shithub: femtolisp

ref: 6bc3fe5f11dde87433d417b30a8f669b22642092
dir: /todo/

View raw version
* setf
* plists
* backquote
* symbol< (make < generic), generic compare function
? (cdr nil) should be nil
* multiple-argument mapcar
? multi-argument apply. for builtins, just push them. for lambdas, must
  cons together the evaluated arguments.
? option *print-shared*. if nil, it still handles circular references
  but does not specially print non-circular shared structure
? option *print-circle*
* read support for #' for compatibility
* #\c read character as code (including UTF-8 support!)
* #| |# block comments
? here-data for binary serialization. proposed syntax:
  #>size:data, e.g. #>6:000000
? better read syntax for packed arrays, e.g. #double[3 1 4]
* use syntax environment concept for user-defined macros to plug
  that hole in the semantics
* make more builtins generic. if typecheck fails, call out to the
  generic version to try supporting more types.
  compare/equal
  +-*/<       for all numeric types
  length      for all sequences
  ? aref/aset for all sequences (vector, list, c-array)
  ? copy
* fixnump, all numeric types should pass numberp
- make sure all uses of symbols don't assume symbols are unmovable without
  checking ismanaged()
* eliminate compiler warnings
* fix printing nan and inf
* move to "2.5-bit" type tags
? builtin abs()
* try adding optional arguments, (lambda (x (opt 0)) ...), see if performance
  is acceptable
* (syntax-environment) to return it as an assoc list
* (environment) for variables, constantp
* prettier printing

* readable gensyms and #:
  . #:n reads similar to #n=#.(gensym) the first time, and #n# after
* circular equal
* integer/truncate function
? car-circularp, cdr-circularp, circularp
- hashtable. plan as equal-hash, over three stages:
  1. first support symbol and fixnum keys, use ptrhash. only values get
     relocated on GC.
  2. create a version of ptrhash that uses equal() and hash(). if a key is
     inserted requiring this, switch vtable pointer to use these functions.
     both keys and values get relocated on GC.
  3. write hash() for pairs and vectors. now everything works.
- expose eq-hashtable to user
- other backquote optimizations:
  * (nconc x) => x  for any x
  . (copy-list (list|append|nconc ...)) => (list|append|nconc ...)
  * (apply vector (list ...)) => (vector ...)
  * (nconc (cons x nil) y) => (cons x y)
* let form without initializers (let (a b) ...), defaults to nil
* print (quote a) as 'a, same for ` etc.

- template keyword arguments. you write
(template (:test eq) (:key caar)
  (defun assoc (item lst)
    (cond ((atom lst) ())
          ((:test (:key lst) item) (car lst))
          (t (assoc item (cdr lst))))))

This writes assoc as a macro that produces a call to a pre-specialized
version of the function. For example
  (assoc x l :test equal)
first tries to look up the variant '(equal caar) in the dictionary for assoc.
If it doesn't exist it gets generated and stored. The result is a lambda
expression.
The macro returns ((lambda (item lst) <code for assoc>) x l).
We might have to require different syntax for template invocations inside
template definitions, such as
  ((t-instance assoc eq :key) item lst)
which passes along the same key but always uses eq.
Alternatively, we could use the keysyms without colons to name the values
of the template arguments, so the keysyms are always used as markers and
never appear to have values:
(template (:test eq) (:key caar)
  (defun assoc? (item lst)
    (cond ((atom lst) ())
          ((test (key lst) item) ...
          ...
           (assoc x y :test test :key key)
This would be even easier if the keyword syntax were something like
  (: test eq)


possible optimizations:
* delay environment creation. represent environment on the stack as
  alternating symbols/values, or if cons instead of symbol then traverse
  as assoc list. only explicitly cons the whole thing when making a closure
* cons_reserve(n) interface, guarantees n conses available without gc.
  it could even link them together for you more efficiently
* assoc builtin
* special check for constant symbol when evaluating head since that's likely
* remove the loop from cons_reserve. move all initialization to the loops
  that follow calls to cons_reserve.
- case of lambda expression in head (as produced by let), can just modify
  env in-place in tail position
- allocate memory by mmap'ing a large uncommitted block that we cut
  in half. then each half heap can be grown without moving addresses.
* try making (list ...) a builtin by moving the list-building code to
  a static function, see if vararg call performance is affected.
- try making foldl a builtin, implement table iterator as table.foldl
  . not great, since then it can't be CPS converted
* represent lambda environment as a vector (in lispv)
x setq builtin (didn't help)
* list builtin, to use cons_reserve
unconventional interpreter builtins that can be used as a compilation
target without moving away from s-expressions:
- (*global* . a)  ; special form, don't look in local env first
- (*local* . 2)   ; direct stackframe access
for internal use:
* a special version of apply that takes arguments on the stack, to avoid
  consing when implementing "call-with" style primitives like trycatch,
  hashtable-foreach, or the fl_apply API
- partial_apply, reapply interface so other iterators can use the same
  fast mechanism as for
* try this environment representation:
 for all kinds of functions (except maybe builtin special forms) push
 all arguments on the stack, either evaluated or not.
 for lambdas, push the lambda list and next-env pointers.
 to capture, save the n+2 pointers to a vector
 . this uses n+2 heap or stack words per environment instead of 2n+1 words
 . argument handling is more uniform which could lead to simplifications,
   and a more efficient apply() entry point
 . disadvantage is looking through the lambda list on every lookup. maybe
   improve by making lambda lists vectors somehow?
* fast builtin bounded iteration construct (for lo hi (lambda (x) ...))
* represent guest function as a tagged function pointer; allocate nothing
- when an instance of (array type n) is requested, use (array type)
  instead, unless the value is part of an aggregate (e.g. struct).
  . this avoids allocating a new type for every size.
  . and/or add function array.alloc
x preallocate all byte,int8,uint8 values, and some wchars (up to 0x31B7?)
  . this made no difference in a string.map microbenchmark
- use faster hash/compare in tables where the keys are eq-comparable
- a way to do open-input-string without copying

