#functional-programming #lazy-evaluation #interpreter #language #functional #haskell #programming-language

bin+lib timi

A visual template instantiation machine interpreter to understand how lazy functional programming languages evaluate

3 unstable releases

Uses old Rust 2015

0.2.1 Jan 31, 2017
0.2.0 Jan 31, 2017
0.1.0 Jan 24, 2017

#187 in Visualization

MIT license

130KB
2K SLoC

Build Status Coverage Status Crates.io

TIMi - Template Instantiation Machine Interpreter

A visual, user-friendly implementation of a template instantiation machine. Built to understand how lazily evaluate programming language evaluates.

asciicast

Table of Contents

Quickstart

Binary from cargo

To get the interpreter timi with cargo (Rust's package manager), run

$ cargo install timi && timi

Build from source

Run

$ git clone https://github.com/bollu/timi.git && cd timi && cargo run

to download and build from source.

Using the interpreter

Evaluating Expressions

Type out expressions to evaluate. For example:

> 1 + 1

will cause 1 + 1 to be evaluated

Creating Definitions

Use define <name> [<param>]* = <expr> to create new supercombinators.

> define plus x y = x + y

Will create a function called plus that takes two parameters x and y. To run this function, call

> plus 1 1

Interpreter Options and Usage

>:step

To go through the execution step-by-step, use

>:step
enabled stepping through execution
>1 + 1
*** ITERATION: 1
Stack - 1 items
## top ##
 0x21  ->  ((+ 1) 1)  H-Ap(0x1F $ 0x20)
## bottom ##
Heap - 34 items
 0x21  ->  ((+ 1) 1)  H-Ap(0x1F $ 0x20)
Dump
  Empty
Globals - 30 items
None in Use
===///===
1>>

Notice that the prompt has changed to to >>.

Step commands
  • >>n (for next) to go to the next step

>:nostep

to enable continuous execution of the entire program, use

>:nostep

Executing .tim files

The interpreter can be invoked on a separate file by passing the file name as a command line parameter.

Example: standalone file

create a file called standalone.tim

#standalone.tim
main = 1

Run this using

$ timi standalone.tim

This will print out the program trace:

*** ITERATION: 1
Stack - 1 items
## top ##
 0x1E  ->  1  H-Num(1)
## bottom ##
Heap - 31 items
 0x1E  ->  1  H-Num(1)
Dump
  Empty
Globals - 30 items
None in Use
===///===

=== Final Value: 1 ===

Language Introduction

The language is a small, lazily evaluated language. Lazy evaluation means that evaluation is delayed till a value is needed.

Top level (Supercombinators)

Top level declarations (which are also called supercombinators) are of the form:

<name> [args]* = <core expr>

#####Example:

K x y = x

Multiple top-level declarations are separated by use of ;

I x = x;
K x y = x;
K1 x y = y

notice that the last expression does not have a ;

The main value

when writing a program (not an expression that is run in the interpreter), the execution starts with a top level function (supercombinator) called as main.

Expressions

Expressions can be one of the given alternatives. Note that lambda and case are missing, since they are difficult to implement in this style of machine. More is talked about this in the section Lack of Lambda and Case.

  • Let

    let <...bindings...> in <expr>
    

    Let bindings can be recursive and can refer to each other.

    #####Example: simple let

    > let y = 10; x = 20 in x + y
    ...
    === Final Value: 30 ===
    
    Example: mututally recursive let
    # keep in mind that K x y = x
    > let x = K 10 y; y = K x x in x
    ...
    === Final Value: 10 ===
    

    Even though x and y are defined in terms of each other, they do not refer to each other directly. Rather, they refer to each other as components of some larger function. This is allowed, since it is possible to "resolve" x and y.

    Hence, this example will evaluate to 10.

    Example: code disallowed because of strict let binding

    NOTE: Let binds strictly, not lazily, so this code will not work

    > let y = x; x = y in 10
    *** ITERATION: 1
    step error: variable contains cyclic definition: y
    

    Here, it is impossible to even resolve y and x. So, this will be rejected by the interpreter.

  • Function application

    function <args>
    

    Like Haskell's function application. The <args> are primitive values or variables.

    All n-ary application are represented by nested 1-ary applications. Functions are curried by default.

    f x y z == (((f x) y) z)
    
  • Data Declaration

    Pack{<tag>, <arity>}
    

    The Pack primitive operation takes a tag and an arity. When used, it packages up an arity number of expressions into a single object and tags it with tag.

