; break
            fi
          done
          if [ -z "$PERF" ]; then
            echo "::error::no working perf binary found"; find / -name perf -type f 2>/dev/null | head; exit 1
          fi
          echo "PERF=$PERF" >> "$GITHUB_ENV"
          "$PERF" --version
          sudo sysctl -w kernel.perf_event_paranoid=-1 kernel.kptr_restrict=0

      - name: Build replay
        run: cargo build --release --example replay_workload

      - name: Unprofiled baseline (wall time, max RSS)
        run: |
          BIN=target/release/examples/replay_workload
          OUT="$RUNNER_TEMP/out"; mkdir -p "$OUT"
          cp -r "$RUNNER_TEMP/fixtures" "$RUNNER_TEMP/base"
          for phase in decompose recompose; do
            /usr/bin/time -v "$BIN" "$phase" "$RUNNER_TEMP/base" --format "$FORMAT" \
              2> "$OUT/time-$phase.txt"
          done
          {
            echo "## Unprofiled baseline (profile=$PROFILE, format=$FORMAT)"
            echo
            echo "Compare with sf-decomposer's perf comment rows \`$PROFILE.$FORMAT.decompose\` / \`.recompose\`."
            echo
            echo '| Phase | Wall | User CPU (s) | Sys CPU (s) | Max RSS (MB) |'
            echo '|---|---:|---:|---:|---:|'
            for phase in decompose recompose; do
              f="$OUT/time-$phase.txt"
              wall=$(grep 'Elapsed (wall clock)' "$f" | awk '{print $NF}')
              user=$(grep 'User time' "$f" | awk '{print $NF}')
              sys=$(grep 'System time' "$f" | awk '{print $NF}')
              rss=$(grep 'Maximum resident' "$f" | awk '{printf "%.1f", $NF/1024}')
              echo "| $phase | $wall | $user | $sys | $rss |"
            done
            echo
          } > "$OUT/summary-baseline.md"
          # Job summaries aren't readable through the API; print to the log too.
          cat "$OUT/summary-baseline.md" | tee -a "$GITHUB_STEP_SUMMARY"

      - name: CPU profile (perf)
        run: |
          BIN=target/release/examples/replay_workload
          OUT="$RUNNER_TEMP/out"
          cp -r "$RUNNER_TEMP/fixtures" "$RUNNER_TEMP/cpu"
          for phase in decompose recompose; do
            "$PERF" record -F 499 -g --call-graph fp -o "$OUT/perf-$phase.data" -- \
              "$BIN" "$phase" "$RUNNER_TEMP/cpu" --format "$FORMAT" 2>/dev/null
            "$PERF" report -i "$OUT/perf-$phase.data" --stdio --no-children --sort symbol \
              --percent-limit 0.5 2>/dev/null > "$OUT/top-self-$phase.txt"
            "$PERF" report -i "$OUT/perf-$phase.data" --stdio --children --sort symbol \
              --percent-limit 2 -g none 2>/dev/null > "$OUT/top-inclusive-$phase.txt"
            "$PERF" script -i "$OUT/perf-$phase.data" 2>/dev/null \
              | inferno-collapse-perf > "$OUT/$phase.folded"
            inferno-flamegraph --title "$phase ($PROFILE, $FORMAT)" < "$OUT/$phase.folded" \
              > "$OUT/flamegraph-$phase.svg"
          done
          {
            for phase in decompose recompose; do
              echo "## $phase: top functions by self time"
              echo
              echo '```'
              grep -E '^\s+[0-9.]+%' "$OUT/top-self-$phase.txt" | head -n 30
              echo '```'
              echo
              echo "## $phase: top functions by inclusive time"
              echo
              echo '```'
              grep -E '^\s+[0-9.]+%' "$OUT/top-inclusive-$phase.txt" | head -n 30
              echo '```'
              echo
            done
          } > "$OUT/summary-cpu.md"
          cat "$OUT/summary-cpu.md" | tee -a "$GITHUB_STEP_SUMMARY"

      - name: Upload CPU profiles
        if: always()
        uses: actions/upload-artifact@v7
        with:
          name: profile-cpu-${{ env.PROFILE }}-${{ env.FORMAT }}
          path: |
            ${{ runner.temp }}/out/*.svg
            ${{ runner.temp }}/out/*.txt
            ${{ runner.temp }}/out/*.md
            ${{ runner.temp }}/out/*.folded
          retention-days: 30
          if-no-files-found: warn

  ab:
    name: Same-runner A/B (${{ inputs.compare_ref }} vs ${{ github.ref_name }}, ${{ inputs.profile || 'large' }}, ${{ inputs.format || 'xml' }})
    if: ${{ inputs.compare_ref }}
    runs-on: ubuntu-latest
    timeout-minutes: 60
    env:
      ROUNDS: ${{ inputs.rounds || '6' }}
      BASE_REF: ${{ inputs.compare_ref }}
      HEAD_REF: ${{ github.ref_name }}
    steps:
      - name: Checkout head (dispatched ref)
        uses: actions/checkout@v7
        with:
          path: head
          persist-credentials: false

