Field access within methods

[leonerd]

How should the optree inside a method body access the fields of the instance?

In Object::Pad I couldn't easily influence the way the parser would build, so about the only thing I could do was pretend that the lexical scope of every method in fact contained a whole bunch of extra lexicals, then arrange for all those extra pad slots to be SV-aliased to the relevant instance fields as part of method startup. It makes entry of every method somewhat slow, because it has to do all that aliasing.

A real core implementation ought to do something a lot nicer.

[leonerd]

In particular, I'm thinking of adding three new opcodes, mirrors of OP_PAD{SV,AV,HV}, which fetch the SV pointer of an instance field at that index. These would likely be named OP_FIELD{SV,AV,HV}.

[leonerd]

An initial experiment in Object::Pad at doing some optree-rewriting to test this idea out, has found mixed success. While overall it works, there are two tests that fail.

The two failing cases are essentially:

class Ca { field $x; method test { $x = 123; eval('$x') } }

i.e. string-eval relying on being able to see the field at runtime, and

class Cb { field $x; method test { $x = 123; return sub { $x } } }

i.e. capture of a field into a non-method closure.

I can't think of a solution to both of these that doesn't basically turn out to just be using the same lexical-pad trick as before. So it may turn out that the OP_FIELDxV ops are not a good fit here, and the core implementation does have to do something closer to what Object::Pad does after all.

[leonerd]

Current approach is that OP_METHSTART is an UNOP_AUX, whose aux list stores a list of pad+field index pairs to bind at the start of the method. The behavior of whatObject::Pad calls pp_fieldpad is now rolled into a big while loop inside core's pp_methstart. (and a similar change is now present in Object::Pad as well)

[leonerd]

This is now entirely solved by the "field binding" part of OP_METHSTART, as described by #8. This issue can now be closed (if I could find the button to do that...)