The Amforth Compiler is based upon immediate words. They are always executed, regardless of the value in the state variable. All non-immediate words get compiled verbatim with their respective execution token. It is simply appended to the current DP location.

Immediate words are usually executed (unless some special action such as postpone is applied). The immediate words do usually generate some data or compile it to the dictionary. They are not compiled with their execution token.

There are no optimization steps involved. The XT are written immediately into the dictionary (flash).

The inner interpreter, the forth virtual machine, can, just like a real CPU, only execute words, one after the next. This linear control flow is usually not sufficient to do real work. The Forth VM needs to be redirected to other places instead of the next one, often depending on runtime decisions.

Since Edsgar Dijkstra the structured programming is the preferred way to do it. AmForth provides all kinds of them: sequences, selections and repetitions. Sequences are the simple, linear execution of consecutive words. Selections provide a conditional jump over code segments. They are usually implemented with the ìf command. Multiple selections can be made with case. Repetitions can be unlimited or limited. Limited Repetitions can use flags and counter/limits to leave the loop.

There is also support for out-of-band control flow: Exceptions. They provide some kind of emergency exits to solve hard problems. They can be catched at any level up to the outer text interpreter. It will print a message on the command terminal and will wait for commands.

Building Blocks

All control structures can be implemented using jumps and conditional jumps. Every control operation results in either a forward or a backward jump. Thus 6 building blocks are needed to create them all: (branch), (0branch), >mark, <mark, >resolve and <resolve. None of them are directly accessible however. Most of these words are used in pairs. The data stack is used as the control flow stack. At runtime the top-of-stack element is the flag. All words are used in immediate words. They are executed at compile time and produce code for the runtime action.

(branch) is a unconditional jump. It reads the flash cell after the command and takes it as the jump destination. Jumps can be at any distance in any direction. (0branch) reads the Top-Of-Stack element and jumps if it is zero (e.g. logically FALSE). If it is non-zero, the jump is not made and execution continues with the next XT in the dictionary. In this case, the branch destination field is ignored. These two words are implemented in assembly. A equivalent forth implementation would be

: (branch) r> 1+ @i >r ;
: (0branch) if (branch) else r> 1+ >r then ;

Note the chicken-and-egg problem with the conditional branch operation.

Contrary the MSP430. Its inner interpreter uses relative branches instead. That influences the next higher level word internally, but does not affect words using them.

The mark words put the jump destination onto the data stack. This information is used by the resolve words to actually complete the operation. The <mark additionally reserves one flash cell. The <resolve stores the information for the backward jump at the current location of the dictionary pointer, the >resolve places the information at the place the >mark has reserved and completes the forward jump. Every mark needs to be paired with the right resolve.

: >mark dp -1 , ;
: >resolve ?stack dp swap !i ;

: <mark dp ;
: <resolve ?stack , ;

The place holder -1 in >mark prevents a flash erase cycle when the jump is resolved using the !i in >resolve. The ?stack checks for the existence of a data stack entry, not for a plausible value. It the data stack is empty, an exception -4 is thrown.

: ?stack depth 0< if -4 throw then ;

Highlevel Structures

The building blocks described above create the standard control structures: conditional execution and various loop constructs.

The conditional execution compiles a forward jump to another location. The jump destination is resolved with then. An else terminates the first jump and starts a new one for the final then. This way an alternate code block is executed at runtime depending on the flag given to the if.

: if   postpone (0branch) >mark ; immediate
: else postpone (branch)  >mark swap >resolve ; immediate
: then >resolve ; immediate

There is a rarely used variant of the if command, that compiles an unconditional forward branch: ahead. It needs to be paired with a then to resolve the branch destination too. An else would not make any sense, but is syntactically ok.

: ahead postpone (branch) >mark ; immediate

There are more variants of multiple selections possible. The case structure is based upon nested if‘s. Computed goto’s can be implemented with jump tables whith execution tokens as code blocks. Examples are in the lib directory.

The loop commands create a structure for repeated execution of code blocks. A loop starts with a begin to which the program flow can jump back any time.

: begin <mark ; immediate

The first group of loop command are created with again and until. They basically differ from each with the branch command they compile:

: until postpone (0branch) <resolve ; immediate
: again postpone (branch) <resolve ; immediate

The other loop construct starts with begin too. The control flow is further organized with while and repeat. while checks wether a flag is true and leaves the loop while repeat unconditionally repeats it. Multiple while ‘s are possible, they have to be terminated properly with a then for each of them (except the one, which is terminated with the repeat.

: while postpone (0branch) >mark swap ; immediate
: repeat again >resolve ; immediate

Counted loops repeat a sequence of words for some predefined number of iterations. It is possible to exit prematurely. The standard loop checks for the exit condition after the loop body has been executed. A special variant (?DO) does it once at the beginning and may skip the loop body completely. To actually implement the loop and its possible exit points a separate LEAVE stack (named after the LEAVE forth word) is used at compile time. It receives all premature exit points which are resolved when compiling LOOP (or +LOOP).

: endloop
 <resolve \ standard backward loop
 \ now resolve the premature exits from the leave stack
 begin l> ?dup while postpone then repeat ;

: do postpone (do) <mark 0 >l ; immediate
: loop postpone (loop) endloop ;  immediate
: +loop postpone (+loop) endloop ; immediate
: leave postpone unloop postpone ahead >l ; immediate

unloop is an assembly word dropping the loop counter and loop limit information from the return stack.

The ?do works differently. It uses the do and the leave stack to achieve its goals.

... ?docheck if do ... loop then ....

The helper word ?docheck checks the loop numbers and creates a well prepared stack content.

\ helper word
: ?docheck ( count limit -- count limit true | false )
    2dup = dup >r if 2drop then r> invert ;

: ?do postpone ?docheck
    postpone if \ here we create the forward branch
    postpone do \ initialite leave stack
    swap >l     \ put the IF destination on the leave stack
; immediate

The runtime action of do (the (do)) puts two information onto the return stack: The modified loop counter abd the loop limit. The loop index and the loop limit are modified by adding 0x8000 to both numbers. That makes it easy to check the boundary cross required by Forth by simply checking the controller overflow check. The price to pay is a slightly slower access to the loop index (I and J).

The runtime of loop (the (loop)) checks the limits and with 0branch decides whether to repeat the loop body with the next loop counter value or to exit the loop body. If the loop has terminated, it cleans up the return stack. The +loop works almost identically, except that it reads the loop counter increment from the data stack.

The access to the loop counters within the loops is done with i and j. Since the return stack is used to manage the loop runtime, it is necessary to clean it up. This is done with either unloop or leave. Note that unloop does not leave the loop!


DOES> is used to change the runtime action of a word that create has already defined. Since the dictionary is in flash which may only be written once, the use of create is should be replaced with the command <builds. This commands works exactly the same way but enables does> to work properly.

Its working is described best using a simple example: defining a constant. The standard word constant does exactly the same.

> : con <builds , does> @i ;
> 42 con answer
> answer .
 42 ok

The first command creates a new command con. With it a new word gets defined, in this example answer. con calls create, that parses the source buffer and creates a wordlist entry answer. After that, within con the top-of-stack element (42) is compiled into the newly defined word. The does> changes the runtime of the newly defined word answer to the code that follows does>.

does> is an immediate word. That means, it is not compiled into the new word (con) but executed when con gets compiled. This compile time action creates a small data structure similar to the wordlist entry for a noname: word. The address of this data structure is an execution token. This execution token replaces the standard XT that a wrongly defined con (using create instead of builds) would have written already. This leads inevitably to a flash erase cycle, that may not be available on all platforms.