Header Fields and Terms

Part of the LGPL:ed PFE


There is a set of established terminology about describing the way that the forth system registers words and executions. A forth word does have a few parts that are sometimes referred to with different words coming from a different historic background.

NAME-field, NFA
A forth word does have a name usually which can be used in source text to refer to the other parts associated with that forth name. Note however that ":NONAME" creates a word without a name. The starting adress of the NAME-field is called the NFA, the Name-Field Adress. Ambiguities arise when a variable holds an NFA being an NFA-pointer but such a thing is often called an NFA-field.
CODE-field, CFA, CODE-area, Execution
A forth word does have an execution behavior usually which is coded in the native instruction set of the local CPU. This is the primary code and the starting adress is usally called CFA, the Code-Field Adress. There is an important ambiguity with ITC, the indirect threaded code where CFA-pointer itself is called the Code-Field, so its adress is a pointer to a pointer.
XT, Execution-Token, Synonym
Following from above, it is better to speak of an Execution Adress here. Actually, there exist implementations where the indirection is done through a table with pointers to code, and the execution field carries an integer. That brought up the tradition to call it an Execution-Token. The forth standard describes it as the primary accessor to a word being also valid for nameless words. An exception is implemented in PFE where the forth dictionary may carry Synonym entries being a name-field pointing to the execution of another word which makes the execution token not a unique thing for a word entry.
PFA, BODY, Primitives
LFA, LINK, Threads
FFA, Flags, Immediate

The list above contains the usual fields that are put into a word entry in the forth dictionary. Some forth system have other fields. A common approach would be to split the execution field into multiple fields that could be used for optimizations, e.g. different compile-state and interpret-state executions, specialized compile-executions for native code threading, or similar stuff - often these get hinted with additional flag bits that denote which part shall be used and how those field values shall be interpreted.

Traditional Word Entry Layout

The Forth Interest Group (FIG) did publish a Forth reference implementation in the late 70s which were allowed to be picked up even by commercial entities. Consequently, a lot of real Forth systems in the 80s happened actually derivates of that FIG-Forth. Even the first Forth Standard (Forth83) transpired a lot of the implementation specifics of the FIG-Forth model.

That FIG-Forth has a known layout of Word Entries in the Forth Wordlists, and it did happen that a lot of Forth applications were to make assumptions on the exact parts of that layout. The first Forth83 was trying to make up a dedicated Wordset that were to cover those specifics in the hope that applications would atleast not make assumptions on the exact size and order of the fields in a word entry.

Still, many specialty applications exist that did not pick up these Forth83 words, and PFE does try to follow the FIG word layout quite closely. In this section, we describe the original layout and semantics of a word entry and its fields to give you an idea about the extensions that were developed later on.

 NFA       LFA      CFA       PFA
 | name[n] | link*  | code*   | data....
            nextNFA  C call    body(only for non-prims)

Above is the traditional layout: each word entry starts with the name field, and the name is put inline, so that field has a variable length. The name string uses the counted string representation where the first byte holds the length of the actual string followed by the characters of that string. It is therefore easy to convert an NFA into an LFA by just fetching the length byte and adding it.

After the name field follows the link field. In FIG-Forth it points to the start of the elder word entry which happens to be a name field adress as well. As a pointer it is exactly one CELL long followed by the execution field. In FIG-Forth that field has a pointer directly to the CODE execution being associated with the word, and it was quite usually called CFA back then (even tho that is not quite correct as it is the adress of a pointer to CODE - but there were no other execution representations at that time). After that CFA pointer follows the body of the word (if not primitive). In FIG-Forth, the start adress was called the PFA and the area known as the Parameter-Field.

The last thing not noted so far: where are the Flag bits of the word entry. Actually, in FIG-Forth it was decided to cut off a few bits in the count byte of the Name-Field. The upper three bits are reserved for usage as Flag-Bits, which has two consequences. First, the NFA and FFA are exactly the same addess. Secondly, the length of a name field length was limited to 31 being the maximum number hat can be expressed in the 5bits remaining in an octet byte.