    Example:

    False = Pack{0, 0}
    True = Pack{1, 0}
    

    True and False are represented as 1 tagged and 0 tagged objects that have arity 0.

    MkPair = Pack{2, 2}
    my_tuple = MkPair 42 -42
    

    MkPair, a function used to create tuples uses a tag of 2 and requires two arguments. my_tuple is now a data node that holds the values 42 and -42.

    NOTE: using custom tags will not be very beneficial since the language does not have case expressions. Rather, List and Tuple are created as language inbuilts with custom de-structuring functions called caseList and casePair respectively.

  • Primitive application

    <arg1> primop <arg2>
    

    Primitive operation on integers. The following operations are supported:

    - Arithmetic
      - `+`: addition
      - `-`: subtraction
      - `*`: multiplication
      - `/`: integer division
    
    - Boolean, returning `True` (`Pack{1, 0}`) for truth and `False` (`Pack{0, 0}`)
      for falsehood:
      `<`, `<=`, `==`, `/=`, `>=`, `>`
    
  • Primitive literal An integer declaration.

    >3
    
  • Booleans

    True = Pack{1, 0}
    False = Pack{0, 0}
    

    True and False are represented by 1 tagged and 0 tagged data types.

  • Tuples Tuples are a language inbuilt and are constructed by using MkPair.

    MkPair <left> <right>
    

    Tuples are pattern matched on by using casePair

    casePair (MkPair a b) f = f a b
    

    Note that the default fst and snd are defined as follows

    K x y = x;
    K1 x y = y;
    fst t = casePair t K;
    snd t = casePair t K1;
    
  • ** Lists Lists are language inbuilts and have two constructors: Nil and Cons

    Nil
    Cons <value> <list>
    

    Lists are pattern matched by using

    caseList <nil-handler> <cons-handler>
    
    • nil-handler is a value
    • cons-handler is a function that takes 2 parameters, the value in the
    • Cons cell and the rest of the list.
  • Comments

    # anything after a # till the end of the line is commented
    main = 1 # this is a comment as well
    

    Comment style is like python, where # is used to comment till the end of the line. There are no multiline comments.

Lack of Lambda and Case

Case

Case requires us to have some notion of pattern matching / destructuring which is not present in this machine.

Lambda

As a simplification, the language assumes that all lambdas have been converted to top level definitions. This process is called as lambda lifting and TIMi assumes that all lambdas have been lifted.

Runtime

functions are curried by default. Thus (f x y z) is actually (((f x) y) z)

Components of the machine

The runtime has 4 components:

  • Heap: a map from addresses to Heap Nodes
  • Stack: a stack of Heap Addresses
  • Dump: a stack of stacks to hold intermediate evaluations
  • Globals: a map from names to addresses

Everything the machine uses during runtime must be allocated on the heap before the machine starts executing. So, we need a way to convert a CoreExpr into a Heap. This conversion process is called as instantiation.

Example of instantiation: Sample program 1 + 1

Consider the program 1 + 1. The initial state of the machine is

>1 + 1
*** ITERATION: 1
Stack - 1 items
## top ##
 0x21  ->  ((+ 1) 1)  H-Ap(0x1F $ 0x20)
## bottom ##
Heap - 34 items
 0x21  ->  ((+ 1) 1)  H-Ap(0x1F $ 0x20)
 0x1F  ->  (+ 1)      H-Ap(0xE $ 0x1E)
 0x1E  ->  1          H-Num(1)
 0xE   ->  +          H-Primitive(+)
 0x20  ->  1          H-Num(1)
Dump
  Empty
Globals - 30 items
 +  ->  0xE

Notice that every single part of the expression 1 + 1 is on the heap, and the symbol + is mapped to its address 0xE in the Globals section. The whole expression sits on top of the stack, waiting to be evaluated.

Evaluation

We will explain how code evaluates by considering an explanation of how 1+1 is evaluated

Example of evaluation: (((S K) K) 3)

Consider the definitions:

S f g x = f x (g x)
K x y = x

(these are the S and K combinators from lambda calculus)

Now, let us understand how the program S K K 3 evaluates.

*** ITERATION: 1
Stack - 1 items
## top ##
 0x21  ->  (((S K) K) 3)  H-Ap(0x1F $ 0x20)
## bottom ##
...
===///===

Initially, the code that we want to run (((S K) K) 3) is on the top of the stack.

