Merge branch 'main' into match-class-opcodes

Author: cdce8p

.gitattributes

@@ -94,6 +94,8 @@ Lib/test/test_stable_abi_ctypes.py                  generated
 Lib/test/test_zoneinfo/data/*.json                  generated
 Lib/token.py                                        generated
 Misc/sbom.spdx.json                                 generated
+Modules/_testinternalcapi/test_cases.c.h            generated
+Modules/_testinternalcapi/test_targets.h            generated
 Objects/typeslots.inc                               generated
 PC/python3dll.c                                     generated
 Parser/parser.c                                     generated

Include/internal/pycore_opcode_metadata.h

@@ -76,6 +76,8 @@ typedef struct {
 typedef enum {
     JIT_PRED_IS,
     JIT_PRED_IS_NOT,
+    JIT_PRED_EQ,
+    JIT_PRED_NE,
 } JitOptPredicateKind;
 
 typedef struct {

Include/internal/pycore_optimizer_types.h

@@ -890,6 +890,138 @@ def testfunc(n):
         self.assertLessEqual(len(guard_nos_unicode_count), 1)
         self.assertIn("_COMPARE_OP_STR", uops)
 
+    def test_compare_int_eq_narrows_to_constant(self):
+        def f(n):
+            def return_1():
+                return 1
+
+            hits = 0
+            v = return_1()
+            for _ in range(n):
+                if v == 1:
+                    if v == 1:
+                        hits += 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_INT"), 1)
+
+    def test_compare_int_ne_narrows_to_constant(self):
+        def f(n):
+            def return_1():
+                return 1
+
+            hits = 0
+            v = return_1()
+            for _ in range(n):
+                if v != 1:
+                    hits += 1000
+                else:
+                    if v == 1:
+                        hits += v + 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD * 2)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_INT"), 1)
+
+    def test_compare_float_eq_narrows_to_constant(self):
+        def f(n):
+            def return_tenth():
+                return 0.1
+
+            hits = 0
+            v = return_tenth()
+            for _ in range(n):
+                if v == 0.1:
+                    if v == 0.1:
+                        hits += 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_FLOAT"), 1)
+
+    def test_compare_float_ne_narrows_to_constant(self):
+        def f(n):
+            def return_tenth():
+                return 0.1
+
+            hits = 0
+            v = return_tenth()
+            for _ in range(n):
+                if v != 0.1:
+                    hits += 1000
+                else:
+                    if v == 0.1:
+                        hits += 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_FLOAT"), 1)
+
+    def test_compare_str_eq_narrows_to_constant(self):
+        def f(n):
+            def return_hello():
+                return "hello"
+
+            hits = 0
+            v = return_hello()
+            for _ in range(n):
+                if v == "hello":
+                    if v == "hello":
+                        hits += 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_STR"), 1)
+
+    def test_compare_str_ne_narrows_to_constant(self):
+        def f(n):
+            def return_hello():
+                return "hello"
+
+            hits = 0
+            v = return_hello()
+            for _ in range(n):
+                if v != "hello":
+                    hits += 1000
+                else:
+                    if v == "hello":
+                        hits += 1
+            return hits
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        # Constant narrowing allows constant folding for second comparison
+        self.assertLessEqual(count_ops(ex, "_COMPARE_OP_STR"), 1)
+
     @unittest.skip("gh-139109 WIP")
     def test_combine_stack_space_checks_sequential(self):
         def dummy12(x):
@@ -2897,6 +3029,29 @@ def testfunc(n):
         self.assertIn("_POP_TOP_NOP", uops)
         self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
 
+    def test_binary_op_refcount_elimination(self):
+        class CustomAdder:
+            def __init__(self, val):
+                self.val = val
+            def __add__(self, other):
+                return CustomAdder(self.val + other.val)
+
+        def testfunc(n):
+            a = CustomAdder(1)
+            b = CustomAdder(2)
+            res = None
+            for _ in range(n):
+                res = a + b
+            return res.val if res else 0
+
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, 3)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP", uops)
+        self.assertIn("_POP_TOP_NOP", uops)
+        self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
+
     def test_binary_op_extend_float_long_add_refcount_elimination(self):
         def testfunc(n):
             a = 1.5
@@ -3866,6 +4021,29 @@ def __next__(self):
         """), PYTHON_JIT="1", PYTHON_JIT_STRESS="1")
         self.assertEqual(result[0].rc, 0, result)
 
+    def test_144068_daemon_thread_jit_cleanup(self):
+        result = script_helper.run_python_until_end('-c', textwrap.dedent("""
+        import threading
+        import time
+
+        def hot_loop():
+            end = time.time() + 5.0
+            while time.time() < end:
+                pass
+
+        # Create a daemon thread that will be abandoned at shutdown
+        t = threading.Thread(target=hot_loop, daemon=True)
+        t.start()
+
+        time.sleep(0.1)
+        """), PYTHON_JIT="1", ASAN_OPTIONS="detect_leaks=1")
+        self.assertEqual(result[0].rc, 0, result)
+        stderr = result[0].err.decode('utf-8', errors='replace')
+        self.assertNotIn('LeakSanitizer', stderr,
+                         f"Memory leak detected by ASan:\n{stderr}")
+        self.assertNotIn('_PyJit_TryInitializeTracing', stderr,
+                         f"JIT tracer memory leak detected:\n{stderr}")
+
 def global_identity(x):
     return x
 

Include/internal/pycore_uop_ids.h

@@ -0,0 +1 @@
+Eliminate redundant refcounting in the JIT for ``BINARY_OP``.

Include/internal/pycore_uop_metadata.h

@@ -0,0 +1 @@
+Fix JIT tracer memory leak, ensure the JIT tracer state is freed when daemon threads are cleaned up during interpreter shutdown.

Lib/test/test_capi/test_opt.py

@@ -5152,7 +5152,7 @@ dummy_func(
             assert(oparg <= NB_OPARG_LAST);
         }
 
-        op(_BINARY_OP, (lhs, rhs -- res)) {
+        op(_BINARY_OP, (lhs, rhs -- res, l, r)) {
             PyObject *lhs_o = PyStackRef_AsPyObjectBorrow(lhs);
             PyObject *rhs_o = PyStackRef_AsPyObjectBorrow(rhs);
 
@@ -5162,10 +5162,13 @@ dummy_func(
                 ERROR_NO_POP();
             }
             res = PyStackRef_FromPyObjectSteal(res_o);
-            DECREF_INPUTS();
+            l = lhs;
+            r = rhs;
+            DEAD(lhs);
+            DEAD(rhs);
         }
 
-        macro(BINARY_OP) = _SPECIALIZE_BINARY_OP + unused/4 + _BINARY_OP;
+        macro(BINARY_OP) = _SPECIALIZE_BINARY_OP + unused/4 + _BINARY_OP + POP_TOP + POP_TOP;
 
         pure replicate(2:4) inst(SWAP, (bottom, unused[oparg-2], top --
                     bottom, unused[oparg-2], top)) {