-*- indented-text -*-


	Introduction
	============

Hedgehog is a very concise implementation of a LISP-like language for
low-end and embedded devices, primarily for machine-to-machine
systems. It consists of a compiler to byte code and a corresponding
interpreter.  The byte code interpreter is written in standard
conforming C, is efficient and easily portable, and can be compiled to
a very small executable or library of only some 20-30 kilobytes.

The Hedgehog LISP dialect has proper support for local and lambda
functions, lexical scoping, variable argument functions, garbage
collection, exceptions, macros, and over a hundred predefined functions
or special forms.  The built-in types are lists, symbols, strings, 32-bit
integers, AVL-trees, and tuples up to 16 elements wide.  Proper 32-bit
wide integers are necessary for various bit-level operations in embedded
systems.

The Hedgehog compiler and byte code interpreter, essentially all code
written in C, is distributed under the "Lesser" GNU General Public
License (GPL).  In layman's terms, you have to ship the source code of
the Hedgehog compiler and interpreter whenever you ship the
corresponding binary executable.  However, all files written in
Hedgehog Lisp itself, including the standard library in the prelude.d
directory, are distributed the revised BSD license, which allows them
to be included also in programs that are distributed without full
source code.  See the files LICENSE.LGPL and LICENSE.BSD.

This public version of the lisp compiler is routinely exercised on
Linux and Cygwin, it has been briefly tested on FreeBSD and SunOS, and
should be easily portable to any Posix system.  The interpreter part
has additionally been ported to several embedded operating systems.
Support, alternative licenses, and ports can be discussed with
Oliotalo Ltd.  The ports have support for various protocols and
devices specific for that platform.

For further information or to download, see
http://hedgehog.oliotalo.fi/ or mail [email protected].

Authors: Kenneth Oksanen <[email protected]>
	 Lars Wirzenius <[email protected]>


	Version-specific Note
	=====================

This series of versions, hedgehog-2.*, has been extracted and cleaned
up from a significantly more featureful hedgehog-1.6, which is used by
Oliotalo in several embedded machine-to-machine solutions.  While many
parts of the code are mature and well-tested, this version undoubtedly
contains also immature and incorrect pieces of code.  Also some parts
are missing, such as a few good demo applications and various features
planned for the state machine library (which hedgehog-1.6 has but we
wish to improve for 2.0).  Therefore we consider this as an
beta-quality at best.  But we will gradually improve.


	Running HedgeHog
	================

To compile and install Hedgehog on a typical Linux installation,
unpack the sources and change to the source directory and then do the
following:
    	mkdir o
	cd o
	../configure linux /usr/local
	make
where `make' refers to GNU make.  This builds to executables the lisp
compiler `hhc' and the byte code interpreter `hhi', and some lisp
library files in the directory `prelude.d'.  Hedgehog has been
compiled and tested on a number of other platforms, including FreeBSD
and SunOS, but you will need GNU make and in many cases, especially if
you wish to edit the develop the code further, also the GNU awk and
lex replacements gawk and flex.

Assuming the above completed successfully, you should now be able to
compile tests in the `tests'-directory, for example by issuing
	./hhc -g -p prelude.d ../tests/fib.hl
in order to compile the ubiquitous Fibonacci-test.  This produces two
output files, `fib.hlo' and `fib.hls'.  The latter is a symbolic
assembler file and used mostly for debugging after a possible error
reported by the interpreter.  It is particularly useful if you didn't
pass `hhc' the flag `-g' telling it to add a little debugging
information to the `fib.hlo'.  

The former is the actual byte code program, augmented with a small
header and the program's constant data.  You should now be able to run
the compiled Fibonacci-test in `fib.hlo' by by issuing
	./hhi fib.hlo
Depending on your machine's speed this will last a few seconds, but
eventually you should see the result:
	fib(30) = 1346269
or
	fib(22) = 28657
depending on which distribution you have.

The flag `-p prelude.d' you passed to `hhc' tells it to use the lisp
library in the subdirectory `prelude.d'.  Conventionally you should 
	make install
with sufficient priviledges.  This places `hhc' and `hhi' in
`/usr/local/bin' and the standard library `/usr/local/lib/hh', as you
had instructed `configure' to do.  Now that the prelude is in its
predefined place, you can refrain from passing `-p prelude.d' to `hhc'.

