Theory of Operation

This chapter discusses how Leo’s code works, paying particular attention to topics that have caused difficulties in design or implementation. This chapter will be of use primarily to those wanting to change Leo’s code.

Contents

• Theory of Operation
  • Autocompletion
  • Clones
  • Drawing and events
  • Find and change commands
  • Key handling
  • Nodes
  • Overview
  • Unicode
  • Unlimited undo
  • Leo’s load process

Autocompletion

UI notes

Both the legacy and new completer now work exactly the same way, because they both use the AutoCompleterClass to compute the list of completions.

The strict “stateless” requirement means that the “intermediate” completions must be entered into the body pane while completion is active. It works well as a visual cue when using the tabbed completer: indeed, the tabbed completer would be difficult to use without this cue.

The situation is slightly different with the qcompleter. Adding code before the user accepts the completion might be considered an “advanced” feature. However, it does have two important advantages, especially when “chaining” across periods: it indicates the status of the chaining and it limits what must appear in the qcompleter window.

Appearance

There is little change to the legacy completer, except that no text is highlighted in the body pane during completion. This is calmer than before. Furthermore, there is no longer any need for highlighting, because when the user types a backspace the legacy completer now simply deletes a single character instead of the highlighted text.

One minor change: the legacy completer now does insert characters that do not match the start of any possible completion. This is an experimental feature, but it might play well with using codewise completions as a fallback to Leo-related completions.

Performance

Performance of Leo-related completions is much better than before. The old code used Python’s inspect module and was horribly complex. The new code uses eval and is perfectly straightforward.

The present codewise-related code caches completions for all previously-seen prefixes. This dramatically speeds up backspacing. Global caching is possible because completions depend only on the present prefix, not on the presently selected node. If ContextSniffer were used, completions would depend on the selected node and caching would likely be impractical. Despite these improvements, the performance of codewise-oriented completions is noticeably slower than Leo-related completions.

The ac.get_cached_options cuts back the prefix until it finds a cached prefix. ac.compute_completion_list then uses this (perhaps-way-too-long-list) as a starting point, and computes the final completion list by calling g.itemsMatchingPrefixInList.

This may not be absolutely the fastest way, but it is much simpler and more robust than attempting to do “prefix AI” based on comparing old and new prefixes. Furthermore, this scheme is completely independent of the how completions are actually computed. The autocompleter now caches options lists, regardless of whether using eval or codewise.

In most cases the scheme is extremely fast: calls to get_completions replace calls to g.itemsMatchingPrefixInList. However, for short prefixes, the list that g.g.itemsMatchingPrefixInList scans can have thousands of items. Scanning large lists can’t be helped in any case for short prefixes.

Happily, the new scheme is still completely stateless: the completionDict does not define state (it is valid everywhere) and no state variables had to be added. In short, the new caching scheme is much better than before, and it probably is close to optimal in most situations.

Clones

New in Leo 4.7. All clones of a node are the same node. This is the so-called one-node world. In this world, vnodes represent data, generators and positions represent the location of the data in an outline. This is a much simpler world than all previous data representations.

In Leo versions 4.2 to 4.6 clones were represented by sharing tnodes. Cloned vnodes shared the same tnode. This shared tnode represented the entire shared subtree of both clones. Thus, the _firstChild link had to reside in tnodes, not vnodes.

Prior to Leo version 4.2, Leo duplicated all the descendants of vnode v when cloning v. This created many complications that were removed in the shared tnode world. In particular, in the shared tnode scheme a vnode v is cloned if and only if len(v.vnodeList) > 1.

Drawing and events

Leo must redraw the outline pane when commands are executed and as the result of mouse and keyboard events. The main challenges are eliminating flicker and handling events properly. These topics are interrelated.

Eliminating flicker. Leo must update the outline pane with minimum flicker. Various versions of Leo have approached this problem in different ways. The drawing code in leo.py is robust, flexible, relatively simple and should work in almost any conceivable environment. Leo assumes that all code that changes the outline pane will be enclosed in matching calls to the c.beginUpdate and c.endUpdate methods of the Commands class. c.beginUpdate() inhibits drawing until the matching c.endUpdate(). These calls may be nested; only the outermost call to c.endUpdate() calls c.redraw() to force a redraw of the outline pane.

