Back to index

Under the hood

This chapter describes some internals of LispMe. You don't need this to work with LispMe, but some people asked me about the implementation, so here you are:

General

The latest version 3.2 of LispMe was developed with GCC 2.95.2-kgpd, prc-tools 2.0, PilRC v2.5b3, and the PalmOS 4.0 SDK on a PC running Windows XP and tested with Poser 3.2 and the PalmOS5 Simulator.

Altogether LispMe consists of about 26000 lines of highly compact C code. All error handling is done using Pilot's ErrCatch and ErrThrow macros/functions and not a single function from the C library has been used.

The trick I found to create executables >32k code with GCC is fortunately no more necessary prc-tools 2.0, since multi-segment applications are supported now. If you're interested in these topics, have a look here.

The implementation techniques of LispMe (V3.0) have been presented at the Scheme workshop of the PLI 2001 conference in Florence in September 2001. A detailed article is available as DVI (80 kB) or as compressed PostScript (99 kB).

SECD machine model

LispMe is based on the SECD virtual machine described in

Peter Henderson
Functional Programming - Application and Implementation
Prentice Hall International
ISBN 0-13-331579-7

The name stems from the 4 main registers this machine operates on:

S	stack	used to hold intermediate values while evaluating an expression
E	environment	used to hold the current environment
C	control	the machine-language program being executed
D	dump	a stack the other registers are pushed on when a new function is called

SECD virtual machine

In fact, all registers hold pointers to list structures in the heap, so for example going to the next machine instruction is accomplished by an assignment of the form C = cdr(C);

The virtual machine currently has 93 different instructions (much less than prior versions, since many builtin procedures have been re-implemented as native C calls). The main extensions to Henderson's machine are

error checking
including function arity in closures
closures with variable number of arguments
support for tail-recursion
continuation support
support for sequencing (begin)
real, complex, char, string, vector and port support
compile quasiquote templates to SECD code
support for macros and eval
call native functions

The runtime environment is represented as a list of frames, where each frame is a list of values. Variables are always accessed by lexical adresses (a pair indicating the frame and the offset of the variable). This is exactly the form you see when tapping the names button. In fact, the first step in each evaluation is assigning the value list to the E register.

SECD compiler

In contrast to Henderson's approach, the compiler is not bootstrapped, it's implemented in C like the rest of LispMe. The reasons for that are compilation speed and saving heap memory for user programs.

A LispMe expression is compiled by traversing its tree structure and consing up a list of machine instructions backwards to avoid using append. There's a simple peephole optimizer mainly used for removing unnecessary stack operations and for propagating returns through conditionals to provide proper tail recursion elimination.

There's also a compile time environment (represented as a list of lists, mirroring the structure of the runtime environment) to determine lexical adresses for all variables. You can view the SECD code of a closure with disasm.

The compiler is extensible by registering keyword/C-function pairs.

Macros are expanded by calling the VM with the expansion procedure and the original source expression as argument.

Parsing expressions

Parsing is divided into 2 steps:

The scanner groups characters into tokens by using a finite automaton consisting of 28 states. Most of the states deal with real and complex numbers and escape sequences in strings.
The parser is recursive-descent

Scanning and formatting floating point was particularly nasty, as the Pilot ROM routines FlpFToA and FlpAToF are unusable (limited precision, limited exponent range, no error support), so all this stuff had to be written from scratch.

Memory management

Each LispMe session is kept in a separate database bearing the same name as the session. All LispMe session databases use creator ID fbLM so that the total memory used by them is displayed correctly in the memory application.

