retroforth/literate/Rx.md
crc d7ae8b1497 bump version number
FossilOrigin-Name: 6876a8d27a690f634cd8b81e77050167b1c09ef3da0f390107c2cf6f0a2cd8d3
2019-01-11 03:11:15 +00:00

30 KiB

____  _   _
|| \\ \\ //
||_//  )x(
|| \\ // \\ 2019.6
a minimalist forth for nga

Rx (retro experimental) is a minimal Forth implementation for the Nga virtual machine. Like Nga this is intended to be used within a larger supporting framework adding I/O and other desired functionality.

General Notes on the Source

Rx is developed using a literate tool called unu. This allows easy extraction of fenced code blocks into a separate file for later compilation. I've found the use of a literate style to be very beneficial as it makes it easier for me to keep the code and commentary in sync, and helps me to approach development in a more structured manner.

This source is written in Muri, an assembler for Nga.

Before going on, I should explain a bit about Nga and Muri.

Nga provides a MISC inspired virtual machine for a dual stack architecture. There are 27 instructions, with up to four packed into each memory location (cell). The instructions are:

0  nop        7  jump      14  gt        21  and
1  lit <v>    8  call      15  fetch     22  or
2  dup        9  ccall     16  store     23  xor
3  drop      10  return    17  add       24  shift
4  swap      11  eq        18  sub       25  zret
5  push      12  neq       19  mul       26  end
6  pop       13  lt        20  divmod

I won't explain them here, but if you're familiar with Forth, it should be pretty easy to figure out.

Packing of instructions lets me save space, but does require a little care. Instructions that modify the instruction pointer should be followed by NOP. These are: JUMP, CALL, CCALL, RETURN, and ZRET. Additionally, if the instruction bundle contains a LIT, a value must be in the following cell. (One for each LIT in the bundle)

The reason for this relates to how Nga processes the opcodes. To illustrate, assume a stack with a couple of values:

#1 #2

And a function that consumes two values before returning a some new ones:

:function * #3 ;

If we were to use an instruction bundle like:

lit call add nop
function

Nga will:

(1) push a pointer to the function to the stack
(2) setup a call to the function
(3) add the top values on the stack (#1 #2),
    leaving a single value (#3)
(4) do nothing

At this point the bundle is done, so control goes to the called function. But we now have only one value on the stack, so the stack underflows and Nga will crash.

It's not forbidden, and this can be useful to improve code density, but exercise caution and be sure to keep track of this behavior to avoid hard to identify bugs.

Muri uses the first two characters of each instruction name when composing the bundles, with NOP being named as two dots.

So:

lit lit add nop

Is a bundle named:

liliad..

And with two li instructions, must be followed by two values.

Muri uses a directive in the first line to tell it what to expect. Directives are:

i   instruction bundle
d   decimal value
r   reference to label
:   label
s   zero terminated string

In the Beginning...

Nga expects code to start with a jump to the main entry point. Rx doesn't really have a main entry point (the top level loop is assumed to be part of the interface layer), but I allocate the space for a jump here anyway. This makes it possible to patch the entry point later, if using an interface that adds the appropriate I/O functionality.

i liju....
d -1

With this, it's time to allocate some data elements. These are always kept in known locations after the initial jump to ensure that they can be easily identified and interfaced with external tools. This is important as Nga allows for a variety of I/O models to be implemented and I don't want to tie Rx into any one specific model.

Here's the initial memory map:

Offset Contains
0 lit call nop nop
1 Pointer to main entry point
2 Dictionary
3 Heap
4 RETRO version
: Dictionary
r 9999

: Heap
d 1536

: Version
d 201906

Both of these are pointers. Dictionary points to the most recent dictionary entry. (See the Dictionary section at the end of this file.) Heap points to the next free address. This is hard coded to an address beyond the end of the Rx kernel. I adjust this as needed if the kernel grows or shinks significantly. See the Interpreter & Compiler section for more on this.

Nga Instruction Set

As mentioned earlier, Nga provides 27 instructions. Rx begins the actual coding by assigning each to a separate function. These are not intended for direct use; the compiler will fetch the opcode values to use from these functions when compiling. Many will also be exposed in the initial dictionary.

: _nop
d 0
i re......

: _lit
d 1
i re......

: _dup
d 2
i re......

: _drop
d 3
i re......

: _swap
d 4
i re......

: _push
d 5
i re......

: _pop
d 6
i re......

