Wednesday, December 15, 2010

The Woes of Change

To get to native string support, I ran into a set of mundane problems, the most pressing libffi support.

Since my language is bootstrapped, moving to string support is a non-trivial exercise since it changes a fundamental assumption on how data is represented.

In order to get it to work, I need to advance by going through the following steps:

  1. Devise a runtime which supports native strings
  2. Build a compiler which targets that runtime
  3. Build a compiler which uses that runtime
  4. Compile 2 with the regular compiler
  5. Compile 3 with 2
  6. Compile 3 with 3
Whoever taught me that bootstrapped languages are easily extended?

One problem is in the change of the runtime. To keep the implementation minimal, all Hi FFI type structures are compiled down to libffi type structures, and all subsequent data conversions are done with the libffi type structure as the normative reference. The latter choice I made because working with the libffi type structure is a bit faster and more easily expressible in C than working with the Hi FFI type structure, it also keeps the number of invariants down.

Problem now is that since I want to treat character data as values, but must pass a pointer to libffi, I want to discriminate between ordinary pointers and pointers to character data. Woe on me, I need an extra type tag in libffi types, which seems to be impossible. I ran into the expression problem in C.

There are two options:
  1. A quick-and-dirty fix, encode character pointers as something else such as a structure containing just one pointer, or an unsigned integer of pointer size. Not nice, and it will break if an ABI treats such an encoding different as a pointer.
  2. Take a step back, and use the Hi FFI type structure as the basis of all data conversions, which is a bit less nice, since I need to write some extra code for that, and it is probably somewhat slower.
Ah well. The better solution is to take a step back.

No comments:

Post a Comment