Pass `hhc' and `hhc' the flag `--help' or `-h' to see their full
command line interface.  Note that the heap and stack sizes are fixed
during each execution, but are adjustable from the command line.


        The prelude
        ===========
        
The Hedgehog standard library, or prelude, consists of all the files
in the `prelude.d' directory that have a `.hl' suffix. The filenames
are expected to be sorted according to ASCII character codes (in the
"C" locale, that is: some other locales have weird rules for sorting,
e.g., the fi_FI one).  The naming convention is, therefore, to begin
each filename with three digits and a dash. This will allow ordering
between the files.

XXX Here comes a little tour of the prelude.


	Porting and Adaptation
	======================

XXX Design and write a guide.

If memcmp performs signed comparison of bytes on your platform, you
must define HH_MEMCMP to some function which performs unsigned
comparison.  This is necessary for Lisp-level string comparisons and
compression to work correctly.

The macros HH_UNIX, HH_SUNOS, HH_BSD, HH_LINUX etc. refer to the
target platform where we run hhi, but not necessarily to hhc.

But while nothing is written yet, a few words of advice:
  - In general, the code is supposed to be touched.
  - Note that the configure script is not GNU autoconf, so you are
    supposed to add definitions relevant to your target in it.
  - The interpreter is able to change byte order of the `.hlo'-files
    automatically as needed, so that you needn't worry about this.

If your target looks like unix:
  - No or at least small editions in hh_common.h should be sufficient
    to compile hhi.
  - By default `make' (cross-)compiles and (cross-)executes a small C
    program which digs out various flag values and struct layouts of
    the system call interface into the prelude file `300-unix.hl'.
    Ensure the cross-execution succeeded.

If your target does not look like unix:
  - You have to write a new main program file.  You can take
    `hh_interp_unix.c' as your starting point.
  - Search for occurrences of HH_UNIX in the code (including the
    compiler), and adapt the code for your target in similar respects.


	Byte Code Interpreter
	=====================

	Instruction Format
	------------------

The byte code instruction set is designed to be concise, allow a small
implementation of the interpreter, but still allow a reasonably
efficient implementation.

The instructions are of variable instruction length.  The uppermost
two bits in the first byte of the instruction tell the byte code type.
Bits 00 indicate an instruction without an immediate field, bits 01
indicate that the second byte of the instruction is a one-byte signed
immediate field.  Bits 10 indicate a two-byte immediate value and bits
11 indicate a four-byte immediate value.  A handful of instructions
use more than a single immediate field.

The lowermost six bits of the first byte of each instruction indicate
the mnemonic, such as `push' or `load_imm'.  The byte code
instructions with an immediate field and the ones without are
separate.  This means there are at most 64 instructions with an
immediate field and 64 instructions without.  The latter is often too
few, and therefore rarely used non-immediate instructions are encoded
as "ext insns" a special immediate instruction plus an immediate field
that specifies the instructions.  All instructions are defined in the
file `hh_insn.def', which with some C macro trickery is included in
several places for various purposes.

Some instructions, such `load_imm' or `pick', have variants which
first push the value of accu to stack before actual computation.  This
both reduces byte code program lengths and improves performance
because of reduced instruction dispatching.  For some other
instructions, for example for jumps and many non-atomic expression,
such variants would be rarely used, and are therefore omitted in order
to save instruction space.


	Virtual Machine
	---------------

The virtual machine is a stack machine but so that the least recently
computed value cached in a special "register" `accu'.  The benefit of
`accu' is twofold: Firstly, if the result is not used, there is no
need to pop it off the stack.  Secondly, `accu' reduces the traffic to
and from the stack, thereby somewhat improving the performance.

The stack pointer `sp' points to the position in stack where the next
pushed value goes to.  The stack grows upward.

A function is a tuple whose first word refers to the beginning of the
function's byte code and the second word refers to the function's
environment, or HH_NIL (i.e., the Lisp nil) if there is only the
global environment.  The first byte code instruction of each function
is actually not an instruction at all, but a byte of information of
the arity of the function.  The seven lowermost bits tell the number
of arguments the function expects, and if the uppermost bit is set,
then the function is a variable argument function.