To get a bit into details here, be noted that the operation of NFA into LFA is slightly modified here with fetching the byte value from the NFA, and then using an AND31 to get at the length. Effectivly, in later FIG-Forth derived forth systems the following word and implementation exists:

 : N>LINK ( nfa -- lfa ) COUNT 31 AND + ;

The inverse operation is a bit more complex. If you do only have an LFA then there is no count byte nearby that tells of the length of the Name-Field. Therefore one can not just add a length value to the LFA to get to the NFA. Instead a trick was used under the assumption that the Name-Field contains only ASCII characters which use only the lower 7bits of an octet byte. Therefore the highest bit of each character is zero, and it was defined that the highest bit in the count byte a.k.a. the higest bit in the Flag-Field would be set to one. Which leads to the following word and implementation:

 : L>NAME ( lfa -- nfa ) REPEAT 1- DUP C@ 128 AND UNTIL ;

The other fields have a fixed size in FIG-Forth. Let's define the other conversion words here - the names stem from the Forth-83 Standard that was trying to allow forth system implementations with different sizes for each field or simply allow a different order of the fields.

 : LINK> ( lfa -- cfa ) CELL + ;       \ "name-from"
 : >BODY ( cfa -- pfa ) CELL + ;       \ "to-link"
 : >LINK ( cfa -- lfa ) CELL - ;       \ "to-link"
 : BODY> ( pfa -- cfa ) CELL - ;       \ "body-from"
 : >NAME ( cfa -- nfa ) >LINK L>NAME ; \ "to-name"
 : NAME> ( nfa -- cfa ) N>LINK LINK> ; \ "name-from"

Searching A Word

As an introductory implementation example, we show here a procedure to walk a wordlist and print all WORDS to the screen.

  : WORDS ( -- )
     CONTEXT @ ( get the most current wordlist in the ORDER )
     @ ( get the pointer to the LATEST word in the wordlist )
        DUP COUNT 31 AND TYPE CR ( print the name field of the word )
        N>LINK @           ( fetch the next name field in the chain )

This example shows how each word chains to the next via the value in the Link-Field, and each Name-Field is a counted string with a maxium of 31 characters that we want to print to the screen. A common operation however is not to print just the word name but to look for the definition hints associated with a given name needed by the application, so we search the wordlist for the word entry of that name and a return a reference to it.

The early FIG-Forth system was case-insensitive and it was not using hashed wordlist threads. A wordlist was simply a pointer to the latest word entry in the chain. Each word entry has a link field pointing to the next name, the last word in the chain has simply a null. That makes for the following implementation valid for FIG-Forth systems:

  : SEARCH-WORDLIST ( name-str name-len wordlist-ptr -- 0 | xt 1 )
    @ ( fetch the adress of the first name in the chain )
      2DUP C@ 31 AND = ( the two names must have atleast the same length )
      IF ( then do the slower operation to mem compare the two name areas )
        3DUP OVER COMPARE  0= IF ( found a match ... to be returned )
           >R 2DROP R> NAME> TRUE   
      THEN ( otherwise - go to the next word entry )
      >LINK @

Note how that implementation makes a number of assumptions: the first field in the wordlist reference has the pointer to the word chain. A word pointer is identical with a name adress. A name is a counted string with a maximum of 31 characters. The link field points to the next word entry being the name adress of that word. And finally, the example is somewhat wrong since it does not return the +1/-1 categorization of the word being found, it should be -1 for immediate words. We would have to return a hint based on the Flag-Field content.

The immediate-flag is one of the three bits reserved for the Flag-Field. By tradition the highest bit (decimal 128) is used for the name-field-start detection algorithm, the middle bit (decimal 64) is used for the immediate-flag, and the lowest flag (decimal 32) is used for the smudge-bit. The latter bit denotes names that are invalidated and which should not be returned from a SEARCH-WORDLIST. A more realistic implementation might look like this:

  : SEARCH-WORDLIST ( name-str name-len wordlist-ptr -- 0 | xt +1 | xt -1 )
    @ ( fetch the adress of the first name in the chain )
      C@ 32 AND 0= IF ( the name is not smudged )
        2DUP C@ 31 AND = ( the two names must have atleast the same length )
        IF ( then do the slower operation to mem compare the two name areas )
          3DUP OVER COMPARE  0= IF ( found a match ... to be returned )
            >R 2DROP R@ NAME> R> C@ 64 AND IF -1 ELSE +1 THEN   
        THEN ( otherwise - go to the next word entry )
      >LINK @

The SEARCH-WORDLIST procedure is used most often in forth system trying to lookup the word information for a name currently found in the input stream. In the usual case the system is trying to find the execution associated with a given which is the reason why the standard word SEARCH-WORDLIST returns the execution token instead of the name-field that matches - note that in older forth systems a similar word FIND was used that would take a counted string as the input and return the NFA that matches it. A lot of FIG-Forth sourcecode uses that word instead.