*** ITERATION: 2
Stack - 2 items
## top ##
 0x1F  ->  ((S K) K)      H-Ap(0x1E $ 0x1)
 0x21  ->  (((S K) K) 3)  H-Ap(0x1F $ 0x20)
## bottom ##
...
===///===

Remember that all application is always curried. That is (((S K) K) 3) is thought of as (((S K) K) applied on 3.

The LHS of the function application ((S K) K) is pushed on top of the current stack. This process continues till there is a supercombinator on the top of the stack.

*** ITERATION: 3
Stack - 3 items
## top ##
 0x1E  ->  (S K)          H-Ap(0x3 $ 0x1)
 0x1F  ->  ((S K) K)      H-Ap(0x1E $ 0x1)
 0x21  ->  (((S K) K) 3)  H-Ap(0x1F $ 0x20)
## bottom ##
...
===///===
*** ITERATION: 4
Stack - 4 items
## top ##
 0x3   ->  S              H-Supercombinator(S f g x = { ((f $ x) $ (g $ x)) })
 0x1E  ->  (S K)          H-Ap(0x3 $ 0x1)
 0x1F  ->  ((S K) K)      H-Ap(0x1E $ 0x1)
 0x21  ->  (((S K) K) 3)  H-Ap(0x1F $ 0x20)
## bottom ##
Heap - 34 items
 0x1F  ->  ((S K) K)      H-Ap(0x1E $ 0x1)
 0x1E  ->  (S K)          H-Ap(0x3 $ 0x1)
 0x1   ->  K              H-Supercombinator(K x y = { x })
 0x3   ->  S              H-Supercombinator(S f g x = { ((f $ x) $ (g $ x)) })
 0x20  ->  3              H-Num(3)
 0x21  ->  (((S K) K) 3)  H-Ap(0x1F $ 0x20)
===///===

Look at the current state of the stack. T he left argument of the application keeps getting pushed onto the stack. This continues till there is a supercombinator on the top of the stack.

This process is called as unwinding the spine of the function call.

Instantiation

Now that a supercombinator (S) is on the top of the stack, we need to actually apply it by passing the arguments. At this stage, the "spine is unwound".

  • The top entry of the stack (S) is popped off to be evaluated.
  • Since S takes 3 parameters (f, g, and x), 3 more entries are popped off
  • The arguments to S are taken from the popped off elements.
    • The argument of the 1st application((S K) at 0x1E) becomes f
    • The argument of the 2nd application ((S K) K at 0x1F) becomes g
    • The argument of the 3rd application (((S K) K ) 3) at 0x21) becomes 3

So, summarizing the current stage:

  • S: supercombinator to unwind
  • K: first parameter, f
  • K: second parameter, g
  • 3: third parameter, x

Next, in iteration 5 we push onto an empty stack the body of the supercombinator, with variables replaced.

*** ITERATION: 5
Stack - 1 items
## top ##
 0x24  ->  ((K 3) (K 3))  H-Ap(0x22 $ 0x23)
## bottom ##
Heap - 37 items
 0x24  ->  ((K 3) (K 3))  H-Ap(0x22 $ 0x23)
 0x1   ->  K              H-Supercombinator(K x y = { x })
 0x20  ->  3              H-Num(3)
 0x23  ->  (K 3)          H-Ap(0x1 $ 0x20)
 0x22  ->  (K 3)          H-Ap(0x1 $ 0x20)
 ...
 ===///===

Notice that the parameters for S have now been instantiated on the heap. This is why it is called as an "instantiation machine" - it expands supercombinators by instantiating parameters on the heap.

  • f (K) is at 0x22
  • g (K) is at 0x23
  • x (3) is at 0x20

How does evaluation provide laziness?

First, we shall make some observations about the evaluation process:

  • During supercombinator expansion, only variables that are used are instantiated
  • parameters are not evaluated, only replaced in function bodies.

Hence, we can state that:

  • Evaluation occurs from the outside in

this is true because of the way in which application is unwound:

*** ITERATION: 7
Stack - 3 items
## top ##
 0x1   ->  K              H-Supercombinator(K x y = { x })
 0x22  ->  (K 3)          H-Ap(0x1 $ 0x20)
 0x24  ->  ((K 3) (K 3))  H-Ap(0x22 $ 0x23)
## bottom ##

Notice that the K which is the most "outside" part of the expression ((K 3) (K 3)) gets evaluated first.

When K is expanded, it expands like so:

*** ITERATION: 8
Stack - 1 items
## top ##
 0x20  ->  3  H-Num(3)
## bottom ##

The second parameter to ((K 3 (K 3)), the (K 3) is never even evaluated! the 3 is replaced as x in the body of K x y = x.

