From c98c5b344941c45e50dded708177f7ec2e225b2b Mon Sep 17 00:00:00 2001 From: Neil Schemenauer Date: Mon, 3 Nov 2025 11:36:37 -0800 Subject: [PATCH] gh-131253: free-threaded build support for pystats (gh-137189) Allow the --enable-pystats build option to be used with free-threading. The stats are now stored on a per-interpreter basis, rather than process global. For free-threaded builds, the stats structure is allocated per-thread and then periodically merged into the per-interpreter stats structure (on thread exit or when the reporting function is called). Most of the pystats related code has be moved into the file Python/pystats.c. --- Include/cpython/pystate.h | 24 + Include/cpython/pystats.h | 64 +- Include/internal/pycore_interp_structs.h | 14 +- Include/internal/pycore_pystats.h | 2 +- Include/internal/pycore_stats.h | 77 +- Include/internal/pycore_tstate.h | 6 + Lib/test/test_pystats.py | 215 +++++ Makefile.pre.in | 1 + ...-07-29-17-51-14.gh-issue-131253.GpRjWy.rst | 1 + Modules/_xxtestfuzz/fuzzer.c | 4 +- PCbuild/_freeze_module.vcxproj | 1 + PCbuild/_freeze_module.vcxproj.filters | 3 + PCbuild/pythoncore.vcxproj | 1 + Python/ceval_macros.h | 3 +- Python/gc.c | 7 +- Python/gc_free_threading.c | 16 +- Python/initconfig.c | 6 - Python/lock.c | 3 + Python/pylifecycle.c | 17 + Python/pystate.c | 29 + Python/pystats.c | 819 ++++++++++++++++++ Python/qsbr.c | 3 +- Python/specialize.c | 426 +-------- Python/sysmodule.c | 4 +- 24 files changed, 1265 insertions(+), 481 deletions(-) create mode 100644 Lib/test/test_pystats.py create mode 100644 Misc/NEWS.d/next/Core_and_Builtins/2025-07-29-17-51-14.gh-issue-131253.GpRjWy.rst create mode 100644 Python/pystats.c diff --git a/Include/cpython/pystate.h b/Include/cpython/pystate.h index ac8798ff6129..dd2ea1202b37 100644 --- a/Include/cpython/pystate.h +++ b/Include/cpython/pystate.h @@ -217,6 +217,15 @@ struct _ts { */ PyObject *threading_local_sentinel; _PyRemoteDebuggerSupport remote_debugger_support; + +#ifdef Py_STATS + // Pointer to PyStats structure, NULL if recording is off. For the + // free-threaded build, the structure is per-thread (stored as a pointer + // in _PyThreadStateImpl). For the default build, the structure is stored + // in the PyInterpreterState structure (threads do not have their own + // structure and all share the same per-interpreter structure). + PyStats *pystats; +#endif }; /* other API */ @@ -239,6 +248,21 @@ PyAPI_FUNC(void) PyThreadState_EnterTracing(PyThreadState *tstate); // function is set, otherwise disable them. PyAPI_FUNC(void) PyThreadState_LeaveTracing(PyThreadState *tstate); +#ifdef Py_STATS +#if defined(HAVE_THREAD_LOCAL) && !defined(Py_BUILD_CORE_MODULE) +extern _Py_thread_local PyThreadState *_Py_tss_tstate; + +static inline PyStats* +_PyThreadState_GetStatsFast(void) +{ + if (_Py_tss_tstate == NULL) { + return NULL; // no attached thread state + } + return _Py_tss_tstate->pystats; +} +#endif +#endif // Py_STATS + /* PyGILState */ /* Helper/diagnostic function - return 1 if the current thread diff --git a/Include/cpython/pystats.h b/Include/cpython/pystats.h index cf830b6066f4..d0a925a30554 100644 --- a/Include/cpython/pystats.h +++ b/Include/cpython/pystats.h @@ -4,7 +4,7 @@ // // - _Py_INCREF_STAT_INC() and _Py_DECREF_STAT_INC() used by Py_INCREF() // and Py_DECREF(). -// - _Py_stats variable +// - _PyStats_GET() // // Functions of the sys module: // @@ -14,7 +14,7 @@ // - sys._stats_dump() // // Python must be built with ./configure --enable-pystats to define the -// Py_STATS macro. +// _PyStats_GET() macro. // // Define _PY_INTERPRETER macro to increment interpreter_increfs and // interpreter_decrefs. Otherwise, increment increfs and decrefs. @@ -109,6 +109,18 @@ typedef struct _gc_stats { uint64_t objects_not_transitively_reachable; } GCStats; +#ifdef Py_GIL_DISABLED +// stats specific to free-threaded build +typedef struct _ft_stats { + // number of times interpreter had to spin or park when trying to acquire a mutex + uint64_t mutex_sleeps; + // number of times that the QSBR mechanism polled (compute read sequence value) + uint64_t qsbr_polls; + // number of times stop-the-world mechanism was used + uint64_t world_stops; +} FTStats; +#endif + typedef struct _uop_stats { uint64_t execution_count; uint64_t miss; @@ -173,22 +185,48 @@ typedef struct _stats { CallStats call_stats; ObjectStats object_stats; OptimizationStats optimization_stats; +#ifdef Py_GIL_DISABLED + FTStats ft_stats; +#endif RareEventStats rare_event_stats; - GCStats *gc_stats; + GCStats gc_stats[3]; // must match NUM_GENERATIONS } PyStats; +// Export for most shared extensions +PyAPI_FUNC(PyStats *) _PyStats_GetLocal(void); + +#if defined(HAVE_THREAD_LOCAL) && !defined(Py_BUILD_CORE_MODULE) +// use inline function version defined in cpython/pystate.h +static inline PyStats *_PyThreadState_GetStatsFast(void); +#define _PyStats_GET _PyThreadState_GetStatsFast +#else +#define _PyStats_GET _PyStats_GetLocal +#endif -// Export for shared extensions like 'math' -PyAPI_DATA(PyStats*) _Py_stats; +#define _Py_STATS_EXPR(expr) \ + do { \ + PyStats *s = _PyStats_GET(); \ + if (s != NULL) { \ + s->expr; \ + } \ + } while (0) + +#define _Py_STATS_COND_EXPR(cond, expr) \ + do { \ + PyStats *s = _PyStats_GET(); \ + if (s != NULL && (cond)) { \ + s->expr; \ + } \ + } while (0) #ifdef _PY_INTERPRETER -# define _Py_INCREF_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.interpreter_increfs++; } while (0) -# define _Py_DECREF_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.interpreter_decrefs++; } while (0) -# define _Py_INCREF_IMMORTAL_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.interpreter_immortal_increfs++; } while (0) -# define _Py_DECREF_IMMORTAL_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.interpreter_immortal_decrefs++; } while (0) +# define _Py_INCREF_STAT_INC() _Py_STATS_EXPR(object_stats.interpreter_increfs++) +# define _Py_DECREF_STAT_INC() _Py_STATS_EXPR(object_stats.interpreter_decrefs++) +# define _Py_INCREF_IMMORTAL_STAT_INC() _Py_STATS_EXPR(object_stats.interpreter_immortal_increfs++) +# define _Py_DECREF_IMMORTAL_STAT_INC() _Py_STATS_EXPR(object_stats.interpreter_immortal_decrefs++) #else -# define _Py_INCREF_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.increfs++; } while (0) -# define _Py_DECREF_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.decrefs++; } while (0) -# define _Py_INCREF_IMMORTAL_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.immortal_increfs++; } while (0) -# define _Py_DECREF_IMMORTAL_STAT_INC() do { if (_Py_stats) _Py_stats->object_stats.immortal_decrefs++; } while (0) +# define _Py_INCREF_STAT_INC() _Py_STATS_EXPR(object_stats.increfs++) +# define _Py_DECREF_STAT_INC() _Py_STATS_EXPR(object_stats.decrefs++) +# define _Py_INCREF_IMMORTAL_STAT_INC() _Py_STATS_EXPR(object_stats.immortal_increfs++) +# define _Py_DECREF_IMMORTAL_STAT_INC() _Py_STATS_EXPR(object_stats.immortal_decrefs++) #endif diff --git a/Include/internal/pycore_interp_structs.h b/Include/internal/pycore_interp_structs.h index 9cdaa950e347..e8cbe9d894e1 100644 --- a/Include/internal/pycore_interp_structs.h +++ b/Include/internal/pycore_interp_structs.h @@ -199,7 +199,7 @@ enum _GCPhase { }; /* If we change this, we need to change the default value in the - signature of gc.collect. */ + signature of gc.collect and change the size of PyStats.gc_stats */ #define NUM_GENERATIONS 3 struct _gc_runtime_state { @@ -963,6 +963,18 @@ struct _is { # ifdef Py_STACKREF_CLOSE_DEBUG _Py_hashtable_t *closed_stackrefs_table; # endif +#endif + +#ifdef Py_STATS + // true if recording of pystats is on, this is used when new threads + // are created to decide if recording should be on for them + int pystats_enabled; + // allocated when (and if) stats are first enabled + PyStats *pystats_struct; +#ifdef Py_GIL_DISABLED + // held when pystats related interpreter state is being updated + PyMutex pystats_mutex; +#endif #endif /* the initial PyInterpreterState.threads.head */ diff --git a/Include/internal/pycore_pystats.h b/Include/internal/pycore_pystats.h index f8af398a5605..50ab21aa0f19 100644 --- a/Include/internal/pycore_pystats.h +++ b/Include/internal/pycore_pystats.h @@ -9,7 +9,7 @@ extern "C" { #endif #ifdef Py_STATS -extern void _Py_StatsOn(void); +extern int _Py_StatsOn(void); extern void _Py_StatsOff(void); extern void _Py_StatsClear(void); extern int _Py_PrintSpecializationStats(int to_file); diff --git a/Include/internal/pycore_stats.h b/Include/internal/pycore_stats.h index 24f239a2135b..850e6ea45522 100644 --- a/Include/internal/pycore_stats.h +++ b/Include/internal/pycore_stats.h @@ -15,39 +15,56 @@ extern "C" { #include "pycore_bitutils.h" // _Py_bit_length -#define STAT_INC(opname, name) do { if (_Py_stats) _Py_stats->opcode_stats[opname].specialization.name++; } while (0) -#define STAT_DEC(opname, name) do { if (_Py_stats) _Py_stats->opcode_stats[opname].specialization.name--; } while (0) -#define OPCODE_EXE_INC(opname) do { if (_Py_stats) _Py_stats->opcode_stats[opname].execution_count++; } while (0) -#define CALL_STAT_INC(name) do { if (_Py_stats) _Py_stats->call_stats.name++; } while (0) -#define OBJECT_STAT_INC(name) do { if (_Py_stats) _Py_stats->object_stats.name++; } while (0) -#define OBJECT_STAT_INC_COND(name, cond) \ - do { if (_Py_stats && cond) _Py_stats->object_stats.name++; } while (0) -#define EVAL_CALL_STAT_INC(name) do { if (_Py_stats) _Py_stats->call_stats.eval_calls[name]++; } while (0) -#define EVAL_CALL_STAT_INC_IF_FUNCTION(name, callable) \ - do { if (_Py_stats && PyFunction_Check(callable)) _Py_stats->call_stats.eval_calls[name]++; } while (0) -#define GC_STAT_ADD(gen, name, n) do { if (_Py_stats) _Py_stats->gc_stats[(gen)].name += (n); } while (0) -#define OPT_STAT_INC(name) do { if (_Py_stats) _Py_stats->optimization_stats.name++; } while (0) -#define OPT_STAT_ADD(name, n) do { if (_Py_stats) _Py_stats->optimization_stats.name += (n); } while (0) -#define UOP_STAT_INC(opname, name) do { if (_Py_stats) { assert(opname < 512); _Py_stats->optimization_stats.opcode[opname].name++; } } while (0) -#define UOP_PAIR_INC(uopcode, lastuop) \ - do { \ - if (lastuop && _Py_stats) { \ - _Py_stats->optimization_stats.opcode[lastuop].pair_count[uopcode]++; \ - } \ - lastuop = uopcode; \ +#define STAT_INC(opname, name) _Py_STATS_EXPR(opcode_stats[opname].specialization.name++) +#define STAT_DEC(opname, name) _Py_STATS_EXPR(opcode_stats[opname].specialization.name--) +#define OPCODE_EXE_INC(opname) _Py_STATS_EXPR(opcode_stats[opname].execution_count++) +#define CALL_STAT_INC(name) _Py_STATS_EXPR(call_stats.name++) +#define OBJECT_STAT_INC(name) _Py_STATS_EXPR(object_stats.name++) +#define OBJECT_STAT_INC_COND(name, cond) _Py_STATS_COND_EXPR(cond, object_stats.name++) +#define EVAL_CALL_STAT_INC(name) _Py_STATS_EXPR(call_stats.eval_calls[name]++) +#define EVAL_CALL_STAT_INC_IF_FUNCTION(name, callable) _Py_STATS_COND_EXPR(PyFunction_Check(callable), call_stats.eval_calls[name]++) +#define GC_STAT_ADD(gen, name, n) _Py_STATS_EXPR(gc_stats[(gen)].name += (n)) +#define OPT_STAT_INC(name) _Py_STATS_EXPR(optimization_stats.name++) +#define OPT_STAT_ADD(name, n) _Py_STATS_EXPR(optimization_stats.name += (n)) +#define UOP_STAT_INC(opname, name) \ + do { \ + PyStats *s = _PyStats_GET(); \ + if (s) { \ + assert(opname < 512); \ + s->optimization_stats.opcode[opname].name++; \ + } \ + } while (0) +#define UOP_PAIR_INC(uopcode, lastuop) \ + do { \ + PyStats *s = _PyStats_GET(); \ + if (lastuop && s) { \ + s->optimization_stats.opcode[lastuop].pair_count[uopcode]++; \ + } \ + lastuop = uopcode; \ } while (0) -#define OPT_UNSUPPORTED_OPCODE(opname) do { if (_Py_stats) _Py_stats->optimization_stats.unsupported_opcode[opname]++; } while (0) -#define OPT_ERROR_IN_OPCODE(opname) do { if (_Py_stats) _Py_stats->optimization_stats.error_in_opcode[opname]++; } while (0) +#define OPT_UNSUPPORTED_OPCODE(opname) _Py_STATS_EXPR(optimization_stats.unsupported_opcode[opname]++) +#define OPT_ERROR_IN_OPCODE(opname) _Py_STATS_EXPR(optimization_stats.error_in_opcode[opname]++) #define OPT_HIST(length, name) \ do { \ - if (_Py_stats) { \ + PyStats *s = _PyStats_GET(); \ + if (s) { \ int bucket = _Py_bit_length(length >= 1 ? length - 1 : 0); \ bucket = (bucket >= _Py_UOP_HIST_SIZE) ? _Py_UOP_HIST_SIZE - 1 : bucket; \ - _Py_stats->optimization_stats.name[bucket]++; \ + s->optimization_stats.name[bucket]++; \ } \ } while (0) -#define RARE_EVENT_STAT_INC(name) do { if (_Py_stats) _Py_stats->rare_event_stats.name++; } while (0) -#define OPCODE_DEFERRED_INC(opname) do { if (_Py_stats && opcode == opname) _Py_stats->opcode_stats[opname].specialization.deferred++; } while (0) +#define RARE_EVENT_STAT_INC(name) _Py_STATS_EXPR(rare_event_stats.name++) +#define OPCODE_DEFERRED_INC(opname) _Py_STATS_COND_EXPR(opcode==opname, opcode_stats[opname].specialization.deferred++) + +#ifdef Py_GIL_DISABLED +#define FT_STAT_MUTEX_SLEEP_INC() _Py_STATS_EXPR(ft_stats.mutex_sleeps++) +#define FT_STAT_QSBR_POLL_INC() _Py_STATS_EXPR(ft_stats.qsbr_polls++) +#define FT_STAT_WORLD_STOP_INC() _Py_STATS_EXPR(ft_stats.world_stops++) +#else +#define FT_STAT_MUTEX_SLEEP_INC() +#define FT_STAT_QSBR_POLL_INC() +#define FT_STAT_WORLD_STOP_INC() +#endif // Export for '_opcode' shared extension PyAPI_FUNC(PyObject*) _Py_GetSpecializationStats(void); @@ -71,6 +88,9 @@ PyAPI_FUNC(PyObject*) _Py_GetSpecializationStats(void); #define OPT_HIST(length, name) ((void)0) #define RARE_EVENT_STAT_INC(name) ((void)0) #define OPCODE_DEFERRED_INC(opname) ((void)0) +#define FT_STAT_MUTEX_SLEEP_INC() +#define FT_STAT_QSBR_POLL_INC() +#define FT_STAT_WORLD_STOP_INC() #endif // !Py_STATS @@ -90,6 +110,11 @@ PyAPI_FUNC(PyObject*) _Py_GetSpecializationStats(void); RARE_EVENT_INTERP_INC(interp, name); \ } while (0); \ +PyStatus _PyStats_InterpInit(PyInterpreterState *); +bool _PyStats_ThreadInit(PyInterpreterState *, _PyThreadStateImpl *); +void _PyStats_ThreadFini(_PyThreadStateImpl *); +void _PyStats_Attach(_PyThreadStateImpl *); +void _PyStats_Detach(_PyThreadStateImpl *); #ifdef __cplusplus } diff --git a/Include/internal/pycore_tstate.h b/Include/internal/pycore_tstate.h index bad968428c73..29ebdfd7e016 100644 --- a/Include/internal/pycore_tstate.h +++ b/Include/internal/pycore_tstate.h @@ -70,8 +70,14 @@ typedef struct _PyThreadStateImpl { // When >1, code objects do not immortalize their non-string constants. int suppress_co_const_immortalization; + +#ifdef Py_STATS + // per-thread stats, will be merged into interp->pystats_struct + PyStats *pystats_struct; // allocated by _PyStats_ThreadInit() #endif +#endif // Py_GIL_DISABLED + #if defined(Py_REF_DEBUG) && defined(Py_GIL_DISABLED) Py_ssize_t reftotal; // this thread's total refcount operations #endif diff --git a/Lib/test/test_pystats.py b/Lib/test/test_pystats.py new file mode 100644 index 000000000000..c50cecfcfddf --- /dev/null +++ b/Lib/test/test_pystats.py @@ -0,0 +1,215 @@ +import sys +import textwrap +import unittest +from test.support import script_helper + +# This function is available for the --enable-pystats config. +HAVE_PYSTATS = hasattr(sys, '_stats_on') + +TEST_TEMPLATE = """ + import sys + import threading + import time + + THREADS = 2 + + class A: + pass + + class B: + pass + + def modify_class(): + # This is used as a rare event we can assume doesn't happen unless we do it. + # It increments the "Rare event (set_class)" count. + a = A() + a.__class__ = B + + TURNED_ON = False + def stats_on(): + global TURNED_ON + sys._stats_on() + TURNED_ON = True + + TURNED_OFF = False + def stats_off(): + global TURNED_OFF + sys._stats_off() + TURNED_OFF = True + + CLEARED = False + def stats_clear(): + global CLEARED + sys._stats_clear() + CLEARED = True + + def func_start(): + pass + + def func_end(): + pass + + def func_test(thread_id): + pass + + _TEST_CODE_ + + func_start() + threads = [] + for i in range(THREADS): + t = threading.Thread(target=func_test, args=(i,)) + threads.append(t) + t.start() + for t in threads: + t.join() + func_end() + """ + + +def run_test_code( + test_code, + args=[], + env_vars=None, +): + """Run test code and return the value of the "set_class" stats counter. + """ + code = textwrap.dedent(TEST_TEMPLATE) + code = code.replace('_TEST_CODE_', textwrap.dedent(test_code)) + script_args = args + ['-c', code] + env_vars = env_vars or {} + res, _ = script_helper.run_python_until_end(*script_args, **env_vars) + stderr = res.err.decode("ascii", "backslashreplace") + for line in stderr.split('\n'): + if 'Rare event (set_class)' in line: + label, _, value = line.partition(':') + return value.strip() + return '' + + +@unittest.skipUnless(HAVE_PYSTATS, "requires pystats build option") +class TestPyStats(unittest.TestCase): + """Tests for pystats functionality (requires --enable-pystats build + option). + """ + + def test_stats_toggle_on(self): + """Check the toggle on functionality. + """ + code = """ + def func_start(): + modify_class() + """ + + # If turned on with command line flag, should get one count. + stat_count = run_test_code(code, args=['-X', 'pystats']) + self.assertEqual(stat_count, '1') + + # If turned on with env var, should get one count. + stat_count = run_test_code(code, env_vars={'PYTHONSTATS': '1'}) + self.assertEqual(stat_count, '1') + + # If not turned on, should be no counts. + stat_count = run_test_code(code) + self.assertEqual(stat_count, '') + + code = """ + def func_start(): + modify_class() + sys._stats_on() + modify_class() + """ + # Not initially turned on but enabled by sys._stats_on(), should get + # one count. + stat_count = run_test_code(code) + self.assertEqual(stat_count, '1') + + def test_stats_toggle_on_thread(self): + """Check the toggle on functionality when threads are used. + """ + code = """ + def func_test(thread_id): + if thread_id == 0: + modify_class() + stats_on() + modify_class() + else: + while not TURNED_ON: + pass + modify_class() + """ + # Turning on in one thread will count in other thread. + stat_count = run_test_code(code) + self.assertEqual(stat_count, '2') + + code = """ + def func_test(thread_id): + if thread_id == 0: + modify_class() + stats_off() + modify_class() + else: + while not TURNED_OFF: + pass + modify_class() + """ + # Turning off in one thread will not count in other threads. + stat_count = run_test_code(code, args=['-X', 'pystats']) + self.assertEqual(stat_count, '1') + + def