Code may call c.endUpdate(flag) instead of c.endUpdate(). Leo redraws the screen only if flag is true. This allows code to suppress redrawing entirely when needed. For example, here is how the idle_body_key event handler in leoTree.py conditionally redraws the outline pane:

redraw_flag = false
c.beginUpdate()
val = v.computeIcon()
if val != v.iconVal:
        v.iconVal = val
        redraw_flag = true
c.endUpdate(redraw_flag) # redraw only if necessary

The leoTree class redraws all icons automatically when c.redraw() is called. This is a major simplification compared to previous versions of Leo. The entire machinery of drawing icons in the vnode class has been eliminated. The v.computeIcon method tells what the icon should be. The v.iconVal ivar tells what the present icon is. The event handler simply compares these two values and sets redraw_flag if they don’t match.

Handling events. Besides redrawing the screen, Leo must handle events or commands that change the text in the outline or body panes.

The leoTree class contains all the event handlers for the body and outline panes. The actual work is done in the idle_head_key and idle_body_key methods. These routines are surprisingly complex; they must handle all the tasks mentioned above, as well as others. The idle_head_key and idle_body_key methods should not be called outside the leoTree class. However, it often happens that code that handles user commands must simulate an event. That is, the code needs to indicate that headline or body text has changed so that the screen may be redrawn properly. The leoTree class defines the following simplified event handlers: onBodyChanged, onBodyWillChange, onBodyKey, onHeadChanged and onHeadlineKey. Commanders and subcommanders call these event handlers to indicate that a command has changed, or will change, the headline or body text. Calling event handlers rather than c.beginUpdate and c.endUpdate ensures that the outline pane is redrawn only when needed.

Find and change commands

The find and change commands are tricky; there are many details that must be handled properly. The following principles govern the LeoFind class:

1. Find and Change commands initialize themselves using only the state of the present Leo window. In particular, the Find class must not save internal state information from one invocation to the next. This means that when the user changes the nodes, or selects new text in headline or body text, those changes will affect the next invocation of any Find or Change command. Failure to follow this principle caused all kinds of problems in the Borland and Macintosh codes. There is one exception to this rule: we must remember where interactive wrapped searches start. This principle simplifies the code because most ivars do not persist. However, each command must ensure that the Leo window is left in a state suitable for restarting the incremental (interactive) Find and Change commands. Details of initialization are discussed below.
2. The Find and Change commands must not change the state of the outline or body pane during execution. That would cause severe flashing and slow down the commands a great deal. In particular, the c.selectPosition and c.editPosition methods must not be called while looking for matches.
3. When incremental Find or Change commands succeed they must leave the Leo window in the proper state to execute another incremental command. We restore the Leo window as it was on entry whenever an incremental search fails and after any Find All and Change All command. Initialization involves setting the self.c, self.v, self.in_headline, self.wrapping and self.s_text ivars.

Setting self.in_headline is tricky; we must be sure to retain the state of the outline pane until initialization is complete. Initializing the Find All and Change All commands is much easier because such initialization does not depend on the state of the Leo window. Using the same kind of text widget for both headlines and body text results in a huge simplification of the code.

Indeed, the searching code does not know whether it is searching headline or body text. The search code knows only that self.s_text is a text widget that contains the text to be searched or changed and the insert and sel attributes of self.search_text indicate the range of text to be searched. Searching headline and body text simultaneously is complicated. The selectNextVnode() method handles the many details involved by setting self.s_text and its insert and sel attributes.

Key handling

The following three sections deal with different aspects of how Leo handle’s keystrokes that the user types. This is the most complex code in Leo.

Key domains

Leo’s key-handling code has almost nothing to do with supporting multiple guis. Rather, the key-handling code is complex because it must deal with the following four fundamentally different problem domains.

Domain 1: Parsing user bindings in @keys, @mode and @shortcut nodes

In this domain, complexity arises from allowing the user a variety of equivalent ways of specifying bindings. Furthermore, this domain allows the user to specify modes and per-pane bindings. Thus, all these complexities are unavoidable.

