Difference between revisions of "Haskell in web browser"
(All sections here)
(Author's contact information)
|Line 20:||Line 20:|
<small>For any questions, suggestions, concerns please contact Dmitry Golubovsky (golubovsky at gmail dot com).</small>
Revision as of 19:57, 26 March 2008
- 1 Preface
- 2 Basics of programming for web browser
- 3 Haskell web toolkit
- 3.1 Widgets
- 3.2 Activators
- 3.3 Conclusion: all together
Basics of programming for web browser
It was necessary to convert these definitions to Haskell function declarations to make them available to Haskell programs for Web browser. The special utility,
DOM interfaces vs. Haskell type classes
Web Consortium's DOM definitions are presented as a hierarchy of interfaces. For example, the Node interface is a parent to the majority of other interfaces, such as Document (direct ancestor), or HTMLElement (not a direct ancestor, but HTMLElement should inherit all properties and methods of Node).
This is achieved by defining Haskell type classes whose hierarchy repeats the hierarchy of DOM interfaces. Thus, we have the CNode and CDocument classes. For each DOM interface, also a phantom data type is defined: TNode, and TDocument correspondingly. Phantom types are assigned to concrete values (references to DOM objects) while type classes are used to constrain types of parameters of functions working with those DOM objects. The CDocument class is defined as:
class CNode a => CDocument a data TNode data TDocument instance CNode TNode instance CDocument TDocument instance CNode TDocument
to reflect inheritance of Document from Node. Accordingly, continuing our example, for HTMLElement, we have:
class CNode a => CElement a class CElement a => CHTMLElement a data THTMLElement instance CElement THTMLElement instance CHTMLElement THTMLElement instance CNode THTMLElement
Below is an example of such type constrained function:
hasChildNodes :: CNode this => this -> CPS c Bool
which corresponds to the hasChildNodes function defined within the Node interface. Any DOM object which is a Node can be passed to this function (by reference) as the
Attributes vs. getters and setters
Within interfaces, DOM specification defines attributes and methods. Attributes are either read-only (such as nodeName of the Node interface) or read-write (such as nodeValue of the same interface). In Haskell bindings, getter (for read-only attributes), and both getter and setter (for read-write attributes) functions are defined in straightforward manner:
get'nodeName :: CNode this => this -> CPS c String set'nodeValue :: CNode zz => String -> zz -> CPS c zz get'nodeValue :: CNode this => this -> CPS c String
Getters always take the object containing an attribute as the first argument,
this, and it is always constrained to the type class corresponding to the DOM interface. Setters always take the value to be set as the first argument, and the object containing the attribute as the second argument. Setters always return reference to the same object where an attribute was set. The latter property allows to concatenate multiple setters in Continuation-passing style, such as:
........$ \he -> (set'id "myid") (set'lang "en") (set'title "Hello")
This whole construction will pass the same object (
he) to the continuation, but continuation will deal with updated object.
The setters in the example above are defined in the DOM.Level2.HTMLElement module.
Methods vs. functions
Interface methods are translated to Haskell functions whose type signatures have proper type coetraints. Thus, the getElementById function defined in the Document interface as
Element getElementById(in DOMString elementId);
translates to Haskell function:
getElementById :: (CDocument this, CElement zz) => this -> String -> CPS c zz
as follows from its type, getElementById does not return a value of concrete type, but rather a type-constrained value. Values of types corresponding to DOM interfaces, translate to type-constrained rather than concrete values. This sometimes makes it necessary to supply explicit type signatures unless a function receiving the returned constrained value has a type signature that brings a constrained type down to a concrete type.
The IDL conversion utility
domconv auto-creates convenient functions that serve as constructors of DOM objects corresponding to HTML tags. An example of such maker function is:
mkDiv :: CHTMLDocument a => a -> CPS c THTMLDivElement
which creates a DOM node tagged with <DIV>. Such maker functions are defined for most of HTML elements. Maker functions always return values of concrete type.
CPS and threads
Continuation Passing Style allows for very efficient implementation of threads: in fact, switching context between threads is merely saving a continuation of currently executing thread in some static memory object, and evaluating continuation of (resuming) a thread that was similarly suspended earlier.