: _jump
d 7
i re......

: _call
d 8
i re......

: _ccall
d 9
i re......

: _ret
d 10
i re......

: _eq
d 11
i re......

: _neq
d 12
i re......

: _lt
d 13
i re......

: _gt
d 14
i re......

: _fetch
d 15
i re......

: _store
d 16
i re......

: _add
d 17
i re......

: _sub
d 18
i re......

: _mul
d 19
i re......

: _divmod
d 20
i re......

: _and
d 21
i re......

: _or
d 22
i re......

: _xor
d 23
i re......

: _shift
d 24
i re......

: _zret
d 25
i re......

: _end
d 26
i re......

Though Nga allows for multiple instructions to be packed into a single memory location (called a cell), Rx only packs a few specific combinations.

Since calls and jumps take a value from the stack, a typical call (in Muri assembly) would look like:

i lica....
r bye

Without packing this takes three cells: one for the lit, one for the address, and one for the call. Packing drops it to two since the lit/call combination can be fit into a single cell. Likewise, I use a packed jump for use with quotations. These saves several hundred cells (and thus fetch/decode cycles) when loading the standard library.

The raw values for these are:

2049  lica....
1793  liju....

These are hardcoded in a few places later. I had previously used a lookup, but this proved costly in processing time, so hard coding proved better. (These places are clearly marked)

Memory

Memory is a big, flat, linear array. The addressing starts at zero and counts upwards towards a fixed upper limit (set by the VM).

The basic memory accesses are handled via fetch and store.

The next two functions provide easier access to sequences of data by fetching or storing a value and returning the next address.

fetch-next takes an address and fetches the stored value. It returns the next address and the stored value.

: fetch-next
i duliadsw
d 1
i fere....

store-next takes a value and an address. It stores the value to the address and returns the next address.

: store-next
i duliadpu
d 1
i stpore..

Conditionals

The Rx kernel provides three conditional forms:

flag true-pointer false-pointer choose
flag true-pointer   if
flag false-pointer -if

choose is a conditional combinator which will execute one of two functions, depending on the state of a flag. I use a little hack here. I store the pointers into a jump table with two fields, and use the flag as the index. Defaults to the false entry, since a true flag is -1.

Note that this requires that the flags be -1 (for TRUE) and 0 (for FALSE). It's possible to make this more flexible, but at a significant performance hit, so I'm leaving it this way.

: choice:true
d 0

: choice:false
d 0

: choose
i listlist
r choice:false
r choice:true
i liadfeca
r choice:false
i re......

Next the two if forms. Note that -if falls into if. This saves two cells of memory.

: -if
i pulieqpo
d 0
: if
i cc......
i re......

Strings

The kernel needs two basic string operations for dictionary searches: obtaining the length and comparing for equality.

Strings in Rx are zero terminated. This is a bit less elegant than counted strings, but the implementation is quick and easy.

First up, string length. The process here is trivial:

  • Make a copy of the starting point

  • Fetch each character, comparing to zero

    • If zero, break the loop
    • Otherwise discard and repeat
  • When done subtract the original address from the current one

  • Then subtract one (to account for the zero terminator)

: count
i lica....
r fetch-next
i zr......
i drliju..
r count

: s:length
i dulica..
r count
i lisuswsu
d 1
i re......

String comparisons are harder. In high level code this is:

dup fetch push n:inc swap dup fetch push n:inc pop dup pop -eq? [ drop-pair drop #0 pop pop drop drop ] [ 0; drop s:eq? pop pop drop drop ] choose drop-pair #-1 ;

I've rewritten this a few times. The current implementation is fast enough, and not overly long. It may be worth looking into a hash based comparsion in the future.

: mismatch
i drdrdrli
d 0
i popodrdr
i re......

: matched
i zr......
i drlica..
r s:eq
i popodrdr
i re......

: s:eq
i dufepuli
d 1
i adswdufe
i puliadpo
d 1
i duponeli
r mismatch
i lilica..
r matched
r choose
i drdrlire
d -1

Interpreter & Compiler

Compiler Core

The heart of the compiler is comma which stores a value into memory and increments a variable (Heap) pointing to the next free address.

: comma
i lifelica
r Heap
r store-next
i listre..
r Heap

I also add a couple of additional forms. comma:opcode is used to compile VM instructions into the current defintion. This is where those functions starting with an underscore come into play. Each wraps a single instruction. Using this I can avoid hard coding the opcodes.