bugs:
* with the fully recursive (simpler) relocate(), the size of cons chains
  is limited by the process stack size. with the iterative version we can
  have unlimited cdr-deep structures.
* in #n='e, the case that makes the cons for 'e needs to use label fixup
* symbol token |.| does not work
* ltable realloc not multiplying by sizeof(unsigned long)
* not relocating final cdr in iterative version if it is a vector
- (setf (car x) y) doesn't return y
* reader needs to check errno in isnumtok
* prettyprint size measuring is not utf-8 correct
* stack is too limited.
  . add extra heap-allocated stack segments as needed.
* argument list length is too limited.
  need to fix it for: +,-,*,/,&,|,$,list,vector,apply,string,array
  . for builtins, make Nth argument list of rest args
  . write a function to evaluate directly from list to list, use it for
    Nth arg and for user function rest args
  . modify vararg builtins accordingly
* filter should be stable. right now it reverses.


femtoLisp3...with symbolic C interface

c values are builtins with value > N_BUILTINS
((u_int32_t*)cvalue)[0] & 0x3 must always be 2 to distinguish from vectors

typedef struct _cvtable_t {
	void (*relocate)(struct _cvalue_t *);
	void (*free)(struct _cvalue_t *);
	void (*print)(struct _cvalue_t *, FILE *);
} cvtable_t;

c type representations:
symbols void, [u]int[8,16,32,64], float, double, [u]char, [u]short,
[u]int, [u]long, lispvalue
(c-function ret-type (argtype ...))
(array type[ N])
(struct ((name type) (name type) ...))
(union ((name type) (name type) ...))
(mlayout ((name type offset) (name type offset) ...))
(enum (name1 name2 ...))
(pointer type)

constructors:
([u]int[8,16] n)
([u]int32 hi lo)
([u]int64 b3 b2 b1 b0)
(float hi lo) or (float "3.14")
(double b3 b2 b1 b0) or (double "3.14")
(array ctype val ...)
(struct ((name type) ...) val ...)
(pointer ctype)      ; null pointer
(pointer cvalue)     ; constructs pointer to the given value
                     ; same as (pointer (typeof x) x)
(pointer ctype cvalue)  ; pointer of given type, to given value
(pointer ctype cvalue addr)  ; (ctype*)((char*)cvalue + addr)
(c-function ret-type (argtype ...) ld-symbol-name)

? struct/enum tag:
  (struct 'tag <initializer>) or (pointer (struct tag))
  where tag is a global var with a value ((name type) ...)


representing c data from lisp is the tricky part to make really elegant and
efficient. the most elegant but too inefficient option is not to have opaque
C values at all and always marshal to/from native lisp values like #int16[10].
the next option is to have opaque values "sometimes", for example returning
them from C functions but printing them using their lisp representations.
the next option is to relax the idea that C values of a certain type have a
specific lisp structure, and use a coercion system that "tries" to translate
a lisp value to a specified C type. for example [0 1 2], (0 1 2),
#string[0 1 2], etc. might all be accepted by a C function taking int8_t*.
you could say (c-coerce <lispvalue> <typedesc>) and get a cvalue back or
an error if the conversion fails.