To get a better understanding of what the SEARCH-WORDLIST procedure does, have a look at the following visual representation of the chain. Note how beneficial it is to put the characters of the name-field and the link-pointer very close to each other, so they will usually end up on the same cache-line of the CPU. An indirect reference of the name field is a rare implementation option in forth systems, instead we find optimizations where the a combination of name-field, link-field and flag-field are using a separate block in memory (called the header-space) being set aside from the memory area containing code and data of a word. That will ensure all header parts for SEARCH-WORDLIST to be strictly adjascent (without interleaved CFA and PFA) and therefore increase the number of times that a header field is in the cpu cache when being accessed.

WWWW | XXXX |....
     +------+----                                        +--------------
      the latest NFA ----\              /--------------->| DUP CODE.....
        /----------------/              |                +--------------
        V                               |
     +---------+---------+--------+--------+             a normal word with
XXXX | 83 | D  | U  | P  |  YYYY  |  CCCC  |             a 3 character name
                             |                           +--------------
        /--------------------/          /--------------->| THEN CODE....
        V                               |                +--------------
XXXX | C4 | T  | H  | E  | N  |  ZZZZ  |  CCCC  | pppp   an immediate word
     +---------+---------+--------+--------+---------+   with name length 4

Mechanics Of VARIABLE and VALUE

A common thing of firsthand misunderstanding is the fact that the two words VARIABLE and VALUE are no data words. They are primitives which have no BODY-field at all. Instead they do CREATE other words that are non-primitives. The standard word "CREATE" parses the next word from the input stream and creates a header block as shown above including a default CFA value being equivalent to that of a VARIABLE. Let's assume a user did type "VARIABLE MYVAR" then it may result in the following memory layout:

                          |                               +--------------
     /--------------------/                               | "make variable"
     |                                                    +--------------
     V                                                            ^
  +---------+---------+---------+---------+---------+---------    |
  | 88 | V  | A  | R  | I  | A  | B  | L  | E  | XXXX   | cccccc -/
  NFA                                          LFA      CFA

                                              /----->  | "variable runtime"
                                              |        +-----------------
  | 85 | M  | Y  | V  | A  | R  |  YYYY   | cccccc  |  0000   <--- initial
  +---------+---------+---------+---------+---------+--------+     value 
  NFA                           LFA      CFA        PFA

As soon as the forth outer interpreter will find the standard word "VARIABLE" in the input stream, it will call SEARCH-WORDLIST to retrieve the execution of that word. That execution is run and create a variable instance - it parses the next part from the input stream ("MYVAR") and creates a new header block with a specific execution for all variable instances. So - if the outer interpreter does later find "MYVAR" in the input stream then will call SEARCH-WORDLIST to retrieve the execution of that word. That execution is a very simple one, as it will return the PFA on the stack, so we can use "@"-fetch and "!"-store on the data body of the variable.

If we would have been using "VALUE MYVAL" instead then the resulting memory layout would be the same with just a few bits changed. Apart from the slightly different name, it would get another code pointer into its execution field that would point to a "value runtime" instead of "variable runtime". The value runtime does slightly more, it does not return the adress of its data body but the actual value within its data body. To change the data body, a special word "TO" must be used. However, with all the knowledge about word layout that can be achieved differently as well:

  VARIABLE MYVAR        \ a VARIABLE instance does return the
  2 MYVAR !             \ adress of its body - to be stored to.
  MYVAR @ .
> 2
  4 ' MYVAR >BODY !     \ setting the data body of MYVAR to 4
  MYVAR @ .             \ with another variant to store into.
> 4

  6 VALUE MYVAL         \ a VALUE returns the value not  
  MYVAL .               \ just the adress of its data body
> 6                     
  7 TO MYVAL            \ changing needs a dedicated word
> 7
  8 ' MYVAL >BODY !     \ or the same sequence as above where
  MYVAL .               \ we did access data body adress too.
> 8