In each session database, at least 7 records are used:

Record 0: Global vars, root pointers
Record 1: The atom store; all identifier names are entered here by the reader. The names are simply stored one after another separated by '\0' bytes.
Record 2: The floating point store; all floating point numbers are stored here (8 bytes each). Unused cells are linked into a free list.
Record 3: car pointers of the heap; 32 bit each

Record 4: cdr pointers of the heap; 32 bit each In contrast to versions before 3.2, the heap has been split in two blocks of the same size. The car and the cdr pointer at the same relative address form a "virtual" heap cell. The cells do not carry type tags, instead the pointers do according to this table:

`sddd dddd dddd ddd0`	16 bit signed offset from virtual heap base, which is 32k above actual heap begin (no scale neccessary)
`sddd dddd dddd dd01`	14 bit signed integer (-8192 .. 8191)
`dddd dddd dddd 0011`	unsigned 12 bit index into atom table (shift right 4 bits), 4kb atom space
`dddd dddd dddd 0111`	unsigned 12 bit index into double table (shift right 1 bit and unmask 3 low bits), 4096 reals
`ffff ffss sss0 1011`	Primitive symbol, 64 frames of 32 symbols
`ffff ffss sss1 1011`	Primitive value, 64 frames of 32 values
`0000 0000 0000 1111`	vector header in heap, upper 8 bit unused, cdr field of this cell contains 16 bit record index (untagged!)
`0000 0000 0001 1111`	string header in heap, similar to vector
`0000 0000 001s 1111`	bigint header in heap, similar to vector, s is sign
`0000 tttt 0100 1111`	Foreign data, 4 type bits, cdr field of this cell points to 32 bit value (untagged!)
`cccc cccc 1000 1111`	8 bit ANSI character

Primitives (symbols and the associated values) are stored in a two-dimensional (ragged) table consisting of upto 64 frames of upto 32 entries each. Currently, 21 frames are used, extension modules can use the rest. The very first frame (SPECIAL_MODULE) contains special values, like empty list, #f, #t, or a closure tag.

All pointers into the heap are relative pointers (offsets) to two virtual heap base pointers lying 32k behind the actual beginning of the heap (one for the cars and one for the cdrs). This allows relocation of the heap by the operating system. Additionally, this saves memory (16 bit relative pointers vs. 32 bit absolute pointers) and maps nicely to the 68000 adressing mode adress register indirect with index and offset, which uses signed offsets .

Record 5: The input field
Record 6: The output field
Each vector and string occupies a single DB record starting with record number 6. Starting in version 3.1, a vector or string is identified by its record index stored (instead of the unique ID in earlier versions) in a cons cell in the heap. Each record has a PTR back to its header cell as its first element to allow garbage collection care for changing record indices when deleting a record.

Starting with version 1.7, LispMe mo more uses the dynamic heap for its memory, instead it works directly on DB blocks, gaining write access via MemSemaphoreReserve. This is a much discussed technique, as accessing memory out of its bounds can overwrite other databases and even require a hard reset. On the other hand, LispMe has been out for some months now and several hundred people have tested it and I never got a report of an "invalid pointer access". Furthermore, there's no way in LispMe to create an invalid pointer by will or accident (like in other Scheme implementations, too). All pointers are created in the memory subsystem which almost certainly is bugfree. So I finally decided to use this technique to make a bigger heap available on all Pilots.

Garbage collection

LispMe uses mark/scan garbage collector for heap cells and floating point cells. The reasons for preferring mark/scan to copying garbage collection are:

Memory usage has top priority on the Pilot. Wasting half of the memory with copying GC isn't affordable.
All heap cells have the same size, so fragmentation can't happen
Heap compaction (provided by copying GC) doesn't affect performance, as there's no virtual memory or cache on the Pilot.
The disadvantage that mark/scan GC has to touch every memory cell in the heap is ignorable with heap sizes possible on the Pilot

All in all, a typical garbage collection of 16k heap takes about 0.2 seconds.

Foreign types can register garbage collection hooks for finalization and marking other referenced cells.

Strings, vectors, and bigints are garbage collected without having to move their actual contents. Instead, DmAttachRecord() is used to swap the memoryblock/record number association and unused records are moved to the end of the workspace database where they can be deleted without having to shift intermediate records.