Message passing between threads
Threads may pass messages to each other, using Message Boxes. Sending messages is asynchronous (although rescheduling of threads execution occurs to resume the receiving thread). Sending a message may fail if there is a message in the Message Box (no message buffering). Receiving messages is always a blocking operation. Receiving a message may fail if there is already a thread waiting on the Message Box. Therefore more than one thread may send messages to the same Message Box (but sending all messages is not guaranteed unless result of sending is checked, and appropriate action taken if
sendMsg fails), but more than one thread may not receive messages from the same Message Box.
The JSON API provided to Haskell programs is based on the opaque type JsonNode. Operations over JSON nodes are monadic as they may fail. For example, the getValueByName function fails if the JSON node queried does not have a value with given name. Monadic interface allows to write code that retrieves values from JSON node in
do notation and compose operations using
>>= (monadic bind).
Thus, the following code:
uri = fromMaybe "----" $ do buri <- splitURI loc prot <- getValueByName "protocol" buri >>= getString auth <- getValueByName "authority" buri >>= getString anchor <- getValueByName "anchor" buri >>= getString let uri = prot ++ "://" ++ auth ++ "/" ++ anchor return uri
retrieves parts of an URI (see below) represented as a JSON node, and composes a new URI out of them. If the URI JSON node does not contain any of the values requested, the whole monadic sequence fails, and the fall-back value "----" will be returned.
Refer to the Data.JsonNode module for more information.
Haskell web toolkit
HsWTK is no longer actively developed or maintained.
Haskell Web Toolkit (further referred to as HsWTK) is a thin layer built on top of DOM interfaces. It provides program interfaces to compose static layout of a web application page, and to hook up visual elements of an application to event handlers and XML HTTP communication means. HsWTK hides the low-level DOM APIs where possible; however their knowledge may be necessary to develop certain types of visual components and event handlers.
To build a web-based GUI, HsWTK defines the following type:
type Widget = THTMLDocument -> THTMLElement -> Bool
That is, Widget is a function, or, to be a more precise, an action (as evaluation of this function does assume side effects). This action's first argument, of type THTMLDocument, refers to the HTML document containing the GUI elements. The action's second argument, of type THTMLElement, refers to a parent HTML element. This makes perfect sense from the DOM standpoint, as in order to create a visible element on a Web page, it is at least necessary to create an element by calling
Document.createElement method (which needs a Document), and next to insert the newly created element into some (parent) node by calling
HTML elements as widgets
So, inside the Widget action, some work may be done to create a HTML element, and make it a child of some other Widget. This is generalized by HsWTK as Element Creation Function, and defined as
type ECRF n = (THTMLDocument -> CPS Bool n)
Now, if we refer to Maker functions defined for HTML-tagged Elements we may see some similarity:
mkDiv :: CHTMLDocument a => a -> CPS c THTMLDivElement
If we substitute n in ECRF definition with THTMLDivElement, and remember that THTMLDocument is an instance of CHTMLDocument, we get a perfect match. From this, it may be concluded that Maker functions may serve as Element creation functions.
The simpliest form of Widget is passive widget. Passive widgets only display themselves as part of Web GUI, but are not capable of nesting other widgets. A good example of such passive widget is text label or non-clickable image.
HsWTK defines a function
passive which given an Element Creation Function, returns a Widget:
passive :: CNode n => ECRF n -> Widget
Thus, given a Maker function (e. g. mkImg), applying passive to it creates an image which will appear in the proper place of a Web page.
For a text element itself, there is a Maker function
mkText which produces a text node to be inserted into a < DIV > or < SPAN > or any other element with closing tag. mkText is not a pure Maker function though; due to its type signature,
flip has to be applied to it.
One important passive widget not based on a HTML element is
nowidget. It may be used as a placeholder for any passive widget, and does not produce any kind of effects.
HTML elements with closing tags contain other elements in between. To reflect this, HsWTK defines another function,
container. This function, applied to a Maker function, produces a widget capable of nesting other widgets.
