The MLKit Standard ML Compiler Tool Kit Martin Elsman Department of Computer Science University of Copenhagen DIKU

March 15, 2018

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

1 / 35

Overview of the MLKit

Outline

1 Overview of the MLKit

History Compiler Overview Modules and Recompilation 2 SMLserver

Overview RDBMS Interfacing The Region Model Measurements 3 SMLtoJs

Features JavaScript Integration Tail Calls Composing Js Fragments Compiling in a Browser Other Uses Related Work and Conclusion Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

2 / 35

Overview of the MLKit

History

History of the MLKit 1 MLKit version 1

Test-bed for language development; 1990 Definition (static and dynamic semantics) [Tofte et al. 1991] 2 MLKit with Regions [1993–...]

Region-inference and region-based runtime system for the Core language [Tofte et al. 1993–1999] HPPA and X86 backends [1996] Static interpretation of Modules and smart recompilation [1999] Combination with GC [1999–2003] Support for the entire Standard ML Basis Library [1999–2003] 3 SMLserver [2003–...]

Region-based bytecode interpreter combined with Apache web server with efficient DB integration [2003] Used at ITU (course mngt., course eval., diplomas, ...) [2003–...] 4 SMLtoJs [2005–...]

MLKit bootstrapped in a browser; good support for tail-calls [2010] Used with SMLserver for single-language multi-tier applications Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

3 / 35

Overview of the MLKit

History

Many Contributors 1 MLKit version 1

Lars Birkedal, Nick Rothwell, Mads Tofte, David N. Turner 2 MLKit with Regions [1993–...]

Lars Birkedal, Peter Bertelsen, Martin Elsman, Niels Hallenberg, Tommy Højfeld Olesen, Peter Sestoft, Mads Tofte, Magnus Vejlstrup 3 SMLserver [2003–...]

Martin Elsman, Niels Hallenberg, Ken Friis Larsen, Carsten Varming 4 SMLtoJs [2005–...]

Martin Elsman

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

4 / 35

Overview of the MLKit

Compiler Overview

MLKit Overview file.{sml,mlb} Frontend lexing, parsing, type-checking static interpretation of modules pattern-match compilation

Typed Lambda

Region Lambda :

Javascript

3-address code

Bytecode backend

X86 backend

file.bc

file.o

Martin Elsman (DIKU)

Optimisations: - function inlining - recursive function specialisations - constant propagation - dead code elim - let-floating -…

file.js

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

5 / 35

Overview of the MLKit

Modules and Recompilation

Static Interpretation of Modules and Smart Recompilation Programmers organise source code in so-called MLB-files (same as for MLton) Modules (signatures, structures, and functors) are eliminated entirely at compile time... A smart recompilation system allows for cross-module optimisations while avoiding unnecessary recompilation when source code changes.

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

6 / 35

SMLserver

Overview

SMLserver (http://www.smlserver.org/) A web server platform for Standard ML programs

Features Access to a variety of RDBMSs through an efficient generic interface that supports reuse of database connections Support for type safe data caching, HTTP requests, filtering and script scheduling Programs are compiled into bytecode files, which are loaded only once but may be executed many times A multi-threaded execution model allows multiple requests to be served simultaneously Integrated with the Apache web server

Why MLKit? No reference tracing GC and no tags, but region based memory management

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

7 / 35

SMLserver

RDBMS Interfacing

Interfacing to an RDBMS Support for Oracle, Postgresql, and MySQL Database pooling A handle identifies a connection to an RDBMS SMLserver maintains a configurable number of pools (of handles) A database handle is owned by at most one script at a time Handles are requested and released by the database functions in such a way that no deadlocks appears The programmer needs not know about handles, unless transactions are used

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

8 / 35

SMLserver

RDBMS Interfacing

The RDBMS Interface signature WEB_DB = sig val dml : quot -> unit val fold : ((string->string) * ’a -> ’a) -> ’a -> quot -> ’a val qqq : string -> string ... end

