#!/usr/bin/env bash
set -euo pipefail

# M3: Generate flamegraphs + top hotspots for Manifest decode+profile (Ours vs Routinator).
#
# Outputs under:
#   specs/develop/20260224/flamegraph/
#   specs/develop/20260224/hotspots/
#   specs/develop/20260224/perf/
#
# Notes:
# - On WSL2, /usr/bin/perf is often a wrapper that fails. This script uses a real perf binary
#   from /usr/lib/linux-tools/*/perf (if present).
# - Ours profiling uses perf + flamegraph --perfdata to avoid rebuilding the whole crate graph
#   with RocksDB.

ROOT_REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
RPKI_DIR="$ROOT_REPO/rpki"

DATE_TAG="${DATE_TAG:-20260224}"
OUT_BASE="$ROOT_REPO/specs/develop/${DATE_TAG}"
OUT_FLAME="$OUT_BASE/flamegraph"
OUT_HOT="$OUT_BASE/hotspots"
OUT_PERF="$OUT_BASE/perf"

RUN_TAG="${RUN_TAG:-p2}"

OURS_BENCH_DIR="$RPKI_DIR/benchmark/ours_manifest_bench"
OURS_BIN="$OURS_BENCH_DIR/target/release/ours-manifest-bench"

ROUT_BENCH_DIR="${ROUT_BENCH_DIR:-/home/yuyr/dev/rust_playground/routinator/benchmark}"
ROUT_BIN="$ROUT_BENCH_DIR/target/release/routinator-manifest-benchmark"
ROUT_ISSUER="$ROUT_BENCH_DIR/fixtures/ta.cer"

PROFILE_HZ="${PROFILE_HZ:-99}"

mkdir -p "$OUT_FLAME" "$OUT_HOT" "$OUT_PERF"

PERF_WRAPPER_OUT="$(perf --version 2>&1 || true)"
PERF_REAL=""
if echo "${PERF_WRAPPER_OUT}" | grep -q "WARNING: perf not found for kernel"; then
  PERF_REAL="$(ls -1 /usr/lib/linux-tools/*/perf 2>/dev/null | head -n 1 || true)"
else
  PERF_REAL="$(command -v perf || true)"
fi

if [[ -z "${PERF_REAL}" ]]; then
  echo "ERROR: usable perf binary not found (wrapper detected and no /usr/lib/linux-tools/*/perf)." >&2
  exit 2
fi

SHIM_DIR="$RPKI_DIR/target/bench/tools"
mkdir -p "$SHIM_DIR"
cat > "$SHIM_DIR/perf" <<EOF
#!/usr/bin/env bash
exec "${PERF_REAL}" "\$@"
EOF
chmod +x "$SHIM_DIR/perf"
export PATH="$SHIM_DIR:$PATH"

echo "Using perf: $PERF_REAL"

echo "[1/3] Build ours benchmark with frame pointers..."
(cd "$OURS_BENCH_DIR" && RUSTFLAGS="-C force-frame-pointers=yes" cargo build --release -q)

echo "[2/3] Build routinator benchmark (release)..."
(cd "$ROUT_BENCH_DIR" && cargo build --release -q)

taskset_prefix=""
if command -v taskset >/dev/null 2>&1; then
  taskset_prefix="taskset -c 0"
fi

profile_ours() {
  local sample="$1"
  local iters="$2"
  local warm="$3"
  local fixture="$RPKI_DIR/tests/benchmark/selected_der/${sample}.mft"
  if [[ ! -f "$fixture" ]]; then
    echo "ERROR: ours fixture not found: $fixture" >&2
    exit 1
  fi

  local perfdata="$OUT_PERF/ours_${sample}_${RUN_TAG}.perf.data"
  local svg="$OUT_FLAME/ours_${sample}_${RUN_TAG}.svg"
  local tsv="$OUT_HOT/ours_${sample}_${RUN_TAG}.tsv"

  echo "== ours $sample (iters=$iters warmup=$warm hz=$PROFILE_HZ)"
  $taskset_prefix perf record -o "$perfdata" -F "$PROFILE_HZ" -g -- \
    "$OURS_BIN" --manifest "$fixture" --iterations "$iters" --warmup-iterations "$warm" --repeats 1 >/dev/null

  flamegraph --perfdata "$perfdata" --output "$svg" --title "ours ${sample} ManifestObject::decode_der" --deterministic >/dev/null

  perf report -i "$perfdata" --stdio --no-children --sort symbol --percent-limit 0.5 \
    | awk '/^[[:space:]]*[0-9.]+%/ {pct=$1; sub(/%/,"",pct); $1=""; sub(/^[[:space:]]+/,""); print pct "\t" $0}' \
    > "$tsv"
}

profile_routinator() {
  local sample="$1"
  local iters="$2"
  local warm="$3"
  local fixture="$ROUT_BENCH_DIR/fixtures/selected_der/${sample}.mft"
  if [[ ! -f "$fixture" ]]; then
    echo "ERROR: routinator fixture not found: $fixture" >&2
    exit 1
  fi

  local svg="$OUT_FLAME/routinator_${sample}_${RUN_TAG}.svg"
  local tsv="$OUT_HOT/routinator_${sample}_${RUN_TAG}.tsv"

  echo "== routinator $sample (iters=$iters warmup=$warm hz=$PROFILE_HZ)"
  $taskset_prefix "$ROUT_BIN" \
    --target decode_only \
    --manifest "$fixture" \
    --issuer "$ROUT_ISSUER" \
    --iterations "$iters" \
    --repeats 1 \
    --warmup-iterations "$warm" \
    --strict false \
    --profile-hz "$PROFILE_HZ" \
    --flamegraph "$svg" \
    --hotspots "$tsv" \
    >/dev/null
}

echo "[3/3] Profile samples..."

# Choose iterations so each capture runs ~10-20s serially.
profile_ours small-01 200000 0
profile_routinator small-01 200000 0

profile_ours large-02 50000 0
profile_routinator large-02 50000 0

profile_ours xlarge-02 5000 0
profile_routinator xlarge-02 5000 0

echo "Done."
echo "- Flamegraphs: $OUT_FLAME/"
echo "- Hotspots:    $OUT_HOT/"
echo "- Perf data:   $OUT_PERF/"