Switch to perf for profiling from gperftools

benchmark.py

@@ -5,19 +5,74 @@ import argparse
 import fcntl
 import os
 import select
+import shutil
 import signal
 import struct
+import subprocess
 import sys
 import termios
 import time
 from pty import CHILD, fork
 
-from kitty.constants import kitten_exe
+from kitty.constants import kitten_exe, kitty_exe
 from kitty.fast_data_types import ChildMonitor, Screen, safe_pipe
 from kitty.utils import read_screen_size
 
 BENCHMARK_WINDOW_ID = 1
 ALL_BENCHMARKS = ('ascii', 'unicode', 'unique_unicode', 'csi', 'images', 'long_escape_codes')
+PERF_OUTPUT = '/tmp/kitty-benchmark.perf'
+
+# Set by the re-exec wrapper so we don't recurse when --perf is in argv.
+_UNDER_PERF_ENV = '_KITTY_BENCHMARK_UNDER_PERF'
+
+
+def find_perf() -> str | None:
+    return shutil.which('perf')
+
+
+def run_perf_reports(perf_exe: str) -> None:
+    sep = '=' * 70
+    print(f'\n{sep}')
+    print('PERF PROFILING RESULTS')
+    print(sep)
+    print(f'Profile data saved to: {PERF_OUTPUT}')
+    print(f'Re-run interactively:   perf report -i {PERF_OUTPUT}\n')
+
+    print('--- Top CPU hotspots (call graph, >=0.5% threshold) ---\n')
+    subprocess.run(
+        [
+            perf_exe,
+            'report',
+            '--stdio',
+            '-n',
+            '--call-graph',
+            'fractal,0.5',
+            '--percent-limit',
+            '0.5',
+            '-i',
+            PERF_OUTPUT,
+        ],
+        check=False,
+    )
+
+    print('\n--- Per-thread CPU breakdown ---\n')
+    subprocess.run(
+        [
+            perf_exe,
+            'report',
+            '--stdio',
+            '-n',
+            '--sort',
+            'overhead,tid,comm,symbol',
+            '--percent-limit',
+            '1.0',
+            '-i',
+            PERF_OUTPUT,
+        ],
+        check=False,
+    )
+
+    print(f'\n{sep}\n')
 
 
 def run_parsing_benchmark(
@@ -90,6 +145,35 @@ def run_parsing_benchmark(
         sys.stdout.write(str(screen.linebuf))
 
 
+def exec_under_perf(perf_exe: str) -> None:
+    """Re-exec this script as a child of perf record.
+
+    perf becomes the outer process so it can profile the entire benchmark
+    run without any subprocess/SIGCHLD conflicts with ChildMonitor.
+    After the benchmark exits perf finalises its output, then we run
+    perf report to print the results.
+    """
+    script = os.path.abspath(__file__)
+    env = {**os.environ, _UNDER_PERF_ENV: '1'}
+    cmd = [
+        perf_exe,
+        'record',
+        '-g',
+        '-F',
+        '999',
+        '--call-graph',
+        'dwarf',
+        '-o',
+        PERF_OUTPUT,
+        '--',
+        kitty_exe(),
+        '+launch',
+        script,
+    ] + sys.argv[1:]
+    subprocess.run(cmd, env=env, check=False)
+    run_perf_reports(perf_exe)
+
+
 def main() -> None:
     p = argparse.ArgumentParser(description='Run kitty parsing benchmarks')
     p.add_argument(
@@ -106,7 +190,26 @@ def main() -> None:
         default=True,
         help='Use the main screen instead of the alt screen so scrollback speed is also tested (default: enabled)',
     )
+    p.add_argument(
+        '--perf',
+        action='store_true',
+        default=False,
+        help=(
+            'Profile with Linux perf: records at 999 Hz with DWARF call graphs, '
+            'then prints per-thread CPU breakdown and call-graph hotspots before benchmark results. '
+            'Requires perf in PATH with setcap cap_sys_admin,cap_sys_ptrace,cap_syslog=ep /usr/bin/perf'
+        ),
+    )
     args = p.parse_args()
+
+    if args.perf and not os.environ.get(_UNDER_PERF_ENV):
+        perf_exe = find_perf()
+        if perf_exe is None:
+            print('Warning: perf not found in PATH, running without profiling', file=sys.stderr)
+        else:
+            exec_under_perf(perf_exe)
+            return
+
     benchmarks = tuple(args.benchmarks) if args.benchmarks else ALL_BENCHMARKS
     run_parsing_benchmark(benchmarks=benchmarks, with_scrollback=args.with_scrollback)
 

docs/performance.rst

@@ -142,9 +142,15 @@ admittedly biased, eyes).
 Instrumenting kitty
 -----------------------
 
