Re: An operator syntax for bitfields using metatables

[Jay Carlson <nop@...>]

On Jul 13, 2013, at 7:28 AM, John Hind wrote:

>> From: John Hind [mailto:john.hind@zen.co.uk]
> 
>> It is straightforward to add a syntax for this in
>> a C library by providing a metatable for the number type with 'index'
>> and 'newindex' metamethods. Then:
>> 
>> b = n[3]        -- Extract bit 3 as a boolean.
>> n[23] = true    -- Set bit 23.
>> n[6] = not n[6] -- Toggle bit 6.
>> 
> 
> Arghh! This does not work! (But I still think it should.)
> 
> The problem is that Lua does not honour a return value from the 'newindex'
> metamethod so the second and third lines in the above example do not
> actually change n. This is fine for tables and userdata, but makes the
> 'newindex' metamethod pretty useless for any other type!

More broadly, this is an issue for any immutable type. Consider tuples:

  tpl = Tuple(1, 2, 3)
  orig_tpl = tpl
  assert( tpl[2] == 2 )

  tpl[2] = 0   ==> error("that's not really an lvalue")

  s = "123"
  s[2]="0"   ==> error("that's not an lvalue either")

So for __newindex-like-behavior to have any meaning, the assignment had to be syntactic sugar for

  tpl = tuple__copy_replacing_index(tpl, 2, 0)
  s = string__copy_replacing_index(s, 2, "0")

and note that definitionally tpl~=orig_tpl.

I think this is a scary transformation to do conditionally at runtime. I suppose the answer is to turn *all* t[k]=v statements into sugar; the table implementation doesn't need to copy of course. I'm still scared of it though....

Jay