Domain 2: Maintaining and using binding tables

The result of parsing user bindings are a set of binding tables. These tables are complex, but we need not go into details here because only one method, k.masterKeyHandler (and its helper, getPaneBinding) uses the tables.

The only thing we have to remember about the binding tables is that bindings are expressed in terms of so-called strokes. Strokes are the “official” form of every user binding. The essential property of a stroke is that it contains all the information required to handle the stroke: Leo can unambiguously determine exactly what a stroke means and what bindings are in effect for a stroke. If necessary, Leo can correctly insert the proper character corresponding to the stroke into any text widget.

This correspondence (association) between the stroke and the actual character to be inserted into text widgets is crucial. This correspondence is created in the next domain.

Anticipating a bit, for any incoming key event, event.stroke is the stroke, and event.char is the character (if any) that might (depending on bindings) be inserted into text widgets.

New in Leo 4.9: k.stroke2char calculates the to-be-inserted char from any stroke.

Domain 3: Translating incoming key events into standard events

The eventFilter method in qtGui.py creates leoKeyEvent objects. Turning “raw” Qt key events into leoKeyEvents is unavoidably complicated because eventFilter (and its allies) must carefully compute the stroke corresponding to the raw key event. There is no way around this requirement if Leo’s binding machinery in domain 2 is to work. This code has been stable for a long time.

Domain 4: Printing key bindings in a human-readable format

It’s important not to forget this domain: there are some situations in which we want to represent ‘b’,’r’,’n’ and ‘t’ as ‘BackSpace’,’Linefeed’,’Return’ and ‘Tab’.

Key bindings

There are two kinds of bindings, gui bindings and pane bindings.

Gui bindings are the actual binding as seen by whatever gui is in effect. Leo binds every key that has a binding to k.masterKeyHander.

At present Leo makes gui bindings in several places, all equivalent. Bindings are made to callbacks, all of which have this form:

def callback(event=None,k=k,stroke=stroke):
   return k.masterKeyHandler(event,stroke)

As a result, changing gui bindings actually has no effect whatever. It would be clearer to have a single place to make these bindings...

In any case, the purpose of these callbacks is to capture the value of ‘stroke’ so that it can be passed to k.masterKeyHandler. This relieves k.masterKeyHandler of the impossible task of computing the stroke from the event.

Important: No function argument is ever passed to k.masterKeyHandler from these callbacks, because k.masterKeyHandler binds keys to command handlers as described next.

Pane bindings are bindings represented by various Python dictionaries in the keyHandlerClass (see below). k.masterKeyHandler and its helpers use these dictionaries to call the proper command or mode handler. This logic is hairy, but it is completely separate from the gui binding logic.

Here are the dictionaries that k.masterKeyHandler uses:

• c.commandsDict: Keys are minibuffer command names; values are functions f.
• k.inverseCommandsDict: Keys are f.__name__l values are emacs command names.
• k.bindingsDict: Keys are shortcuts; values are lists of g.bunch(func,name,warningGiven).
• k.masterBindingsDict: Keys are pane names: ‘all’,’text’,etc. or mode names. Values are dicts: keys are strokes; values are g.Bunch(commandName,func,pane,stroke).
• k.modeWidgetsDict: Keys are mode names; values are lists of widgets to which bindings have been made.
• k.settingsNameDict: Keys are lowercase settings; values are ‘real’ Tk key specifiers. Important: this table has no inverse.
• inverseBindingDict: This is not an ivar; it is computed by k.computeInverseBindingDict(). Keys are emacs command names; values are lists of shortcuts.

Handling key events

All events are key events

All event objects passed around Leo are key event objects. Taking a look at the eventFilter method, we see clearly see that only key events ever get passed to Leo’s core. All other events are handled by Qt-specific event handlers.

As can be seen, these non-key events can be passed to Leo, but only as the event arg in g.doHook (!) At present, no plugin ever calls k.masterKeyHandler. The only call to k.masterKeyHandler in qtGui.py is the expected call in eventFilter.