-You can generate detailed per-function performance data using
-`gperftools <https://github.com/gperftools/gperftools>`__. Build |kitty| with
-``make profile``. Run kitty and perform the task you want to analyse, for
-example, scrolling a large file with :program:`less`. After you quit, function
-call statistics will be displayed in *KCachegrind*. Hence, profiling is best done
-on Linux which has these tools easily available.
+To profile kitty performance on Linux, first install ``perf`` and run::
+
+    sudo setcap cap_sys_admin,cap_sys_ptrace,cap_syslog=ep /usr/bin/perf
+
+Then build kitty with::
+
+    make profile
+
+Finally get benchmark and profiling results using::
+
+    ./benchmark.py --perf
+

kitty/data-types.c

@@ -31,10 +31,6 @@
 #include <stdio.h>
 #include <locale.h>
 
-#ifdef WITH_PROFILER
-#include <gperftools/profiler.h>
-#endif
-
 #include "monotonic.h"
 
 #ifdef __APPLE__
@@ -351,22 +347,6 @@ pyset_iutf8(PyObject UNUSED *self, PyObject *args) {
     Py_RETURN_NONE;
 }
 
-#ifdef WITH_PROFILER
-static PyObject *
-start_profiler(PyObject UNUSED *self, PyObject *args) {
-    char *path;
-    if (!PyArg_ParseTuple(args, "s", &path)) return NULL;
-    ProfilerStart(path);
-    Py_RETURN_NONE;
-}
-
-static PyObject *
-stop_profiler(PyObject UNUSED *self, PyObject *args UNUSED) {
-    ProfilerStop();
-    Py_RETURN_NONE;
-}
-#endif
-
 static bool
 put_tty_in_raw_mode(int fd, const struct termios *termios_p, bool read_with_timeout, int optional_actions) {
     struct termios raw_termios = *termios_p;
@@ -880,10 +860,6 @@ static PyMethodDef module_methods[] = {
 #ifdef __APPLE__
     METHODB(user_cache_dir, METH_NOARGS),
     METHODB(process_group_map, METH_NOARGS),
-#endif
-#ifdef WITH_PROFILER
-    {"start_profiler", (PyCFunction)start_profiler, METH_VARARGS, ""},
-    {"stop_profiler", (PyCFunction)stop_profiler, METH_NOARGS, ""},
 #endif
     {NULL, NULL, 0, NULL} /* Sentinel */
 };

setup.py

@@ -675,7 +675,6 @@ def init_env(
     if profile:
         cppflags.append('-DWITH_PROFILER')
         cflags.append('-g3')
-        ldflags.append('-lprofiler')
 
     if debug or profile:
         cflags.append('-fno-omit-frame-pointer')
@@ -1602,8 +1601,6 @@ def build_launcher(args: Options, launcher_dir: str = '.', bundle_type: str = 's
             cflags.extend(sanitize_args)
             ldflags.extend(sanitize_args)
             libs += ['-lasan'] if not is_macos and env.compiler_type is not CompilerType.clang else []
-        if args.profile:
-            libs.append('-lprofiler')
     else:
         cflags.append('-g3' if args.debug else '-O3')
     if bundle_type.endswith('-freeze'):
@@ -2299,7 +2296,7 @@ def option_parser() -> argparse.ArgumentParser:  # {{{
         ' the Python used to run setup.py is queried for these.',
     )
     p.add_argument('--full', dest='incremental', default=Options.incremental, action='store_false', help='Do a full build, even for unchanged files')
-    p.add_argument('--profile', default=Options.profile, action='store_true', help='Use the -pg compile flag to add profiling information')
+    p.add_argument('--profile', default=Options.profile, action='store_true', help='Use compile flags to add profiling information')
     p.add_argument(
         '--libdir-name', default=Options.libdir_name, help='The name of the directory inside --prefix in which to store compiled files. Defaults to "lib"'
     )