Notes Quotations are used for embedding SQL The function qqq escapes quotes (’) for SQL string embedding The function dml can be used for executing insert and update statements The function fold folds over the rows returned by a SQL statement (similar to List.foldr)

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

9 / 35

SMLserver

The Region Model

The Region based memory model

Memory allocation and deallocation directives are inserted in the program at compile time Memory is allocated from a free list of region pages A region is a list of region pages — a free-list contains region pages currently unused Activation records stored on the runtime stack contain: temporary variables and region descriptors

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

10 / 35

SMLserver

The Region Model

The Region based memory model 6

6

6

Region Heap

6

r1

6

6

r2

6

6

r3

6

-

6 free list

Machine stack with activation records

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

11 / 35

SMLserver

The Region Model

The efficient thread-safe SMLserver memory model Code Caching

Cached library code

Cached script code

Thread 1 Data Caching

Library cache Library regions

Cached script code Thread N Library cache

Script regions

When server boots

Library regions

Shared region page free list

Script regions

When executing script

load library code

load script code

execute library code

execute script code

copy “Library regions” into “Library cache”

deallocate script regions

Martin Elsman (DIKU)

restore “Library regions”

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

12 / 35

SMLserver

The Region Model

Consequences of the Execution Model Execution starts in an initial heap each time a request is served. Thus, it is not possible to maintain state implicitly in web applications using Standard ML references or arrays. Instead, state must be maintained explicitly using an RDBMS, perhaps combined with SMLserver cache primitives. Alternative: emulate state with form variables or cookies. Region-inference works very well in environments where programs run shortly but are executed often!

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

13 / 35

SMLserver

Measurements

Measurements with ApacheBench Requests / second Program hello date db guest calendar mul table log

MosML MSP

AOLserver TCL

Apache PHP

SMLserver MSP

55 54 27 25 36 50 21 8

724 855 558 382 27 185 59 12

489 495 331 274 37 214 0.7 0.4

1326 1113 689 543 101 455 93 31

ApacheBench (v. 1.3d) uses eight threads (60 seconds each) Old experiments: 850Mhz Pentium 3 Linux box (384Mb RAM) Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

14 / 35

SMLtoJs

SMLtoJs: Higher-Order and Typed (HOT) Web browsing http://www.smlserver.org/smltojs/ Easy development and maintainance of advanced Web browser libraries (e.g., Reactive Web Programming) Allow developers to build client-based web applications in a HOT language

Allow for existing code to execute in browsers Programs (e.g., SMLtoJs itself — it is itself written in SML) Libraries (e.g., The IntInf Basis Library module) Support all of SML and (almost) all of the SML Basis Library

Web programming without tiers Allow the same code to run both in the browser and on the server (e.g., complex serialisation code) Type-safe multi-tier applications

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

15 / 35

SMLtoJs

Features

Features of SMLtoJs Supports all Browsers SMLtoJs compiles Standard ML programs to JavaScript for execution in all main Internet browsers.

Compiles all of Standard ML SMLtoJs compiles all of SML, including higher-order functions, pattern matching, generative exceptions, and modules.

Basis Library Support Supports most of the Standard ML Basis Library, including: Array2 ArraySlice Array Bool Byte Char CharArray CharArraySlice CharVector CharVectorSlice Date General Int Int31 Int32 IntInf LargeWord ListPair List Math Option OS.Path Pack32Big Pack32Little PackReal Random Real StringCvt String Substring Text Time Timer Vector VectorSlice Word Word31 Word32 Word8 Word8Array Word8ArraySlice Word8Vector Word8VectorSlice

Additional Libraries: Martin Elsman (DIKU)