the final option is to have cvalues be the only officially-sanctioned
representation of c data, and make them via constructors, like
(int32 hi lo) returns an int32 cvalue
(struct '((name type) (name type) ...) a b ...) makes a struct
there is a constructor function for each primitive C type.
you can print these by brute force as e.g. #.(int32 hi lo)
then all checking just looks like functions checking their arguments

this option seems almost ideal. what's wrong with it?
. to construct cvalues from lisp you have to build code instead of data
. it seems like it should take more explicit advantage of tagged vectors
. should you accept multiple forms? for example
  (array 'int8 0 1 2) or (array 'int8 [0 1 2])
  if you're going to be that permissive, why not allow [0 1 2] to be passed
  directly to a function that expects int8_t* and do the conversion
  implicitly?
  . even if these c-primitive-constructor functions exist, you can still
    write things like c-coerce (in lisp, even) and hack in implicit
    conversion attempts when something other than a cvalue is passed.
. the printing code is annoying, because it's not enough to print readably,
  you have to print evaluably.
  . solution: constructor notation, #int32(hi lo)

in any case, "opaque" cvalues will not really be opaque because we want to
know their types and be able to take them apart on the byte level from lisp.
C code can get references to lisp values and manipulate them using lisp
operations like car, so to be fair it should work vice-versa; give
c references to lisp code and let it use c operations like * on them.
you can write lisp in c and c in lisp, though of course you don't usually
want to. however, c written in lisp can be generated by a macro, printed,
and fed to TCC for compilation.


for a struct the names and types are parameters of the type, not the
constructor, so it seems more correct to do

((struct (name type) (name type) ...) (val val ...))

where struct returns a constructor. but this isn't practical because it
can't be printed in constructor notation and the type is a lambda rather
than a more sensible expression.


notice constructor calls and type representations are "similar". they
should be related formally:

(define (new type)
  (if (symbolp type) (apply (eval type) ())
    (apply (eval (car type)) (cdr type))))

NOTE: this relationship is no longer true. we don't want to have to
construct 1 cvalue from 1 lisp value every time, since that could
require allocating a totally redundant list or vector. it should be
possible to make a cvalue from a series of lisp arguments. for
example there are now 2 different ways to make an array:

1) from series of arguments: (array type val0 val1 ...)
2) from 1 (optional) value: (c-value '(array int8[ size])[ V])

constructors will internally use the second form to initialize elements
of aggregates. e.g. 'array' in the first case will conceptually call
  (c-value type val0)
  (c-value type val1)
  ...


for aggregate types, you can keep a variable referring to the relevant
piece:

(setq point '((x int) (y int)))
(struct point 2 3)   ; looks like c declaration 'struct point x;'

a type is a function, so something similar to typedef is achieved by:

(define (point_t vals) (struct point vals))

design points:
. type constructors will all be able to take 1 or 0 arguments, so i could say
  (new (typeof val))   ; construct similar
  (define (new type)
    (if (symbolp type) (apply (eval type) ())
      (apply (eval (car type)) (cdr type))))
. values can be marked as autorelease (1) if user says so, (2) if we can
  prove that it's ok (e.g. we only allocated the value using malloc because
  it is too large to move on every GC).
  in the future you should be able to specify an arbitrary finalization
  function, not just free().
. when calling a C function, a value of type_t can be passed to something
  expecting a type_t* by taking the address of the representation. BUT
  this is dangerous if the C function might save a reference.
  a type_t* can be passed as a type_t by copying the representation.
. you can use (pointer v) to switch v to "malloc'd representation", in
  which case the value is no longer autoreleased, but you can do whatever
  you want with the pointer. (other option is to COPY v when making a
  pointer to it, but this still doesn't prevent C from holding a reference
  too long)


add a cfunction binding to symbols. you register in C simply by setting
this binding to a function pointer, then

(defun open (path flags)
    ; could insert type checks here
    (ccall 'int32 'open path flags))

(setq fd (open "path" 0))

using libdl you could even omit the registration step and extra binding

this is possible:
(defun malloc (size)
    (ccall `(array int8 ,size) 'malloc   size))
           ;ret type          ;f name   ; . args


vtable:
we'd like to be able to define new lisp "types", like vectors
and hash tables, using this. there needs to be a standard value interface
you can implement in C and attach a vtable to some c values.
interface: relocate, finalize, print(, copy)

implementation plan:
- write cvalue constructors
- if a head evaluates to a cvalue, call the pointer directly with the arg array
  . this is the "guest function" interface, a C function written specifically
    to the femtolisp API. its type must be
    '(c-function lispvalue ((pointer lispvalue) uint32))
    which corresponds to
    value_t func(value_t *args, u_int32_t nargs);
  . this interface is useful for writing additional builtins, types,
    interpreter extensions, etc. more efficient.
  . one of these functions could also be called with
    (defun func args
      (ccall 'func 'lispvalue (array 'lispvalue args) (length args)))
  - these functions are effectively builtins and should have names so they
    can be printed as such.
    . have a registration function
      void guest_function(value_t (*f)(value_t*,u_int32_t), const char *name);
      so at least the function type can be checked from C
    . set a flags bit for functions registered this way so we can identify
      them quickly

- ccall lisp builtin, (ccall rettype name . args). if name has no cfunc
  binding, looks it up lazily with dlsym and stores the result.
  this is a guest function that handles type checking, translation, and
  invocation of foreign c functions.