This performs a jump to the comma word instead of using a call/ret to save a cell and slightly improve performance. I will use this technique frequently.

: comma:opcode
i feliju..
r comma

comma:string is used to compile a string into the current definition. As with comma:opcode, this uses a jump to eliminate the final tail call.

: ($)
i lica....
r fetch-next
i zr......
i lica....
r comma
i liju....
r ($)

: comma:string
i lica....
r ($)
i drliliju
d 0
r comma

With the core functions above it's now possible to setup a few more things that make compilation at runtime more practical.

First, a variable indicating whether we should compile or run a function. In traditional Forth this would be STATE; I call it Compiler.

This will be used by the word classes.

: Compiler
d 0

Next is semicolon; which compiles the code to terminate a function and sets the Compiler to an off state (0). This just needs to compile in a RET.

: t-;
i lilica..
r _ret
r comma:opcode
i lilistre
d 0
r Compiler

Word Classes

Rx is built over the concept of word classes. Word classes are a way to group related words, based on their compilation and execution behaviors. A class handler function is defined to handle an execution token passed to it on the stack.

Rx provides several classes with differing behaviors:

class:data provides for dealing with data structures.

interpret compile
leave value on stack compile value into definition
: class:data
i lifezr..
r Compiler
i drlilica
r _lit
r comma:opcode
i liju....
r comma

class:word handles most functions.

interpret compile
call a function compile a call to a function
: class:word:interpret
i ju......

: class:word:compile
i lilica..
d 2049
r comma
i liju....
r comma

: class:word
i lifelili
r Compiler
r class:word:compile
r class:word:interpret
i liju....
r choose

class:primitive is a special class handler for functions that correspond to Nga instructions.

interpret compile
call the function compile an instruction
: class:primitive
i lifelili
r Compiler
r comma:opcode
r class:word:interpret
i liju....
r choose

class:macro is the class handler for compiler macros. These are functions that always get called. They can be used to extend the language in interesting ways.

interpret compile
call the function call the function
: class:macro
i ju......

The class mechanism is not limited to these classes. You can write custom classes at any time. On entry the custom handler should take the XT passed on the stack and do something with it. Generally the handler should also check the Compiler state to determine what to do in either interpretation or compilation.

Dictionary

Rx has a single dictionary consisting of a linked list of headers. The current form of a header is shown in the chart below.

field holds accessor
link link to the previous entry d:link
xt link to start of the function d:xt
class link to the class handler function d:class
name zero terminated string d:name

The initial dictionary is constructed at the end of this file. It'll take a form like this:

: 0000
d 0
r _dup
r class:primitive
s dup

: 0001
r 0000
r _drop
r class:primitive
s drop

: 0002
r 0001
r _swap
r class:primitive
s swap

Each entry starts with a pointer to the prior entry (with a pointer to zero marking the first entry in the dictionary), a pointer to the start of the function, a pointer to the class handler, and a null terminated string indicating the name exposed to the Rx interpreter.

Rx stores the pointer to the most recent entry in a variable called Dictionary. For simplicity, I just assign the last entry an arbitrary label of 9999. This is set at the start of the source. (See In the Beginning...)

Rx provides accessor functions for each field. Since the number of fields (or their ordering) may change over time, using these reduces the number of places where field offsets are hard coded.

: d:link
i re......

: d:xt
i liadre..
d 1

: d:class
i liadre..
d 2

: d:name
i liadre..
d 3

A traditional Forth has create to make a new dictionary entry pointing to the next free location in Heap. Rx has newentry which serves as a slightly more flexible base. You provide a string for the name, a pointer to the class handler, and a pointer to the start of the function. Rx does the rest.

In actual practice, I never use this outside of Rx. New words are made using the : prefix, or d:create (once defined in the standard library). At some point I may simplify this by moving d:create into Rx and using it in place of newentry.

: newentry
i lifepuli
r Heap
r Dictionary
i felica..
r comma
i lica....
r comma
i lica....
r comma
i lica....
r comma:string
i polistre
r Dictionary

Rx doesn't provide a traditional create as it's designed to avoid assuming a normal input stream and prefers to take its data from the stack.

: Which
d 0

: Needle
d 0

: found
i listlire
r Which
r _nop

: find
i lilistli
d 0
r Which
r Dictionary
i fe......

