gh-143732: Add tier2 specialization for TO_BOOL

Lib/test/test_capi/test_opt.py

@@ -4028,6 +4028,148 @@ class A:
         uops = get_opnames(ex)
         self.assertNotIn("_REPLACE_WITH_TRUE", uops)
 
+    def test_to_bool_kwargs_dict(self):
+        """**kwargs is known to be dict, so TO_BOOL specializes to _TO_BOOL_DICT."""
+        def inner(**kwargs):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if kwargs:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner(a=1, b=2)
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_DICT", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_kwargs_empty_dict(self):
+        """**kwargs is known to be dict even when empty."""
+        def inner(**kwargs):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if not kwargs:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner()
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_DICT", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_varargs_tuple(self):
+        """*args is known to be tuple, so TO_BOOL specializes to _TO_BOOL_SIZED."""
+        def inner(*args):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if args:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner(1, 2, 3)
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_SIZED", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_varargs_empty_tuple(self):
+        """*args is known to be tuple even when empty."""
+        def inner(*args):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if not args:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner()
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_SIZED", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_args_and_kwargs(self):
+        """Combined *args and **kwargs both get correct types."""
+        def inner(*args, **kwargs):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if args and kwargs:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner(1, 2, a=3)
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_SIZED", uops)
+        self.assertIn("_TO_BOOL_DICT", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_args_kwargs_with_regular_params(self):
+        """*args/**kwargs slot calculation is correct with regular params."""
+        def inner(x, y, *args, key=None, **kwargs):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if args and kwargs:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner(1, 2, 3, 4, key="v", extra=5)
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_SIZED", uops)
+        self.assertIn("_TO_BOOL_DICT", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
+    def test_to_bool_kwargs_only_no_varargs(self):
+        """**kwargs without *args gets correct dict type."""
+        def inner(x, **kwargs):
+            cnt = 0
+            for i in range(TIER2_THRESHOLD):
+                if kwargs:
+                    cnt += 1
+            return cnt
+
+        def f(n):
+            return inner(1, a=2)
+
+        res, ex = self._run_with_optimizer(f, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        ex_inner = get_first_executor(inner)
+        self.assertIsNotNone(ex_inner)
+        uops = get_opnames(ex_inner)
+        self.assertIn("_TO_BOOL_DICT", uops)
+        self.assertNotIn("_TO_BOOL", uops)
+
     def test_attr_promotion_failure(self):
         # We're not testing for any specific uops here, just
         # testing it doesn't crash.

Python/bytecodes.c

@@ -601,6 +601,23 @@ dummy_func(
             _REPLACE_WITH_TRUE +
             POP_TOP;
 
+        tier2 op(_TO_BOOL_DICT, (value -- res)) {
+            PyObject *value_o = PyStackRef_AsPyObjectBorrow(value);
+            assert(PyDict_CheckExact(value_o));
+            STAT_INC(TO_BOOL, hit);
+            res = PyDict_GET_SIZE(value_o) ? PyStackRef_True : PyStackRef_False;
+            PyStackRef_CLOSE(value);
+        }
+
+        tier2 op(_TO_BOOL_SIZED, (value -- res)) {
+            /* Covers types whose truthiness is Py_SIZE(obj) != 0:
+               tuple, set, frozenset, bytes, bytearray. */
+            PyObject *value_o = PyStackRef_AsPyObjectBorrow(value);
+            STAT_INC(TO_BOOL, hit);
+            res = Py_SIZE(value_o) ? PyStackRef_True : PyStackRef_False;
+            PyStackRef_CLOSE(value);
+        }
+
         macro(UNARY_INVERT) = _UNARY_INVERT + POP_TOP;
 
         op(_UNARY_INVERT, (value -- res, v)) {

Python/optimizer_bytecodes.c

@@ -538,6 +538,16 @@ dummy_func(void) {
     op(_TO_BOOL, (value -- res)) {
         int already_bool = optimize_to_bool(this_instr, ctx, value, &res, false);
         if (!already_bool) {
+            if (sym_matches_type(value, &PyDict_Type)) {
+                REPLACE_OP(this_instr, _TO_BOOL_DICT, 0, 0);
+            }
+            else if (sym_matches_type(value, &PyTuple_Type) ||
+                     sym_matches_type(value, &PySet_Type) ||
+                     sym_matches_type(value, &PyFrozenSet_Type) ||
+                     sym_matches_type(value, &PyBytes_Type) ||
+                     sym_matches_type(value, &PyByteArray_Type)) {
+                REPLACE_OP(this_instr, _TO_BOOL_SIZED, 0, 0);
+            }
             res = sym_new_truthiness(ctx, value, true);
         }
     }
@@ -600,6 +610,21 @@ dummy_func(void) {
         }
     }
 
+    op(_TO_BOOL_DICT, (value -- res)) {
+        int already_bool = optimize_to_bool(this_instr, ctx, value, &res, false);
+        if (!already_bool) {
+            sym_set_type(value, &PyDict_Type);
+            res = sym_new_truthiness(ctx, value, true);
+        }
+    }
+
+    op(_TO_BOOL_SIZED, (value -- res)) {
+        int already_bool = optimize_to_bool(this_instr, ctx, value, &res, false);
+        if (!already_bool) {
+            res = sym_new_truthiness(ctx, value, true);
+        }
+    }
+
     op(_UNARY_NOT, (value -- res)) {
         REPLACE_OPCODE_IF_EVALUATES_PURE(value, res);
         sym_set_type(value, &PyBool_Type);