In a conventional concatenative language, every symbol represents a function which takes program states to program states, and the program is a concatenation (sequential composition) of such functions. This is fine when it comes to manipulating state, but what about control?
One can allow individual functions to be defined and named seperately, then applied, perhaps conditionally and perhaps recursively. While this is conventional in the world of concatenative languages, such languages are arguably no longer purely concatenative (a subject explored by Carriage and Equipage).
If one wishes the language to remain purely concatenative, it seems one must store control structures in the state (for example, Equipage stores functions on the stack) or, more drastically, allow the function that is being constructed, to be examined somehow during the construction process. Functions typically aren't considered examinable in this way, but even so, examining it this way is not very different from putting a copy of the program itself in the state.
Wagon is an experiment with a third option: second-order functions. Instead of functions which take states to states, the symbols of the language represent functions which take functions from states to states, to functions that take states to states. The state is merely a stack of integers, and the program is purely a concatenation of functions.
My hope was that these second-order functions could express control in addition to expressing state manipulation. This does turn out to be the case, but only to a degree. As far as I have been able to determine, they can express some control structures, but not arbitrary ones.
Fundamentals of Wagon
Talking about functions that themselves work on functions is admittedly somewhat awkward, so let's define some terms. Let's call a function that takes states to states an operation. Let's call a function that takes operations to operations a macro. A Wagon program, then, consists of a sequence of symbols, each of which represents a macro. These individual macros are concatenated (sequentially composed) to form a single macro.
Since a macro takes operations to operations, it's not possible to "run" a macro directly. So, when asked to "run" or "evaluate" a Wagon program, we take the following convention: apply the macro that the Wagon program represents to the identity function, and apply the resulting function to an initial stack. We may allow a Wagon program to accept input by supplying it on the initial stack, but usually the initial stack is just empty. The output is usually a depiction of the final state of the stack.
Many of the primitive macros in Wagon are based directly on operations. There are two main ways we "lift" an operation k to a macro. The first is a macro which takes an operation o and returns an operation which performs o then performs k — we call this sort of macro an after-lifting of k. The second is a macro which takes an operation o and returns an operation which performs k then performs o — we call this sort of macro a before-lifting of k.
Wagon's convention is that after-lifted operations are represented by lowercase letters, and before-lifted operations are represented by uppercase letters. Macros which do not clearly fall into either of these two categories are represented by punctuation symbols.
Primitive Macros of Wagon
-> Tests for functionality "Evaluate Wagon Program" -> Functionality "Evaluate Wagon Program" is implemented by -> shell command "bin/wagon run %(test-body-file)" -> Functionality "Evaluate Wagon Program" is implemented by -> shell command "bin/wagon eval %(test-body-file)"
Note that in the following examples, the
===> line is not part of the program;
instead it shows the expected result of running each program. In the expected
result, the resulting stack is depicted top-to-bottom.
i is a macro which takes an operation o and returns an operation which
performs o then pushes a 1 onto the stack.
s is a macro which takes an operation o and returns an operation which
performs o then pops a from the stack then pops b from the stack and
pushes b - a.
With these we can construct some numbers.
i ===>  iis ===>  iis is ===> [-1] i iis is s ===> 
As you can see,
s are after-lifted operations. There are also
before-lifted counterparts of them:
I is a macro which takes an operation o and returns an operation which
pushes a 1 onto the stack then performs o.
S is a macro which takes an operation o and returns and operation which
pops a from the stack then pops b from the stack and pushes b - a
then performs o.
SII ===> 
Note also that whitespace maps to the identity function (a macro which takes an operation and returns that same operation) so is effectively insignificant in a Wagon program.
Rudimentary Stack Manipulation
p is an after-lifting of the operation: pop a value from the stack and discard it.
i iis iis iis ppp ===> 
P is the before-lifted counterpart to
PI ===> 
d is an after-lifting of the operation: duplicate the top value on the stack.
iis ddd ===> [0,0,0,0]
D is the before-lifted counterpart to
DDDI ===> [1,1,1,1]
Sophisticated Stack Manipulation
r is an after-lifted operation: it pops a value n from the stack. Then it
pops n values from the stack and temporarily remembers them. Then it
reverses the remainder of the stack. Then it pushes those n remembered
values back onto the stack. n must be zero or one.
iis i iiisiss ===> [2,1,0] iis i iiisiss iis r ===> [0,1,2] iis i iiisiss i r ===> [2,0,1]
R is the before-lifted counterpart to
I SII ===> [1,0] R SII I SII ===> [0,1]
Rudimentary Control Flow
@ (pronounced "while") is a macro which takes an operation o and returns
an operation that repeatedly performs o as long as there are elements on
the stack and the top element of the stack is non-zero.
p@ I I I SII SII ===> [0,0]
An "if" can be constructed by writing a "while" that, first thing it does is, pop the non-zero it detected, and last thing it does is, push a 0 onto the stack. Then immediately afterwards, pop the top element of the stack (which we know must be zero because we know the loop just exited, whether it was executed once, or not at all.)
When Wagon was first formed, it was not clear if it would be Turing-complete or not. The author observed it was possible to translate many programs written in Loose Circular Brainfuck into Wagon, using the following correspondence:
+ iisiss - is > iisrir < iriisr x[y]z y@Xz
Loose Circular Brainfuck is Turing-complete, so if this correspondence was total, Wagon would be shown Turing-complete too. However, the correspondence is not total.
@ that's the problem. We can construct a "while" loop
with contents, and with operations that happen before it,
and with operations that happen after it. And the contents can themselves
contain a nested "while" loop. But we cannot place a second "while" loop
after an already-given "while" loop. That is, we cannot have more than
one "while" loop on the same nesting level.
This might be best illustrated with a depiction of the nested structure that a Wagon program represents.
-> Tests for functionality "Depict Wagon Program" -> Functionality "Depict Wagon Program" is implemented by -> shell command "bin/wagon depict %(test-body-file)" p@ I I I SII SII ===> Push1 Push1 Sub Push1 Push1 Sub Push1 Push1 Push1 (while Pop) is@I is@I ===> Push1 (while Push1 (while Push1 Sub) Push1 Sub) isis@I @I ===> Push1 (while Push1 (while Push1 Sub Push1 Sub)) i@Dp ===> Dup (while Push1) Pop i@Dp i@Dp ===> Dup (while Dup (while Push1) Pop Push1) Pop
While it is possible to simulate a universal Turing machine with only a single top-level "while" loop and a series of "if" statements inside the body of the "while" (see, for example, Burro), it is not known to me if it is possible to simulate a universal Turing machine with only strictly-singly-nested "while" loops as constructible in Wagon.
Any inner "while" loop can be turned into an "if" as described above, but a conventional "if/else" is not possible, because it would normally require two consecutive "while"s, one to check the condition, and one to check the inverse of the condition. Similarly, it would not be possible to check if the finite control of a Turing machine is in one state, or another state. This would seem to be a fairly serious restriction.
However, shortly after being announced in the
#esoteric IRC channel, it was
shown by int-e
that it is possible to compile a Tag system into a Wagon program.
Since Tag systems are Turing-complete, Wagon is as well.
As of this writing, it remains unclear if Wagon is able to simulate a Turing machine or Loose Circular Brainfuck program directly rather than via a Tag system.
August 13th or maybe August 16th or maybe even September 4th, 2019