JsCore Js Html Rwp The MLKit Standard ML Compiler Tool Kit

March 15, 2018

16 / 35

SMLtoJs

Features

JavaScript Integration and DOM Access ML code may call JavaScript functions and execute JavaScript statements. SMLtoJs has support for simple DOM access and for installing ML functions as DOM event handlers and timer call back functions. Applies MLKit’s optimisations Static interpretation of Modules. Function inlining and constant propagation Specialisation of higher-order recursive functions (map, foldl) JavaScript-specific optimisations Tail-call optimisation of so-called straight tail calls Unboxing of certain datatypes (lists, certain trees, etc.) Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

17 / 35

SMLtoJs

Features

Example: Compiling the Fibonacci Function (fib.sml) fun fib n = if n < 2 then 1 else fib(n-1) + fib(n-2) val _ = print(Int.toString(fib 23))

Resulting JavaScript Code var fib$45 = function fib$45(n$48){ if (n$48 ’a1 * ’a2 -> ’b ... end Phantom types are used to ensure proper interfacing: fun documentWrite d s = J.exec2 {stmt="return d.write(s);", arg1=("d",J.fptr), arg2=("s",J.string), res=J.unit} (d,s) SMLtoJs inlines stmt if it is known statically; otherwise a Function object is created and stmt resolved and executed at runtime. Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

19 / 35

SMLtoJs

JavaScript Integration

Library (Js) for Manipulating the DOM and Element Events signature JS = sig eqtype win and doc and elem (* dom *) val openWindow : string -> string -> win val document : doc val windowDocument : win -> doc val documentElement : doc -> elem val getElementById : doc -> string -> elem option val value : elem -> string val innerHTML : elem -> string -> unit datatype eventType = onclick | onchange (* events *) val installEventHandler : elem -> eventType -> (unit->bool) -> unit type intervalId val setInterval : int -> (unit->unit) -> intervalId val clearInterval : intervalId -> unit val onMouseMove : (int*int -> unit) -> unit ... end Notice: implemented using the JsCore module. Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

20 / 35

SMLtoJs

JavaScript Integration

The Inner Workings of SMLtoJs (no tail calls) SMLtoJs compiles SML to JavaScript through an MLKit IL. SML reals, integers, words, and chars are implemented as JavaScript numbers with explicit checks for overflow. SML variables are compiled into JavaScript variables. SML functions are compiled into JavaScript functions: [[fn x → e]]exp = function(x){[[e]]stmt } SML variable bindings compiles to JS function applications: [[let val x = e in e0 end]]exp = function(x){[[e0 ]]exp }([[e]]exp ) When compilation naturally results in a JavaScript statement, the statement is converted into an expression: [[e]]stmt = stmt

(1)

[[e]]exp = function(){stmt; }()

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

21 / 35

SMLtoJs

Tail Calls

Optimizing Straight Tail Calls A straight tail call is a tail call to the nearest enclosing function with a tail call context containing no function abstractions. Example SML code: fun sum (n,acc) = if n ’’c) -> val >>> : (’’b,’’c,’k)arr val fst : (’’b,’’c,’k)arr (* derived combinators *) val snd : (’’b,’’c,’k)arr val *** : (’’b,’’c,’k)arr arr val &&& : (’’b,’’c,’k)arr end

(’’b,’’c,’k) arr * (’’c,’’d,’k)arr -> (’’b,’’d,’k)arr -> (’’b*’’d,’’c*’’d,’k)arr -> (’’d*’’b,’’d*’’c,’k)arr * (’’d,’’e,’k)arr -> (’’b*’’d,’’c*’’e,’k) * (’’b,’’d,’k)arr -> (’’b,’’c*’’d,’k)arr

Notice The ARROW signature specifies combinators for creating basic arrows and for composing arrows. Specifically, we model behavior transformers and event stream transformers as arrows. The ’k’s are instantiated either to B (behavior) or to E (events). Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