test_thread_exit_merge(self): + """Check that per-thread stats (when free-threading enabled) are merged. + """ + code = """ + def func_test(thread_id): + modify_class() + if thread_id == 0: + raise SystemExit + """ + # Stats from a thread exiting early should still be counted. + stat_count = run_test_code(code, args=['-X', 'pystats']) + self.assertEqual(stat_count, '2') + + def test_stats_dump(self): + """Check that sys._stats_dump() works. + """ + code = """ + def func_test(thread_id): + if thread_id == 0: + stats_on() + else: + while not TURNED_ON: + pass + modify_class() + sys._stats_dump() + stats_off() + """ + # Stats from a thread exiting early should still be counted. + stat_count = run_test_code(code) + self.assertEqual(stat_count, '1') + + def test_stats_clear(self): + """Check that sys._stats_clear() works. + """ + code = """ + ready = False + def func_test(thread_id): + global ready + if thread_id == 0: + stats_on() + modify_class() + while not ready: + pass # wait until other thread has called modify_class() + stats_clear() # clears stats for all threads + else: + while not TURNED_ON: + pass + modify_class() + ready = True + """ + # Clearing stats will clear for all threads + stat_count = run_test_code(code) + self.assertEqual(stat_count, '0') + + +if __name__ == "__main__": + unittest.main() diff --git a/Makefile.pre.in b/Makefile.pre.in index 656d9dacd962..dd28ff5d2a3e 100644 --- a/Makefile.pre.in +++ b/Makefile.pre.in @@ -483,6 +483,7 @@ PYTHON_OBJS= \ Python/pylifecycle.o \ Python/pymath.o \ Python/pystate.o \ + Python/pystats.o \ Python/pythonrun.o \ Python/pytime.o \ Python/qsbr.o \ diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2025-07-29-17-51-14.gh-issue-131253.GpRjWy.rst b/Misc/NEWS.d/next/Core_and_Builtins/2025-07-29-17-51-14.gh-issue-131253.GpRjWy.rst new file mode 100644 index 000000000000..2826fad23305 --- /dev/null +++ b/Misc/NEWS.d/next/Core_and_Builtins/2025-07-29-17-51-14.gh-issue-131253.GpRjWy.rst @@ -0,0 +1 @@ +Support the ``--enable-pystats`` build option for the free-threaded build. diff --git a/Modules/_xxtestfuzz/fuzzer.c b/Modules/_xxtestfuzz/fuzzer.c index a04f1412eefd..0cbe10c79ab4 100644 --- a/Modules/_xxtestfuzz/fuzzer.c +++ b/Modules/_xxtestfuzz/fuzzer.c @@ -10,8 +10,8 @@ See the source code for LLVMFuzzerTestOneInput for details. */ -#ifndef Py_BUILD_CORE -# define Py_BUILD_CORE 1 +#ifndef Py_BUILD_CORE_MODULE +# define Py_BUILD_CORE_MODULE 1 #endif #include diff --git a/PCbuild/_freeze_module.vcxproj b/PCbuild/_freeze_module.vcxproj index e65f201623fb..605861ad3fd0 100644 --- a/PCbuild/_freeze_module.vcxproj +++ b/PCbuild/_freeze_module.vcxproj @@ -257,6 +257,7 @@ + diff --git a/PCbuild/_freeze_module.vcxproj.filters b/PCbuild/_freeze_module.vcxproj.filters index a9fb6f2328ad..c67fe53363ee 100644 --- a/PCbuild/_freeze_module.vcxproj.filters +++ b/PCbuild/_freeze_module.vcxproj.filters @@ -367,6 +367,9 @@ Source Files + + Source Files + Source Files diff --git a/PCbuild/pythoncore.vcxproj b/PCbuild/pythoncore.vcxproj index a50ffb120bc0..359a47fbfc4f 100644 --- a/PCbuild/pythoncore.vcxproj +++ b/PCbuild/pythoncore.vcxproj @@ -658,6 +658,7 @@ + diff --git a/Python/ceval_macros.h b/Python/ceval_macros.h index 868ab6f75587..afdcbc563b2c 100644 --- a/Python/ceval_macros.h +++ b/Python/ceval_macros.h @@ -62,8 +62,9 @@ #ifdef Py_STATS #define INSTRUCTION_STATS(op) \ do { \ + PyStats *s = _PyStats_GET(); \ OPCODE_EXE_INC(op); \ - if (_Py_stats) _Py_stats->opcode_stats[lastopcode].pair_count[op]++; \ + if (s) s->opcode_stats[lastopcode].pair_count[op]++; \ lastopcode = op; \ } while (0) #else diff --git a/Python/gc.c b/Python/gc.c index a1f3d86d9103..03a5d7366ea6 100644 --- a/Python/gc.c +++ b/Python/gc.c @@ -2111,10 +2111,11 @@ _PyGC_Collect(PyThreadState *tstate, int generation, _PyGC_Reason reason) _PyErr_SetRaisedException(tstate, exc); GC_STAT_ADD(generation, objects_collected, stats.collected); #ifdef Py_STATS - if (_Py_stats) { + PyStats *s = _PyStats_GET(); + if (s) { GC_STAT_ADD(generation, object_visits, - _Py_stats->object_stats.object_visits); - _Py_stats->object_stats.object_visits = 0; + s->object_stats.object_visits); + s->object_stats.object_visits = 0; } #endif validate_spaces(gcstate); diff --git a/Python/gc_free_threading.c b/Python/gc_free_threading.c index 842aa3401548..f39793c3eeb5 100644 --- a/Python/gc_free_threading.c +++ b/Python/gc_free_threading.c @@ -2362,8 +2362,9 @@ gc_collect_main(PyThreadState *tstate, int generation, _PyGC_Reason reason) assert(generation >= 0 && generation < NUM_GENERATIONS); #ifdef Py_STATS - if (_Py_stats) { - _Py_stats->object_stats.object_visits = 0; + PyStats *s = _PyStats_GET(); + if (s) { + s->object_stats.object_visits = 0; } #endif GC_STAT_ADD(generation, collections, 1); @@ -2426,10 +2427,13 @@ gc_collect_main(PyThreadState *tstate, int generation, _PyGC_Reason reason) GC_STAT_ADD(generation, objects_collected, m); #ifdef Py_STATS - if (_Py_stats) { - GC_STAT_ADD(generation, object_visits, - _Py_stats->object_stats.object_visits); - _Py_stats->object_stats.object_visits = 0; + { + PyStats *s = _PyStats_GET(); + if (s) { + GC_STAT_ADD(generation, object_visits, + s->object_stats.object_visits); + s->object_stats.object_visits = 0; + } } #endif diff --git a/Python/initconfig.c b/Python/initconfig.c index 5dc68eb4ec2c..7176670c110d 100644 --- a/Python/initconfig.c +++ b/Python/initconfig.c @@ -2810,12 +2810,6 @@ _PyConfig_Write(const PyConfig *config, _PyRuntimeState *runtime) return _PyStatus_NO_MEMORY(); } -#ifdef Py_STATS - if (config->_pystats) { - _Py_StatsOn(); - } -#endif - return _PyStatus_OK(); } diff --git a/Python/lock.c b/Python/lock.c index 98f3f89c201f..789065d81627 100644 --- a/Python/lock.c +++ b/Python/lock.c @@ -6,6 +6,7 @@ #include "pycore_parking_lot.h" #include "pycore_semaphore.h" #include "pycore_time.h" // _PyTime_Add() +#include "pycore_stats.h" // FT_STAT_MUTEX_SLEEP_INC() #ifdef MS_WINDOWS # ifndef WIN32_LEAN_AND_MEAN @@ -62,6 +63,8 @@ _PyMutex_LockTimed(PyMutex *m, PyTime_t timeout, _PyLockFlags flags) return PY_LOCK_FAILURE; } + FT_STAT_MUTEX_SLEEP_INC(); + PyTime_t now; // silently ignore error: cannot report error to the caller (void)PyTime_MonotonicRaw(&now); diff --git a/Python/pylifecycle.c b/Python/pylifecycle.c index 8fcb31cfd122..805805ef188e 100644 --- a/Python/pylifecycle.c +++ b/Python/pylifecycle.c @@ -26,6 +26,7 @@ #include "pycore_runtime.h" // _Py_ID() #include "pycore_runtime_init.h" // _PyRuntimeState_INIT #include "pycore_setobject.h" // _PySet_NextEntry() +#include "pycore_stats.h" // _PyStats_InterpInit() #include "pycore_sysmodule.h" // _PySys_ClearAttrString() #include "pycore_traceback.h" // _Py_DumpTracebackThreads() #include "pycore_typeobject.h" // _PyTypes_InitTypes() @@ -656,6 +657,14 @@ pycore_create_interpreter(_PyRuntimeState *runtime, return status; } +#ifdef Py_STATS + // initialize pystats. This must be done after the settings are loaded. + status = _PyStats_InterpInit(interp); + if (_PyStatus_EXCEPTION(status)) { + return status; + } +#endif + // initialize the interp->obmalloc state. This must be done after // the settings are loaded (so that feature_flags are set) but before // any calls are made to obmalloc functions. @@ -2469,6 +2478,14 @@ new_interpreter(PyThreadState **tstate_p, return status; } +#ifdef Py_STATS + // initialize pystats. This must be done after the settings are loaded. + status = _PyStats_InterpInit(interp); + if (_PyStatus_EXCEPTION(status)) { + return status; + } +#endif + // initialize the interp->obmalloc state. This must be done after // the settings are loaded (so that feature_flags are set) but before // any calls are made to obmalloc functions. diff --git a/Python/pystate.c b/Python/pystate.c index 24681536797f..cf251c120d75 100644 --- a/Python/pystate.c +++ b/Python/pystate.c @@ -21,6 +21,7 @@ #include "pycore_runtime.h" // _PyRuntime #include "pycore_runtime_init.h" // _PyRuntimeState_INIT #include "pycore_stackref.h" // Py_STACKREF_DEBUG +#include "pycore_stats.h" // FT_STAT_WORLD_STOP_INC() #include "pycore_time.h" // _PyTime_Init() #include "pycore_uop.h" // UOP_BUFFER_SIZE #include "pycore_uniqueid.h" // _PyObject_FinalizePerThreadRefcounts() @@ -465,6 +466,12 @@ alloc_interpreter(void) static void free_interpreter(PyInterpreterState *interp) { +#ifdef Py_STATS + if (interp->pystats_struct) { + PyMem_RawFree(interp->pystats_struct); + interp->pystats_struct = NULL; + } +#endif // The main interpreter is statically allocated so // should not be freed. if (interp != &_PyRuntime._main_interpreter) { @@ -1407,6 +1414,9 @@ static void free_threadstate(_PyThreadStateImpl *tstate) { PyInterpreterState *interp = tstate->base.interp; +#ifdef Py_STATS + _PyStats_ThreadFini(tstate); +#endif // The initial thread state of the interpreter is allocated // as part of the interpreter state so should not be freed. if (tstate == &interp->_initial_thread) { @@ -1535,6 +1545,13 @@ new_threadstate(PyInterpreterState *interp, int whence) return NULL; } #endif +#ifdef Py_STATS + // The PyStats structure is quite large and is allocated separated from tstate. + if (!