Forth – is it still relevant?

With all the advantages, it is unfortunate that Forth lost out to C language over the years and have been reduced to a niche. Per ChatGPT: due to C's broader appeal, standardization, and support ecosystem likely contributed to its greater adoption and use in mainstream computing.

So, the question is, how to encourage today's world of C programmers to take a look at Forth. How do we convince them that Forth can be 10 times more productive? Well, we do know that by keep saying how elegant Forth is or even bashing how bad C can be probably won't get us anywhere.

Bill Muench created eForth for simplicity and educational purpose. Dr. Ting, ported to many processors, described Forth in his well-written eForth genesis and overview. I like the idea and decided to pick it up.

1. 100% C/C++ with multi-platform support. Though classic implementation of primitives in assembly language and scripted high-level words gave the power to Forth, it also became the hurtle for newbies. Because they have to learn the assembly and Forth syntax before peeking into the internal beauty of Forth.

2. Dictionary is just an array. It's remodeled from linear memory linked-list to an array (or a vector in C++'s term) of words.

   • To search for a word, simply scan the name string of dictionary entries. So, to define a new word during compile time is just to append those found word pointers to the its parameter array one by one.
   • To execute become just a walk of the word pointers in the array. This is our inner interpreter.
   • Hashtables might go even faster but we'll try that later.

3. Data and Return Stacks are also arrays. With push, pop and [] methods to clarify intentions.

4. Parameter fields are all arrays. Why not!

5. No vocabulary, or meta-compilation. Except CREATE..DOES>, and POSTPONE, these black-belt skills of Forth greatness are dropped to keep the focus on core concepts.

6. Multi-threading and message passing are available From v5.0 and on, multi-core platform can utilize Forth VMs running in parallel. see the multi-threading section below for details

   • A thread pool is built-in. Size is defaults to number of cores.
   • Message Passing send/recv with pthread mutex waiting.
   • IO and memory update can be synchronized with lock/unlock.

If you are fluent in C/C++ and in the process of building your own Forth, skipping the verbage, the easiest path to gain understanding of how things work together is to download release v4.2 and work from there.

In the release, a heavily commented ceforth.cpp, the companion ceforth.h, and a config.h. Altogether, about 800 lines. Check them out!

The core of current implementation of eForth is the dictionary composed of an array of Code objects that represent each of Forth words.

1. Code - the heart of eForth, depends on the constructor called, the following fields are populated accordingly

   + name - a string that holds primitive word's name, i.e. NFA in classic FORTH, can also holds branching mnemonic for compound words which classic FORTH keeps on parameter memory + xt - pointer to a lambda function for primitive words i.e. XT in classic FORTH + pf, p1, p2 - parameter arrays of Code objects for compound words, i.e. PFA in classic FORTH + q - holds the literal value which classic FORTH keep on parameter memory

2. Lit, Var, Str, Bran, Tmp - the polymorphic classes extended from the base class Code which serve the functionalities of primitive words of classic Forth.

   + Lit - numeric literals + Var - variable or constant + Str - string for dostr or dotstr + Bran - Branching opcode + Tmp - temp storage for branching word

3. Dictionary - an array of Code objects

   + build-it words - constructed by initializer_list at start up, before main is called, degenerated lambdas become function pointers stored in Code.xt dict[0].xt ------> lambda[0] <== These function pointers can be converted dict[1].xt ------> lambda[1] into indices to a jump table ... which is exactly what WASM does dict[N-1].xt ----> lambda[N-1] <== N is number of built-in words + colon (user defined) words - collection of word pointers during compile time dict[N].pf = [ *Code, *Code, ... ] <== These are called the 'threads' in Forth's term dict[N+1].pf = [ *Code, *Code, ... ] So, instead of subroutine threading ... this is 'object' threading. dict[-1].pf = [ *Code, *Code, ... ] It can be further compacted into token (i.e. dict index) threading if desired

4. Inner Interpreter - Code.exec() is self-explanatory

   if (xt) { xt(this); return; } // run primitive word for (Code *w : pf) { // run colon word try { w->exec(); } // execute recursively catch (...) { break; } // handle exception if any }

   i.e. either we call a built-in word's lambda function or walk the Code.pf array recursively like a depth-first tree search.