29 / 35

SMLtoJs

Other Uses

The Rwp library: Building Basic Behaviors and Event Streams signature RWP = sig type B type E (* kinds: Behaviors (B) and Events (E) *) type (’a,’k)t type ’a b = (’a, B)t type ’a e = (’a, E)t include ARROW where type (’a,’b,’k)arr = (’a,’k)t -> (’b,’k)t val timer : int -> Time.time b val textField : string -> string b val mouseOver : string -> bool b val mouse : unit -> (int*int) b val pair : ’’a b * ’’b b -> (’’a * ’’b) b val merge : ’’a e * ’’a e -> ’’a e val delay : int -> (’’a,’’a,B)arr val calm : int -> (’’a,’’a,B)arr val fold : (’’a * ’’b -> ’’b) -> ’’b -> ’’a e -> ’’b e val click : string -> ’’a -> ’’a e val changes : ’’a b -> ’’a e val hold : ’’a -> ’’a e -> ’’a b val const : ’’a -> ’’a b val flatten : ’’a b b -> ’’a b val insertDOM : string -> string b -> unit end Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

30 / 35

SMLtoJs

Other Uses

Example: Adding the Content of Fields open Rwp infix *** &&& >>> val _ = print("

Add Content of Fields

" ˆ " + " ˆ " = ?") val si_t : (string,int,B)arr = arr (Option.valOf o Int.fromString) val form = pair( textField "a", textField "b" ) val t = (si_t *** si_t) >>> (arr op +) >>> (arr Int.toString) val _ = insertDOM "c" (t form)

Notice t takes a behavior of pairs of integers and returns an integer behavior.

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

31 / 35

SMLtoJs

Other Uses

Example: Reporting the Mouse Position val _ = print ("

Mouse Position

" ˆ "?
" ˆ "?
" ˆ "?
") val t : (int*int,string,B) arr = arr (fn (x,y) => ("[" ˆ Int.toString x ˆ "," ˆ Int.toString y ˆ "]")) val bm = mouse() val t10 : (int*int,int*int,B) arr = arr (fn (x,y) => (x div 10 * 10, y div 10 * 10)) val bm2 = (t10 >>> t) bm val _ = insertDOM "mouse0" (t bm) val _ = insertDOM "mouse1" (calm 400 bm2) val _ = insertDOM "mouse2" (delay 400 bm2)

Notice calm waits for the underlying behavior to be stable. delay transforms the underlying behavior in time. Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

32 / 35

SMLtoJs

Other Uses

Implementation Issues Behaviors and event streams are implemented using “listeners”: type (’a,’k) t = {nid : int, listeners : (LId.t * (’a->unit)) list ref, current : ’a ref option} Behaviors (of type (’a,B)t) always have a current value, whereas event streams do not. Installing a behavior b in the DOM tree involves adding a listener to b that updates the element using Js.innerHTML. The implementations of calm and delay make use of Js.setTimeout. The implementation of textField makes use of Js.installEventHandler. The implementation of mouse makes use of Js.onMouseMove. A heap is used to guarantee that when some node change value, other nodes are at most updated once.

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

33 / 35

SMLtoJs

Related Work and Conclusion

Related Work The Google Web Toolkit project (GWT). The Scm2Js and Hop projects by Loitsch and Serrano, TFL’2007. The Links project. Wadler et al. 2006. The AFAX F# project by Syme and Petricek, 2007. The ML5 project by Murphy, Crary, and Harper, 2007. O’Browser by Canou, Balat, and Chailloux, ML’2008. The js_of_ocaml project by Vouillon, 2011.

Related Reactive Programming Work The Flapjax language and JavaScript library by Shriram Krishnamurthi et al. John Hughes. Generalising Monads to Arrows. Science of Computer Programming 37. Elsevier 2000. The Fruit Haskell library by Courtney and Elliott.

Martin Elsman (DIKU)

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

34 / 35

SMLtoJs

Martin Elsman (DIKU)

Related Work and Conclusion

The MLKit Standard ML Compiler Tool Kit

March 15, 2018

35 / 35