source: project/chicken/trunk/TODO @ 13337

TODO                                                            -*- org -*-
====


* bugs

** compiler
*** pre-optimization 
**** changes call-sites and makes them invalid for later pre-optimization
*** check profiling (-profile vs. profile-declaration)
**** this doesn't always seem to work, in particular with library units
**** ##sys#finish-profile should always be invoked
     is it in `exit'?
*** when re-defining intrinsics, the compiler should warn and disable re-writes
    add declaration to keep re-writes enabled for core library files
*** check in foreign.import.scm and compiler.import.scm whether the import
    took place in the compiler

** expander
*** TODO unstripped quoted constants                                       :BUG:
    if macro expands into the name of an exported macro, ##sys#strip-syntax will 
    not strip the symbol (the se contains a macro-def). Possibly just use the
    macro name, then.
*** expansion of r5rs_pitfall 4.2 incorrect
    possibly due to a non-aliased implicit "begin" somewhere
*** dirty-macros.scm loops when using `defile'
    possibly due to unrenamed special forms
*** extended lambda-lists refer to `optional' and `let-optionals[*]'
    this will break, when these macros are renamed on import

** modules
*** `require-extension' fails in interpreter when extension is .so without import library
    probably because load and import is in same toplevel form, and the
    load is performed after the import (at runtime). A possibly fix
    would be to load the extension immediately (at
    compile-/expansion-time), when import is #t and o import lib
    exists.
*** mark import-source (module) on plist of imported symbols and use in re-import warning
    doesn't work that easily, try to find another solution

** libraries
*** library/runtime: cyclic list checks for assq/assv/assoc/memq/memv/member
    and C_i_list_tail
*** finalizers on lexically ref'd data not working in interpreter (reported
    by Jim Ursetto)


* tasks

** Convert this file to org mode

** branches
*** try to improve performance in lazy-gensyms

** expander
*** test new implementation of `define-for-syntax'
**** test "numbers" egg
**** is s48-modules still working?
*** at some stage remove debug-output in expand.scm

** module issues
*** need a way to omit generating module registration without generating import lib
    for example when compiling static version, where implib already exists
*** code-duplication in compiler and evaluator for ##core#module
*** "scheme" module does not include some special forms ("define-syntax", etc.)
*** import-for-syntax imports value bindings into import-env
    should actually be a distinct meta-import-env.
    (does it work at all?)
    example: we need `(import-for-syntax chicken)' to have access to
    `receive' in a procedural syntax definition.
**** could be tested with xlib egg
*** curried define performs expansion in empty se - problem?
    (as comment in expand.scm indicated (##sys#register-export))
*** check phase separation and module access
**** see "expander" above
**** interaction for define-for-syntax, begin-for-syntax, and macros
***** check for different execution modes (compile-time, run-time, csi, ...)
***** figure out visibility

** compiler
*** test define-rewrite-rule
*** (csc) generate object-files in /tmp (or TMPDIR)?

** setup/install
*** setup-download
**** use "HTTP_PROXY"/"http_proxy" env. var
**** handle redirects in http-fetch
*** automatically update db after extension installation?
*** test sudo

** library units
*** read-mark list should be stored in read-table
*** ports should have indicator slot about exact type (tcp, file, ...)
*** normalize-pathname: return short name on windows? (mingw/msvc)

** syntax-error
*** if ##sys#current-module is set, add name to error message?
*** fluidly keep track of expanded forms (extend meaning of culprit) to pprint pruned expr on error

** build
*** need script to process import libraries for generating indices for doc.callcc.org
    then tell Toby Butzon about it
*** test DESTDIR and proper linking
*** test special status of setup-api.import.scm and setup-download.import.scm
    touching .c files caused chicken-install to be rebuilt; did using
    .SECONDARY help?
*** using "touch" with WINDOWS_SHELL won't work (need alternative)

** documentation
*** document qs, normalize-pathname


* tests

** optional longer run (env. var)
*** syntax-rules-stress-test
*** fully compiled ec-tests
** 3-stage bootstrap with compiler-output comparison


* documentation

** document new .meta entries in tutorials on wiki
   depends, test-depends
** chicken.texi needs to be regenerated


* optimizations

** self-recursion optimization
   what MacScheme called "benchmark-mode" (assume self-calls are recursion)
*** needs declaration or option, >= -O2

** when inlining, consing arg-list with "list" may make get-keyword possible foldable

** compiler-support for get-keyword ?

** lambda-fusion / "fuse-and-dispatch" (suggested by Alex Shinn)
   convert groups of local lambdas referenced to only in operator-position into
   looping lambda + dispatch (static variable can be used), otherwise similar to
   a conditional
*** new forms (after optimization, prepared language)
    [##core#dispatch LAMBDABODY1 ... BODY]
    [##core#goto {INDEX} ARGUMENT1 ...}

** lazy gensyms (see "lazy-gensyms" branch)

** handle optional args primitively 
   for example, set to distinguished val (checking argc) on C level, core
   primitive for checking this - should be quite simple, but llist processing
   (decompose-lambda-list) will be more expensive and cumbersome.

** pre-hashed symbols (extra symbol slot)
   The memory usage should be acceptable, performance gain is hard to guess.
   Some experiments indicate that hashing the string is cheaper than it appears,
   but low-level hashtables should get the most of this (and thus speed up
   the compiler)

** Lisp2-hack by Dybvig
*** add function-cell to symbol, call on direct call without checks
*** initialized to trap function on set!/define, which does fixup
*** would also allow inline-caching hacks