blog-static/02_cs325_languages_hw3.md at a026e67a3be5665a8f1d5b95bcc9aa81c517063a

15 KiB

Raw Blame History

title

date

Homework 3

Let's take a look at the three homework problems. The first two are related, but are solved using a different technique:

This problem requires us to find the k numbers closest to some query (which I will call n) from a list xs. The list isn't sorted, and the problem must run in linear time. Sorting the list would require the standard {{< sidenote "right" "n-note" "O(n\log n) time." >}} The n in this expression is not the same as the query n, but rather the length of the list. In fact, I have not yet assigned the length of the input xs to any variable. If we say that m is a number that denotes that length, the proper expression for the complexity is O(m \log m). {{< /sidenote >}} Thus, we have to take another route, which should already be familiar: quickselect. Using quickselect, we can find the kth closest number, and then collect all the numbers that are closer than the kth closest number. So, we need a language that:

Supports quickselect (and thus, list partitioning and recursion).
Supports iteration, {{< sidenote "left" "iteration-note" "multiple times." >}} Why would we need to iterate multiple times? Note that we could have a list of numbers that are all the same, [1,1,1,1,1]. Then, we'll need to know how many of the numbers equally close as the kth element we need to include, which will require another pass through the list. {{< /sidenote >}}

That's a good start. Let's take a look at the second problem:

This problem really is easier. We have to find the position of the closest element, and then try expand towards either the left or right, depending on which end is better. This expansion will take several steps, and will likely require a way to "look" at a given part of the list. So let's add two more rules. We need a language that also:

Supports looping control flow, such as while.
{{< sidenote "right" "view-note" "Allows for a "view" into the list" >}} We could, of course, simply use list indexing. But then, we'd just be making a simple imperative language, and that's boring. So let's play around with our design a little, and experimentally add such a "list view" component. {{< /sidenote >}} (like an abstraction over indexing).

This is shaping up to be a fun language. Let's take a look at the last problem: {{< codelines "text" "cs325-langs/hws/hw3.txt" 50 64 >}}

This problem requires more iterations of a list. We have several {{< sidenote "right" "cursor-note" ""cursors"" >}} I always make the language before I write the post, since a lot of design decisions change mid-implementation. I realize now that "cursors" would've been a better name for this language feature, but alas, it is too late. {{< /sidenote >}} looking into the list, and depending if the values at each of the cursors add up, we do or do not add a new tuple to a list. So, two more requirements:

The "cursors" must be able to interact.
The language can represent {{< sidenote "left" "tuple-note" "tuples." >}} We could, of course, hack some other way to return a list of tuples, but it turns out tuples are pretty simple to implement, and help make for nicer programming in our language. {{< /sidenote >}}

I think we've gathered what we want from the homework. Let's move on to the language!

A Language

As is now usual, let's envision a solution to the problems in our language. There are actually quite a lot of functions to look at, so let's see them one by one. First, let's look at qselect.

After the early return, the first interesting part of the language is the use of what I have decided to call a list traverser. The list traverser is a generalization of a list index. Whenever we use a list index variable, we generally use the following operations:

Initialize: we set the list index to some initial value, such as 0.
Step: If we're walking the list from left to right, we increment the index. If we're walking the list from right to left, we decrement the index.
Validity Check: We check if the index is still valid (that is, we haven't gone past the edge of the list).
Access: Get the element the cursor is pointing to.

A {{< sidenote "right" "cpp-note" "traverser declaration" >}} A fun fact is that we've just rediscovered C++ iterators. C++ containers and their iterators provide us with the operations I described:

We can initialize an iterator like auto it = list.begin(). We can step the iterator using it++. We can check its validity using it != list.end(), and access what it's pointing to using *it. While C++ uses templates and inheritance for this, we define a language feature specifically for lists.

{{< /sidenote >}} describes these operations. The declartion for the bisector traverser creates a "cursor" over the list xs, that goes between the 0th and last elements of xs. The declaration for the pivot traverser creates a "cursor" over the list xs that jumps around random locations in the list.

The next interesting part of the language is a traverser macro. This thing, that looks like a function call (but isn't), performs an operation on the cursor. For instance, pop! removes the element at the cursor from the list, whereas bisect! categorizes the remaining elements in the cursor's list into two lists, using a boolean-returning lambda (written in Java syntax).

Note that this implementation of qselect takes a function c, which it uses to judge the actual value of the number. This is because our qselect won't be finding the smallest number, but the number with the smallest difference with n. n will be factored in via the function.

Next up, let's take a look at the function that uses qselect, closestUnsorted:

Like we discussed, it finds the kth closest element (calling it min), and counts how many elements that are equal need to be included, by setting the number to k at first, and subtracting 1 for every number it encounters that's closer than min. Notice that we use the valid! and step! macros, which implement the opertions we described above. Notice that the user doesn't deal with adding and subtracting numbers, and doing comparisons. All they have to do is ask "am I still good to iterate?"

Next, let's take a look at closestSorted, which will require more traverser macros.

The first new macro is canstep!. This macro just verifies that the traverser can make another step. We need this for the "reverse" iterator, which indicates the lower bound of the range of numbers we want to return, because subset! (which itself is just Python's slice, like xs[a:b]), uses an inclusive bottom index, and thus, we can't afford to step it before knowing that we can, and that it's a better choice after the step.

Similarly, we have the at!(t, i) macro, which looks at the traverser t, with offset i.

We have two loops. The first loop runs as long as we can expand the range in both directions, and picks the better direction at each iteration. The second loop runs as long as we still want more numbers, but have already hit the edge of the list on the left or on the right.