: find_next
i zr......
i dulica..
r d:name
i lifelica
r Needle
r s:eq
i licc....
r found
i feliju..
r find_next

: d:lookup
i listlica
r Needle
r find
i lifere..
r Which

Number Conversion

This code converts a zero terminated string into a number. The approach is very simple:

  • Store an internal multiplier value (-1 for negative, 1 for positive)

  • Clear an internal accumulator value

  • Loop:

    • Fetch the accumulator value
    • Multiply by 10
    • For each character, convert to a numeric value and add to the accumulator
    • Store the updated accumulator
  • When done, take the accumulator value and the modifier and multiply them to get the final result

Rx only supports decimal numbers. If you want more bases, it's pretty easy to add them later, but it's not needed in the base kernel.

: next
i lica....
r fetch-next
i zr......
i lisuswpu
d 48
i swlimuad
d 10
i poliju..
r next

: check
i dufelieq
d 45
i zr......
i drswdrli
d -1
i swliadre
d 1

: s:to-number
i liswlica
d 1
r check
i liswlica
d 0
r next
i drmure..

Token Processing

An input token has a form like:

<prefix-char>string

Rx will check the first character to see if it matches a known prefix. If it does, it will pass the string (sans prefix) to the prefix handler. If not, it will attempt to find the token in the dictionary.

Prefixes are handled by functions with specific naming conventions. A prefix name should be:

prefix:<prefix-char>

Where is the character for the prefix. These are compiler macros (using the class:macro class) and watch the Compiler to decide how to deal with the token. To find a prefix, Rx stores the prefix character into a string named prefixed. It then searches for this string in the dictionary. If found, it sets an internal variable (prefix:handler) to the dictionary entry for the handler function. If not found, prefix:handler is set to zero. The check, done by prefix?, also returns a flag.

: prefix:no
d 32
d 0

: prefix:handler
d 0

: prefixed
s prefix:_

: prefix:prepare
i feliliad
r prefixed
d 7
i stre....

: prefix:has-token?
i dulica..
r s:length
i lieqzr..
d 1
i drdrlire
r prefix:no

: prefix?
i lica....
r prefix:has-token?
i lica....
r prefix:prepare
i lilica..
r prefixed
r d:lookup
i dulistli
r prefix:handler
d 0
i nere....

Rx makes extensive use of prefixes for implementing major parts of the language, including parsing numbers (prefix with #), obtaining pointers (prefix with &), and defining functions (using the : prefix).

prefix used for example
# numbers #100
$ ASCII characters $e
& pointers &swap
: definitions :foo
( Comments (n-)
: prefix:(
i drre....

: prefix:#
i lica....
r s:to-number
i liju....
r class:data

: prefix:$
i feliju..
r class:data

: prefix::
i lilifeli
r class:word
r Heap
r newentry
i ca......
i lifelife
r Heap
r Dictionary
i lica....
r d:xt
i stlilist
d -1
r Compiler
i re......

: prefix:&
i lica....
r d:lookup
i lica....
r d:xt
i feliju..
r class:data

Quotations

Quotations are anonymous, nestable blocks of code. Rx uses them for control structures and some aspects of data flow. A quote takes a form like:

[ #1 #2 ]
#12 [ square #144 eq? [ #123 ] [ #456 ] choose ] call

Begin a quotation with [ and end it with ]. The code here is slightly complicated by the fact that these have to be nestable, and so must compile the appropriate jumps around the nested blocks, in addition to properly setting and restoring the Compiler state.

: t-[
i lifeliad
r Heap
d 2
i lifelili
r Compiler
d -1
r Compiler
i stlilica
d 1793
r comma
i lifelili
r Heap
d 0
r comma
i ca......
i lifere..
r Heap

: t-]
i lilica..
r _ret
r comma:opcode
i lifeswli
r Heap
r _lit
i lica....
r comma:opcode
i lica....
r comma
i swstlist
r Compiler
i lifezr..
r Compiler
i drdrre..

Lightweight Control Structures

Rx provides a couple of functions for simple flow control apart from using quotations. These are repeat, again, and 0;. An example of using them:

:s:length
  dup [ repeat fetch-next 0; drop again ] call
  swap - #1 - ;

These can only be used within a definition or quotation. If you need to use them interactively, wrap them in a quote and call it.

