gh-146393: Optimize float division operations by mutating uniquely-referenced operands in place (JIT only)

Lib/test/test_capi/test_opt.py

@@ -3348,6 +3348,129 @@ def testfunc(args):
         uops = get_opnames(ex)
         self.assertNotIn("_UNARY_NEGATIVE_FLOAT_INPLACE", uops)
 
+    def test_float_truediv_inplace_unique_lhs(self):
+        # (a + b) / (c + d): LHS is unique float from add, RHS is unique
+        # float from add. The division reuses the LHS in place.
+        def testfunc(args):
+            a, b, c, d, n = args
+            total = 0.0
+            for _ in range(n):
+                total += (a + b) / (c + d)
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc, (2.0, 3.0, 1.0, 3.0, TIER2_THRESHOLD))
+        self.assertAlmostEqual(res, TIER2_THRESHOLD * 1.25)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP_TRUEDIV_FLOAT_INPLACE", uops)
+
+    def test_float_truediv_inplace_unique_rhs(self):
+        # x = c + d stores to a local (not unique when reloaded).
+        # (a + b) is unique. The division should use inplace on the RHS.
+        def testfunc(args):
+            a, b, c, d, n = args
+            total = 0.0
+            for _ in range(n):
+                x = c + d
+                total += x / (a + b)
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc, (2.0, 3.0, 4.0, 5.0, TIER2_THRESHOLD))
+        self.assertAlmostEqual(res, TIER2_THRESHOLD * (9.0 / 5.0))
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP_TRUEDIV_FLOAT_INPLACE_RIGHT", uops)
+
+    def test_float_truediv_type_propagation(self):
+        # Test the _BINARY_OP_TRUEDIV_FLOAT propagates type information
+        def testfunc(args):
+            a, b, n = args
+            total = 0.0
+            for _ in range(n):
+                x = (a + b) # type of x will specialize to float
+                total += x / x - x / x
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc,
+            (2.0, 3.0, TIER2_THRESHOLD))
+        expected = TIER2_THRESHOLD * ((2.0 + 3.0) / (2.0 + 3.0) - (2.0 + 3.0) / (2.0 + 3.0))
+        self.assertAlmostEqual(res, expected)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP_TRUEDIV_FLOAT", uops)
+        self.assertIn("_BINARY_OP_SUBTRACT_FLOAT_INPLACE", uops)
+
+    def test_float_truediv_unique_result_enables_inplace(self):
+        # (a+b) / (c+d) / (e+f): chained divisions where each result
+        # is unique, enabling inplace for subsequent divisions.
+        def testfunc(args):
+            a, b, c, d, e, f, n = args
+            total = 0.0
+            for _ in range(n):
+                total += (a + b) / (c + d) / (e + f)
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc,
+            (2.0, 3.0, 1.0, 1.0, 1.0, 1.0, TIER2_THRESHOLD))
+        expected = TIER2_THRESHOLD * ((2.0 + 3.0) / (1.0 + 1.0) / (1.0 + 1.0))
+        self.assertAlmostEqual(res, expected)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP_TRUEDIV_FLOAT_INPLACE", uops)
+
+    def test_float_add_chain_both_unique(self):
+        # (a+b) + (c+d): both sub-additions produce unique floats.
+        # The outer + should use inplace on one of them.
+        def testfunc(args):
+            a, b, c, d, n = args
+            total = 0.0
+            for _ in range(n):
+                total += (a + b) + (c + d)
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc, (1.0, 2.0, 3.0, 4.0, TIER2_THRESHOLD))
+        self.assertAlmostEqual(res, TIER2_THRESHOLD * 10.0)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        # The outer + should use inplace (at least one operand is unique)
+        inplace = (
+            "_BINARY_OP_ADD_FLOAT_INPLACE" in uops
+            or "_BINARY_OP_ADD_FLOAT_INPLACE_RIGHT" in uops
+        )
+        self.assertTrue(inplace, "Expected inplace add for unique sub-results")
+
+    def test_float_truediv_non_float_type_no_crash(self):
+        # Fraction / Fraction goes through _BINARY_OP with NB_TRUE_DIVIDE
+        # but returns Fraction, not float. The optimizer must not assume
+        # the result is float for non-int/float operands. See gh-146306.
+        from fractions import Fraction
+        def testfunc(args):
+            a, b, n = args
+            total = Fraction(0)
+            for _ in range(n):
+                total += a / b
+            return float(total)
+
+        res, ex = self._run_with_optimizer(testfunc, (Fraction(10), Fraction(3), TIER2_THRESHOLD))
+        expected = float(TIER2_THRESHOLD * Fraction(10, 3))
+        self.assertAlmostEqual(res, expected)
+
+    def test_float_truediv_mixed_float_fraction_no_crash(self):
+        # float / Fraction: lhs is known float from a prior guard,
+        # but rhs is Fraction. The guard insertion for rhs should
+        # deopt cleanly at runtime, not crash.
+        from fractions import Fraction
+        def testfunc(args):
+            a, b, c, n = args
+            total = 0.0
+            for _ in range(n):
+                total += (a + b) / c  # (a+b) is float, c is Fraction
+            return total
+
+        res, ex = self._run_with_optimizer(testfunc, (2.0, 3.0, Fraction(4), TIER2_THRESHOLD))
+        expected = TIER2_THRESHOLD * (5.0 / Fraction(4))
+        self.assertAlmostEqual(res, float(expected))
+
     def test_load_attr_instance_value(self):
         def testfunc(n):
             class C():

