archives

Making a one-pass compiler by generating fexprs that generate code

I'm starting to write a simple compiler that transcompiles a simple scripting language for general game playing (at least for chess-like and a few other board games) into C which will be compiled in memory with the Tiny C library.

I noticed that there's a mismatch between the order in which parser generators trigger actions and the order in which tree nodes need to be visited in order that identifiers can be type checked and used to generate code.

Parser generators trigger actions from the bottom up. Ideally when you generate code, you visit nodes in the tree in whatever order gives you the type information before you see the identifiers used in an expression. Since fexprs let you control what order you visit parts of the inner expression, they're perfect for that.

So my parser is being written so that the parse generates an s-expression that contains fexprs that when run semantically checks the program and transcompiles in a single pass.

This also suggests a new version of Greenspun's 10th rule:
A sufficiently complex C++ program will contain an informally specified, buggy and incomplete implementation of John Shutt's Kernel language.

A refutation of Gödel's first incompleteness theorem

Contradictions in Gödel's First Incompleteness Theorem.

Notes for the second edition. I have edited this post a bit, sorry if it is against the etiquette. The original post can be found here. I have removed a couple of -periopheral- mistakes. However I want to also warn that the argument is a lot clearer when adapted to a modern rendition of Gödel's theorem, based on Turing machines, as suggested in i this comment. That adaptation is carried out in the comments that hang from the previous referenced comment, this an onwards.

Intro

This refers to Kurt Gödel's "On formally undecidable propositions of principia mathematica and related systems". The notation used here will be the same as that used by Gödel in that paper.

In that work, Gödel starts with a description of the formal system P, which, according to himself, "is essentially the system obtained by superimposing on the Peano axioms the logic of PM".

Then he goes on to define a map Phi, which is an arithmetization of system P. Phi is a one-to-one correspondence that assigns a natural number, not only to every basic sign in P, but also to every finite series of such signs.

One

There are alternative arithmetizations of system P. I will later delve on how many.

This is obvious from simply considering a different order in the basic signs when they are assigned numbers. For example, if we assign the number 13 to "(" and the number 11 to ")", we obtain a different Phi.

If we want Gödel's proof to be well founded, it should obviously be independent of which arithmetization is chosen to carry it out. The procedure should be correct for any valid Phi chosen. otherwise it would **not** apply to system P, but to system P **and** some particular Phi.

To take care of this, in Gödel's proof we may use a symbol for Phi that represents abstractly any possible valid choice of Phi, and that we can later substitute for a particular Phi when we want to actually get to numbers. This is so that we can show that substituting for any random Phi will produce the same result.

The common way to do this is to add an index i to Phi, coming from some set I with the same cardinality as the set of all possible valid Phi's, so we can establish a bijection among them - an index. Thus Phi becomes here Phi^i.

Two

Later on, Gödel proceeds to spell out Phi, his Phi, which we might call Phi^0, with his correspondence of signs and numbers and his rules to combine them.

And then Gödel proceeds to define a number of metamathematical concepts about system P, that are arithmetizable with Phi^0, with 45 consecutive definitions, culminating with the definition of provable formula.

Definition of provable formula means, in this context, definition of a subset of the natural numbers, so that each number in this set corresponds biunivocally with a provable formula in P.

Let's now stop at his definition number (10):


  E(x) === R(11) * x * R(13)

Here Gödel defines "bracketing" of an expression x, and this is the first time Gödel makes use of Phi^0, since:


  Phi^0( '(' ) = 11

  Phi^0( ')' ) = 13

If we want to remain general, we may rather do:


   E^i(x) === R(Phi^i( '(' )) * x * R(Phi^i( ')' ))

Two little modifications are made in this definition. First, we substitute 11 and 13 for Phi^i acting on "(" and ")". 11 and 13 would be the case if we instantiate the definition with Phi^0.

And second, E inherits an index i; obviously, different Phi^i will define different E^i. And so do most definitions afterwards.

Since, for the moment, in the RHS of definitions from (10) onwards, we are not encoding in Phi^i the index i, which has sprouted on top of all defined symbols, we cease to have an actual number there (in the RHS); we now have an expresion that, given a particular Phi^i, will produce a number.

So far, none of this means that any of Gödel's 45 deffinitions are in any way inconsistent; we are just building a generalization of his argument.

Three

There is something to be said of the propositions Gödel labels as (3) and (4), immediately after his 1-45 definitions. With them, he establishes that, in his own words, "every recursive relation [among natural numbers] is definable in the [just arithmetized] system P", i.e., with Phi^0.