_PyStats_ThreadInit(interp, tstate)) { + free_threadstate(tstate); + return NULL; + } +#endif /* We serialize concurrent creation to protect global state. */ HEAD_LOCK(interp->runtime); @@ -1846,6 +1863,9 @@ _PyThreadState_DeleteCurrent(PyThreadState *tstate) _Py_EnsureTstateNotNULL(tstate); #ifdef Py_GIL_DISABLED _Py_qsbr_detach(((_PyThreadStateImpl *)tstate)->qsbr); +#endif +#ifdef Py_STATS + _PyStats_Detach((_PyThreadStateImpl *)tstate); #endif current_fast_clear(tstate->interp->runtime); tstate_delete_common(tstate, 1); // release GIL as part of call @@ -2020,6 +2040,10 @@ tstate_deactivate(PyThreadState *tstate) assert(tstate_is_bound(tstate)); assert(tstate->_status.active); +#if Py_STATS + _PyStats_Detach((_PyThreadStateImpl *)tstate); +#endif + tstate->_status.active = 0; // We do not unbind the gilstate tstate here. @@ -2123,6 +2147,10 @@ _PyThreadState_Attach(PyThreadState *tstate) _PyCriticalSection_Resume(tstate); } +#ifdef Py_STATS + _PyStats_Attach((_PyThreadStateImpl *)tstate); +#endif + #if defined(Py_DEBUG) errno = err; #endif @@ -2272,6 +2300,7 @@ stop_the_world(struct _stoptheworld_state *stw) stw->thread_countdown = 0; stw->stop_event = (PyEvent){0}; // zero-initialize (unset) stw->requester = _PyThreadState_GET(); // may be NULL + FT_STAT_WORLD_STOP_INC(); _Py_FOR_EACH_STW_INTERP(stw, i) { _Py_FOR_EACH_TSTATE_UNLOCKED(i, t) { diff --git a/Python/pystats.c b/Python/pystats.c new file mode 100644 index 000000000000..2e377b8e6da3 --- /dev/null +++ b/Python/pystats.c @@ -0,0 +1,819 @@ +#include "Python.h" + +#include "pycore_opcode_metadata.h" // _PyOpcode_Caches +#include "pycore_pyatomic_ft_wrappers.h" +#include "pycore_pylifecycle.h" // _PyOS_URandomNonblock() +#include "pycore_tstate.h" +#include "pycore_initconfig.h" // _PyStatus_OK() +#include "pycore_uop_metadata.h" // _PyOpcode_uop_name +#include "pycore_uop_ids.h" // MAX_UOP_ID +#include "pycore_pystate.h" // _PyThreadState_GET() +#include "pycore_runtime.h" // NUM_GENERATIONS + +#include // rand() + +#ifdef Py_STATS + +PyStats * +_PyStats_GetLocal(void) +{ + PyThreadState *tstate = _PyThreadState_GET(); + if (tstate) { + return tstate->pystats; + } + return NULL; +} + +#ifdef Py_GIL_DISABLED +#define STATS_LOCK(interp) PyMutex_Lock(&interp->pystats_mutex) +#define STATS_UNLOCK(interp) PyMutex_Unlock(&interp->pystats_mutex) +#else +#define STATS_LOCK(interp) +#define STATS_UNLOCK(interp) +#endif + + +#if PYSTATS_MAX_UOP_ID < MAX_UOP_ID +#error "Not enough space allocated for pystats. Increase PYSTATS_MAX_UOP_ID to at least MAX_UOP_ID" +#endif + +#define ADD_STAT_TO_DICT(res, field) \ + do { \ + PyObject *val = PyLong_FromUnsignedLongLong(stats->field); \ + if (val == NULL) { \ + Py_DECREF(res); \ + return NULL; \ + } \ + if (PyDict_SetItemString(res, #field, val) == -1) { \ + Py_DECREF(res); \ + Py_DECREF(val); \ + return NULL; \ + } \ + Py_DECREF(val); \ + } while(0); + +static PyObject* +stats_to_dict(SpecializationStats *stats) +{ + PyObject *res = PyDict_New(); + if (res == NULL) { + return NULL; + } + ADD_STAT_TO_DICT(res, success); + ADD_STAT_TO_DICT(res, failure); + ADD_STAT_TO_DICT(res, hit); + ADD_STAT_TO_DICT(res, deferred); + ADD_STAT_TO_DICT(res, miss); + ADD_STAT_TO_DICT(res, deopt); + PyObject *failure_kinds = PyTuple_New(SPECIALIZATION_FAILURE_KINDS); + if (failure_kinds == NULL) { + Py_DECREF(res); + return NULL; + } + for (int i = 0; i < SPECIALIZATION_FAILURE_KINDS; i++) { + PyObject *stat = PyLong_FromUnsignedLongLong(stats->failure_kinds[i]); + if (stat == NULL) { + Py_DECREF(res); + Py_DECREF(failure_kinds); + return NULL; + } + PyTuple_SET_ITEM(failure_kinds, i, stat); + } + if (PyDict_SetItemString(res, "failure_kinds", failure_kinds)) { + Py_DECREF(res); + Py_DECREF(failure_kinds); + return NULL; + } + Py_DECREF(failure_kinds); + return res; +} +#undef ADD_STAT_TO_DICT + +static int +add_stat_dict( + PyStats *src, + PyObject *res, + int opcode, + const char *name) { + + SpecializationStats *stats = &src->opcode_stats[opcode].specialization; + PyObject *d = stats_to_dict(stats); + if (d == NULL) { + return -1; + } + int err = PyDict_SetItemString(res, name, d); + Py_DECREF(d); + return err; +} + +PyObject* +_Py_GetSpecializationStats(void) { + PyThreadState *tstate = _PyThreadState_GET(); + PyStats *src = FT_ATOMIC_LOAD_PTR_RELAXED(tstate->interp->pystats_struct); + if (src == NULL) { + Py_RETURN_NONE; + } + PyObject *stats = PyDict_New(); + if (stats == NULL) { + return NULL; + } + int err = 0; + err += add_stat_dict(src, stats, CONTAINS_OP, "contains_op"); + err += add_stat_dict(src, stats, LOAD_SUPER_ATTR, "load_super_attr"); + err += add_stat_dict(src, stats, LOAD_ATTR, "load_attr"); + err += add_stat_dict(src, stats, LOAD_GLOBAL, "load_global"); + err += add_stat_dict(src, stats, STORE_SUBSCR, "store_subscr"); + err += add_stat_dict(src, stats, STORE_ATTR, "store_attr"); + err += add_stat_dict(src, stats, JUMP_BACKWARD, "jump_backward"); + err += add_stat_dict(src, stats, CALL, "call"); + err += add_stat_dict(src, stats, CALL_KW, "call_kw"); + err += add_stat_dict(src, stats, BINARY_OP, "binary_op"); + err += add_stat_dict(src, stats, COMPARE_OP, "compare_op"); + err += add_stat_dict(src, stats, UNPACK_SEQUENCE, "unpack_sequence"); + err += add_stat_dict(src, stats, FOR_ITER, "for_iter"); + err += add_stat_dict(src, stats, TO_BOOL, "to_bool"); + err += add_stat_dict(src, stats, SEND, "send"); + if (err < 0) { + Py_DECREF(stats); + return NULL; + } + return stats; +} + + +#define PRINT_STAT(i, field) \ + if (stats[i].field) { \ + fprintf(out, " opcode[%s]." #field " : %" PRIu64 "\n", _PyOpcode_OpName[i], stats[i].field); \ + } + +static void +print_spec_stats(FILE *out, OpcodeStats *stats) +{ + /* Mark some opcodes as specializable for stats, + * even though we don't specialize them yet. */ + fprintf(out, "opcode[BINARY_SLICE].specializable : 1\n"); + fprintf(out, "opcode[STORE_SLICE].specializable : 1\n"); + fprintf(out, "opcode[GET_ITER].specializable : 1\n"); + for (int i = 0; i < 256; i++) { + if (_PyOpcode_Caches[i]) { + /* Ignore jumps as they cannot be specialized */ + switch (i) { + case POP_JUMP_IF_FALSE: + case POP_JUMP_IF_TRUE: + case POP_JUMP_IF_NONE: + case POP_JUMP_IF_NOT_NONE: + case JUMP_BACKWARD: + break; + default: + fprintf(out, "opcode[%s].specializable : 1\n", _PyOpcode_OpName[i]); + } + } + PRINT_STAT(i, specialization.success); + PRINT_STAT(i, specialization.failure); + PRINT_STAT(i, specialization.hit); + PRINT_STAT(i, specialization.deferred); + PRINT_STAT(i, specialization.miss); + PRINT_STAT(i, specialization.deopt); + PRINT_STAT(i, execution_count); + for (int j = 0; j < SPECIALIZATION_FAILURE_KINDS; j++) { + uint64_t val = stats[i].specialization.failure_kinds[j]; + if (val) { + fprintf(out, " opcode[%s].specialization.failure_kinds[%d] : %" + PRIu64 "\n", _PyOpcode_OpName[i], j, val); + } + } + for (int j = 0; j < 256; j++) { + if (stats[i].pair_count[j]) { + fprintf(out, "opcode[%s].pair_count[%s] : %" PRIu64 "\n", + _PyOpcode_OpName[i], _PyOpcode_OpName[j], stats[i].pair_count[j]); + } + } + } +} +#undef PRINT_STAT + + +static void +print_call_stats(FILE *out, CallStats *stats) +{ + fprintf(out, "Calls to PyEval_EvalDefault: %" PRIu64 "\n", stats->pyeval_calls); + fprintf(out, "Calls to Python functions inlined: %" PRIu64 "\n", stats->inlined_py_calls); + fprintf(out, "Frames pushed: %" PRIu64 "\n", stats->frames_pushed); + fprintf(out, "Frame objects created: %" PRIu64 "\n", stats->frame_objects_created); + for (int i = 0; i < EVAL_CALL_KINDS; i++) { + fprintf(out, "Calls via PyEval_EvalFrame[%d] : %" PRIu64 "\n", i, stats->eval_calls[i]); + } +} + +static void +print_object_stats(FILE *out, ObjectStats *stats) +{ + fprintf(out, "Object allocations from freelist: %" PRIu64 "\n", stats->from_freelist); + fprintf(out, "Object frees to freelist: %" PRIu64 "\n", stats->to_freelist); + fprintf(out, "Object allocations: %" PRIu64 "\n", stats->allocations); + fprintf(out, "Object allocations to 512 bytes: %" PRIu64 "\n", stats->allocations512); + fprintf(out, "Object allocations to 4 kbytes: %" PRIu64 "\n", stats->allocations4k); + fprintf(out, "Object allocations over 4 kbytes: %" PRIu64 "\n", stats->allocations_big); + fprintf(out, "Object frees: %" PRIu64 "\n", stats->frees); + fprintf(out, "Object inline values: %" PRIu64 "\n", stats->inline_values); + fprintf(out, "Object interpreter mortal increfs: %" PRIu64 "\n", stats->interpreter_increfs); + fprintf(out, "Object interpreter mortal decrefs: %" PRIu64 "\n", stats->interpreter_decrefs); + fprintf(out, "Object mortal increfs: %" PRIu64 "\n", stats->increfs); + fprintf(out, "Object mortal decrefs: %" PRIu64 "\n", stats->decrefs); + fprintf(out, "Object interpreter immortal increfs: %" PRIu64 "\n", stats->interpreter_immortal_increfs); + fprintf(out, "Object interpreter immortal decrefs: %" PRIu64 "\n", stats->interpreter_immortal_decrefs); + fprintf(out, "Object immortal increfs: %" PRIu64 "\n", stats->immortal_increfs); + fprintf(out, "Object immortal decrefs: %" PRIu64 "\n", stats->immortal_decrefs); + fprintf(out, "Object materialize dict (on request): %" PRIu64 "\n", stats->dict_materialized_on_request); + fprintf(out, "Object materialize dict (new key): %" PRIu64 "\n", stats->dict_materialized_new_key); + fprintf(out, "Object