There are other calls to k.masterKeyHandler in Leo’s core, but we can prove (by induction, if you will), that all events passed to k.masterKeyHandler are proper leoKeyEvent objects.

c.check_event

The essential invariant is that the events passed to Leo’s core methods really are leoKeyEvent objects created by qtGui.py. Rather than asserting this invariant, the code will contains calls to c.check_event in essential places. c.check_event is a “relaxed” place to do as much error checking is needed. In particular, running the unit tests calls c.check_event many times.

c.check_event is a happy “accident”. It turns out to be the essential consistency check that continually verifies that the Qt event methods are delivering the expected keys to k.masterKeyHandler.

About the KeyStroke class

The KeyStroke class distinguish between “raw” user settings and the “canonicalized” form used throughout Leo. Indeed, the ability to explicitly distinguish between the two, using type checking, substantially clarifies and simplifies the code.

The KeyStroke class makes possible vital type-related assertions. Knowing for sure exactly what crucial data is and what it means is a huge step forward.

Objects of the KeyStroke class can be used exactly as a strings may be used:

A. KeyStroke objects may be used as dictionary keys, because they have __hash__ methods and all the so-called rich comparison methods: __eq__, __ne__, __ge__, __gt__, __le__ and __lt__. Note that KeyStroke objects may be compared with other KeyStroke objects, strings and None.

B. At present, KeyStroke objects supports the find, lower and startswith methods. This simplifies the code substantially: we can apply these methods to either strings or KeyStroke objects, so there is no need to create different versions of the code depending on the value of g.new_strokes.

However, having the KeyStroke class support string methods is bad design. Indeed, it is a symptom that the client code that uses KeyStroke objects knows too much about the internals of KeyStroke objects. Instead, the KeyStroke class should have higher-level methods that use s.find, s.lower and s.startswith internally.

You could say that the fact that code in leoKeys.py calls s.find, s.lower and s.startswith is a symptom of non OO programming. The internal details of settings and strokes “pollutes” the code. This must be fixed. This will likely create opportunities for further simplifications.

> Why not just have .s attribute in KeyStroke, that contains the string version?

It is truly impossible to understand the key code without knowing whether an object is a string representing a user setting or the canonicalized version used in Leo’s core, that is, a KeyStroke object. Using ks.s instead of ks destroys precisely the information needed to understand the code.

Again, this is not a theoretical concern. The key code now contains assertions of the form:

assert g.isStroke(stroke)

or:

assert g.isStrokeOrNone(stroke)

Getting these assertions to pass in all situations required several important revisions of the code. The code that makes the assertions pass is “innocuous”, that is, almost invisible in the mass of code, but obviously, these small pieces of code are vital.

This is in no way a violation of OO principles. The code is not dispatching on the type of objects, it is merely enforcing vital consistency checks. This code is complex: confusion about the types of objects is intolerable. Happily, the resulting clarity allows the code to be substantially simpler than it would otherwise be, which in turn clarifies the code further, and so on...

Simplifying the Qt input code

The Qt key input code can be greatly simplified by calling a new k.makeKeyStrokeFromData factory method. At present, the Qt key input code knows all the details of the format of canonicalized settings. This is absolutely wretched design.

Instead, the Qt input key code should simply pass the key modifiers and other key information to k.makeKeyStrokeFromData, in a some kind of “easy” format. For example, the Qt input key code would represent the internal Qt modifiers as lists of strings like “alt”, “ctrl”, “meta”, “shift”. k.makeKeyStrokeFromData would then create a user setting from the components, and then call k.strokeFromSetting to complete the transformation.

Typed dicts

Leo’s key dictionaries will always be complex, but basing them on the TypedDict class is a major improvement.

The g.TypedDict and g.TypedDictOfLists classes are useful for more than type checking: they have unique names and a dump method that dumps the dict in an easy-to-read format that includes the name, and valid types for keys and values.

Plain dicts do have their uses, but for “long-lived” dicts, and dicts passed around between methods, plain dicts are as ill-advised as g.Bunches.

Nodes

The vnode class is Leo’s fundamental model class. A vnode represents the data represented by headlines. As of Leo 4.7, all clones of a node are in fact exactly the same node.

The vnode contains all data associated with a node. A vnode contains headline text, body text, and user attributes, uA’s for short.

Because Leo has unlimited Undo commands, Leo deletes vnodes only when a window closes. Leo deletes nodes indirectly using destroy methods. Several classes, including the vnode, leoFrame and leoTree classes, have destroy methods. destroy methods merely clear links so that Python’s reference counting mechanisms will eventually delete vnodes and other data when a window closes.