So in the LHS of these two propositions we have a relation among natural numbers, and in the RHS we have a "number", constructed from Phi^0 and his 45 definitions. Between them, we have an arrow from LHS to RHS. It is not clear to me from the text what Gödel was meaning that arrow to be. But it clearly contains an implicit Phi^0.

If we make it explicit and generalized, we must add indexes to all the mathematical and metamathematical symbols he uses: All Bew, Sb, r, Z, u1... must be generalized with an index i.

Then, if we instantiate with some particular Phi^i, it must somehow be added in both sides: in the RHS to reduce the given expression to an actual number, and in the LHS to indicate that the arrow now goes from the relation in the LHS **and** the particular Phi^i chosen, to that actual number.

Obviously, if we want to produce valid statements about system P, we must use indexes, otherwise the resulting numbers are just talking about P and some chosen Phi^i, together.

Only after we have reached some statement about system P that we want to corroborate, should we instantiate some random Phi^i and see whether it still holds, irrespective of any particularity of that map.

These considerations still do not introduce contradiction in Gödel's reasoning.

Four

So we need to keep the indexes in Gödel's proof. And having indexes provides much trouble in (8.1).

In (8.1), Gödel introduces a trick that plays a central role in his proof. He uses the arithmetization of a formula y to substitute free variables in that same formula, thereby creating a self reference within the resulting expression.

However, given all previous considerations, we must now have an index in y, we need y^i, and so, it ceases to be a number. But Z^i is some function that takes a natural number and produces its representation in Phi^i. It needs a number.

Therefore, to be able to do the trick of expressing y^i with itself within itself, we need to convert y^i to a number, and so, we must also encode the index i with our 45 definitions.

The question is that if we choose some Phi^i, and remove the indexes in the RHS to obtain a number, we should also add Phi^i to the LHS, for it is now the arithmetic relattion **plus** some arizmetization Phi^i which determine the number in the RHS, and this is not wanted.

Five

But to encode the index, we ultimately need to encode the actual Phi^i. In (3) and (4), If in the RHS we are to have a number, in the LHS we need the actual Phi^i to determine that number. If we use a reference to the arithmetization as index, we'll also need the "reference map" providing the concrete arithmetizations that correspond to each index. Otherwise we won't be able to reach the number in the RHS.

Thus, if we want definitions 1-45 to serve for Gödel's proof, we need an arithmetization of Phi^i itself -with itself.

This may seem simple enough, since, after all, the Phi^i are just maps, But it leads to all sorts of problems.

Five one

Now, suppose that we can actually arithmetize any Phi^i with itself, and that we pick some random Phi^i, let's call it Phi^0: we can define Phi^0 with Phi^0, and we can use that definition to further define 10-45.

But since Phi^0 is just a random arithmetization of system P, if it suffices to arithmetize Phi^0, then it must also suffice to arithmetize any other Phi^i equally well. However, with Phi^0, we can only use the arithmetization of Phi^0 as index to build defns 10-45.

This means that, as arithmetizations of system P, the different Phi^i are not identical among them, because each one treats differently the arithmetization of itself from the arithmetization of other Phi^i.

Exactly identical arithmetical statements, such as definition (10) instatiated with some particular Phi^i, acquire different meaning and truth value when expressed in one or another Phi^i.

Among those statements, Gödel's theorem.

Five two [This argument does not hold if we have to consider recursively enumerable arithmetiazations.]

A further argument that shows inconsistency in Gödel's theorem comes from considering that if we are going to somehow encode the index with Phi^i, we should first consider what entropy must that index have, since it will correspond to the size of the numbers that we will need to encode them. And that entropy corresponds to the logarithm of the cardinality of I, i.e., of the number of valid Phi^i.

To get an idea about the magnitude of this entropy, it may suffice to think that variables have 2 degrees of freedom, both with countably many choices. Gödel very conveniently establishes a natural correspondence between the indexes of the variables and the indexes of primes and of their consecutive exponentiations, but in fact any correspondence between both (indexes of vars, and indexes of primes and exponents) should do. For example, we can clearly have a Phi^i that maps the first variable of the first order to the 1000th prime number exponentiated to the 29th power.

This gives us all permutations of pairs of independent natural numbers, and so, uncountably many choices for Phi^i; so I must have at least the same cardinality as the real line. Therefore y^i doesn't correspond to a natural number, since it needs more entropy than a natural number can contain, and cannot be fed into Z^i, terminating the proof.