Misc/NEWS.d/next/Core_and_Builtins/2026-03-25-12-00-00.gh-issue-146306.B9f62e.rst

@@ -0,0 +1,2 @@
+Specialize float true division in the tier 2 optimizer with inplace
+mutation for uniquely-referenced operands.

Python/bytecodes.c

@@ -838,6 +838,53 @@ dummy_func(
             INPUTS_DEAD();
         }
 
+        // Float true division — not specialized at tier 1, emitted by the
+        // tier 2 optimizer when both operands are known floats.
+        tier2 op(_BINARY_OP_TRUEDIV_FLOAT, (left, right -- res, l, r)) {
+            PyObject *left_o = PyStackRef_AsPyObjectBorrow(left);
+            PyObject *right_o = PyStackRef_AsPyObjectBorrow(right);
+            assert(PyFloat_CheckExact(left_o));
+            assert(PyFloat_CheckExact(right_o));
+            STAT_INC(BINARY_OP, hit);
+            double divisor = ((PyFloatObject *)right_o)->ob_fval;
+            if (divisor == 0.0) {
+                PyErr_SetString(PyExc_ZeroDivisionError,
+                                "float division by zero");
+                ERROR_NO_POP();
+            }
+            double dres = ((PyFloatObject *)left_o)->ob_fval / divisor;
+            PyObject *d = PyFloat_FromDouble(dres);
+            if (d == NULL) {
+                ERROR_NO_POP();
+            }
+            res = PyStackRef_FromPyObjectSteal(d);
+            l = left;
+            r = right;
+            INPUTS_DEAD();
+        }
+
+        tier2 op(_BINARY_OP_TRUEDIV_FLOAT_INPLACE, (left, right -- res, l, r)) {
+            FLOAT_INPLACE_DIVOP(left, right, left);
+            if (_divop_err) {
+                ERROR_NO_POP();
+            }
+            res = left;
+            l = PyStackRef_NULL;
+            r = right;
+            INPUTS_DEAD();
+        }
+
+        tier2 op(_BINARY_OP_TRUEDIV_FLOAT_INPLACE_RIGHT, (left, right -- res, l, r)) {
+            FLOAT_INPLACE_DIVOP(left, right, right);
+            if (_divop_err) {
+                ERROR_NO_POP();
+            }
+            res = right;
+            l = left;
+            r = PyStackRef_NULL;
+            INPUTS_DEAD();
+        }
+
         pure op(_BINARY_OP_ADD_UNICODE, (left, right -- res, l, r)) {
             PyObject *left_o = PyStackRef_AsPyObjectBorrow(left);
             PyObject *right_o = PyStackRef_AsPyObjectBorrow(right);

Python/ceval_macros.h

@@ -562,3 +562,27 @@ gen_try_set_executing(PyGenObject *gen)
         ((PyFloatObject *)PyStackRef_AsPyObjectBorrow(TARGET))           \
             ->ob_fval = _dres;                                           \
     } while (0)
+
+// Inplace float true division. Sets _divop_err to 1 on zero division.
+// Caller must check _divop_err and call ERROR_NO_POP() if set.
+#define FLOAT_INPLACE_DIVOP(left, right, TARGET)                         \
+    int _divop_err = 0;                                                  \
+    do {                                                                 \
+        PyObject *left_o = PyStackRef_AsPyObjectBorrow(left);            \
+        PyObject *right_o = PyStackRef_AsPyObjectBorrow(right);          \
+        assert(PyFloat_CheckExact(left_o));                              \
+        assert(PyFloat_CheckExact(right_o));                             \
+        assert(_PyObject_IsUniquelyReferenced(                           \
+            PyStackRef_AsPyObjectBorrow(TARGET)));                       \
+        STAT_INC(BINARY_OP, hit);                                        \
+        double _divisor = ((PyFloatObject *)right_o)->ob_fval;           \
+        if (_divisor == 0.0) {                                           \
+            PyErr_SetString(PyExc_ZeroDivisionError,                     \
+                            "float division by zero");                   \
+            _divop_err = 1;                                              \
+            break;                                                       \
+        }                                                                \
+        double _dres = ((PyFloatObject *)left_o)->ob_fval / _divisor;    \
+        ((PyFloatObject *)PyStackRef_AsPyObjectBorrow(TARGET))           \
+            ->ob_fval = _dres;                                           \
+    } while (0)