Finally, let's look at the solution to xyz:

I won't go in depth, but notice that the expression in the span part of the traverser declaration can access another traverser. We treat as a feature the fact that this expression isn't immediately evaluated at the place of the traverser declaration. Rather, every time that a comparison for a traverser operation is performed, this expression is re-evaluated. This allows us to put dynamic bounds on traversers y and z, one of which must not exceed the other.

This is more than enough to work with. Let's move on to the implementation.

Implementation

Again, let's not go too far into the details of implementing the language from scratch. Instead, let's take a look into specific parts of the language that deserve attention.

Revenge of the State Monad

Our previous language was, indeed, a respite from complexity. Translation was straightforward, and the resulting expressions and statements were plugged straight into a handwritten AST. We cannot get away with this here; the language is powerful enough to implement three list-based problems, which comes at the cost of increased complexity.

We need, once again, to generate temporary variables. We also need to keep track of which variables are traversers, and the properties of these traversers, throughout each function of the language. We thus fall back to using Control.Monad.State:

{{< todo >}}Code for Translator Monad{{< /todo >}}

There's one part of the state tuple that we haven't yet explained: the list of statements.

Generating Statements

Recall that our translation function for expressions in the first homework had the type:

translateExpr :: Expr -> Translator ([Py.PyStmt], Py.PyExpr)

We then had to use do-notation, and explicitly concatenate lists of emitted statements. In this language, I took an alternative route: I made the statements part of the state. They are thus implicitly generated and stored in the monad, and expression generators don't have to worry about concatenating them. When the program is ready to use the generated statements (say, when an if-statement needs to use the statements emitted by the condition expression), we retrieve them from the monad:

{{< todo >}}Code for getting statements{{< /todo >}}

Validating Traverser Declarations

We declare two separate types that hold traverser data. The first is a kind of "draft" type, TraverserData. This record holds all possible configurations of a traverser that occur as the program is iterating through the various key: value pairs in the declaration. For instance, at the very beginning of processing a traverser declaration, our program will use a "default" TraverserData, with all fields set to Nothing or their default value. This value will then be modified by the first key/value pair, changing, for instance, the list that the traverser operates on. This new modified TraverserData will then be modified by the next key/value pair, and so on. This is, effectively, a fold operation.

{{< todo >}}Code for TraverserData{{< /todo >}} {{< todo >}}Maybe sidenote about fold?{{< /todo >}}

The data may not have all the required fields until the very end, and its type reflects that: Maybe String here, Maybe TraverserBounds there. We don't want to deal with unwrapping the Maybe a values every time we use the traverser, especially if we've done so before. So, we define a ValidTraverserData record, that does not have Maybe arguments, and thus, has all the required data. At the end of a traverser declaration, we attempt to translate a TraverserData into a ValidTraverserData, invoking fail if we can't, and storing the ValidTraverserData into the state otherwise. Then, every time we retrieve a traverser from the state, it's guaranteed to be valid, and we have to spend no extra work unpacking it. We define a lookup monadic operation like this:

{{< todo >}}Code for getting ValidTraverserData{{< /todo >}}

Compiling Macros

I didn't call them macros for no reason. Clearly, we don't want to generate code that {{< sidenote "right" "increment-note" "calls functions only to increment an index." >}} In fact, there's no easy way to do this at all. Python's integers (if we choose to represent our traversers using integers), are immutable. Furthermore, unlike C++, where passing by reference allows a function to change its parameters "outside" the call, Python offers no way to reassign a different value to a variable given to a function.

For an example use of C++'s pass-by-reference mechanic, consider std::swap: it's a function, but it modifies the two variables given to it. There's no way to generically implement such a function in Python. {{< /sidenote >}} We also can't allow arbitrary expressions to serve as traversers: our translator keeps some context about which variables are traversers, what their bounds are, and how they behave. Thus, calls to traverser macros are very much macros: they operate on AST nodes, and require that their first argument is a variable, named like the traverser. We use the requireTraverser monadic operation to get the traverser associated with the given variable name, and then perform the operation as intended. The at!(t) operation is straightforward:

{{< todo >}}Code for at!{{< /todo >}}

The at!(t,i) is less so, since it deals with the intricacies of accessing the list at either a positive of negative offset, depending on the direction of the traverser. We implement a function to properly generate an expression for the offset:

{{< todo >}}Code for traverserIncrement{{< /todo >}}

We then implement at!(t,i) as follows:

{{< todo >}}Code for at!{{< /todo >}}

The most complicated macro is bisect!. It must be able to step the traverser, and also return a tuple of two lists that the bisection yields. We also prefer that it didn't pollute the environment with extra variables. To achieve this, we want bisect! to be a function call. We want this function to implement the iteration and list construction.

bisect!, by definition, takes a lambda. This lambda, in our language, is declared in the lexical scope in which bisect! is called. Thus, to guarantee correct translation, we must do one of two things:

Translate 1-to-1, and create a lambda, passing it to a fixed bisect function declared elsewhere.
Translate to a nested function declaration, inlining the lambda.

{{< todo >}}Maybe sidenote about inline?{{< /todo >}}

Since I quite like the idea of inlining a lambda, let's settle for that. To do this, we pull a fresh temporary variable and declare a function, into which we place the traverser iteration code, as well as the body of the lambda, with the variable substituted for the list access expression. Here's the code:

{{< todo >}}Code for bisect!{{< /todo >}}

15 KiB Raw Blame History