pypy / pypy / doc / discussion / improve-rpython.rst

Full commit

Possible improvements of the rpython language

Improve the interpreter API

  • Rationalize the modules, and the names, of the different functions needed to implement a pypy module. A typical rpython file is likely to contain many import statements:

    from pypy.interpreter.baseobjspace import W_Root
    from pypy.interpreter.gateway import ObjSpace, W_Root
    from pypy.interpreter.argument import Arguments
    from pypy.interpreter.typedef import TypeDef, GetSetProperty
    from pypy.interpreter.typedef import interp_attrproperty, interp_attrproperty_w
    from pypy.interpreter.gateway import interp2app
    from pypy.interpreter.error import OperationError
    from rpython.rtyper.lltypesystem import rffi, lltype
  • A more direct declarative way to write Typedef:

    class W_Socket(W_Root):
        _typedef_name_ = 'socket'
        _typedef_base_ = W_EventualBaseClass
        @interp2app_method("connect", ['self', ObjSpace, W_Root])
        def connect_w(self, space, w_addr):
  • Support for metaclasses written in rpython. For a sample, see the skipped test pypy.objspace.std.test.TestTypeObject.test_metaclass_typedef

RPython language

  • Arithmetic with unsigned integer, and between integer of different signedness, when this is not ambiguous. At least, comparison and assignment with constants should be allowed.
  • Allocate variables on the stack, and pass their address ("by reference") to llexternal functions. For a typical usage, see rpython.rlib.rsocket.RSocket.getsockopt_int.

Extensible type system for llexternal

llexternal allows the description of a C function, and conveys the same information about the arguments as a C header. But this is often not enough. For example, a parameter of type int* is converted to rffi.CArrayPtr(rffi.INT), but this information is not enough to use the function. The parameter could be an array of int, a reference to a single value, for input or output...

A "type system" could hold this additional information, and automatically generate some conversion code to ease the usage of the function from rpython. For example:

# double frexp(double x, int *exp);
frexp = llexternal("frexp", [rffi.DOUBLE, OutPtr(], rffi.DOUBLE)

OutPtr indicates that the parameter is output-only, which need not to be initialized, and which value is returned to the caller. In rpython the call becomes:

fraction, exponent = frexp(value)

Also, we could imagine that one item in the llexternal argument list corresponds to two parameters in C. Here, OutCharBufferN indicates that the caller will pass a rpython string; the framework will pass buffer and length to the function:

# ssize_t write(int fd, const void *buf, size_t count);
write = llexternal("write", [rffi.INT, CharBufferAndSize], rffi.SSIZE_T)

The rpython code that calls this function is very simple:

written = write(fd, data)

compared with the present:

count = len(data)
buf = rffi.get_nonmovingbuffer(data)
    written = rffi.cast(lltype.Signed, os_write(
        rffi.cast(rffi.INT, fd),
        buf, rffi.cast(rffi.SIZE_T, count)))
    rffi.free_nonmovingbuffer(data, buf)

Typemaps are very useful for large APIs where the same conversions are needed in many places. XXX example