Internal Decisions

Part of the LGPL:ed PFE

The following section about the Head.aux field is outdated since the 32.x generation of PFE. Please skip to the next section labelled "non-const >body"

The Head.aux Field

this section applies up to and including the 31.x generation of pfe.

The ANSI standard on forth requires that a DOES> code changes the CFA-vector of the latest CREATE to the colon's token-list right after the DOES>. However, PFE does actually regard the value in the CFA as the address of an actuall C-routine, so it does simply jump there - the bytes at the DOES> address need to be actual cpu asm code, but PFE can not compile cpu specific code as it would require knowledge about the cpu beyond that of the C compiler.

Actually, the CFA of the last CREATE word is changed to a C routine in the standard .text body that shall start to execute the colon's token-list right after DOES>. However, if the CFA of CREATE points to a common routine the .text section, how shall the DODOES code know where to start execution?

In the forth-83 and earlier forth dialects, the address where to actually jump to was stored in the field directly following the CREATE address, i.e. in the first PFA cell. Therefore they used the word <BUILDS (defined as : <BUILDS CREATE 0 , ;) to ALLOT the extra field, and the sequence after DOES> did receive the address of CREATE plus one cell. The forth'94 standard does not allow this - the sequence after DOES> gets the actual PFA address, and <BUILDS is not defined in the standard - just use CREATE.

All non-asm implementation of forth do the same trick - they add the extra field before the CFA, in PFE called the AUX-field. It is only used by DOES-words but by its implementation issues it will be allocated for every word. The header size of a word in a non-asm implementation is therefore bigger than in in asm implementations of forth just for these DOES-words.

To sum it up - between the LINK-field and the CODE-field is an AUX-field. All forth languages base on C do this.

(the name aux is from older times where this field was also used for some other information instead of the does-address. it isn't now and the field should perhaps be just named Head.does but feel free to use this field if there is a special need)

The Flag.byte Field

The forth'94 does not do any requirements on the structure of a name-field, it does not even need to exists. The forth'83 implementations had a atleast a header field, and the traditional structure of a name field from FIG-times is a counted string whose upper bits have been (mis-)used for the flag-bits. The highest bit marks it as the the flagfield (it is always set), followed by the immediate-bit and the smudge-bit. This leaves just 5 bit for the actual count of the name in the lower bits, and the names were hence limited to 31 chars.

However, 31 chars is clearly inappropriate for names more than twenty years later - the PFE is often used to interface to C-defined API and whereas old implementations of C had a 32-char limit too, they do not now anymore. The limit had to be raised for PFE, and instead of making extremely radical changes to the name-structure, the trick of F-PC was used - to move the the flags just before the count-byte of the name.

Due to some other implemenation issues (the link-to-name routine looks for the highbit that marks the flag-field) the limit is now 128 chars. As an advantage, the structure of the namefield is now a plain counted string without any flags mangled in. As always, the PFE allows to change back to the traditional ways, and in the source code a macro is used to enhance readability - "_FFA". And the other macro is called "NFACNT" which will return (*nfa&31) for the traditional mode to mask out the flag-bits from the count-value - in the default mode of pfe however, it is just (*nfa). Use these macros to calm down the #ifdef-noise.

To sum it up - the flags live in an extra byte just before the name field which is a normal counted string.

The Head Layout

  structure Head                   structure Head
    1 chars field flags              1 chars field FFA
    x chars field name               x chars field NFA
    1 cells field link               1 cells field LFA
    1 cells field aux                1 cells field AUX
    1 cells field code               1 cells field CFA
    ( followed by body)              ( followed by PFA)
  endstructure                     endstructure

Using `configure` options, both the flags-field and the aux-field can be cut out again to arrive at a header-structure that is FIG-forth compatible. The use of `with-fig` will cut out both fields.


(there are a lot more internal decisions to be explained, however they are of not much interest to forth programmers - they will be added to this document later)

Non-Const >BODY

this section applies to PFE generation 32.x or later.

The first PFE generation took the option to add an extra aux field into the header of a PFE word, and each word in pfe had the exact same layout which did closely follow the FIG model in earlier times. Just the flag.byte was not folded anymore with the count.byte of the name.

As explained above, there is a fundamental problem between the FIG forth and ANS forth systems about the usage of <BUILDS and CREATE related to DOES words. The ANS Forth expects that the code-field of a CREATed word is modified to point to the DOES-field which then contains a trampolin to execute the colon XTs just following it. However, this is impossible to achieve with a traditional indirect-threaded model using no assumptions about the target processor ISA. That's why the PFE does follow the habit of FIG forth to place a second field near the CFA of the CREATEd word in ANS Forth, which is simply the address of the DOES code. Therefore, the DOES-address does effectivly not need to carry an additional code address.

However, there is a problem with compatibility about the second argument, as this one is neither the actual runtime code value of the CREATEd word and it is not data parameter value to be used by the execution of the word. It constitutes its own categorization as a helper field.