materialize dict (too big): %" PRIu64 "\n", stats->dict_materialized_too_big); + fprintf(out, "Object materialize dict (str subclass): %" PRIu64 "\n", stats->dict_materialized_str_subclass); + fprintf(out, "Object method cache hits: %" PRIu64 "\n", stats->type_cache_hits); + fprintf(out, "Object method cache misses: %" PRIu64 "\n", stats->type_cache_misses); + fprintf(out, "Object method cache collisions: %" PRIu64 "\n", stats->type_cache_collisions); + fprintf(out, "Object method cache dunder hits: %" PRIu64 "\n", stats->type_cache_dunder_hits); + fprintf(out, "Object method cache dunder misses: %" PRIu64 "\n", stats->type_cache_dunder_misses); +} + +static void +print_gc_stats(FILE *out, GCStats *stats) +{ + for (int i = 0; i < NUM_GENERATIONS; i++) { + fprintf(out, "GC[%d] collections: %" PRIu64 "\n", i, stats[i].collections); + fprintf(out, "GC[%d] object visits: %" PRIu64 "\n", i, stats[i].object_visits); + fprintf(out, "GC[%d] objects collected: %" PRIu64 "\n", i, stats[i].objects_collected); + fprintf(out, "GC[%d] objects reachable from roots: %" PRIu64 "\n", i, stats[i].objects_transitively_reachable); + fprintf(out, "GC[%d] objects not reachable from roots: %" PRIu64 "\n", i, stats[i].objects_not_transitively_reachable); + } +} + +#ifdef _Py_TIER2 +static void +print_histogram(FILE *out, const char *name, uint64_t hist[_Py_UOP_HIST_SIZE]) +{ + for (int i = 0; i < _Py_UOP_HIST_SIZE; i++) { + fprintf(out, "%s[%" PRIu64"]: %" PRIu64 "\n", name, (uint64_t)1 << i, hist[i]); + } +} + +extern const char *_PyUOpName(int index); + +static void +print_optimization_stats(FILE *out, OptimizationStats *stats) +{ + fprintf(out, "Optimization attempts: %" PRIu64 "\n", stats->attempts); + fprintf(out, "Optimization traces created: %" PRIu64 "\n", stats->traces_created); + fprintf(out, "Optimization traces executed: %" PRIu64 "\n", stats->traces_executed); + fprintf(out, "Optimization uops executed: %" PRIu64 "\n", stats->uops_executed); + fprintf(out, "Optimization trace stack overflow: %" PRIu64 "\n", stats->trace_stack_overflow); + fprintf(out, "Optimization trace stack underflow: %" PRIu64 "\n", stats->trace_stack_underflow); + fprintf(out, "Optimization trace too long: %" PRIu64 "\n", stats->trace_too_long); + fprintf(out, "Optimization trace too short: %" PRIu64 "\n", stats->trace_too_short); + fprintf(out, "Optimization inner loop: %" PRIu64 "\n", stats->inner_loop); + fprintf(out, "Optimization recursive call: %" PRIu64 "\n", stats->recursive_call); + fprintf(out, "Optimization low confidence: %" PRIu64 "\n", stats->low_confidence); + fprintf(out, "Optimization unknown callee: %" PRIu64 "\n", stats->unknown_callee); + fprintf(out, "Executors invalidated: %" PRIu64 "\n", stats->executors_invalidated); + + print_histogram(out, "Trace length", stats->trace_length_hist); + print_histogram(out, "Trace run length", stats->trace_run_length_hist); + print_histogram(out, "Optimized trace length", stats->optimized_trace_length_hist); + + fprintf(out, "Optimization optimizer attempts: %" PRIu64 "\n", stats->optimizer_attempts); + fprintf(out, "Optimization optimizer successes: %" PRIu64 "\n", stats->optimizer_successes); + fprintf(out, "Optimization optimizer failure no memory: %" PRIu64 "\n", + stats->optimizer_failure_reason_no_memory); + fprintf(out, "Optimizer remove globals builtins changed: %" PRIu64 "\n", stats->remove_globals_builtins_changed); + fprintf(out, "Optimizer remove globals incorrect keys: %" PRIu64 "\n", stats->remove_globals_incorrect_keys); + for (int i = 0; i <= MAX_UOP_ID; i++) { + if (stats->opcode[i].execution_count) { + fprintf(out, "uops[%s].execution_count : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].execution_count); + } + if (stats->opcode[i].miss) { + fprintf(out, "uops[%s].specialization.miss : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].miss); + } + } + for (int i = 0; i < 256; i++) { + if (stats->unsupported_opcode[i]) { + fprintf( + out, + "unsupported_opcode[%s].count : %" PRIu64 "\n", + _PyOpcode_OpName[i], + stats->unsupported_opcode[i] + ); + } + } + + for (int i = 1; i <= MAX_UOP_ID; i++){ + for (int j = 1; j <= MAX_UOP_ID; j++) { + if (stats->opcode[i].pair_count[j]) { + fprintf(out, "uop[%s].pair_count[%s] : %" PRIu64 "\n", + _PyOpcode_uop_name[i], _PyOpcode_uop_name[j], stats->opcode[i].pair_count[j]); + } + } + } + for (int i = 0; i < MAX_UOP_ID; i++) { + if (stats->error_in_opcode[i]) { + fprintf( + out, + "error_in_opcode[%s].count : %" PRIu64 "\n", + _PyUOpName(i), + stats->error_in_opcode[i] + ); + } + } + fprintf(out, "JIT total memory size: %" PRIu64 "\n", stats->jit_total_memory_size); + fprintf(out, "JIT code size: %" PRIu64 "\n", stats->jit_code_size); + fprintf(out, "JIT trampoline size: %" PRIu64 "\n", stats->jit_trampoline_size); + fprintf(out, "JIT data size: %" PRIu64 "\n", stats->jit_data_size); + fprintf(out, "JIT padding size: %" PRIu64 "\n", stats->jit_padding_size); + fprintf(out, "JIT freed memory size: %" PRIu64 "\n", stats->jit_freed_memory_size); + + print_histogram(out, "Trace total memory size", stats->trace_total_memory_hist); +} +#endif + +#ifdef Py_GIL_DISABLED +static void +print_ft_stats(FILE *out, FTStats *stats) +{ + fprintf(out, "Mutex sleeps (mutex_sleeps): %" PRIu64 "\n", stats->mutex_sleeps); + fprintf(out, "QSBR polls (qsbr_polls): %" PRIu64 "\n", stats->qsbr_polls); + fprintf(out, "World stops (world_stops): %" PRIu64 "\n", stats->world_stops); +} +#endif + +static void +print_rare_event_stats(FILE *out, RareEventStats *stats) +{ + fprintf(out, "Rare event (set_class): %" PRIu64 "\n", stats->set_class); + fprintf(out, "Rare event (set_bases): %" PRIu64 "\n", stats->set_bases); + fprintf(out, "Rare event (set_eval_frame_func): %" PRIu64 "\n", stats->set_eval_frame_func); + fprintf(out, "Rare event (builtin_dict): %" PRIu64 "\n", stats->builtin_dict); + fprintf(out, "Rare event (func_modification): %" PRIu64 "\n", stats->func_modification); + fprintf(out, "Rare event (watched_dict_modification): %" PRIu64 "\n", stats->watched_dict_modification); + fprintf(out, "Rare event (watched_globals_modification): %" PRIu64 "\n", stats->watched_globals_modification); +} + +static void +print_stats(FILE *out, PyStats *stats) +{ + print_spec_stats(out, stats->opcode_stats); + print_call_stats(out, &stats->call_stats); + print_object_stats(out, &stats->object_stats); + print_gc_stats(out, stats->gc_stats); +#ifdef _Py_TIER2 + print_optimization_stats(out, &stats->optimization_stats); +#endif +#ifdef Py_GIL_DISABLED + print_ft_stats(out, &stats->ft_stats); +#endif + print_rare_event_stats(out, &stats->rare_event_stats); +} + +#ifdef Py_GIL_DISABLED + +static void +merge_specialization_stats(SpecializationStats *dest, const SpecializationStats *src) +{ + dest->success += src->success; + dest->failure += src->failure; + dest->hit += src->hit; + dest->deferred += src->deferred; + dest->miss += src->miss; + dest->deopt += src->deopt; + for (int i = 0; i < SPECIALIZATION_FAILURE_KINDS; i++) { + dest->failure_kinds[i] += src->failure_kinds[i]; + } +} + +static void +merge_opcode_stats_array(OpcodeStats *dest, const OpcodeStats *src) +{ + for (int i = 0; i < 256; i++) { + merge_specialization_stats(&dest[i].specialization, &src[i].specialization); + dest[i].execution_count += src[i].execution_count; + for (int j = 0; j < 256; j++) { + dest[i].pair_count[j] += src[i].pair_count[j]; + } + } +} + +static void +merge_call_stats(CallStats *dest, const CallStats *src) +{ + dest->inlined_py_calls += src->inlined_py_calls; + dest->pyeval_calls += src->pyeval_calls; + dest->frames_pushed += src->frames_pushed; + dest->frame_objects_created += src->frame_objects_created; + for (int i = 0; i < EVAL_CALL_KINDS; i++) { + dest->eval_calls[i] += src->eval_calls[i]; + } +} + +static void +merge_object_stats(ObjectStats *dest, const ObjectStats *src) +{ + dest->increfs += src->increfs; + dest->decrefs += src->decrefs; + dest->interpreter_increfs += src->interpreter_increfs; + dest->interpreter_decrefs += src->interpreter_decrefs; + dest->immortal_increfs += src->immortal_increfs; + dest->immortal_decrefs += src->immortal_decrefs; + dest->interpreter_immortal_increfs += src->interpreter_immortal_increfs; + dest->interpreter_immortal_decrefs += src->interpreter_immortal_decrefs; + dest->allocations += src->allocations; + dest->allocations512 += src->allocations512; + dest->allocations4k += src->allocations4k; + dest->allocations_big += src->allocations_big; + dest->frees += src->frees; + dest->to_freelist += src->to_freelist; + dest->from_freelist += src->from_freelist; + dest->inline_values += src->inline_values; + dest->dict_materialized_on_request += src->dict_materialized_on_request; + dest->dict_materialized_new_key += src->dict_materialized_new_key; + dest->dict_materialized_too_big += src->dict_materialized_too_big; + dest->dict_materialized_str_subclass += src->dict_materialized_str_subclass; + dest->type_cache_hits += src->type_cache_hits; + dest->type_cache_misses += src->type_cache_misses; + dest->type_cache_dunder_hits += src->type_cache_dunder_hits; + dest->type_cache_dunder_misses += src->type_cache_dunder_misses; + dest->type_cache_collisions += src->type_cache_collisions; + dest->object_visits += src->object_visits; +} + +static void +merge_uop_stats_array(UOpStats *dest, const UOpStats *src) +{ + for (int i = 0; i <= PYSTATS_MAX_UOP_ID; i++) { + dest[i].execution_count += src[i].execution_count; + dest[i].miss += src[i].miss; + for (int j = 0; j <= PYSTATS_MAX_UOP_ID; j++) { + dest[i].pair_count[j] += src[i].pair_count[j]; + } + } +} + +static void +merge_optimization_stats(OptimizationStats *dest, const OptimizationStats *src) +{ + dest->attempts += src->attempts; + dest->traces_created += src->traces_created; + dest->traces_executed += src->traces_executed; + dest->uops_executed += src->uops_executed; + dest->trace_stack_overflow += src->trace_stack_overflow; + dest->trace_stack_underflow += src->trace_stack_underflow; + dest->trace_too_long += src->trace_too_long; + dest->trace_too_short += src->trace_too_short; + dest->inner_loop += src->inner_loop; + dest->recursive_call += src->recursive_call; + dest->low_confidence += src->low_confidence; + dest->unknown_callee += src->unknown_callee; + dest->executors_invalidated += src->executors_invalidated; + dest->optimizer_attempts += src->optimizer_attempts; + dest->optimizer_successes += src->optimizer_successes; + dest->optimizer_failure_reason_no_memory += src->optimizer_failure_reason_no_memory; + dest->remove_globals_builtins_changed += src->remove_globals_builtins_changed; + dest->remove_globals_incorrect_keys += src->remove_globals_incorrect_keys; + dest->jit_total_memory_size += src->jit_total_memory_size; + dest->jit_code_size += src->jit_code_size; + dest->jit_trampoline_size += src->jit_trampoline_size; + dest->jit_data_size += src->jit_data_size; + dest->jit_padding_size += src->jit_padding_size; + dest->jit_freed_memory_size += src->jit_freed_memory_size; + + merge_uop_stats_array(dest->opcode, src->opcode); + + for (int i = 0; i < 256; i++) { + dest->unsupported_opcode[i] += src->unsupported_opcode[i]; + } + for (int i = 0; i < _Py_UOP_HIST_SIZE; i++) { + dest->trace_length_hist[i] += src->trace_length_hist[i]; + dest->trace_run_length_hist[i] += src->trace_run_length_hist[i]; + dest->optimized_trace_length_hist[i] += src->optimized_trace_length_hist[i]; + dest->trace_total_memory_hist[i] += src->trace_total_memory_hist[i]; + } + for (int i = 0; i <= PYSTATS_MAX_UOP_ID; i++) { + dest->error_in_opcode[i] += src->error_in_opcode[i]; + } +} + +static void +merge_ft_stats(FTStats *dest, const FTStats *src) +{ + dest->mutex_sleeps = src->mutex_sleeps; + dest->qsbr_polls = src->qsbr_polls; + dest->world_stops = src->world_stops; +} + +static void +merge_rare_event_stats(RareEventStats *dest, const RareEventStats *src) +{ + dest->set_class += src->set_class; + dest->set_bases += src->set_bases; + dest->set_eval_frame_func += src->set_eval_frame_func; + dest->builtin_dict += src->builtin_dict; + dest->func_modification += src->func_modification; + dest->watched_dict_modification += src->watched_dict_modification; + dest->watched_globals_modification += src->watched_globals_modification; +} + +static void +merge_gc_stats_array(GCStats *dest, const GCStats *src) +{ + for (int i = 0; i < NUM_GENERATIONS; i++) { + dest[i].collections += src[i].collections; + dest[i].object_visits += src[i].object_visits; + dest[i].objects_collected += src[i].objects_collected; + dest[i].objects_transitively_reachable += src[i].objects_transitively_reachable; + dest[i].objects_not_transitively_reachable += src[i].objects_not_transitively_reachable; + } +} + +void +stats_zero_thread(_PyThreadStateImpl *tstate) +{ + // Zero the thread local stat counters + if (tstate->pystats_struct) { + memset(tstate->pystats_struct, 0, sizeof(PyStats)); + } +} + +// merge stats for a single thread into the global structure +void +stats_merge_thread(_PyThreadStateImpl *tstate) +{ + PyStats *src = tstate->pystats_struct; + PyStats *dest = ((PyThreadState *)tstate)->interp->pystats_struct; + + if (src == NULL || dest == NULL) { + return; + } + + // Merge each category of stats using the helper functions. + merge_opcode_stats_array(dest->opcode_stats, src->opcode_stats); + merge_call_stats(&dest->call_stats, &src->call_stats); + merge_object_stats(&dest->object_stats, &src->object_stats); + merge_optimization_stats(&dest->optimization_stats, &src->optimization_stats); + merge_ft_stats(&dest->ft_stats, &src->ft_stats); + merge_rare_event_stats(&dest->rare_event_stats, &src->rare_event_stats); + merge_gc_stats_array(dest->gc_stats, src->gc_stats); +} +#endif // Py_GIL_DISABLED + +// toggle stats collection on or off for all threads +static int +stats_toggle_on_off(PyThreadState *tstate, int on) +{ + bool changed = false; + PyInterpreterState *interp = tstate->interp; + STATS_LOCK(interp); + if (on && interp->pystats_struct == NULL) { + PyStats *s = PyMem_RawCalloc(1, sizeof(PyStats)); + if (s == NULL) { + STATS_UNLOCK(interp); + return -1; + } + FT_ATOMIC_STORE_PTR_RELAXED(interp->pystats_struct, s); + } + if (tstate->interp->pystats_enabled != on) { + FT_ATOMIC_STORE_INT_RELAXED(interp->pystats_enabled, on); + changed = true; + } + STATS_UNLOCK(interp); + if (!changed) { + return 0; + } + _PyEval_StopTheWorld(interp); + _Py_FOR_EACH_TSTATE_UNLOCKED(interp, ts) { + PyStats *s = NULL; + if (interp->pystats_enabled) { +#ifdef Py_GIL_DISABLED + _PyThreadStateImpl *ts_impl = (_PyThreadStateImpl *)ts; + if (ts_impl->pystats_struct == NULL) { + // first activation for this thread, allocate structure + ts_impl->pystats_struct = PyMem_RawCalloc(1, sizeof(PyStats)); + } + s = ts_impl->pystats_struct; +#else + s = ts->interp->pystats_struct; +#endif + } + ts->pystats = s; + } + _PyEval_StartTheWorld(interp); + return 0; +} + +// zero stats for all threads and for the interpreter +static void +stats_zero_all(void) +{ + PyThreadState *tstate = _PyThreadState_GET(); + if (tstate == NULL) { + return; + } + if (FT_ATOMIC_LOAD_PTR_RELAXED(tstate->interp->pystats_struct) == NULL) { + return; + } + PyInterpreterState *interp = tstate->interp; + _PyEval_StopTheWorld(interp); +#ifdef Py_GIL_DISABLED + _Py_FOR_EACH_TSTATE_UNLOCKED(interp, ts) { + stats_zero_thread((_PyThreadStateImpl *)ts); + } +#endif + if (interp->pystats_struct) { + memset(interp->pystats_struct, 0, sizeof(PyStats)); + } + _PyEval_StartTheWorld(interp); +} + +// merge stats for all threads into the per-interpreter structure +static void +stats_merge_all(void) +{ + PyThreadState *tstate = _PyThreadState_GET(); + if (tstate == NULL) { + return; + } + if (FT_ATOMIC_LOAD_PTR_RELAXED(tstate->interp->pystats_struct) == NULL) { + return; + } + PyInterpreterState *interp = tstate->interp; + _PyEval_StopTheWorld(interp); +#ifdef Py_GIL_DISABLED + _Py_FOR_EACH_TSTATE_UNLOCKED(interp, ts) { + stats_merge_thread((_PyThreadStateImpl *)ts); + stats_zero_thread((_PyThreadStateImpl *)ts); + } +#endif + _PyEval_StartTheWorld(interp); +} + +int +_Py_StatsOn(void) +{ + PyThreadState *tstate = _PyThreadState_GET(); + return stats_toggle_on_off(tstate, 1); +} + +void +_Py_StatsOff(void) +{ + PyThreadState *tstate = _PyThreadState_GET(); + stats_toggle_on_off(tstate, 0); +} + +void +_Py_StatsClear(void) +{ + stats_zero_all(); +} + +static int +mem_is_zero(unsigned char *ptr, size_t size) +{ + for (size_t i=0; i < size; i++) { + if (*ptr != 0) { + return 0; + } + ptr++; + } + return 1; +} + +int +_Py_PrintSpecializationStats(int to_file) +{ + assert(to_file); + stats_merge_all(); + PyThreadState *tstate = _PyThreadState_GET(); + STATS_LOCK(tstate->interp); + PyStats *stats = tstate->interp->pystats_struct; + if (stats == NULL) { + STATS_UNLOCK(tstate->interp); + return 0; + } +#define MEM_IS_ZERO(DATA) mem_is_zero((unsigned char*)DATA, sizeof(*(DATA))) + int is_zero = ( + MEM_IS_ZERO(stats->gc_stats) // is a pointer + && MEM_IS_ZERO(&stats->opcode_stats) + && MEM_IS_ZERO(&stats->call_stats) + && MEM_IS_ZERO(&stats->object_stats) + ); +#undef MEM_IS_ZERO + STATS_UNLOCK(tstate->interp); + if (is_zero) { + // gh-108753: -X pystats command line was used, but then _stats_off() + // and _stats_clear() have been called: in this case, avoid printing + // useless "all zeros" statistics. + return 0; + } + + FILE *out = stderr; + if (to_file) { + /* Write to a file instead of stderr. */ +# ifdef MS_WINDOWS + const char *dirname = "c:\\temp\\py_stats\\"; +# else + const char *dirname = "/tmp/py_stats/"; +# endif + /* Use random 160 bit number as file name, + * to avoid both accidental collisions and + * symlink attacks. */ + unsigned char rand[20]; + char hex_name[41]; + _PyOS_URandomNonblock(rand, 20); + for (int i = 0; i < 20; i++) { + hex_name[2*i] = Py_hexdigits[rand[i]&15]; + hex_name[2*i+1] = Py_hexdigits[(rand[i]>>4)&15]; + } + hex_name[40] = '\0'; + char buf[64]; + assert(strlen(dirname) + 40 + strlen(".txt") < 64); + sprintf(buf, "%s%s.txt", dirname, hex_name); + FILE *fout = fopen(buf, "w"); + if (fout) { + out = fout; + } + } + else { + fprintf(out, "Specialization stats:\n"); + } + STATS_LOCK(tstate->interp); + print_stats(out, stats); + STATS_UNLOCK(tstate->interp); + if (out != stderr) { + fclose(out); + } + return 1; +} + +PyStatus +_PyStats_InterpInit(PyInterpreterState *interp) +{ + if (interp->config._pystats) { + // start with pystats