For the purpose of sequencing widgets within a container, or nesting widgets in containers, HsWTK defines combinators
++|. Please refer to the appropriate section of the Graphics.UI.HsWTK module documentation.
+++ combinator sequences two widgets inside a container. Widgets (
a +++ b) appear second (b) at the right of the first (a). Result of such composition is also a widget, and it may be sequenced with other widgets.
(a +++ b) +++ c is same as
a +++ b +++ c.
<< combinator nests a widget (or composition of widgets) in a container.
So, the following fragment of code creates a widget consisting of a < DIV > with a text inside:
mkTextFl = flip mkText container mkDiv << passive (mkTextFl "Hello World")
To simplity the code, HsWTK defines these two functions:
textPwhich is equivalent to
passive (flip mkText txt)
|<<which is defined as
c |<< d = container c << d, that is the word
containermay be omitted.
The toplevel widget (one that is not nested in any other widget) has to be inserted into the HTML document body. This is acieved by applying the function
docBodyC to the toplevel widget. This application often becomes the
main function of a Web application. Please note that document body is not a Widget.
Code example - Hello Web
At this point, we are able to code our first HsWTK example. By tradition, this is a Hello ... program. Paste the code below into Yhc Web Service New Entry Form and press the Submit button (don't forget to fill out the Author and Title field).
After the compilation is finished, load the generated Web page. Haskell says Hello to the new environment it just has started to explore.
The example shows all the facilities of HsWTK we recently discussed:
- widgets sequencing (
+++combines two text elements one after another)
- creation of a HTML Element based widget (mkSpan)
- nesting composition of widgets in a container (
- inserting the toplevel widget into the document body (docBodyC)
Decorators are not Widgets: they are tools to change properties of a widget they are applied to, at the moment of that widget creation.
Earlier we discussed setters defined for HTML Elements' writable attributes. HsWTK defines a function
decorate which when applied to a setter, produces a decorator. An example of such a decorator is
withTitle defined as shown below:
withTitle :: (CHTMLElement b) => ((a -> ((b -> (CPS c b)) -> d)) -> (String -> (a -> d))) withTitle = decorate set'title
set'title function is a setter for the title attribute of HTMLElement.
Decorators are binary functions taking the Widget to decorate as the first argument, and the value used to decorate as the second. In our case, this looks like this:
... mkDiv `withTitle` "This div has a title" |<< textP "titled div"
A special decorator,
withStyle, represents a more complicated case. It does not alter attributes/properties of a widget element it is used with; instead it affects inline style properties of the widget element.
While arttribute-based decorators take simple values (strings, less often numbers) to decorate widget elements, withStyle takes a list of name-value pairs. Both names and values are strings, and should conform to the Cascading Style Sheets Specification.
Property names and values are paired using the
:= data constructor. The latter is just a binary infix data constructor, and it is used just to improve the code appearance, so CSS property assignments look more natural.
Our next pasteable examples deal with widget elements' inline style manipulations. The following example displays a centered white bold text on green background spanning over 45% of browser window width floating to the right side of browser window.
Our next example is more contrived. It displays a line of text with each character colored differently, and colors are taken from an infinitely repeating sequence. Note the use of nowidget: it is used with
foldr as a "starting value".
At this point, we are able to create static layout of Widgets. No way has been shown yet how to make GUI elements interact.
Activators serve exactly this purpose. Activators are functions, usually endlessly looped which are executed each in its own thread. These functions repeatedly receive messages from Message Boxes, and also intercept events induced by user. Activators also are able to modify on-the-fly properties of Widgets to which they are "attached". Some Activators also perform data exchange over XMLHTTP.
Method of inter-widget communication described here is derived from one discussed in the following paper:
HsWTK defines a type for these functions:
type ACTF = THTMLElement -> Bool
HsWTK also defines a function
active which being applied to an Activator turns it into a Widget, that is, an Activator may be nested within a container: that's what was meant above for an Activator "attached" to a Widget.
Our next example consists of two <INPUT> elements: when something is typed in the left one, it is repeated in the right one:
Let's go line by line. As new things are introduced, they will be discussed in appropriate sections.