: repeat
i lifere..
r Heap

: again
i lilica..
r _lit
r comma:opcode
i lica....
r comma
i liliju..
r _jump
r comma:opcode

: t-0;
i liliju..
r _zret
r comma:opcode

I take a brief aside here to implement push and pop, which move a value to/from the address stack. These are compiler macros.

: t-push
i liliju..
r _push
r comma:opcode

: t-pop
i liliju..
r _pop
r comma:opcode

Interpreter

The interpreter is what processes input. What it does is:

  • Take a string

  • See if the first character has a prefix handler

    • Yes: pass the rest of the string to the prefix handler

    • No: lookup in the dictionary

      • Found: pass xt of word to the class handler
      • Not found: report error via err:notfound

First, the handler for dealing with words that are not found. This is defined here as a jump to the handler for the Nga NOP instruction. It is intended that this be hooked into and changed.

As an example, in Rx code, assuming an I/O interface with some support for strings and output:

[ $? putc space 'word not found' puts ]
&err:notfound #1 + store

An interface should either patch the jump, or catch it and do something to report the error.

: err:notfound
i liju....
r _nop

call:dt takes a dictionary token and pushes the contents of the d:xt field to the stack. It then calls the class handler stored in d:class.

: call:dt
i dulica..
r d:xt
i feswlica
r d:class
i feju....
: input:source
d 0

: interpret:prefix
i lifezr..
r prefix:handler
i lifeliad
r input:source
d 1
i swliju..
r call:dt

: interpret:word
i lifeliju
r Which
r call:dt

: interpret:noprefix
i lifelica
r input:source
r d:lookup
i linelili
d 0
r interpret:word
r err:notfound
i liju....
r choose

: interpret
i dulistli
r input:source
r prefix?
i ca......
i lililiju
r interpret:prefix
r interpret:noprefix
r choose

The Initial Dictionary

This sets up the initial dictionary. Maintenance of this bit is annoying, but it isn't necessary to change this unless you add or remove new functions in the kernel.

: 0000
d 0
r _dup
r class:primitive
s dup
: 0001
r 0000
r _drop
r class:primitive
s drop
: 0002
r 0001
r _swap
r class:primitive
s swap
: 0003
r 0002
r _call
r class:primitive
s call
: 0004
r 0003
r _eq
r class:primitive
s eq?
: 0005
r 0004
r _neq
r class:primitive
s -eq?
: 0006
r 0005
r _lt
r class:primitive
s lt?
: 0007
r 0006
r _gt
r class:primitive
s gt?
: 0008
r 0007
r _fetch
r class:primitive
s fetch
: 0009
r 0008
r _store
r class:primitive
s store
: 0010
r 0009
r _add
r class:primitive
s +
: 0011
r 0010
r _sub
r class:primitive
s -
: 0012
r 0011
r _mul
r class:primitive
s *
: 0013
r 0012
r _divmod
r class:primitive
s /mod
: 0014
r 0013
r _and
r class:primitive
s and
: 0015
r 0014
r _or
r class:primitive
s or
: 0016
r 0015
r _xor
r class:primitive
s xor
: 0017
r 0016
r _shift
r class:primitive
s shift
: 0018
r 0017
r t-push
r class:macro
s push
: 0019
r 0018
r t-pop
r class:macro
s pop
: 0020
r 0019
r t-0;
r class:macro
s 0;
: 0021
r 0020
r fetch-next
r class:word
s fetch-next
: 0022
r 0021
r store-next
r class:word
s store-next
: 0023
r 0022
r s:to-number
r class:word
s s:to-number
: 0024
r 0023
r s:eq
r class:word
s s:eq?
: 0025
r 0024
r s:length
r class:word
s s:length
: 0026
r 0025
r choose
r class:word
s choose
: 0027
r 0026
r if
r class:primitive
s if
: 0028
r 0027
r -if
r class:word
s -if
: 0029
r 0028
r prefix:(
r class:macro
s prefix:(
: 0030
r 0029
r Compiler
r class:data
s Compiler
: 0031
r 0030
r Heap
r class:data
s Heap
: 0032
r 0031
r comma
r class:word
s ,
: 0033
r 0032
r comma:string
r class:word
s s,
: 0034
r 0033
r t-;
r class:macro
s ;
: 0035
r 0034
r t-[
r class:macro
s [
: 0036
r 0035
r t-]
r class:macro
s ]
: 0037
r 0036
r Dictionary
r class:data
s Dictionary
: 0038
r 0037
r d:link
r class:word
s d:link
: 0039
r 0038
r d:xt
r class:word
s d:xt
: 0040
r 0039
r d:class
r class:word
s d:class
: 0041
r 0040
r d:name
r class:word
s d:name
: 0042
r 0041
r class:word
r class:word
s class:word
: 0043
r 0042
r class:macro
r class:word
s class:macro
: 0044
r 0043
r class:data
r class:word
s class:data
: 0045
r 0044
r newentry
r class:word
s d:add-header
: 0046
r 0045
r prefix:#
r class:macro
s prefix:#
: 0047
r 0046
r prefix::
r class:macro
s prefix::
: 0048
r 0047
r prefix:&
r class:macro
s prefix:&
: 0049
r 0048
r prefix:$
r class:macro
s prefix:$
: 0050
r 0049
r repeat
r class:macro
s repeat
: 0051
r 0050
r again
r class:macro
s again
: 0052
r 0051
r interpret
r class:word
s interpret
: 0053
r 0052
r d:lookup
r class:word
s d:lookup
: 0054
r 0053
r class:primitive
r class:word
s class:primitive
: 0055
r 0054
r Version
r class:data
s Version
: 9999
r 0055
r err:notfound
r class:word
s err:notfound

