Fix the perf period calculation

Author: mdboom

bench_runner/scripts/profiling_plot.py

@@ -9,6 +9,7 @@
 from collections import defaultdict
 import csv
 import functools
+import json
 from operator import itemgetter
 from pathlib import Path
 import re
@@ -40,6 +41,7 @@
         "_PyPegen_.+",
         "_PyStack_.+",
         "_PyVectorcall_.+",
+        "_TAIL_CALL_.+",
         "advance",
         "call_instrumentation_vector.*",
         "initialize_locals",
@@ -414,6 +416,52 @@ def plot_pie(categories: list[tuple[float, str]], output_filename: PathLike):
     fig.savefig(output_filename, dpi=200)
 
 
+def handle_tail_call_stats(
+    categories: defaultdict[str, defaultdict[tuple[str, str], float]],
+    output_prefix: Path,
+):
+    tail_call_stats = defaultdict(float)
+    total_time = 0.0
+    for (_, sym), self_time in categories["interpreter"].items():
+        if (bytecode := sym.removeprefix("_TAIL_CALL_")) != sym:
+            tail_call_stats[bytecode] += self_time
+            total_time += self_time
+
+    with Path("pystats.json").open() as fd:
+        pystats = json.load(fd)
+
+    pystats_bytecodes = defaultdict(int)
+    total_count = 0
+    for key, val in pystats.items():
+        if match := re.match(r"opcode\[(.+)\]\.execution_count", key):
+            pystats_bytecodes[match.group(1)] += val
+            total_count += val
+
+    if len(tail_call_stats) == 0:
+        return
+
+    with open(output_prefix.with_suffix(".tail_calls.csv"), "w") as csvfile:
+        writer = csv.writer(csvfile, dialect="unix")
+        writer.writerow(
+            ["Bytecode", "% time", "count", "% count", "time per count (ns)"]
+        )
+        for bytecode, seconds in sorted(
+            tail_call_stats.items(), key=itemgetter(1), reverse=True
+        ):
+            count = pystats_bytecodes[bytecode]
+            if count == 0:
+                continue
+            writer.writerow(
+                [
+                    bytecode,
+                    f"{seconds / total_time:.02%}",
+                    count,
+                    f"{count / total_count:.02%}",
+                    f"{(seconds / count) * 1e9:03f}",
+                ]
+            )
+
+
 def _main(input_dir: PathLike, output_prefix: PathLike):
     input_dir = Path(input_dir)
     output_prefix = Path(output_prefix)
@@ -427,6 +475,8 @@ def _main(input_dir: PathLike, output_prefix: PathLike):
 
     with output_prefix.with_suffix(".md").open("w") as md:
         for csv_path in sorted(input_dir.glob("*.csv")):
+            if "tail_calls.csv" in csv_path.name:
+                continue
             handle_benchmark(csv_path, md, results, categories)
 
         sorted_categories = sorted(
@@ -454,6 +504,8 @@ def _main(input_dir: PathLike, output_prefix: PathLike):
     plot_bargraph(results, sorted_categories, output_prefix.with_suffix(".svg"))
     plot_pie(sorted_categories, output_prefix.with_suffix(".pie.svg"))
 
+    handle_tail_call_stats(categories, output_prefix)
+
 
 def main():
     parser = argparse.ArgumentParser(

bench_runner/scripts/run_benchmarks.py

@@ -189,9 +189,9 @@ def perf_to_csv(lines: Iterable[str], output: PathLike):
         else:
             _, period, command, _, symbol, shared, _ = line.split(maxsplit=6)
             pid, command = command.split(":")
-            self_time = float(int(period)) / total
-            if self_time > 0.0:
-                rows.append([self_time, pid, command, shared, symbol])
+            period = float(period)
+            if period > 0.0:
+                rows.append([period, pid, command, shared, symbol])
 
     rows.sort(key=itemgetter(0), reverse=True)