The same stack is used to store arguments to both byte code
instructions, builtin functions and user-defined functions.
Let's follow through a proper function call.  The caller evaluates all
subexpressions in the s-expression from left (the called function) to
right, pushing all but the last one which is left in `accu'.  Now it
issues the byte code instruction `call' with the immediate value
giving the number of arguments in the call.  This instruction does all
of the following:
  - Find the definition of the function.
  - Check that the number of arguments of the callee matches the
    number of passed arguments.
  - If the callee is a variable argument function, make a list of the
    excess arguments.  The list is left in `accu'.
  - Store the caller's next instruction's program counter into the
    place where the callee function pointer was in stack.
  - Load the callee's pc and environment pointer from the function
    tuple.
And then we're already executing the callee.  Returning occurs with
`return n' which pops `n' values off the stack to find and resume the
caller's stored program counter.

Dynamic extent is achieved by creating small tuples, called
"environments" or "closures", which contain exactly those variables
which do not by their nature have infinite lifetime (such as global
variables) and are referred to from local "lambda" functions created
by `fn's or `def's.  For example 
	(def (mkf2 x)
	     (fn (y) (+ y x)))
is automatically turned into something like
	(def (mkf1 x)
	     (.make_new_env 2)
	     (.put_new_env 0 x)
	     (.bind_env $.0))
	(def ($.0 y)
	     (+ y (.get_env 0)))
The `$.0' is an automatically generated symbol for the local `fn'
lifted to the top-level.  `.bind_env' leaves to `accu' the function
`$.0' bound to the created environment, which contains the value of
variable `x', which the local function refers to.

A callable function is a cons-cell, whose CAR-slot contains a program
counter and CDR-slot contains the environment tuple, or nil if the
function does not refer to local variables defined outside of it.
Calling a function always reads the environment bound to that
function, but otherwise environment pointers are pushed and restored
from the stack on a need -basis.  Since global functions refer only to
values in their argument list or the global environment, there is no
need for environment pointer storage/restoration in global functions.

Catch is implemented by pushing onto stack a special header (see
later) word which indicates the catch tag, the catching program
counter and current environment pointer.  The throw scans the stack
downwards looking for a matching catch tag, and if it sees one, then
it resumes from the program counter and environment after the catch
header.


	Byte Code Program File Format
	-----------------------------

The program is a sequence of bytes so that
  - Bytes 0..3 are a magic cookie and contain the values 0x4E, 0xD6,
    0xE4 and 0x06, respectively.
  - Bytes 4..7 contain a 32-bit checksum of the rest of the program.
    This can be used to check that the program has been received
    correctly, but it is not resilient to malicious attacks.
  - Byte 8 is reserved for future use, most probably a version number
    for the format of constant values, debugging information, and this
    header.  Currently it should be 1.
  - Bytes 9..11 tell `proglen', the length of the program block, in
    bytes.
  - Bytes 12..12+proglen-1 contain the program, the execution starts
    at the 12th byte.
  - Bytes 12+proglen..end contain constant data, the "constant pool",
    most significantly strings and names of quoted symbols for the
    `symboltostring' primitive.

All values are encoded in the most significant byte first order.
There is no general assumption of alignment inside the byte code, but
the first byte of the program must be aligned to a 32-bit word
boundary.


      Garbage Collection
      ------------------

The current garbage collector is a trivial two-semispace stop&copy,
which draws no benefit from the fact that objects are never mutated.

Garbage collection may initiate at the beginning of any instruction,
but not anywhere else.  This means that if the instruction allocates
memory, it must prior to changing the state of the program, such as
manipulating `accu', the stack pointer, or the stack, compute how much
memory it is going to allocate, and issue garbage collection if
necessary.  The macro `HH_RESERVE' is a convenience macro for the last
task.

If garbage collection does not free enough memory for the computation
to proceed, the virtual machine tries to throw an
"out-of-memory-exception".  If it is caught, a second garbage
collection is issued and the execution tries to resume after the
catch.  If the throw is not caught, the virtual machine exits with a
special error code.


	Tagging
	-------

Various bits of each word is reserved to represent type information.
The deduction rules for determining the type of any given 32-bit word
goes as follows:

If the lowest bit is 1, the word is a signed integer in the range from
-2^30 to 2^30-1 and it is obtained by arithmetically shifting the word
one right.