Thus, laziness is achieved by evaluating from the outside-in, and only replacing function bodies without evaluating arguments.

Primitives

+, -, etc. are similar in some ways - they also follow the same model of unwinding the spine of the execution.

>1 + 1
*** ITERATION: 1
Stack - 1 items
## top ##
 0x31  ->  ((+ 1) 1)  H-Ap(0x2F $ 0x30)
## bottom ##
===///===
*** ITERATION: 2
Stack - 2 items
## top ##
 0x2F  ->  (+ 1)      H-Ap(0xE $ 0x2E)
 0x31  ->  ((+ 1) 1)  H-Ap(0x2F $ 0x30)
## bottom ##
===///===
*** ITERATION: 3
Stack - 3 items
## top ##
 0xE   ->  +          H-Primitive(+)
 0x2F  ->  (+ 1)      H-Ap(0xE $ 0x2E)
 0x31  ->  ((+ 1) 1)  H-Ap(0x2F $ 0x30)
## bottom ##
===///===
*** ITERATION: 4
Stack - 1 items
## top ##
 0x31  ->  2  H-Num(2)
## bottom ##
===///===
=== FINAL: "2" ===

Computing something like (I 3) + 1 is not as straightforward, since I 3 now needs to be evaluated before the + can be evaluated. the section The Dump explains this process.

Indirection

When we instantiate a supercombinator, we do not cache the results of an application. Function application is optimized by rewriting the value of the application node with the result obtained. This caches the computation. This is what Indirection nodes do - they redirect a heap address to another address.

We will consider the example where we define x = I 3 where I x = x.

>define x = I 3
>x
*** ITERATION: 1
Stack - 1 items
## top ##
 0x22  ->  x  H-Supercombinator(x = { (I $ n_3) })
## bottom ##
...
===///===
*** ITERATION: 2
Stack - 1 items
## top ##
 0x25  ->  (I 3)  H-Ap(0x0 $ 0x24)
## bottom ##
...
===///===
*** ITERATION: 3
Stack - 2 items
## top ##
 0x0   ->  I      H-Supercombinator(I x = { x })
 0x25  ->  (I 3)  H-Ap(0x0 $ 0x24)
## bottom ##
...
===///===
*** ITERATION: 4
Stack - 1 items
## top ##
 0x24  ->  3  H-Num(3)
## bottom ##
...
===///===
=== FINAL: "3" ===

Now that we have run x once, let us re-run it and see what the value is

>x
*** ITERATION: 1
Stack - 1 items
## top ##
 0x22  ->  indirection(indirection(3))  H-Indirection(0x25)
## bottom ##
Heap - 39 items
 0x22  ->  indirection(indirection(3))  H-Indirection(0x25)
 0x25  ->  indirection(3)               H-Indirection(0x24)
 0x24  ->  3                            H-Num(3)
...
===///===
*** ITERATION: 2
Stack - 1 items
## top ##
 0x25  ->  indirection(3)  H-Indirection(0x24)
## bottom ##
...
===///===
*** ITERATION: 3
Stack - 1 items
## top ##
 0x24  ->  3  H-Num(3)
## bottom ##
Heap - 39 items
 0x24  ->  3  H-Num(3)
...
===///===
=== FINAL: "3" ===
>

Notice that the value of x has now become an indirection to 0x25 that used to hold (I 3).

0x25 is an indirection the value of I 3, which is 3 (at 0x24).

This way, the value of I 3 is not evaluated. It re-routes to 3.

The Dump

Now that we've seen how function application works, we would like to understand how primitives such as +, -, etc. work.

Let us consider the sample code (I 1) + 3 where I x = x (Identity).

>I 1 + 3
*** ITERATION: 1
Stack - 1 items
## top ##
 0x2C  ->  ((+ (I 1)) 3)  H-Ap(0x2A $ 0x2B)
...
===///===
*** ITERATION: 2
Stack - 2 items
## top ##
 0x2A  ->  (+ (I 1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ (I 1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
...
===///===
*** ITERATION: 3
Stack - 3 items
## top ##
 0xE   ->  +              H-Primitive(+)
 0x2A  ->  (+ (I 1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ (I 1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
...
Dump
  Empty
===///===

we now have + on the top of the stack, but the LHS is a computation that needs to be performed. Thus, we need to have some way of performing the computation.

