Amaury Forgeot d'Arc wrote:

Hello,

On Thu, Mar 26, 2009 at 01:41, Ben Mellor wrote:

...

Does anyone know why

lambda x, y: x % y

is being converted to a flow graph as:

v2 = int_mod(x_0, y_0) v1 = int_add(v2, [int_mul([int_and([cast_bool_to_int([int_le([int_xor(x_0, y_0)], (0))])], [cast_bool_to_int([int_ne(v2, (0))])])], y_0)])

(v1 is returned)

What's wrong with just v1 = int_mod(x_0, y_0)?

You are looking at the graph of the function: pypy.rpython.lltypesystem.opimpl.op_int_mod

The difference is certainly because C and python handle modulus differently when one argument is negative: (-5 % 3) returns -2 in C, but 1 with python.

Now, I don't know whether RPython should absolutely implement python semantics here. RPython does not check for overflow for example.

The semantics of % in RPython _are_ the same as in Python. It's just that the semantics of int_mod aren't that of Python, therefore % must be mapped to something more complex. The reason why int_mod has the C semantics is that it makes the life of all backends easier. Cheers, Carl Friedrich