If the lowest bits are 00, the word is valid pointer to a
heap-allocated object.  The next 6 bits give partial type information
of the referred object:
  0000 0000  Forward pointer used by the garbage collector.
  0000 0100  Pointer to an object whose type is in a heap-allocated
	     header word.
  00nn nn00  Pointer to an n-tuple, if nnnn=0010, a cons-cell.
  0100 0x00  Pointer to an integer so that the 32-bit signed integer
	     value can be reconstructed by exclusive or of the
	     referred word and the x-bit shifted right by 2.
  1xxx xx00  Pointer to a cell in the constant pool, `xxxxx' have
	     the same type meaning as described above.
Other cases are reserved for future use.

And finally, if the two lowest bits are 10, the word is a header in a
cell:
  0000 0010  Some special symbols.  The symbol is identified by the 
	     upper 24 bits:
	       0000 0000 0000 0000 0000 0000  Empty list, NIL.
               0000 0000 0000 0000 0000 0001  Symbol t, true.
  0000 0110  Reference to somewhere in the byte code program.
  0000 1010  A heap-allocated string.  The upper 24 bits indicate the
             length of the string in number of bytes.  The following
             bytes after the initial word are used as bytes in the
             string.  Contrary to C-string, these strings may contain
	     null-characters which do not terminate the string.  But
	     for convenience, there is also a redundant null-character
	     after the end of the string so that the string can be
	     passed to normal C routines and system calls without
	     copying it out of the heap.
  0000 1110  A symbol.  The second word contains a pointer to the string.
  0001 0010  A catch cell.  This can exist only in the stack.  The
             upper 24 bits tell the catch/throw tag, the next word is
             catch handler program counter, and the third word is the
             catcher's environment.
  0001 0110  AVL-tree node. The bits 8..15 and 16..23 tell the heights
             of the left and right subtrees, respectively, and bits
             24..31 the height of the the subtree of the current node.
  0001 1010  Debugging info header.  If this exists, then it must exist
             as a second word in the constant pool.  The third word
             contains the filename list and the fourth word contains a
             pointer to the pc to filename-linenumber -mapping.

Other cases are reserved for future use.

XXX Debugging information


	Byte Code Compiler
	==================

	Passes and Data Structures
	--------------------------

The main compiler driver routine is in file `hh_compiler.c'.  It calls
the function `hh_ast_read_file' in file `hh_ast.c' to read in
Lisp-files and construct syntac trees of them.  The lexer in
`hh_lex.l' is implemented with the help of the GNU scanner generator
flex.

The abstract syntax tree of the program is represented as a directed
acyclic graph of `hh_ast_t' structs.  For normal s-expressions it
behaves like a tuple - the i'th element of an s-expression is accessed
with an array lookup - but for potentially extremely large
s-expressions, such as function bodies, do-forms, and top-level
program definitions, the `hh_ast_t's form a NULL-terminated list.
Each `hh_ast_t' node maintains information of its source location.

All strings and symbols are interned in the compiler, i.e. identical
symbols and strings are the same object in the compiler's memory.  The
corresponding types are `hh_string_t' and `hh_symbol_t'.

After reading in the syntax tree of the whole program, it is passed to
`hh_macroexpand.c' for macro expansion.  The macro expansion returns
the expanded program without any macro definitions and macro uses.

The lambda-lifting is one of the most complicated parts of the
compiler.  It "lift" locally defined functions (`fn's and `def's) into
top-level functions which assume to be bound to suitable environment
tuples.  Naturally lambda-lifting also augments the site of the
definition of the local function with code that creates a suitable
environment for the lifted function, and binds the environment to the
lifted function (as a program counter).  Furthermore the lifted
function's references to variables in the environment are replaced
with expressions that read the variables' values from the environment.

After lambda-lifting we perform algebraic optimization passes
implemented in `hh_opt.c'.  Currently the only algebraic optimization
is a simple folding of boolean constants for the primitives `if',
`and', `or', and `not'.  Note that the pass removes code which becomes
unreachable by these boolean constant foldings.