The solution is to migrate the current stack into the Dump, and push I 1 on top of the stack and have it evaluate.

*** ITERATION: 4
Stack - 1 items
## top ##
 0x29  ->  (I 1)  H-Ap(0x0 $ 0x28)
## bottom ##
Heap - 45 items
 0x29  ->  (I 1)      H-Ap(0x0 $ 0x28)
 0x2A  ->  (+ (I 1))  H-Ap(0xE $ 0x29)
 0x0   ->  I          H-Supercombinator(I x = { x })
 0xE   ->  +          H-Primitive(+)
 0x28  ->  1          H-Num(1)
 0x2B  ->  3          H-Num(3)
Dump
## top ##
 0xE   ->  +              H-Primitive(+)
 0x2A  ->  (+ (I 1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ (I 1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
---
===///===

Notice how I 1 is now on top of the stack and the Dump contains the previous stack contents.

We proceed to see I 1 get evaluated.

*** ITERATION: 5
Stack - 2 items
## top ##
 0x0   ->  I      H-Supercombinator(I x = { x })
 0x29  ->  (I 1)  H-Ap(0x0 $ 0x28)
## bottom ##
Dump
## top ##
 0xE   ->  +              H-Primitive(+)
 0x2A  ->  (+ (I 1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ (I 1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
===///===
*** ITERATION: 6
Stack - 1 items
## top ##
 0x28  ->  1  H-Num(1)
## bottom ##
Dump
## top ##
 0xE   ->  +                       H-Primitive(+)
 0x2A  ->  (+ indirection(1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ indirection(1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
===///===

Notice how in Iteration 6, the rewrite of the I 1 at 0x2A also causes the stack to change. The stack now has

0x2A  ->  (+ indirection(1))      H-Ap(0xE $ 0x29)

while at Iteration 5 had

0x2A  ->  (+ (I 1))      H-Ap(0xE $ 0x29)

This allows the + execution to "pick up" the value of 1 later. The rewriting is essential to this evaluation. It allows the dumped stack to get the output of the execution of I 3.

The stack now has one element 1. Nothing is left to be evaluated. So, we know that the value of (I 1) is 1.

We have the rest of the computation in the Dump which we bring back.

*** ITERATION: 7
Stack - 3 items
## top ##
 0xE   ->  +                       H-Primitive(+)
 0x2A  ->  (+ indirection(1))      H-Ap(0xE $ 0x29)
 0x2C  ->  ((+ indirection(1)) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
Dump
===///===
*** ITERATION: 8
Stack - 1 items
## top ##
 0x2C  ->  ((+ 1) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
Dump
  Empty
===///===

At Iteration 7, the stack has

 0x2A  ->  (+ indirection(1))      H-Ap(0xE $ 0x29)

We remove the indirection by "short circuiting" the indirection and replacing it with the value we want.

*** ITERATION: 9
Stack - 2 items
## top ##
 0x2A  ->  (+ 1)      H-Ap(0xE $ 0x28)
 0x2C  ->  ((+ 1) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
===///===
*** ITERATION: 10
Stack - 3 items
## top ##
 0xE   ->  +          H-Primitive(+)
 0x2A  ->  (+ 1)      H-Ap(0xE $ 0x28)
 0x2C  ->  ((+ 1) 3)  H-Ap(0x2A $ 0x2B)
## bottom ##
===///===
*** ITERATION: 11
Stack - 1 items
## top ##
 0x2C  ->  4  H-Num(4)
## bottom ##
Dump
  Empty
===///===
=== FINAL: "4" ===

Now that we have a simple expression, evaluation proceeds as usual, ending with the machine evaluating 1 + 3 on seeing + at the top of the stack.

Roadmap

  • Mark 1 (template instantiation)
  • let, letrec
  • template updates (do not naively instantiate each time)
  • numeric functions
  • Booleans
  • Tuples
  • Lists
  • nicer interface for stepping through execution

Design Decisions

TIMi is written in Rust because:

  • Rust is a systems language, so it allows for more control over memory, references, etc. which I enjoy.
  • Rust has nice libraries for readline, table printing, and a slick stdlib for pretty code

Things Learnt

Difference between lazy recursive bindings and strict recursive bindings

Lazy recursive bindings will let you get away with

let y = x; x = y in 10

while strict recursive bindings will try to instantiate x and y.

Difference between [..] and &[..]

Slice without ref versus Slice with ref

the difference is that the second slice [..] maintains length information which it needs at compile-time.

References

Dependencies

~6.5MB
~102K SLoC