In FIG forth, this was different as one was supposed to use the word <BUILDS to CREATE new words that could be extended with DOES>, and the BUILDS word did boil down to be a combination of a basic CREATE followed by " 0 , " followed by DOES> to put its HERE value into the zero just following the CREATE header. The two words CFA and PFA were simply constant operators that did always add/substract a CELL which is simply the size of the runtime time vector in an indirected forth implementation.

Well, there is a problem attached to this mode, since there might be words that want to adjust the first parameter for the DOES code to see, however the PFA[0] was occupied by the DOES-vector, so the first real parameter for the DOES-code would be in PFA[1]. The FIG programs knew this, so they added another CELL for those words being derived via a <BUILDS/DOES> creation. But - ANS Forth does explicitly disallow this, <BUILDS is identical to CREATE here, and the word >BODY is supposed to convert a CFA (code-field address) into a PFA (parameter-field address), not matter if that is a VARIABLE or a DOES word.

There are effectivly three solutions, each having its own derived problems:

code-field is two-cells, always - the first cell is the runtime vector, the second cell is the does vector, the second field is empty for non-does words.
problem: this is a waste of memory, and quite some old FIG programs made an explicit assumption about the size of the codefield, so they did use CELL- directly to end up at the CFA, which is wrong here.
code-field is one plus one cells, always - the runtime vector is preceded with a does cell that is left empty for non-does words.
problem: again a waste of memory, however old FIG problems can still add/substract a CELL to convert between CFA and PFA. The two words >BODY and PFA are identical and constant. The ANS Forth words will find the first parameter directly in PFA[0], even for DOES words, as it is written there. This breaks however FIG forth programs assuming PFA[1] or simply two cells off the CFA to be the first parameter entry for a DOES word
code-field has variable size, one cell for non-does words, two-cells for the does-words.
problem: this mode does actually follow directly the size assumptions of FIG forth programs, a VARIABLE's parameter field is just one cell after the CFA, and for a DOES-word it is two cells. Since ANS Forth's >BODY is supposed to convert to the first parameter, no matter which CFA it is given, the to-body execution becomes a non-const word, sometimes addding one cell, sometimes two, depending on the target word.

The PFE up to 31.x did use the second option, one-plus-one, which is also the model of many other portable forth systems, e.g. gforth. Since generation 32.x, we use the third model with a variable to-body execution. Among the advantages is the result of some internal code cleanup - so far the FIG variant of forth did have another header layout of the words, so there were some defines and ifdefs needed in the PFE forth implementation. With the non-const to-body variant however, the FIG and ANS header layout is identical, so it makes it easier to port FIG forth sources to the ANS mode of PFE, and it even opens the possibility to mix ANS-converted code with non-converted code, as it just boils down to replace a wordlist on the search-order. The binary result is identical (unless the flag.byte is used, which is an extra option however).

This advantage of the internal binary layout being identical is countered by some disadvantages. First of all, the ANS forth programs may not have made an assumption that the offset of the parameter field is always the same for both variables and does-words. They must have used >BODY in all places, and not memorized the offset of this word somewhere, as if the >BODY is a constant function. So far, there was no incident about this, and the people on comp.lang.forth did give their consent to this model as being fair - perhaps some native code forth systems with direct threading had been using variable size of the code field anyway, to lower memory consumption - that's because in direct threading, the code-field contains native cpu code being either just the complete runtime or a trampolin into an external runtime code. For a few things like a variable-runtime, this can be quite short for may cpu ISAs.

Secondly, the non-const nature of the to-body execution implies some runtime restrictions. First of all, it definitly needs an if-code, making it a bit slower, and the if-value to be checked is the first cell of the CFA on which to decide what size the complete codefield has, and to add this size of the codefield. The latter implies that there is an additional memory access, which is not only slower than a constant to-body offsetword, but constitutes the availability of a SIGBUS in to-body for bad addresses. The value supplied to that to-body must be a valid CFA, and nothing else. (one can however add some backward compatilibity to check for zero. It returns the size of the variable and colonword codefield).

And thirdly, there is some confusion about CREATE - in ANS Forth we must ensure that a DOES vector can be attached, so it must have the two fields ready, while in FIG Forth mode just one cell is needed, and only <BUILDS puts the two-cell type of thing. To get around this, the word CREATE is only a SYNONYM - in an ANS-Forth section, it will resolve to <BUILDS, and show this word on decompilation. The forth application writers are furthermore encouraged to use this word directly in their sources, and add a section like


somewhere in the header of their applications for portability reasons. In a FIG Forth section, the CREATE synonym will resolve to CREATE: which makes a word that is not supposed to be extended with some DOES code later on.

The use of a variable codefield size opens new ways to create optimized code (and codefield areas) for use in native-threading systems, and it seems to be a feature well covered by ANS Forth rules. At the same time, the default pfe internal layout does exactly match that of the FIG, so it makes it a lot easier to convert legacy fig applications to the ANS Forth world. This mode is also recommended to all forth system implementors - do not anymore use the does-field model as it was used in pfe up to 31.x, and adapt your applications to expect the minor restrictions about the usage of words in this mode. It does even save you a few cells in the memory footprint.