Next the compiler passes the whole program through uses-analysis,
which computes the transitive closure of all functions referred from
the top-level main program.  Unreferred functions, ofter a huge
majority if numerous libraries are included, can be discarded from the
final output based on this analysis.

Next it is time for code generation.  The code is generated by a
recursive walk-through in the syntax tree into small `hh_code_t'
structs linked together in a doubly linked list.  Applications of
built-in primitives are recognized and generated according to
definitions in `hh_builtins.def'.  The code generation pass keeps
track of the tail-call position of the currently generated expression
and automatically uses tailcalls wherever possible.  This pass also
keeps track of stack positions of variables in the scope, and
generates corresponding code for the variable references.

Note that code generation uses a "fake" heap (in the fake execution
context `hh_constant_ctx') where it generates all constant data.  This
trick makes it possible to use the definitions in files `hh_data.[hc]'
for both run-time and compile-time data management.

Here I suggest one should add at least a modest peephole optimization,
but currently nothing to that effect is done.

Finally it is time to output the generated code.  This begins with
assigning an absolute position for all byte code instructions.  Note
that immediate fields' width depend on the size of the absolute value
of the immediate field, and the value of the absolute fields in
e.g. relative branch instructions may depend on the width and relative
position of other instructions in the vicinity.  This chicken-and-egg
problem is solved by an approximate "cooling" process which uses
increasingly much unnecessarily large immediate field widths if the
absolute placement of instructions would otherwise seem impossible.
The actual byte code output is then a rather trivial loop over all
generated and placed `hh_code_t's.  The same loop generates both the
actual code file (suffixed `.hlo') and the symbolic assembler file
(suffix `.hls'), which is necessary for any kind of debugging.  The
byte code file is prepended by a small header and prepended by the
contents of the "fake" heap of constant data.

The compiler is admittably somewhat sloppy regarding its memory
management: symbols, string constants, abstract syntax tree nodes and
byte code structs are never actively freed.  But since the memory
consumption and running time of the compiler is negligible in any
case, this shouldn't matter as long as the host's operating system
provides proper process cleanup after the compiler is finished.


	Files and Conventions
	=====================

The byte code interpreter has been designed to be usable as a
stand-alone program as well as a library linked into some larger
system.  It is, for example, possible to run the byte code program for
a while (say, 1000 byte code instructions, or until nothing else can
be done) and then return to do something else in the rest of the
system.  It is also possible to merge select-based event loops of the
interpreted lisp programs and the containing system into one.  The
most essential interface the larger system needs to know is in file
`hh_interp.h', but routines in `hh_data.h' and the ability implement
new byte code instructions in `hh_insn.def' and new Lisp primitives in
`hh_builtins.def' may also be of use when implementing some kind of
communication between the interpreter and the containing system.

Only the files
	hh_interp.c
	hh_data.c
	hh_error.c
	hh_printf.c
	hh_avl.c
and
	hh_insn.def
	hh_error.def
and the headers
	hh_common.h
	hh_interp.h
	hh_data.h
	hh_error.h
	hh_printf.h
	hh_avl.h
are used by the byte code interpreter.

The byte code compiler uses the same `hh_data.[hc]' as the interpreter
in order to construct cells in the identical format to the constant
pool of the generated byte code program.  But there are some minor
allocation direction differences, and hence it defines the macro
`HH_COMPILER'.

All file-, macro and identifier names are prefixed with `hh_' or `HH_'.
This hopefully helps to port HedgeHog to systems with polluted
namespaces.

We have tried to put calls to platform-specific or frequently broken
functions behind macros in the interpreter.  For example, because many
systems have broken `memmove's, all calls go through the macro
`HH_MEMMOVE', which may be defined to something platform-specific.
Other such macros are `HH_ASSERT' and `HH_MEMCMP'.  `malloc' and
`free' (defined to `HH_MALLOC' and `HH_FREE') are used sparingly.

We also use a private version of the `printf' function family, which
not only is much smaller than typical `printf's, but also has a few
useful additional features (and lacks some that are not useful to us).
The interpreter's main program file (`hh_interp_unix.c' in UNIXes)
implements the code that writes the `printf'ed characters to stdout,
system log, serial line, or wherever.  Because `hh_data.c' is used by
both the compiler and interpreter, all calls to the `printf' are
hidden behind the macro `HH_PRINT'.