float() rejects instances of an int subclass #19874

Port, board and/or hardware

Unix port, macOS arm64

MicroPython version

MicroPython v1.30.0-preview-247-gcbbdb6379

Reproduction

class MyInt(int):
    pass

print(float(MyInt(2)))

Expected behaviour

CPython prints (3.11.15):

2.0

Observed behaviour

TypeError: can't convert MyInt to float

The same TypeError comes from complex(MyInt(2)), math.sqrt(MyInt(4)), "%f" % MyInt(2) and 1.5 + MyInt(2), while MyInt(2) * 1.5 returns 3.0.

Additional Information

tests/cpydiff/types_int_subclassconv.py documents "No int conversion for int-derived types available" with an int + A example. It does not mention float(), complex() or math, which reach a different function.

Root Cause

mp_obj_get_float_maybe() recognises an integer only as a small int or an object of exactly type int, so a subclass instance falls to MP_UNARY_OP_FLOAT_MAYBE, which instance_unary_op() forwards to the native int, and int never answers it with a float. CPython's PyNumber_Float() calls the nb_float slot that every int subclass inherits.

Fix Suggestion

When MP_UNARY_OP_FLOAT_MAYBE gives no float, convert mp_obj_cast_to_native_base(arg, &mp_type_int) instead, which keeps a subclass's own __float__. That adds 69 bytes on x64 and 44 bytes on arm64 to the unix build-standard firmware (sum of object-file .text, gcc).

diff --git a/py/obj.c b/py/obj.c
index 252e7372f..603dbdef3 100644
--- a/py/obj.c
+++ b/py/obj.c
@@ -392,9 +392,12 @@ bool mp_obj_get_float_maybe(mp_obj_t arg, mp_float_t *value) {
     } else if (mp_obj_is_float(arg)) {
         val = mp_obj_float_get(arg);
     } else {
-        arg = mp_unary_op(MP_UNARY_OP_FLOAT_MAYBE, (mp_obj_t)arg);
-        if (arg != MP_OBJ_NULL && mp_obj_is_float(arg)) {
-            val = mp_obj_float_get(arg);
+        mp_obj_t res = mp_unary_op(MP_UNARY_OP_FLOAT_MAYBE, arg);
+        if (res != MP_OBJ_NULL && mp_obj_is_float(res)) {
+            val = mp_obj_float_get(res);
+        } else if ((res = mp_obj_cast_to_native_base(arg, MP_OBJ_FROM_PTR(&mp_type_int))) != MP_OBJ_NULL) {
+            // an int subclass without __float__ converts as its int value
+            return mp_obj_get_float_maybe(res, value);
         } else {
             return false;
         }

I can open this as a PR if the size is acceptable.

Code of Conduct

Yes, I agree