Merge pull request #347 from faster-cpython/nogil-profiling-categories

Add new profiling categories, mainly for free-threading

Author: mdboom

bench_runner/scripts/profiling_plot.py

@@ -30,34 +30,38 @@
 # These are matched in-order.
 CATEGORIES: dict[str, list[str]] = {
     "interpreter": [
-        "_Py_GetBaseOpcode",
         "_Py_call_instrumentation_line",
+        "_Py_GetBaseOpcode",
         "_PyCode_.+",
         "_PyEval.+",
         "_PyFrame_ClearExceptCode",
         "_PyFrame_New_NoTrack",
         "_PyFrame_Traverse",
         "_PyPegen_.+",
-        "_PyThreadState_.+",
+        "_PyStack_.+",
+        "_PyVectorcall_.+",
         "advance",
         "call_instrumentation_vector.*",
         "initialize_locals",
         "PyAST_.+",
         "PyEval_.+",
+        "PyInterpreterState.*",
     ],
     "lookup": [
-        "_Py_dict_lookup",
+        "_Py_dict_lookup.*",
         "_Py_hashtable_.+",
         "_Py_type_getattro",
         "_PyType_Lookup",
-        "_PyType_LookupRef",
+        "_PyType_LookupRef.*",
         "builtin_getattr",
         "find_name_in_mro",
         "lookdict_split",
         "lookdict_unicode_nodummy",
         "lookdict_unicode",
+        "lookup_maybe_method",
+        "member_get",
         "PyMember_.*",
-        "unicodekeys_lookup_unicode",
+        "unicodekeys_lookup_unicode.*",
         "update_one_slot",
     ],
     "gc": [
@@ -68,26 +72,36 @@
         ".+_traverse",
         ".+MaybeUntrack",
         "deduce_unreachable",
-        "gc_collect_main",
-        "gc_collect_region.*",
+        "gc_collect.*",
+        "mark_heap_visitor",
         "PyObject_IS_GC",
+        "scan_heap.+",
         "type_is_gc",
+        "update_refs.*",
         "visit_.+",
     ],
     "memory": [
+        "_?mi_.+",
         "_?PyMem_.+",
+        "_Py_DecRefShared",
+        "_Py_MergeZeroLocalRefcount",
         "_Py_NewReference",
+        "_py_qsbr_.+",
+        "_Py_TryIncrefCompare",
         "_PyObject_Free",
         "_PyObject_Malloc",
+        "_PyType_AllocNoTrack",
         ".+_alloc",
         ".+[Nn]ew.*",
         ".+Alloc",
         ".+Calloc",
         ".+dealloc",
         ".+Dealloc",
+        ".+MiMalloc",
         ".+Realloc",
-        "memset.+",
         "memcpy.+",
+        "memset.+",
+        "Py_DECREF",
     ],
     "dynamic": [
         "_?PyMapping_.+",
@@ -124,16 +138,20 @@
         "tuple.+",
     ],
     "dict": [
-        "_?PyDict_.+",
+        "_?PyDict.+",
+        "acquire_key_value",
         "build_indices_unicode",
         "clone_combined_dict_keys.+",
         "dict_.+",
         "dictiter_.+",
+        "dictkeys_.+",
         "dictresize",
         "find_empty_slot",
         "free_keys_object",
+        "insert_combined_dict",
         "insert_to_emptydict",
         "insertdict",
+        "new_dict",
         "new_keys_object",
         "OrderedDict_.+",
     ],
@@ -148,10 +166,11 @@
     ],
     "str": [
         "_?PyUnicode.+",
-        "_copy_characters.+",
+        "_copy_characters.*",
         "ascii_decode",
         "bytes_.+",
-        "intern_common.+",
+        "find_first_nonascii",
+        "intern_common.*",
         "intern_constants",
         "intern_string_constants",
         "PyBytes_.+",
@@ -163,22 +182,25 @@
         "unicode_.+",
     ],
     "miscobj": [
+        "_?PyGen_.+",
+        "_?PySet_.+",
         "_?PySlice_.+",
         "_PyBuildSlice_ConsumeRefs",
         "_PyEval_SliceIndex",
+        "_PyMake_Coro",
         "bytearray_.+",
         "deque_.+",
         "dequeiter_.+",
         "enum_.+",
         "gen_iternext",
+        "get_or_create_weakref",
         "make_gen",
         "PyBool_.+",
         "PyBuffer_.+",
-        "_?PyGen_.+",
-        "_?PySet_.+",
         "range_.+",
         "set_.+",
         "setiter_.+",
+        "weakref_.+",
     ],
     "exceptions": [
         "_?PyErr_.*",
@@ -217,6 +239,13 @@
         "tok_.+",
     ],
     "async": ["async_.+"],
+    "threading": [
+        "_?PyThreadState_.+",
+        "_PyCriticalSection.+",
+        "_PySeqLock_.+",
+        "disable_deferred_refcounting",
+        "PyThread.*",
+    ],
 }
 
 COLOR_ORDER = ["jit", "kernel", "libc", "library"] + list(CATEGORIES.keys())
@@ -249,7 +278,7 @@ def category_for_obj_sym(obj: str, sym: str) -> str:
     if obj == "python":
         for category, patterns in CATEGORIES.items():
             for pattern in patterns:
-                if re.match(f"^{pattern}$", sym.split()[0]):
+                if re.match(f"^{pattern}$", sym.split()[0].split(".")[0]):
                     return category
 
     return "unknown"

bench_runner/scripts/run_benchmarks.py

@@ -248,6 +248,8 @@ def collect_perf(python: PathLike, benchmarks: str):
                 perf_to_csv(
                     output.splitlines(), PROFILING_RESULTS / f"{benchmark}.perf.csv"
                 )
+            else:
+                print(f"No perf.data file generated for {benchmark}", file=sys.stderr)
 
 
 def update_metadata(