- you could register builtins from lisp like this:
  (defun dlopen (name flags) (ccall '(pointer void) 'dlopen name flags))
  (defun dlsym (handle name type) (ccall type 'dlsym handle name))
  (define lisp-process (dlopen nil 0))
  (define vector-sym
    (dlsym lisp-process 'int_vector
           '(function lispvalue (pointer lispvalue) uint32)))
  (ccall 'void 'guest_function vector-sym 'vector)

- write c extensions cref, cset, typeof, sizeof, cvaluep
* read, print, vectorp methods for vectors
- quoted string "" reading, produces #(c c c c ...)
* get rid of primitive builtins read,print,princ,load,exit,
  implement using ccall


other possible design:
- just add two builtins, call and ccall.
  (call 'name arg arg arg)  lisp guest function interface
  we can say e.g.
  (defmacro vector args `(call 'vector ,.args))
- basically the question is whether to introduce a new kind of callable
  object or to do everything through the existing builtin mechanism
  . macros cannot be applied, so without a new kind of callable 'vector'
    would have to be a lisp function, entailing argument consing...
  (defun builtin (name)
    (guest-function name
      (dlsym lisp-process name '(function value (pointer value) uint32))))
  then you can print a guest function as e.g.
  #.(builtin 'vector)

#name(x y z) reads as a tagged vector
#(x y z) is the same as #vector(x y z)
should be internally the same as well, so non-taggedness does not formally
exist.


then we can write the vector clause in backquote as e.g.

(if (vectorp x)
    (let ((body (bq-process (vector-to-list x))))
      (if (eq (tag x) 'vector)
          (list 'list-to-vector body)
        (list 'apply 'tagged-vector
              (list cons (list quote (tag x)) body))))
  (list quote x))


setup plan:
* create source directory and svn repository, move llt sources into it
* write femtolisp.h, definitions for extensions to #include
- add fl_ prefix to all exported functions
* port read and print to llt iostreams
* get rid of flutils; use ptrhash instead
* builtinp needs to be a builtin ;) to distinguish lisp builtins from cvalues
* allocation and gc for cvalues
- interface functions fl_list(...), fl_apply
  e.g. fl_apply(fl_eval(fl_symbol("+")), fl_list(fl_number(2),fl_number(3)))
  and fl_symval("+"), fl_cons, etc.

-----------------------------------------------------------------------------

vector todo:
* compare for vectors
- (aref v i j k) does (reduce aref v '(i j k)); therefore (aref v) => v
- (aref v ... [1 2 3] ...) vectorized indexing
- make (setf (aref v i j k) x) expand to (aset (aref v i j) k x)
these should be done using the ccall interface:
- concatenate
- copy-vec
- (range i j step) to make integer ranges
- (rref v start stop), plus make it settable! (rset v start stop rhs)
lower priority:
- find (strstr)

functions to be generic over vec/list:
* compare, equal, length

constructor notation:

#func(a b c)  does (apply func '(a b c))

-----------------------------------------------------------------------------

how we will allocate cvalues

a vector's size will be a lisp-value number. we will set bit 0x2 to indicate
a resize request, and bit 0x1 to indicate that it's actually a cvalue.

every cvalue will have the following fields, followed by some number of
words according to how much space is needed:

    value_t size;  // | 0x2
    cvtable_t *vtable;
    struct {
#ifdef BITS64
        unsigned pad:32;
#endif
        unsigned whatever:27;
        unsigned mark:1;
        unsigned hasparent:1;
        unsigned islispfunction:1;
        unsigned autorelease:1;
        unsigned inlined:1;
    } flags;
    value_t type;
    size_t len;      // length of *data in bytes
    //void *data;      // present if !inlined
    //value_t parent;  // present if hasparent

size/vtable have the same meaning as vector size/elt[0] for relocation
obviously we only relocate parent and type. if vtable->relocate is present,
we call it at the end of the relocate process, and it must touch every
lisp value reachable from it.

when a cvalue is created with a finalizer, its address is added to a special
list. before GC, everything in that list has its mark bit set. when
we relocate a cvalue, clear the bit. then go through the list to call
finalizers on dead values. this is O(n+m) where n is amt of live data and m
is # of values needing finalization. we expect m << heapsize.

