source: trunk/third/librep/TODO @ 15762

[ This is -*-Indented-Text-*- ]

TODO list for librep
********************

Bugs are marked !, things that should be done soon are marked +,
and more long-term ideas are marked -


Outstanding bugs
================

 ! uses setlocale("LC_NUMERIC", "C") to make floating point I/O work,
   need to reimplement this locally

   [ this is only done temporarily now, but rep-gtk still has some
     instances that can't be removed until librep 0.14 I guess.. ]

 ! modules with `(set-binds)' config shouldn't inline constants?

 ! #!key params can't be inlined by the compiler

 ! non-top-level compiled defvar's aren't quite right

 ! scheme define/lambda doesn't splice begin forms

 ! the scheme module fails some of the guile test.scm cases

 ! can't do numerator/denominator of flonums

 ! the first level lookup of foo#bar isn't cached

 ! interfaces aren't re-parsed when modules are reloaded

 ! environment of macro expanders is not consistent

   interpreted code closes macros in the usual lexical environment,
   the compiler closes them in the *root-strcucture* since the
   lexical environment of the compiled file doesn't exist

   Xerox scheme closes all macros in the `initial environment', this
   would provide consistency, but would break existing code

 ! macro memoization loses

   e.g. if same (eq) expression is expanded in different structures

   OTOH, there is little or no chance of this ever happening

 ! doesn't handle NaN or Inf in floats properly (at all)

 ! Putting a breakpoint in a .jaderc file doesn't work correctly; the
   debugger is entered, but the commands don't work

 ! if load can't find the file, its error message is confusing
   (especially if the named file does exist, but no suffixed file
   exists)

 ! non-fixnum numbers can't be dumped / dump totally broken re: modules

 ! `random' only accepts its range as a fixnum

   GMP 3 has suitable functions, but need to support GMP 2 (e.g. RH 6)

 ! it's legal to do (set foo 42) where the contents of foo has a
   lexical binding

   this breaks: (let ((var 'loop))
                  (let loop ((foo t))
                    (set var print)
                    (loop foo)))

 + document in manual: current-utime, new read syntaxes

 + deprecated:

   * 0xNN and 0NN integer syntaxes [0.13+]
   * define-value [0.13+]
   * &optional and &rest (replaced by #!optional and #!rest) [0.13+]


General programming tasks:
==========================

 + comparison of datums currently compares type id and contents

   should be defined by programmer, e.g.:

        (define-datum-comparer ID FUN)

   by analogy with define-datum-discloser

   (but do callers of rep_value_cmp () assume that it may GC?)

 + avoid spurious init-bind and unbind instructions

   (generated when lexical bindings aren't heap allocated)

 + rename backquote* as quasiquote* -- helps scheme, doesn't affect rep

 + add restricted load to gaol? (same directory only?)

 + %make-structure shouldn't map names to structures itself

   it's done this way to allow mutually recursive structures (debatable
   itself) but really %make-structure should just create the thing
   (with an optional name for printing)

 - allow ML-like functors (parameterized modules)

   the basic support is there. Can do:

   (define (my-functor x)
     (structure (export foo) (open rep)
       (define (foo y) (+ x y))))

   but there's currently no clean way to import the first-class
   structure (without resorting to open-structures)

   I think the way is to require the functor to implement a named
   interface which is specified in the module signature:

   (define-structure foo
       (open rep
             (functor (my-functor x) some-interface))
       ...

   `(my-functor x) would get evaluated in the environment containing
   the define-structure form? Or in the environment created up to that
   point?

   The named-interface trick could also be useful for importing normal
   modules (to avoid having to load them at compile time)

   [ I have a first-stab at this, needs compiler support.. ]

 - add a facility for dumping a set of structures, for later reloading

 - move the gtk-independent parts of the rep-gtk glue generator and
   support code to librep

   [ I've rewritten the glue-code generator in an oo style so that it
   can be easily extended. Need to rewrite the runtime support. Will do
   this in time for GTK 2.0 ]

 - add defadvice (port from elisp? other implementations?)

 - Compiler could annotate output files with their dependences

 - I/O shouldn't block all threads

   wait_for_input () already groks threads, so I think the only problem
   is the use of stdio/fd functions. How does stdio handle streams that
   have been set non-blocking? Maybe reimplement basic stdio?

   (there is now support for waking threads via polling)

 - add regression tests

 - scheme compilation is worse than rep compilation

   should be able to make this a lot better, maybe not as good as rep
   code (since the vm primitives are designed for that), but still a
   lot better..

   [ it's better now, but still room for improvement ]

 - the gc sucks

   is it possible to add a good (generational?) gc?

   could sweeping be sped up by prefetching cache lines?

   do lazy sweeping of block-allocated objects? (problem with cons mark bit?)

   do mostly non-recursive marking? (mark stack)

   [ tried this -- marginally slower than current method ]

 - remove special forms (replacing them with macros) where both
   possible and desirable

   The current (July 2000) list of special forms is:

   cond %define defvar lambda progn quote setq

 - most subrs can't be called tail recursively (apply is special-cased)

 - add a hygienic macro facility

   this may be overkill? capture of introduced bindings can be avoided
   using gensyms, capture of introduced free variables could be avoided
   by introducing a way of tagging variable references with the
   structure they need to be dereferenced in.

   [ I have an experimental low-level hygienic macro implementation,
     but it's a long way from being useful ]

 - do JIT compilation of bytecode where profitable

   there's now GNU lightning, a VCODE-like system, using C macros to do
   portable runtime code generation

   Only do this for _heavily_ used bytecode subrs. Measure this by
   adding an extra vector slot, and counting the number of vm
   instructions executed (and dividing by the length of the
   code-string?)

   Another option is to generate direct-threaded code from the bytecode
   (and cache it). I have an attempt at this but it needs either (1) an
   extra pass to detect labels, or (2) to maintain a strict mapping
   between bytecode addresses and direct-code addresses

   There's an interesting paper about automatically generating meta
   instructions to suit individual instruction sequences, PLDI 98 or
   something (check citeseer for it). Applied with reasonable success
   to Caml interpreter

 - Optimize compilation of case statements

   1. handle constant keys

   2. optimize the search (binary search if all clauses have same type
      and are orderable?)

 - Implement weak-pairs or equivalent

   could use normal cons types for this, but allocate from a `weak'
   freelist, then only gc has to differentiate between weak and
   non-weak pairs

 - Use two passes for `concat'; the first pass calculates the new
   string's length, the second builds it

 - Add more backends for accessing remote files

   Make remote-rcp work properly, and add others (ssh, http, ..?)

 - Make the compiler optimise its output

   now the lisp is mostly lexically scoped, there should be much
   more potential for aggressive optimisation


Manual tasks:
=============

 + Document the error-mode and interrupt-mode variables

 + Document the internals (i.e. the C interface)