5. Outer Interpreter - forth_core() is self-explanatory

   Code *c = find(idiom); // search dictionary if (c) { // word found? if (compile && !c->immd) // are we compiling a new word? dict[-1]->add(c); // then append found code to it else c->exec(); // or, execute the code return; } DU n = parse_number(idiom); // word not found, try as a number if (compile) // are we compiling a new word? dict[-1]->add(new Lit(n)); // append numeric literal to it else PUSH(n); // push onto data stack

With the array implementation, the first difference is in array variable read/write.

While traditional Forths uses narr 2 cells + to get the memory address of narr[2], eforth narr returns its index (or defining order) in the dictionary. So, narr 2 cells + will actually get you the index of the second word defined after narr. You'll be storing the value into that word's empty qf field. To access the nth element of narr, use th instead

With arrays, the doors are open. Dynamically expanding variables as well as storing objects instead of just integers. Parameter fields can be filled in compile time or changed on the fly in runtime i.e. self-morphing code. These can be the "scary" features for Forths to come.

Most classic Forth systems are build with a few low-level primitives in assembly language and bootstrap the high-level words in Forth itself. Over the years, Dr. Ting have implemented many Forth systems using the same model. See here for the detailed list. However, he eventually stated that it was silly trying to explain Forth in Forth to new comers. There are just not many people know Forth, period.

Utilizing modern OS and tool chains, a new generation of Forths implemented in just a few hundreds lines of C code can help someone who did not know Forth to gain the core understanding much quickly. He called the insight Forth without Forth.

In 2021-07-04, I got in touched with Dr. Ting mentioning that he taught at the university when I attended. He, as the usual kind and generous him, included me in his last projects all the way till his passing. I am honored that he considered me one of the frogs living in the bottom of the deep well with him looking up to the small opening of the sky together. With cross-platform portability as our guild-line, we built ooeForth in Java, jeForth in Javascript, wineForth for Windows, and esp32forth for ESP micro-controllers using the same code-base. With his last breath in the hospital, he attempted to build it onto an FPGA using Verilog. see ceForth_403 and eJsv32 for details.

We hope it can serve as a stepping stone for learning Forth to even building their own, one day.

There are two major versions current. eForth. v4 is single-threaded only and v5 default single-threaded but also supports multi-threaded.

Checkout the version you are interested in.

To enable multi-threading, of v5, update the followings in ~/src/config.h

Pretty amazing stuffs! To grasp how they were done, study the individual files (*.fs) under ~/tests/demo.

Note: MacOS added, thanks to Kristopher Johnson's work.

I haven't develop anything useful on Windows for a long time. Just bearly got this compiled on an 2007 Windows7 box. So, take it with a grain of salt. I'm hoping someone can make it more streamlined.

Note: For multi-threading to work, browser needs to receive Cross-Origin policies here for detail in the response header. A Python script ~/tests/cors.py is provided to solve the issue. The same needed to be provided if you use other web server.

Note: Most ESP32 are dual-core. However core0 is dedicated to WiFi and FreeRTOS house keeping. Forth tasks will be tied to core1 only. So, multi-threading is possible but no performance gain. Actually, singled-threaded v4.2 does a bit better.

Forth has been supporting multi-tasking since the 70's. They are single-CPU round-robin/time-slicing systems mostly. Modern system has multiple cores and Forth can certainly take advantage of them. However, unlike most of the matured Forth word sets, multi-threading/processing words are yet to be standardized and there are many ways to do it.

Before we go too far, make sure the following are updated before your build

word stack desc state

task	( xt -- t )	create a task (tid is index to thread pool entry) a free VM from pool is chosen for the task	STOP=>HOLD
rank	( -- t )	fetch current task id	NEST
start	( t -- )	start a task The VM is added to event_queue and kick started when picked up by event_loop	HOLD=>NEST
join	( t -- )	wait until the given task is completed	NEST=>STOP
lock	( -- )	lock (semaphore) IO or memory	NEST
unlock	( -- )	release IO or memory lock	NEST
send	( v1 v2 .. vn n t -- )	send n elements on current stack to designated task's stack (use stack as message queue)	sender NEST receiver HOLD
recv	( -- v1 v2 .. vn )	wait, until message to arrive	HOLD=>NEST
pull	( n t -- )	forced fetch stack elements from a completed task	current NEST target STOP
bcast	( n -- )	not implemented yet, TODO	sender NEST receivers HOLD
clock	( -- n )	fetch microsecond since Epoch, useful for timing