enabled, can be disabled via sys._stats_off() + // this needs to be set before the first tstate is created + interp->pystats_enabled = 1; + interp->pystats_struct = PyMem_RawCalloc(1, sizeof(PyStats)); + if (interp->pystats_struct == NULL) { + return _PyStatus_ERR("out-of-memory while initializing interpreter"); + } + } + return _PyStatus_OK(); +} + +bool +_PyStats_ThreadInit(PyInterpreterState *interp, _PyThreadStateImpl *tstate) +{ +#ifdef Py_GIL_DISABLED + if (FT_ATOMIC_LOAD_INT_RELAXED(interp->pystats_enabled)) { + assert(interp->pystats_struct != NULL); + tstate->pystats_struct = PyMem_RawCalloc(1, sizeof(PyStats)); + if (tstate->pystats_struct == NULL) { + return false; + } + } +#endif + return true; +} + +void +_PyStats_ThreadFini(_PyThreadStateImpl *tstate) +{ +#ifdef Py_GIL_DISABLED + STATS_LOCK(((PyThreadState *)tstate)->interp); + stats_merge_thread(tstate); + STATS_UNLOCK(((PyThreadState *)tstate)->interp); + PyMem_RawFree(tstate->pystats_struct); +#endif +} + +void +_PyStats_Attach(_PyThreadStateImpl *tstate_impl) +{ + PyStats *s; + PyThreadState *tstate = (PyThreadState *)tstate_impl; + PyInterpreterState *interp = tstate->interp; + if (FT_ATOMIC_LOAD_INT_RELAXED(interp->pystats_enabled)) { +#ifdef Py_GIL_DISABLED + s = ((_PyThreadStateImpl *)tstate)->pystats_struct; +#else + s = tstate->interp->pystats_struct; +#endif + } + else { + s = NULL; + } + tstate->pystats = s; +} + +void +_PyStats_Detach(_PyThreadStateImpl *tstate_impl) +{ + ((PyThreadState *)tstate_impl)->pystats = NULL; +} + +#endif // Py_STATS diff --git a/Python/qsbr.c b/Python/qsbr.c index c992c285cb13..b2153bf9d672 100644 --- a/Python/qsbr.c +++ b/Python/qsbr.c @@ -36,6 +36,7 @@ #include "pycore_pystate.h" // _PyThreadState_GET() #include "pycore_qsbr.h" #include "pycore_tstate.h" // _PyThreadStateImpl +#include "pycore_stats.h" // FT_STAT_QSBR_POLL_INC() // Starting size of the array of qsbr thread states @@ -158,7 +159,7 @@ _Py_qsbr_poll(struct _qsbr_thread_state *qsbr, uint64_t goal) if (_Py_qbsr_goal_reached(qsbr, goal)) { return true; } - + FT_STAT_QSBR_POLL_INC(); uint64_t rd_seq = qsbr_poll_scan(qsbr->shared); return QSBR_LEQ(goal, rd_seq); } diff --git a/Python/specialize.c b/Python/specialize.c index a1c5dedd6156..2193596a331d 100644 --- a/Python/specialize.c +++ b/Python/specialize.c @@ -22,437 +22,23 @@ #include // rand() -extern const char *_PyUOpName(int index); - /* For guidance on adding or extending families of instructions see * InternalDocs/interpreter.md `Specialization` section. */ -#ifdef Py_STATS -GCStats _py_gc_stats[NUM_GENERATIONS] = { 0 }; -static PyStats _Py_stats_struct = { .gc_stats = _py_gc_stats }; -PyStats *_Py_stats = NULL; - -#if PYSTATS_MAX_UOP_ID < MAX_UOP_ID -#error "Not enough space allocated for pystats. Increase PYSTATS_MAX_UOP_ID to at least MAX_UOP_ID" -#endif - -#define ADD_STAT_TO_DICT(res, field) \ - do { \ - PyObject *val = PyLong_FromUnsignedLongLong(stats->field); \ - if (val == NULL) { \ - Py_DECREF(res); \ - return NULL; \ - } \ - if (PyDict_SetItemString(res, #field, val) == -1) { \ - Py_DECREF(res); \ - Py_DECREF(val); \ - return NULL; \ - } \ - Py_DECREF(val); \ - } while(0); - -static PyObject* -stats_to_dict(SpecializationStats *stats) -{ - PyObject *res = PyDict_New(); - if (res == NULL) { - return NULL; - } - ADD_STAT_TO_DICT(res, success); - ADD_STAT_TO_DICT(res, failure); - ADD_STAT_TO_DICT(res, hit); - ADD_STAT_TO_DICT(res, deferred); - ADD_STAT_TO_DICT(res, miss); - ADD_STAT_TO_DICT(res, deopt); - PyObject *failure_kinds = PyTuple_New(SPECIALIZATION_FAILURE_KINDS); - if (failure_kinds == NULL) { - Py_DECREF(res); - return NULL; - } - for (int i = 0; i < SPECIALIZATION_FAILURE_KINDS; i++) { - PyObject *stat = PyLong_FromUnsignedLongLong(stats->failure_kinds[i]); - if (stat == NULL) { - Py_DECREF(res); - Py_DECREF(failure_kinds); - return NULL; - } - PyTuple_SET_ITEM(failure_kinds, i, stat); - } - if (PyDict_SetItemString(res, "failure_kinds", failure_kinds)) { - Py_DECREF(res); - Py_DECREF(failure_kinds); - return NULL; - } - Py_DECREF(failure_kinds); - return res; -} -#undef ADD_STAT_TO_DICT - -static int -add_stat_dict( - PyObject *res, - int opcode, - const char *name) { - - SpecializationStats *stats = &_Py_stats_struct.opcode_stats[opcode].specialization; - PyObject *d = stats_to_dict(stats); - if (d == NULL) { - return -1; - } - int err = PyDict_SetItemString(res, name, d); - Py_DECREF(d); - return err; -} - -PyObject* -_Py_GetSpecializationStats(void) { - PyObject *stats = PyDict_New(); - if (stats == NULL) { - return NULL; - } - int err = 0; - err += add_stat_dict(stats, CONTAINS_OP, "contains_op"); - err += add_stat_dict(stats, LOAD_SUPER_ATTR, "load_super_attr"); - err += add_stat_dict(stats, LOAD_ATTR, "load_attr"); - err += add_stat_dict(stats, LOAD_GLOBAL, "load_global"); - err += add_stat_dict(stats, STORE_SUBSCR, "store_subscr"); - err += add_stat_dict(stats, STORE_ATTR, "store_attr"); - err += add_stat_dict(stats, JUMP_BACKWARD, "jump_backward"); - err += add_stat_dict(stats, CALL, "call"); - err += add_stat_dict(stats, CALL_KW, "call_kw"); - err += add_stat_dict(stats, BINARY_OP, "binary_op"); - err += add_stat_dict(stats, COMPARE_OP, "compare_op"); - err += add_stat_dict(stats, UNPACK_SEQUENCE, "unpack_sequence"); - err += add_stat_dict(stats, FOR_ITER, "for_iter"); - err += add_stat_dict(stats, TO_BOOL, "to_bool"); - err += add_stat_dict(stats, SEND, "send"); - if (err < 0) { - Py_DECREF(stats); - return NULL; - } - return stats; -} - - -#define PRINT_STAT(i, field) \ - if (stats[i].field) { \ - fprintf(out, " opcode[%s]." #field " : %" PRIu64 "\n", _PyOpcode_OpName[i], stats[i].field); \ - } - -static void -print_spec_stats(FILE *out, OpcodeStats *stats) -{ - /* Mark some opcodes as specializable for stats, - * even though we don't specialize them yet. */ - fprintf(out, "opcode[BINARY_SLICE].specializable : 1\n"); - fprintf(out, "opcode[STORE_SLICE].specializable : 1\n"); - fprintf(out, "opcode[GET_ITER].specializable : 1\n"); - for (int i = 0; i < 256; i++) { - if (_PyOpcode_Caches[i]) { - /* Ignore jumps as they cannot be specialized */ - switch (i) { - case POP_JUMP_IF_FALSE: - case POP_JUMP_IF_TRUE: - case POP_JUMP_IF_NONE: - case POP_JUMP_IF_NOT_NONE: - case JUMP_BACKWARD: - break; - default: - fprintf(out, "opcode[%s].specializable : 1\n", _PyOpcode_OpName[i]); - } - } - PRINT_STAT(i, specialization.success); - PRINT_STAT(i, specialization.failure); - PRINT_STAT(i, specialization.hit); - PRINT_STAT(i, specialization.deferred); - PRINT_STAT(i, specialization.miss); - PRINT_STAT(i, specialization.deopt); - PRINT_STAT(i, execution_count); - for (int j = 0; j < SPECIALIZATION_FAILURE_KINDS; j++) { - uint64_t val = stats[i].specialization.failure_kinds[j]; - if (val) { - fprintf(out, " opcode[%s].specialization.failure_kinds[%d] : %" - PRIu64 "\n", _PyOpcode_OpName[i], j, val); - } - } - for (int j = 0; j < 256; j++) { - if (stats[i].pair_count[j]) { - fprintf(out, "opcode[%s].pair_count[%s] : %" PRIu64 "\n", - _PyOpcode_OpName[i], _PyOpcode_OpName[j], stats[i].pair_count[j]); - } - } - } -} -#undef PRINT_STAT - - -static void -print_call_stats(FILE *out, CallStats *stats) -{ - fprintf(out, "Calls to PyEval_EvalDefault: %" PRIu64 "\n", stats->pyeval_calls); - fprintf(out, "Calls to Python functions inlined: %" PRIu64 "\n", stats->inlined_py_calls); - fprintf(out, "Frames pushed: %" PRIu64 "\n", stats->frames_pushed); - fprintf(out, "Frame objects created: %" PRIu64 "\n", stats->frame_objects_created); - for (int i = 0; i < EVAL_CALL_KINDS; i++) { - fprintf(out, "Calls via PyEval_EvalFrame[%d] : %" PRIu64 "\n", i, stats->eval_calls[i]); - } -} - -static void -print_object_stats(FILE *out, ObjectStats *stats) -{ - fprintf(out, "Object allocations from freelist: %" PRIu64 "\n", stats->from_freelist); - fprintf(out, "Object frees to freelist: %" PRIu64 "\n", stats->to_freelist); - fprintf(out, "Object allocations: %" PRIu64 "\n", stats->allocations); - fprintf(out, "Object allocations to 512 bytes: %" PRIu64 "\n", stats->allocations512); - fprintf(out, "Object allocations to 4 kbytes: %" PRIu64 "\n", stats->allocations4k); - fprintf(out, "Object allocations over 4 kbytes: %" PRIu64 "\n", stats->allocations_big); - fprintf(out, "Object frees: %" PRIu64 "\n", stats->frees); - fprintf(out, "Object inline values: %" PRIu64 "\n", stats->inline_values); - fprintf(out, "Object interpreter mortal increfs: %" PRIu64 "\n", stats->interpreter_increfs); - fprintf(out, "Object interpreter mortal decrefs: %" PRIu64 "\n", stats->interpreter_decrefs); - fprintf(out, "Object mortal increfs: %" PRIu64 "\n", stats->increfs); - fprintf(out, "Object mortal decrefs: %" PRIu64 "\n", stats->decrefs); - fprintf(out, "Object interpreter immortal increfs: %" PRIu64 "\n", stats->interpreter_immortal_increfs); - fprintf(out, "Object interpreter immortal decrefs: %" PRIu64 "\n", stats->interpreter_immortal_decrefs); - fprintf(out, "Object immortal increfs: %" PRIu64 "\n", stats->immortal_increfs); - fprintf(out, "Object immortal decrefs: %" PRIu64 "\n", stats->immortal_decrefs); - fprintf(out, "Object materialize dict (on request): %" PRIu64 "\n", stats->dict_materialized_on_request); - fprintf(out, "Object