Message boxes and activators
Message boxes were discussed earlier as means for inter-thread communications. Each Activator is represented by a separate thread. Unlike previous code examples, the main Widget code contains two calls to the
msgBox function. The
msgBox function passes reference to newly created Message box to its continuation. Thus, after the following code:
msgBox $ \ibx2 -> msgBox $ \dummy ->
is executed, two Message boxes are created.
Wiring the static layout
The following code of the main Widget would have produced the same static layout of two input elements:
mainW = inputI |<< nowidget +++ inputI |<< nowidget
Note the use of
nowidget here: the
inputI function produces an <INPUT> element, and the
|<< combinator assumes that the <INPUT> element is a container, that is, should have something nested.
This approach may be used during Web GUI design: make Widgets containers where necessary, but nest
nowidget's instead of Activators.
Compare the code of
mainW above with the code from our example:
mainW = ... inputI |<< active (evtBCastA "keyup" rtgt [ibx2]) +++ inputI |<< active (fwdValueA ibx2 dummy)
Each <INPUT> element got an Activator. As it may be guessable, Activator attached to the first declared (showing at the left of Web page) <INPUT> element, transmits its value (whatever is typed in), and Activator attached to the <INPUT> element declared second (showing right of the first <INPUT>) receives the value.
Mapping and broadcasting events
HsWTK defines the following Activator function:
evtBCastA :: CEvent e => String -> (e -> CPS Bool v) -> [MSGBOX Bool v] -> ACTF
The first argument of
String type is the type of event this Activator responds to. Only one event type may be specified. Event types are derived from on- attribute names, e. g. "keyup" corresponds to onkeyup event, etc.
The second argument is a CPS-style function mapping an event to some value of type v. Such a function may retrieve certain parameters of an event (such as code of a key pressed), or event produce some side effects, although this is not recommended.
The third argument is list of Message boxes where the mapped value will be broadcast over. Broadcast will be performed using the
sendMsg_ function, so there is possibility of message non-delivery.
HsWTK defines two functions that can be used as the second argument of
evt2ConstU: maps any event to the given constant value
readTargetU: maps any event to the value (always of type
String) retrieved from its target element. The target element must support the value property.
The latter function is used in our example. Basically the Activator attached to the left <INPUT> element intercepts each onkeyup event, retrieves event target value (that is, the <INPUT> element it is attached to), maps it to the value forwarding message (see below), and sends to the Message box that other <INPUT> element may be listening to by means of its Activator.
Value forwarding protocol
Value forwarding protocol is a convention set by HsWTK regarding the way how Widgets values may be set externally, or queried. A conforming Widget should have an Activator parameterized with two Message boxes. The first Message Box is for receiving messages of type
FwdValueMsg. The second Message Box is for sending messages of type
The Activator should act upon the receipt of FwdValueMsg as follows:
- FwdCurr: forward the current value, not changing it
- FwdCurrSet s: forward the current value to the default Message Box, update with new value s
- FwdUpdSet s: update element's value with new value s, forward the new value to the default Message Box
- FwdCurrTo m: like FwdCurr but to a specified Message Box m
- FwdCurrSetTo m s: like FwdCurrSet but to a specified Message Box m
- FwdUpdSetTo m s: like FwdUpdSet but to a specified Message Box m
Messages FwdCurrTo, FwdCurSetTo, FwdUpdSetTo cause the value of the Widget to be forwarded to a Message Box different from one the Activator is parameterized with.
In our code example, the second <INPUT> element has an Activator
... inputI |<< active (fwdValueA ibx2 dummy)
This is a "standard" Activator provided by HsWTK for Widgets based on HTML elements with attribute value defined. The first Message Box is the same where the first <INPUT> element's Activator broadcasts the value upon each key release.
The second Message Box is not used in any way, but it is there since
fwdValueA requires it.
Now, looking at the event mapping function of the first <INPUT> element:
rtgt e = readTargetU e FwdUpdSet
which means that the constructor
FwdUpdSet is applied to the value retrieved from the event target element (that is, the first <INPUT>), yielding the message per Value forwarding protocol. This value will be received by the second <INPUT> element's Activator, and will subsequently show in the element.