-----------------------------------------------------------------------------

Goal: bootstrap a lisp system where we can do "anything" purely in lisp
starting with the minimal builtins needed for successive levels of
completeness:

1. Turing completeness
quote, if, lambda, eq, atom, cons, car, cdr

2. Naming
set

3. Control flow
progn, prog1, apply, eval
call/cc needed for true completeness, but we'll have attempt, raise

4. Predicate completeness
symbolp, numberp, builtinp

5. Syntax
macro

6. I/O completeness
read, print

7. Mutable state
rplaca, rplacd

8. Arithmetic completeness
+, -, *, /, <

9. The missing data structure(s): vector
alloc, aref, aset, vectorp, length

10. Real-world completeness (escape hatch)
ccall

---
11. Misc unnecessary
while, label, cond, and, or, not, boundp, vector

-----------------------------------------------------------------------------

exception todo:

* silence 'in file' errors when user frame active
* add more useful data to builtin exception types:
  (UnboundError x)
  (BoundsError vec index)
  (TypeError fname expected got)
  (Error v1 v2 v3 ...)
* attempt/raise, rewrite (error) in lisp
* more intelligent exception printers in toplevel handler

-----------------------------------------------------------------------------

lisp variant ideas

- get rid of separate predicates and give every value the same structure
  ala mathematica
  . (tag 'a) => symbol
    (tag '(a b)) => a
    (tag 'symbol 'a) => a
    (tag 'blah 3) => (blah 3)
- have only vectors, not cons cells (sort of like julia)
  . could have a separate tag field as above

- easiest way to add vectors:
  . allocate in same heap with conses, have a tag, size, then elements
    (each elt must be touched on GC for relocation anyway, so might as well
     copy collect it)
  . tag pointers as builtins, we identify them as builtins with big values
  . write (vector) in C, use it from read and eval

8889314663  comcast net #

-----------------------------------------------------------------------------

cvalues reserves the following global symbols:

int8, uint8, int16, uint16, int32, uint32, int64, uint64
char, uchar, wchar, short, ushort, int, uint, long, ulong
float, double
struct, array, enum, union, function, void, pointer, lispvalue

it defines (but doesn't reserve) the following:

typeof, sizeof, autorelease, guestfunction, ccall


user-defined types and typedefs:

the rule is that a type should be viewed as a self-evaluating constant
like a number. if i define a complex_t type of two doubles, then
'complex_t is not a type any more than the symbol 'x could be added to
something just because it happened to have the value 2.

; typedefs from lisp
(define wchar_t 'uint32)
(define complex_t '(struct ((re double) (im double))))

; use them
(new complex_t)
(new `(array ,complex_t 10))
(array complex_t 10)

BUT

(array 'int32 10)

because the primitive types *are* symbols. the fact that they have values is
just a convenient coincidence that lets you do e.g. (int32 0)


; size-annotate a pointer
(setq p  (ccall #c-function((pointer void) (ulong) malloc) n)
(setq a (deref p `(array int8 ,n)))

cvalues todo:

* use uint32_t instead of wchar_t in C code
- make sure empty arrays and 0-byte types really work
* allow int constructors to accept other int cvalues
* array constructor should accept any cvalue of the right size
* make sure cvalues participate well in circular printing
* float, double
- struct, union (may want to start with more general layout type)
- pointer type, function type
* finalizers
- functions autorelease, guestfunction
- cref/cset/byteref/byteset
* wchar type, wide character strings as (array wchar)
* printing and reading strings
- ccall
- anonymous unions
* fix princ for cvalues
* make header size for primitives <= 8 bytes, even on 64-bit arch
- more efficient read for #array(), so it doesn't need to build a pairlist
? lispvalue type
  . keep track of whether a cvalue leads to any lispvalues, so they can
    be automatically relocated (?)

* string constructor/concatenator:
(string 'sym #char(65) #wchar(945) "blah" 23)
   ; gives "symA\u03B1blah23"
"ccc"  reads to (array char)

low-level functions:
; these are type/bounds-checked accesses
- (cref cvalue key)         ; key is field name or index. access by reference.
- (aref cvalue key)         ; access by value, returns fixnums where possible
- (cset cvalue key value)   ; key is field name, index, or struct offset
  . write&use conv_from_long to put fixnums into typed locations
  . aset is the same
* (copy cv)
- (offset type|cvalue field [field ...])
- (eltype type field [field ...])
- (memcpy dest-cv src-cv)
- (memcpy dest doffs src soffs nbytes)
- (bswap cvalue)
- (c2lisp cvalue)  ; convert to sexpr form
* (typeof cvalue)
* (sizeof cvalue|type)
- (autorelease cvalue)     ; mark cvalue as free-on-gc
- (deref pointer[, type])  ; convert an arbitrary pointer to a cvalue
                           ; this is the unsafe operation

; (sizeof '(pointer type)) == sizeof(void*)
; (sizeof '(array type N)) == N * sizeof(type)

(define (reinterpret-cast cv type)
  (if (= (sizeof cv) (sizeof type))
      (deref (pointer 'void cv) type)
      (error "Invalid cast")))

a[n].x looks like (cref (cref a n) 'x), (reduce cref head subs)

things you can do with cvalues:

. call native C functions from lisp code without wrappers
. wrap C functions in pure lisp, automatically inheriting some degree
  of type safety
. use lisp functions as callbacks from C code
. use the lisp garbage collector to reclaim malloc'd storage
. annotate C pointers with size information for bounds checking
. attach symbolic type information to a C data structure, allowing it to
  inherit lisp services such as printing a readable representation
. add datatypes like strings to lisp
. use more efficient represenations for your lisp programs' data


family of cvalue representations.
relevant attributes:
  . large   -- needs full size_t to represent size
  . inline  -- allocated along with metadata
  . prim    -- no stored type; uses primtype bits in flags
  . hasdeps -- depends on other values to stay alive

these attributes have the following dependencies:
  . large -> !inline
  . prim -> !hasdeps && !large

so we have the following possibilities:

large   inline   prim   hasdeps   rep#
  0        0       0       0       0
  0        0       0       1       1

  0        0       1       0       2
  0        1       0       0       3
  0        1       0       1       4
  0        1       1       0       5

  1        0       0       0       6
  1        0       0       1       7

we need to be able to un-inline data, so we need:
change 3 -> 0  (easy; write pointer over data)
change 4 -> 1
change 5 -> 2  (also easy)


rep#0&1:   (!large && !inline && !prim)
typedef struct {
    cvflags_t flags;
    value_t type;
    value_t deps;
    void *data;   /* points to malloc'd buffer */
} cvalue_t;

rep#3&4:   (!large &&  inline && !prim)
typedef struct {
    cvflags_t flags;
    value_t type;
    value_t deps;
    /* data goes here inlined */
} cvalue_t;


rep#2:     (prim && !inline)
typedef struct {
    cvflags_t flags;
    void *data;   /* points to (tiny!) malloc'd buffer */
} cvalue_t;

rep#5:     (prim &&  inline)
typedef struct {
    cvflags_t flags;
    /* data goes here inlined */
} cvalue_t;


rep#6&7:   (large)
typedef struct {
    cvflags_t flags;
    value_t type;
    value_t deps;
    void *data;   /* points to malloc'd buffer */
    size_t len;
} cvalue_t;

-----------------------------------------------------------------------------

times for lispv:

color 2.286s
sort  0.181s
fib34 5.205s
mexpa 0.329s

-----------------------------------------------------------------------------

finalization algorithm that allows finalizers written in lisp:

right after GC, go through finalization list (a weak list) and find objects
that didn't move. relocate them (bring them back to life) and push them
all onto the stack. remove all from finalization list.

call finalizer for each value.

optional: after calling a finalizer, make sure the object didn't get put
back on the finalization list, remove if it did.
if you don't do this, you can make an unkillable object by registering a
finalizer that re-registers itself. this could be considered a feature though.

pop dead values off stack.


-----------------------------------------------------------------------------

femtolisp semantics

eval* is an internal procedure of 2 arguments, expr and env, invoked
implicitly on input.
The user-visible procedure eval performs eval*  e  Env ()

eval*  Symbol s  E     =>  lookup* s E
eval*  Atom a  E       =>  a
... special forms ... quote arg, if a b c, other symbols from syntax env.
eval*  Cons f args  E  =>

First the head expression, f, is evaluated, yielding f-.
Then control is passed to #.apply f- args
  #.apply is the user-visible apply procedure.
  (here we imagine there is a user-invisible environment where f- is
   bound to the value of the car and args is bound to the cdr of the input)


Now (apply b lst) where b is a procedure (i.e. satisfies functionp) is
identical to
(eval (map (lambda (e) `',e) (cons b lst)))

-----------------------------------------------------------------------------

design of new toplevel

system.lsp contains definitions of (load) and (toplevel) and is loaded
from *install-dir* by a bootstrap loader in C. at the end of system.lsp,
we check whether (load) is builtin. if it is, we redefine it and reload
system.lsp with the new loader. the C code then invokes (toplevel).

(toplevel) either runs a script or a repl using (while T (trycatch ...))

(load) reads and evaluates every form, keeping track of defined functions
and macros (at the top level), and grabs a (main ...) form if it sees
one. it applies optimizations to every definition, then invokes main.

an error E during load should rethrow `(load-error ,filename ,E)
such exceptions can be printed recursively

lerror() should make a lisp string S from the result of sprintf, then
raise `(,e ,S). first argument e should be a symbol.


new expansion process:

get rid of macroexpanding versions of define and define-macro
macroexpand doesn't expand (define ...)
  macroexpand implements let-syntax
add lambda-expand which applies f-body to the bodies of lambdas, then
  converts defines to set!
call expand on every form before evaluating
  (define (expand x) (lambda-expand (macroexpand x)))
(define (eval x) (%eval (expand x)))
reload system.lsp with the new eval