[06.1]>> started on T2 [05.1]>> started on T4 [04.1]>> started on T6 [07.1]>> started on T0 18 ms [06.3]>> finished on T2 18 ms [05.3]>> finished on T4 18 ms [04.3]>> finished on T6 18 ms [07.3]>> finished on T0 [06.1]>> started on T1 [06.1]>> waiting [07.1]>> started on T2 [06.1]>> sending 4 items to VM6.1 sent [07.3]>> finished on T2 [00.3]>> VM7 joint [06.3]>> received => state=3 sum=10 [06.3]>> finished on T1 [00.3]>> VM6 joint [00.3]>> joining VM7 [07.1]>> started on T1 [07.3]>> finished on T1 [00.3]>> VM7 joint pulled 1 items from VM7.0 total= 1784293664 -1 -> ok

What is the performance difference?

1. Code *dict[] - where words are dynamically allocated as a collection of pointers, or
2. Code dict[] - where words are statically created as an array of objects.

I have created a git branch 'static' to compare to the 'master. The static version is about 10% slower on 64-bit machine and about 5% slower on 32-bits. This hasn't been carefully analyzed but my guess is because Code is big at 144-bytes on 64-bit. They might get pushed off L1 cache too often.

An array of lambdas vs the classic switch statement, i.e.

Though syntax clarity is pretty much the same, lambda being function pointers takes an extra jump and the cost of stack-frame setup/teardown. It takes more space and about 15% slower in tight loops. However, with the advance of compilers,

1. It does not need a long enum definition, i.e. PLUS, MINUS, ..., which needs to be kept in-sync
2. It is possible to prebuild lambda array as a ROM image or static library that can be transported.
3. A tweak to CODE macro, i.g. adding NEXT, can potentially enable Tail Call Optimization (TCO) which eliminates the stack-frame overhead as did in many functional languages.

Though the use of C++ standard libraries helps us understanding what Forth does but, even on machines with GBs, we still need to be mindful of the followings. It gets expensive especially on MCUs.

The current implementation of ~/src/ceforth.h, a Code node takes 144 bytes on a 64-bit machine. On the other extreme, my ~/orig/40x experimental version, a vector linear-memory hybrid, takes only 16 bytes here. Go figure how the classic Forths needs only 2 or 4 bytes per node via linked-field and the final executable in a just a few KB. You might start to understand why the old Forth builders see C/C++ like plaque.

I try to release allocated blocks before exiting, however due to the dynamic alloc and resizing of std::vector, eForth dictionary hold on to many Code objects and the names string generated with them, valgrind (or similar tool) could reports lost (or leak). Though these memory blocks should all be reclaimed by the OS, it is something to be mindful of.

Current implementation utilize C++ vector as the core storage. Inside a Code object, there are pf, p1, p2 vectors to store branching words similar to that of an AST (Abstract Syntax Tree). The alternative is to stick all words into a single parameter field as done in classic Forth. I have created a branch one_pf doing exactly the same just to check it out. Also, tried polymorphic inner interpreter. So, are they better?

So, what cachegrind said for 100M loop tight loops and chacha.fs a CPU intensive?

Apparently, grown ~30% in all aspects. I think because having branching primitives, i.e. _if/_else/_then, for/next, in C++ prevent the extra fetch of VM branches. Sort of the difference between having hardware and software branchers. However, my gut feeling is the difference shouldn't be so dramatic especially with the recursive nest(). More research on this...

Instead of using vectors (i.e. pf, p1, p2) to keep codes and parameters, this implementation follows classic Forth's model using one big block of parameter memory with words laid down contiguoursly. With 32-bit data, subroutine threaded but hybrid with 16-bit xt offset (to reduce one lookup).

It works better with WASM's memory model. It is used as the foundation for weForth. So far, it is stable but tweaked from time to time and

Hinted by Sean Pringle's Rethinking Forth and Travis Bemann's wornderful zeptoforth. Nested module (or sub-words), simplified control structures are attemped. Now, moved to eForthX

• Dr. Ting's work on eForth between 1995~2011 eForth references and their Source Code Repo

• CC 20210314: Initial

  • Started with ~orig/33b code-base, refactor with enum and VA_ARGS macros targeting 100% C/C++.

