[Python-Dev] Introducing new operators for matrix computation

[Jeremy Hylton]

>>>>> "AMK" == Andrew Kuchling <akuchlin@mems-exchange.org> writes:

  AMK> It should be possible to try out new operators *right now*,
  AMK> with no core changes.  I believe Jeremy's compiler code,
  AMK> written in pure Python, can parse a module and generate the
  AMK> corresponding bytecodes.  Therefore, you could take that code
  AMK> and hack it to support 'a .| b', by generating equivalent code
  AMK> to a.__dotbar__(b), or whatever.  The only visible difference
  AMK> is that some error messages would be obscure; 2 .| 1 would get
  AMK> a '2 has no attribute __dotbar__' exception; this isn't
  AMK> significant.

There is one hitch.  There is a corresponding bytecode op for each of
the builtin operators.  When Python sees 'm * x + b', it generates
BINARY_MULTIPLY and BINARY_ADD opcode.  If you wanted to add a new
binary operator, you couldn't add a new opcode for it.  Instead, you'd
have to generate explicit instructions to check for the attr and then
call it.

This isn't a big problem, although the resulting interpreter is
probably slower than one with builtin opcodes.  Not sure how much
slower. 

  AMK> So, rather than add a whole bunch of new operators to the core,
  AMK> I'd first want to see the alternative syntax implemented
  AMK> through a hacked parser and *in use by a significant
  AMK> community*, and then we can consider what bits from it to add
  AMK> to the core.

This seems like a reasonable approach to me.  You basically build a
frontend that compiles the Python+extensions into standard Python
bytecode.

Jeremy