Appendix: Words, Stack Effects, and Usage

Word Stack Notes
dup n-nn Duplicate the top item on the stack
drop nx-n Discard the top item on the stack
swap nx-xn Switch the top two items on the stack
call p- Call a function (via pointer)
eq? nn-f Compare two values for equality
-eq? nn-f Compare two values for inequality
lt? nn-f Compare two values for less than
gt? nn-f Compare two values for greater than
fetch p-n Fetch a value stored at the pointer
store np- Store a value into the address at pointer
+ nn-n Add two numbers
- nn-n Subtract two numbers
* nn-n Multiply two numbers
/mod nn-mq Divide two numbers, return quotient and remainder
and nn-n Perform bitwise AND operation
or nn-n Perform bitwise OR operation
xor nn-n Perform bitwise XOR operation
shift nn-n Perform bitwise shift
fetch-next a-an Fetch a value and return next address
store-next na-a Store a value to address and return next address
push n- Move value from data stack to address stack
pop -n Move value from address stack to data stack
0; n-n OR n- Exit word (and drop) if TOS is zero
s:to-number s-n Convert a string to a number
s:eq? ss-f Compare two strings for equality
s:length s-n Return length of string
choose fpp-? Execute p1 if f is -1, or p2 if f is 0
if fp-? Execute p if flag f is true (-1)
-if fp-? Execute p if flag f is false (0)
Compiler -p Variable; holds compiler state
Heap -p Variable; points to next free memory address
, n- Compile a value into memory at here
s, s- Compile a string into memory at here
; - End compilation and compile a return instruction
[ - Begin a quotation
] - End a quotation
Dictionary -p Variable; points to most recent header
d:link p-p Given a DT, return the address of the link field
d:xt p-p Given a DT, return the address of the xt field
d:class p-p Given a DT, return the address of the class field
d:name p-p Given a DT, return the address of the name field
class:word p- Class handler for standard functions
class:primitive p- Class handler for Nga primitives
class:macro p- Class handler for immediate functions
class:data p- Class handler for data
d:add-header saa- Add an item to the dictionary
prefix:# s- # prefix for numbers
prefix:: s- : prefix for definitions
prefix:& s- & prefix for pointers
prefix:$ s- $ prefix for ASCII characters
repeat -a Start an unconditional loop
again a- End an unconditional loop
interpret s-? Evaluate a token
d:lookup s-p Given a string, return the DT (or 0 if undefined)
err:notfound - Handler for token not found errors

Legalities

Rx is Copyright (c) 2016-2019, Charles Childers

Permission to use, copy, modify, and/or distribute this software for any purpose with or without fee is hereby granted, provided that the above copyright notice and this permission notice appear in all copies.

THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.

My thanks go out to Michal J Wallace, Luke Parrish, JGL, Marc Simpson, Oleksandr Kozachuk, Jay Skeer, Greg Copeland, Aleksej Saushev, Foucist, Erturk Kocalar, Kenneth Keating, Ashley Feniello, Peter Salvi, Christian Kellermann, Jorge Acereda, Remy Moueza, John M Harrison, and Todd Thomas.

All of these great people helped in the development of RETRO 10 and 11, without which Rx wouldn't have been possible.