      - name: Checkout base (compare_ref)
        uses: actions/checkout@v7
        with:
          ref: ${{ inputs.compare_ref }}
          path: base
          persist-credentials: false

      - name: Checkout sf-decomposer (fixture generator)
        uses: actions/checkout@v7
        with:
          repository: mcarvin8/sf-decomposer
          ref: ${{ inputs.sf_decomposer_ref || 'main' }}
          path: sf-decomposer
          persist-credentials: false

      - name: Setup Node
        uses: actions/setup-node@v7
        with:
          node-version: lts/*

      - name: Generate fixtures
        run: npx --yes tsx scripts/gen-perf-fixtures.ts --profile "$PROFILE" --out "$RUNNER_TEMP/fixtures"
        working-directory: sf-decomposer

      - name: Install Rust toolchain
        uses: actions-rust-lang/setup-rust-toolchain@v2
        with:
          toolchain: stable
          cache: false

      - name: Build both replays
        run: |
          for side in base head; do
            if [ ! -f "$side/examples/replay_workload.rs" ]; then
              echo "::error::$side has no examples/replay_workload.rs (compare_ref must include #155)"; exit 1
            fi
            (cd "$side" && cargo build --release --example replay_workload)
            cp "$side/target/release/examples/replay_workload" "$RUNNER_TEMP/replay-$side"
          done

      - name: Alternating rounds on one runner
        run: |
          OUT="$RUNNER_TEMP/ab"; mkdir -p "$OUT"
          TSV="$OUT/ab.tsv"
          printf 'round\tref\tphase\twall_s\tuser_s\tsys_s\tmaxrss_mb\n' > "$TSV"
          run_one() { # side round
            local side=$1 round=$2 work="$RUNNER_TEMP/work-$1"
            rm -rf "$work"; cp -r "$RUNNER_TEMP/fixtures" "$work"
            sync; echo 3 | sudo tee /proc/sys/vm/drop_caches >/dev/null
            for phase in decompose recompose; do
              /usr/bin/time -f '%e\t%U\t%S\t%M' -o "$RUNNER_TEMP/t" \
                "$RUNNER_TEMP/replay-$side" "$phase" "$work" --format "$FORMAT" 2>/dev/null
              IFS=$'\t' read -r e u s m < "$RUNNER_TEMP/t"
              printf '%s\t%s\t%s\t%s\t%s\t%s\t%.1f\n' "$round" "$side" "$phase" "$e" "$u" "$s" "$(echo "$m/1024" | bc -l)" >> "$TSV"
            done
          }
          # One untimed warm-up each, then alternate which side goes first.
          run_one base 0; run_one head 0
          for r in $(seq 1 "$ROUNDS"); do
            if [ $((r % 2)) -eq 1 ]; then run_one base "$r"; run_one head "$r"; else run_one head "$r"; run_one base "$r"; fi
          done
          cat "$TSV"
          echo
          python3 - "$TSV" > "$OUT/summary-ab.md" <<'PY'
          import csv, os, statistics, sys
          rows = [r for r in csv.DictReader(open(sys.argv[1]), delimiter='\t') if r['round'] != '0']
          env = os.environ
          print(f"## Same-runner A/B: `{env['BASE_REF']}` (base) vs `{env['HEAD_REF']}` (head)\n")
          print(f"profile={env['PROFILE']}, format={env['FORMAT']}, {env['ROUNDS']} alternating rounds after a warm-up; medians.\n")
          print('| phase | metric | base | head | change | head better in |')
          print('|---|---|---:|---:|---:|---:|')
          for phase in ('decompose', 'recompose'):
              for metric in ('wall_s', 'user_s', 'sys_s', 'maxrss_mb'):
                  base = {r['round']: float(r[metric]) for r in rows if r['ref'] == 'base' and r['phase'] == phase}
                  head = {r['round']: float(r[metric]) for r in rows if r['ref'] == 'head' and r['phase'] == phase}
                  b, h = statistics.median(base.values()), statistics.median(head.values())
                  wins = sum(1 for k in base if head[k] < base[k])
                  change = f'{(h / b - 1) * 100:+.1f}%' if b else 'n/a'
                  print(f'| {phase} | {metric} | {b:.2f} | {h:.2f} | {change} | {wins}/{len(base)} |')
          PY
          cat "$OUT/summary-ab.md" | tee -a "$GITHUB_STEP_SUMMARY"

      - name: Upload A/B results
        if: always()
        uses: actions/upload-artifact@v7
        with:
          name: ab-${{ env.PROFILE }}-${{ env.FORMAT }}
          path: ${{ runner.temp }}/ab/
          retention-days: 30
          if-no-files-found: warn