More complex widgets
Our next example shows a Widget which has some functionality encapsulated, while externally this is just an <INPUT> element. The example features an input field which validates its value (as it is typed in) against a regular expression, and changes its border color to green when validation passes, and to red otherwise.
Regular expression in this example matches numbers separated by dots (like 1.2.3). Readers are encouraged to try other regular expressions as well.
Understanding of how this code works is left as reader's excercise. Hint:
evtBCastA is called with empty list of Message Boxes. All event processing work is done on the single HTML element (event target).
Below is an example of code similar to the example with two <INPUT> elements except that user's input is displayed in a non-editable (display-only) Widget.
In this example, both <INPUT> and <DIV> elements are enclosed within left-floating <DIV>'s for proper placement. The character with code 160 (that is,
) is inserted within a framed <DIV> to make sure the frame has full height when nothing is displayed. Finally, a <SPAN> element is nested within the framed <DIV>. The special Activator,
updMapA is attached to the latter. This activator is parameterized with one Message Box to receive messages of any type. Another parameter of
updMapA is a function that maps input messages to
Strings. In this example, the message sent by the <INPUT> element to be displayed, is expected to be wrapped into the
FwdUpdSet constructor. Please note that this is not a full emulation of Value forwarding protocol as nothing would be sent in response to messages like
FwdCurrTo. Please note also that when building a display-only element, avoid nesting anything other than the appopriate Activator within the element that performs actual display. The
updMapA Activator is based on the
updateU function which replaces the first child of given node with given
String value. Other children remain unchanged. With this logic, some interesting effects may be achieved, but this feature should be used with caution.
Conclusion: all together
In conclusion, let's finally develop a program that does something useful. As another "classical" example, this will be a RPN calculator.
The main Widget,
mainW consists of the following pieces:
- Two Message Boxes. One (
calcmb) will be used to transmit keystrokes and mouse clicks received from the browser's input facilities. Another (
dispmb) will be used to update the calculator's display element.
- A <DIV> element with visible border: it encloses all other elements of the calculator. It is also responsible for receiving keyboard input.
The following elements are nested within the above mentioned <DIV>:
- A <DIV> element that contains the calculator's display.
- 16 buttons for mouse input. Note the use of
map. All buttons use the same Message Box
calcmb, so it is passed to each button's creation function. Each button uses its own character taken from the
btxtlist: button creation function is mapped over this list. Finally, the resulting list is folded with
+++to sequence all buttons properly.
- Activator which implements the internal (stateful) logic of the calculator. It is based on
passMapStateAwhich receives messages from its input Message Box, maintains internal state transitions based on what is received, and maps internal state to messages sent to the output Message Box. That's how the two parts (user input and result display) are wired together.
- Activator based on
evtBCastAto catch keystrokes and route them to the same Message Box where mouse clicks go.
- Activator based on
tabIndexA. Although DOM does not define the tabIndex property for most HTML elements, according to this (last comment), it is necessary to initialize this property in order to be able to receive keystrokes on <DIV> elements.
clkey function performs proper mapping of keystrokes to simulate mouse clicks on buttons. It extracts the value of keyCode property from the event received, and remaps certain codes, finally passing a character to the continuation.
CalcST defines the internal state of the calculator. It includes the accumulator (where the user input goes), the stack (a list of numbers), and the reset flag which helps update the display properly when entry of a new number begins.
Internal logic of the calculator (
calc) is pretty straighforward: numbers typed in are pushed down the stack when "^" button is clicked, or the "up arrow" key is pressed. Operation buttons/keys perform binary operations between the accumulator and the topmost stack element. Finally copy of the result is pushed back onto the stack, so next number typed in may be used for the next computation.
btxt list of characters defines what shows on buttons.
The button creation function (
cbtn) defines a <DIV> element nesting a <BUTTON> element. Each button has an Activator based on
evtBCastA that maps each click to the character showing on the button. Note the float: left style and percentage width defined for <DIV>s: no matrix layout logic is necessary; buttons fill the <DIV> with borders with correct placement.
For any questions, suggestions, concerns please contact Dmitry Golubovsky (golubovsky at gmail dot com).