• CC 20210707: Refactor

  • Incorporated list-based dict, ss, rs (i.e. ~orig/ting/ceForth40 and ~orig/802) which I proposed to Dr. Ting in our email exchanges.

• CC 20210816: Code Merge

  • Targeting multi-platform. Common source by consolidating ceForth, wineForth, ESP32forth (kept in ~/orig/*). Officially version 8.0

• CC 20220512: Refactor

  • Though the goal of Dr. Ting's is to demonstrate how a Forth can be easily understood and cleanly constructed. However, the token threading method used is costly (slow) because each call needs 2 indirect lookups (token->dict, dict->xt). On top of that, C/C++ call-frame needs to be setup/teardown. It is worsen by the branch prediction missing every call stalling the CPU pipeline. Bad stuffs!
  • Refactor to subroutine indirect threading. It's not portable but does speed up 25% (see benchmark above).
  • Using 16-bit offsets for pointer arithmetic which speed up another 5% while maintaining 16-bit parameter space consumption.
  • Since C++ code is at least 4-byte aligned and parameter is 2-byte aligned, the LSB of a given parameter is utilized for colon word identification.

• CC 20221118: Refactor

  • WASM function pointer is U32 (index). Token-indirect worked but the two indirect look-up is even slower. Since WASM uses 64K linear memory block, 16-bit pointer offset is a better option. However, the xt "function pointer" in code space is simply an index to the shared _indirect_function_table. Since LSB is used, so we are forced to use MSB to differentiate primitive word from colon word. This left us 15-bit, i.e. 32K, parameter offset available.

• CC 20231011: Review

  • Since the original intention of having a pre-compiled ROM dictionary still end up in C++ static initialization run before main(), moved dictionary compilation into dict_compile as function calls gives a little more debugging control and opportunity for fine tuning.
  • LAMBDA_OK option was originally intended for full VM implementation but 2x slower. Dropped to reduce source clutter.

• CC 20240308: Refactor for multi-platform, accept dynamic vectors

  • Experiment various threading and memory pointer models, archive into ~/orig/40x
  • To support cross-platform, i.g. Linux/Cygwin, Arduino/ESP32, Win32, and WASM, there were many conditional compilation branches which make the code really messy. The following were done
    • Separate cross-platform and configuration into ~/src/config.h
    • Separate platform specific code into ~/platform
    • add included opcode for Forth script loading
    • rename 'next_idiom' to 'word', per Forth standard

• CC 20241001: Add multi-threading support

  • Shared dictionary and code space amount threads.
  • Refactor source into ceforth, ceforth_sys, and ceforth_task for their specific functions.
  • Introduce VM, states
    • local ss, rs, tos, and user area
    • align to cache-line width
    • pass VM& to all lambda and static functions
  • Add thread pool and event_loop with affinity to physical cores.
    • task, start, stop, join for thread life-cycle management
    • add general multi-threading demo
  • Add Inter-task communication
    • pthread mutex and condition variables are used for synchronization
    • rank for task id
    • send, recv, and pull. Use local stack, as queue, for message passing.
    • add producer/consumer demo
  • Add IO sequencing
    • ANSI-Color trace/logging for different cores
    • mutex guard used
    • lock, unlock for output stream synchronization

• CC: 20250610: maintenance and memory leak check

  • Refactor
    • Macros to reduce verbosity i.e. VM referenced TOS, SS, RS, BRAN, BTGT
    • Group IO functions to forth_sys module
    • Macros to clarify intention, i.e. NEST, BASE, ADD_W
    • Code references replace Code pointers
    • Rename ms=>clock, delay=>ms (adhere to Forth Standard)
    • Add destructors to deallocate (reduce valgrind's complaints)
  • Enhance multi-threading
    • Use std::thread instead of pthread (except device specific CPU affinity)
    • Handle recursive include - Save/Restore WP
    • Refined forth_vm state machine transition (QUERY, HOLD, NEST, STOP)
  • Enhance debugging
    • Add dict() to detail dictionary entries
    • Add dump() to show memory/parameter field's content