-----------------------------------------------------------------------------

String API

*string         - append/construct
*string.inc     - (string.inc s i [nchars])
*string.dec
*string.count   - # of chars between 2 byte offsets
*string.char    - char at byte offset
*string.sub     - substring between 2 byte offsets
*string.split   - (string.split s sep-chars)
*string.trim    - (string.trim s chars-at-start chars-at-end)
*string.reverse
*string.find    - (string.find s str|char [offs]), or nil if not found
 string.rfind
*string.encode  - to utf8
*string.decode  - from utf8 to UCS
*string.width   - # columns
*string.map     - (string.map f s)


IOStream API

*read             - (read[ stream]) ; get next sexpr from stream
*print
*princ
*file
 iostream         - (stream[ cvalue-as-bytestream])
*buffer
 fifo
 socket
*io.eof?
*io.flush
*io.close
*io.discardbuffer
*io.write     - (io.write s cvalue [start [count]])
*io.read      - (io.read s ctype [len])
*io.getc      - get utf8 character
*io.putc
 io.peekc
*io.readline
*io.readuntil
*io.copy      - (io.copy to from [nbytes])
*io.copyuntil - (io.copy to from byte)
 io.pos       - (io.pos s [set-pos])
 io.seek      - (io.seek s offset)
 io.seekend   - move to end of stream
 io.trunc
 io.read!     - destructively take data
*io.tostring!
*io.readlines
*io.readall
*print-to-string
*princ-to-string


*path.exists?
 path.dir?
 path.combine
 path.parts
 path.absolute
 path.simplify
 path.tempdir
 path.tempname
 path.homedir
*path.cwd


*time.now
 time.parts
 time.fromparts
*time.string
*time.fromstring


*os.name
*os.getenv
*os.setenv
 os.execv


*rand
*randn
*rand.uint32
*rand.uint64
*rand.double
*rand.float

-----------------------------------------------------------------------------

  * new print algorithm
     1. traverse & tag all conses to be printed. when you encounter a cons
        that is already tagged, add it to a table to give it a #n# index
     2. untag a cons when printing it. if cons is in the table, print
        "#n=" before it in the car, " . #n=" in the cdr. if cons is in the
        table but already untagged, print #n# in car or " . #n#" in the cdr.
  * read macros for #n# and #n= using the same kind of table
    * also need a table of read labels to translate from input indexes to
      normalized indexes (0 for first label, 1 for next, etc.)
  * read macro #. for eval-when-read. use for printing builtins, e.g. "#.eq"

-----------------------------------------------------------------------------

prettyprint notes

* if head of list causes VPOS to increase and HPOS is a bit large, then
switch to miser mode, otherwise default is ok, for example:

> '((lambda (x y) (if (< x y) x y)) (a b c) (d e f) 2 3 (r t y))
((lambda (x y)
   (if (< x y) x y)) (a b c)
                     (d e f) 2 3
                     (r t y))

* (if a b c) should always put newlines before b and c

* write try_predict_len that gives a length for easy cases like
  symbols, else -1. use it to avoid wrapping symbols around lines

* print defun, defmacro, label, for more like lambda (2 spaces)

* *print-pretty* to control it

* if indent gets too large, dedent back to left edge

-----------------------------------------------------------------------------

consolidated todo list as of 7/8:
* new cvalues, types representation
* use the unused tag for TAG_PRIM, add smaller prim representation
* finalizers in gc
* hashtable
* generic aref/aset
* expose io stream object
* new toplevel

* make raising a memory error non-consing
* eliminate string copy in lerror() when possible
* fix printing lists of short strings

* evaluator improvements, perf & debugging (below)
* fix make-system-image to save aliases of builtins
* reading named characters, e.g. #\newline etc.
- #+, #- reader macros
- printing improvements: *print-length*, keep track of horiz. position
  per-stream so indenting works across print calls
- remaining c types
- remaining cvalues functions
- finish ios
* optional arguments
* keyword arguments
- some kind of record, struct, or object system
- improve test coverage

expansion process bugs:
* expand default expressions for opt/keyword args (as if lexically in body)
* make bound identifiers (lambda and toplevel) shadow macro keywords
* to expand a body:
  1. splice begins
  2. add defined vars to env
  3. expand nondefinitions in the new env
     . if one expands to a definition, add the var to the env
  4. expand RHSes of definitions
- add different spellings for builtin versions of core forms, like
  $begin, $define, and $set!. they can be replaced when found during expansion,
  and used when the compiler needs to generate them with known meanings.