materialize dict (new key): %" PRIu64 "\n", stats->dict_materialized_new_key); - fprintf(out, "Object materialize dict (too big): %" PRIu64 "\n", stats->dict_materialized_too_big); - fprintf(out, "Object materialize dict (str subclass): %" PRIu64 "\n", stats->dict_materialized_str_subclass); - fprintf(out, "Object method cache hits: %" PRIu64 "\n", stats->type_cache_hits); - fprintf(out, "Object method cache misses: %" PRIu64 "\n", stats->type_cache_misses); - fprintf(out, "Object method cache collisions: %" PRIu64 "\n", stats->type_cache_collisions); - fprintf(out, "Object method cache dunder hits: %" PRIu64 "\n", stats->type_cache_dunder_hits); - fprintf(out, "Object method cache dunder misses: %" PRIu64 "\n", stats->type_cache_dunder_misses); -} - -static void -print_gc_stats(FILE *out, GCStats *stats) -{ - for (int i = 0; i < NUM_GENERATIONS; i++) { - fprintf(out, "GC[%d] collections: %" PRIu64 "\n", i, stats[i].collections); - fprintf(out, "GC[%d] object visits: %" PRIu64 "\n", i, stats[i].object_visits); - fprintf(out, "GC[%d] objects collected: %" PRIu64 "\n", i, stats[i].objects_collected); - fprintf(out, "GC[%d] objects reachable from roots: %" PRIu64 "\n", i, stats[i].objects_transitively_reachable); - fprintf(out, "GC[%d] objects not reachable from roots: %" PRIu64 "\n", i, stats[i].objects_not_transitively_reachable); - } -} - -#ifdef _Py_TIER2 -static void -print_histogram(FILE *out, const char *name, uint64_t hist[_Py_UOP_HIST_SIZE]) -{ - for (int i = 0; i < _Py_UOP_HIST_SIZE; i++) { - fprintf(out, "%s[%" PRIu64"]: %" PRIu64 "\n", name, (uint64_t)1 << i, hist[i]); - } -} - -static void -print_optimization_stats(FILE *out, OptimizationStats *stats) -{ - fprintf(out, "Optimization attempts: %" PRIu64 "\n", stats->attempts); - fprintf(out, "Optimization traces created: %" PRIu64 "\n", stats->traces_created); - fprintf(out, "Optimization traces executed: %" PRIu64 "\n", stats->traces_executed); - fprintf(out, "Optimization uops executed: %" PRIu64 "\n", stats->uops_executed); - fprintf(out, "Optimization trace stack overflow: %" PRIu64 "\n", stats->trace_stack_overflow); - fprintf(out, "Optimization trace stack underflow: %" PRIu64 "\n", stats->trace_stack_underflow); - fprintf(out, "Optimization trace too long: %" PRIu64 "\n", stats->trace_too_long); - fprintf(out, "Optimization trace too short: %" PRIu64 "\n", stats->trace_too_short); - fprintf(out, "Optimization inner loop: %" PRIu64 "\n", stats->inner_loop); - fprintf(out, "Optimization recursive call: %" PRIu64 "\n", stats->recursive_call); - fprintf(out, "Optimization low confidence: %" PRIu64 "\n", stats->low_confidence); - fprintf(out, "Optimization unknown callee: %" PRIu64 "\n", stats->unknown_callee); - fprintf(out, "Executors invalidated: %" PRIu64 "\n", stats->executors_invalidated); - - print_histogram(out, "Trace length", stats->trace_length_hist); - print_histogram(out, "Trace run length", stats->trace_run_length_hist); - print_histogram(out, "Optimized trace length", stats->optimized_trace_length_hist); - - fprintf(out, "Optimization optimizer attempts: %" PRIu64 "\n", stats->optimizer_attempts); - fprintf(out, "Optimization optimizer successes: %" PRIu64 "\n", stats->optimizer_successes); - fprintf(out, "Optimization optimizer failure no memory: %" PRIu64 "\n", - stats->optimizer_failure_reason_no_memory); - fprintf(out, "Optimizer remove globals builtins changed: %" PRIu64 "\n", stats->remove_globals_builtins_changed); - fprintf(out, "Optimizer remove globals incorrect keys: %" PRIu64 "\n", stats->remove_globals_incorrect_keys); - for (int i = 0; i <= MAX_UOP_ID; i++) { - if (stats->opcode[i].execution_count) { - fprintf(out, "uops[%s].execution_count : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].execution_count); - } - if (stats->opcode[i].miss) { - fprintf(out, "uops[%s].specialization.miss : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].miss); - } - } - for (int i = 0; i < 256; i++) { - if (stats->unsupported_opcode[i]) { - fprintf( - out, - "unsupported_opcode[%s].count : %" PRIu64 "\n", - _PyOpcode_OpName[i], - stats->unsupported_opcode[i] - ); - } - } - - for (int i = 1; i <= MAX_UOP_ID; i++){ - for (int j = 1; j <= MAX_UOP_ID; j++) { - if (stats->opcode[i].pair_count[j]) { - fprintf(out, "uop[%s].pair_count[%s] : %" PRIu64 "\n", - _PyOpcode_uop_name[i], _PyOpcode_uop_name[j], stats->opcode[i].pair_count[j]); - } - } - } - for (int i = 0; i < MAX_UOP_ID; i++) { - if (stats->error_in_opcode[i]) { - fprintf( - out, - "error_in_opcode[%s].count : %" PRIu64 "\n", - _PyUOpName(i), - stats->error_in_opcode[i] - ); - } - } - fprintf(out, "JIT total memory size: %" PRIu64 "\n", stats->jit_total_memory_size); - fprintf(out, "JIT code size: %" PRIu64 "\n", stats->jit_code_size); - fprintf(out, "JIT trampoline size: %" PRIu64 "\n", stats->jit_trampoline_size); - fprintf(out, "JIT data size: %" PRIu64 "\n", stats->jit_data_size); - fprintf(out, "JIT padding size: %" PRIu64 "\n", stats->jit_padding_size); - fprintf(out, "JIT freed memory size: %" PRIu64 "\n", stats->jit_freed_memory_size); - - print_histogram(out, "Trace total memory size", stats->trace_total_memory_hist); -} -#endif - -static void -print_rare_event_stats(FILE *out, RareEventStats *stats) -{ - fprintf(out, "Rare event (set_class): %" PRIu64 "\n", stats->set_class); - fprintf(out, "Rare event (set_bases): %" PRIu64 "\n", stats->set_bases); - fprintf(out, "Rare event (set_eval_frame_func): %" PRIu64 "\n", stats->set_eval_frame_func); - fprintf(out, "Rare event (builtin_dict): %" PRIu64 "\n", stats->builtin_dict); - fprintf(out, "Rare event (func_modification): %" PRIu64 "\n", stats->func_modification); - fprintf(out, "Rare event (watched_dict_modification): %" PRIu64 "\n", stats->watched_dict_modification); - fprintf(out, "Rare event (watched_globals_modification): %" PRIu64 "\n", stats->watched_globals_modification); -} - -static void -print_stats(FILE *out, PyStats *stats) -{ - print_spec_stats(out, stats->opcode_stats); - print_call_stats(out, &stats->call_stats); - print_object_stats(out, &stats->object_stats); - print_gc_stats(out, stats->gc_stats); -#ifdef _Py_TIER2 - print_optimization_stats(out, &stats->optimization_stats); -#endif - print_rare_event_stats(out, &stats->rare_event_stats); -} - -void -_Py_StatsOn(void) -{ - _Py_stats = &_Py_stats_struct; -} - -void -_Py_StatsOff(void) -{ - _Py_stats = NULL; -} - -void -_Py_StatsClear(void) -{ - memset(&_py_gc_stats, 0, sizeof(_py_gc_stats)); - memset(&_Py_stats_struct, 0, sizeof(_Py_stats_struct)); - _Py_stats_struct.gc_stats = _py_gc_stats; -} - -static int -mem_is_zero(unsigned char *ptr, size_t size) -{ - for (size_t i=0; i < size; i++) { - if (*ptr != 0) { - return 0; - } - ptr++; - } - return 1; -} - -int -_Py_PrintSpecializationStats(int to_file) -{ - PyStats *stats = &_Py_stats_struct; -#define MEM_IS_ZERO(DATA) mem_is_zero((unsigned char*)DATA, sizeof(*(DATA))) - int is_zero = ( - MEM_IS_ZERO(stats->gc_stats) // is a pointer - && MEM_IS_ZERO(&stats->opcode_stats) - && MEM_IS_ZERO(&stats->call_stats) - && MEM_IS_ZERO(&stats->object_stats) - ); -#undef MEM_IS_ZERO - if (is_zero) { - // gh-108753: -X pystats command line was used, but then _stats_off() - // and _stats_clear() have been called: in this case, avoid printing - // useless "all zeros" statistics. - return 0; - } - - FILE *out = stderr; - if (to_file) { - /* Write to a file instead of stderr. */ -# ifdef MS_WINDOWS - const char *dirname = "c:\\temp\\py_stats\\"; -# else - const char *dirname = "/tmp/py_stats/"; -# endif - /* Use random 160 bit number as file name, - * to avoid both accidental collisions and - * symlink attacks. */ - unsigned char rand[20]; - char hex_name[41]; - _PyOS_URandomNonblock(rand, 20); - for (int i = 0; i < 20; i++) { - hex_name[2*i] = Py_hexdigits[rand[i]&15]; - hex_name[2*i+1] = Py_hexdigits[(rand[i]>>4)&15]; - } - hex_name[40] = '\0'; - char buf[64]; - assert(strlen(dirname) + 40 + strlen(".txt") < 64); - sprintf(buf, "%s%s.txt", dirname, hex_name); - FILE *fout = fopen(buf, "w"); - if (fout) { - out = fout; - } - } - else { - fprintf(out, "Specialization stats:\n"); - } - print_stats(out, stats); - if (out != stderr) { - fclose(out); - } - return 1; -} - +#if Py_STATS #define SPECIALIZATION_FAIL(opcode, kind) \ do { \ - if (_Py_stats) { \ + PyStats *s = _PyStats_GET(); \ + if (s) { \ int _kind = (kind); \ assert(_kind < SPECIALIZATION_FAILURE_KINDS); \ - _Py_stats->opcode_stats[opcode].specialization.failure_kinds[_kind]++; \ + s->opcode_stats[opcode].specialization.failure_kinds[_kind]++; \ } \ } while (0) - -#endif // Py_STATS - - -#ifndef SPECIALIZATION_FAIL +#else # define SPECIALIZATION_FAIL(opcode, kind) ((void)0) -#endif +#endif // Py_STATS // Initialize warmup counters and optimize instructions. This cannot fail. void diff --git a/Python/sysmodule.c b/Python/sysmodule.c index 59baca26793f..86dd1395cae4 100644 --- a/Python/sysmodule.c +++ b/Python/sysmodule.c @@ -2281,7 +2281,9 @@ static PyObject * sys__stats_on_impl(PyObject *module) /*[clinic end generated code: output=aca53eafcbb4d9fe input=43b5bfe145299e55]*/ { - _Py_StatsOn(); + if (_Py_StatsOn() < 0) { + return NULL; + } Py_RETURN_NONE; } -- 2.47.3