Please add support for EXTENDED_ARG in jumps

[skywodd]

Hi,

Currently, EXTENDED_ARG in jumps are not supported (see: https://github.com/haypo/bytecode/blob/master/bytecode/concrete.py#L395).
Without this feature, it's impossible to craft python object code >64Ko.

Please add support for EXTENDED_ARG in jumps if it's not too difficult to do :)

For example, with this feature, my brainfuck-to-python-bytecode will be able to run any BF program >64Ko (like hanoi.bf).

[vstinner]

Oh. I didn't expect such quick feedback on a project which is a little bit older than a PoC :-)

byteplay doesn't seem to support jump > 2^16-1 neither:

            if jump > 0xFFFF:
                raise NotImplementedError("Extended jumps not implemented")

codetransformers has no specific code to handle long jumps, so I guess that it fails too.

I checked CPython: it implements what I planned to do, it uses a loop. Extract of assemble_jump_offsets() of Python/compile.c:

    /* XXX: This is an awful hack that could hurt performance, but
        on the bright side it should work until we come up
        with a better solution.

        The issue is that in the first loop blocksize() is called
        which calls instrsize() which requires i_oparg be set
        appropriately.          There is a bootstrap problem because
        i_oparg is calculated in the second loop above.

        So we loop until we stop seeing new EXTENDED_ARGs.
        The only EXTENDED_ARGs that could be popping up are
        ones in jump instructions.  So this should converge
        fairly quickly.
    */

[vstinner]

Fixed. Thanks for the report.