- special efficient reader for #array
- reimplement vectors as (array lispvalue)
- implement fast subvectors and subarrays

-----------------------------------------------------------------------------

cvalues redesign

goals:
. allow custom types with vtables
. use less space, share types more
. simplify access to important metadata like length
. unify vectors and arrays

typedef struct {
    fltype_t *type;
    void *data;
    size_t len;      // length of *data in bytes
    union {
        value_t parent;    // optional
        char _space[1];    // variable size
    };
} cvalue_t;

#define owned(cv)      ((cv)->type & 0x1)
#define hasparent(cv)  ((cv)->type & 0x2)
#define isinlined(cv)  ((cv)->data == &(cv)->_space[0])
#define cv_class(cv)   ((fltype_t*)(((uptrint_t)(cv)->type)&~3))
#define cv_type(cv)    (cv_class(cv)->type)
#define cv_len(cv)     ((cv)->len)
#define cv_data(cv)    ((cv)->data)
#define cv_numtype(cv) (cv_class(cv)->numtype)

typedef struct _fltype_t {
    value_t type;
    int numtype;
    size_t sz;
    size_t elsz;
    cvtable_t *vtable;
    struct _fltype_t *eltype;  // for arrays
    struct _fltype_t *artype;  // (array this)
    int marked;
} fltype_t;

-----------------------------------------------------------------------------

new evaluator todo:

* need builtin = to handle nans properly, fix equal? on nans
* builtin quasi-opaque function type
  fields: signature, maxstack, bcode, vals, cloenv
  function->vector
* make (for ...) a special form
* trycatch should require 2nd arg to be a lambda expression
* immediate load int8 instruction
* unlimited lambda lists
  . need 32-bit argument versions of loada, seta, loadc, setc
  . largs instruction to move args after MAX_ARGS from list to stack
* maxstack calculation, make Stack growable
  * stack traces and better debugging support
* improve internal define
* try removing MAX_ARGS trickery
? apply optimization, avoid redundant list copying calling vararg fns
- let eversion
- variable analysis - avoid holding references to values in frames
  captured by closures but not used inside them
* lambda lifting
* let optimization
* fix equal? on functions
* store function name
* have macroexpand use its own global syntax table
* be able to create/load an image file
* fix trace and untrace
* opcodes LOADA0, LOADA1, LOADC00, LOADC01
- opcodes CAAR, CADR, CDAR, CDDR
- EQTO N, compare directly to stored datum N
- peephole opt
  done:
  not brf => brt
  eq brf => brne
  null brf => brnn
  null brt => brn
  null not brf => brn
  cdr car => cadr

  not yet:
  not brt => brf
  constant+pop => nothing, e.g. 2-arg 'if' in statement position
  loadt+brf => nothing
  loadf+brt => nothing
  loadt+brt => jmp
  loadf+brf => jmp

-----------------------------------------------------------------------------

new stack organization:

func
arg1
...
argn
cloenv         |
prev           |
nargs	       |
ip	       |
captured       | 

to call:
push func and arguments
args[nargs+3] = ip    // save my state in my frame
assign nargs
goto top

on entry:
push cloenv
push curr_frame  (a global initialized to 0)
push nargs
SP += 1
curr_frame = SP

to return:
v = POP();
SP = curr_frame
curr_frame = Stack[SP-4]
if (args == top_args) return v;
SP -= (5+nargs);
move Stack[curr_frame-...] back into locals
Stack[SP-1] = v
goto next_op

to relocate stack:
for each segment {
  curr_top = SP
  f = curr_frame
  while (1) {
    for i=f, i<curr_top, i++
      relocate stack[i]
    if (f == 0) break;
    curr_top = f - 4
    f = stack[f - 4]
  }
}

typedef struct {
    value_t *Stack;
    uint32_t size;
    uint32_t SP;
    uint32_t curr_frame;
} stackseg_t;

-----------------------------------------------------------------------------

optional and keyword args:

check nargs >= #required
grow frame by ntotal-nargs   ; ntotal = #req+#opt+#kw
(sort keyword args into their places)
branch if arg bound around initializer for each opt arg

example: (lambda (a (b 0) (c b)))

minargs 1
framesize 3
brbound 1 L1
load0
seta 0
L1:
brbound 2 L2
loada 1
seta 2
L2:

-----------------------------------------------------------------------------

what needs more test coverage:

- more error cases, lerrorf() cases
- printing gensyms
- gensyms with bindings
- listn(), isnumber(), list*, boolean?, function?, add2+ovf, >2arg add,div
- large functions, requiring long versions of branch opcodes
- setal, loadvl, (long arglist and lots of vals cases)
- aref/aset on c array
- printing everything
- reading floats, escaped symbols, multiline comment, octal chars in strs
- equal? on functions
- all cvalue ctors, string_from_cstrn()
- typeof, copy, podp, builtin()
- bitwise and logical ops
- making a closure in a default value expression for an optional arg
- gc during a catch block, then get stack trace

-----------------------------------------------------------------------------

5/4/10 todo:

- flush and close open files on exit
* make function versions of opcode builtins by wrapping in a lambda,
  stored in a table indexed by opcode. use in _applyn