Leo’s XML file format uses tx and t attributes to associate <v> elements with <t> elements. <v> elements represent nodes. <t> elements represent the body text of nodes. This is a (somewhat dubious) space optimization. The values of tx and t attributes are gnx’s (global node indices). These indices do not change once a node has created.

Overview

All versions of Leo are organized as a collection of classes. The general organization of Leo has remained remarkably stable throughout all versions of Leo, although the names of classes are different in different versions. Smalltalk’s Model/View/Controller terminology is a good way think about Leo’s classes. Model classes represent the fundamental data. The vnode class is Leo’s primary model class.

View classes draw the screen. The main view classes are leoFrame.py and leoTree.py. The colorizer class in leoColor.py handles syntax coloring in the body pane. Leo’s view classes know about data stored in the vnode class. Most events (keystrokes and mouse actions) in the outline and body pane are handled in the leoTree class. The leoFrame class also creates the Leo window, including menus, and dispatches the appropriate members of the controller classes in response to menu commands.

Controller classes (aka commanders) control the application. In Leo, controllers mostly handle menu commands. Commanders create subcommanders to handle complex commands. The atFile class reads and writes files derived from @file trees. The LeoFind class handles the Find and Change commands. The leoImportCommands class handles the Import and Export commands, and the undoer class handles the Undo command. Other classes could be considered controller classes.

Each Leo window has its own commander and subcommanders. Subcommanders are not subclasses of their commander. Instead, subcommanders know the commander that created them, and call that commander as needed. Commanders and subcommanders call the model and view classes as needed. For example, the Commands class handles outline commands. To move a headline, the commander for the window calls a vnode move routine to alter the data, then calls the view class to redraw the screen based on the new data.

A singleton instance of the LeoApp class represents the application itself. All code uses the app() global function to gain access to this singleton member. The ivars of the LeoApp object are the equivalent of Leo’s global variables. leo.py uses no global Python variables, except the gApp variable returned by app(). leoGlobals.py defines all application constants. Naturally, most constants are local to the class that uses them.

Several classes combine aspects of model, view and controller. For example, the LeoPrefs class represents user preferences (model), the Preference Panel (view) and the Preferences menu command (controller). Similarly, the LeoFind class represents find settings, the Find/Change dialog, and the Find/Change commands.

We use the following convention throughout this documentation. Any variable named c is a commander, i.e., an instance of the Commands class in leoCommands.py. Variables named v are vnodes. These classes are defined in leoNodes.py.

Unicode

Leo uses unicode objects in vnodes to denote headline and body text. Note that unicode strings have no encoding; only plain strings have encodings. This means that once an (encoded) plain string has been converted to a unicode string it doesn’t matter how the unicode string was created. This is the key that makes Leo’s new code robust: internally Leo never has to worry about encodings. Encoding matter only when encoded strings are converted to and from Unicode. This happens when Leo reads or writes files.

Python expressions that mix unicode strings u and plain strings s, like one of these:

u + s
u == s
u[5] == s[2:]

are promoted to unicode objects using the “system encoding”. This encoding should never be changed, but we can’t assume that we know what it is, so for safety we should assume the most restrictive encoding, namely “ascii”. With this assumption, Leo’s code can’t throw an exception during these promotions provided that:

• Leo converts all text to unicode when Leo reads files or gets text from text widgets.
• All string literals in Leo’s code have only ascii characters.

Unlimited undo

Unlimited undo is straightforward; it merely requires that all commands that affect the outline or body text must be undoable. In other words, everything that affects the outline or body text must be remembered. We may think of all the actions that may be Undone or Redone as a string of beads (undo nodes).

Undoing an operation moves backwards to the next bead; redoing an operation moves forwards to the next bead. A bead pointer points to the present bead. The bead pointer points in front of the first bead when Undo is disabled. The bead pointer points at the last bead when Redo is disabled. An undo node is a Python dictionary containing all information needed to undo or redo the operation. The Undo command uses the present bead to undo the action, then moves the bead pointer backwards.

The Redo command uses the bead after the present bead to redo the action, then moves the bead pointer forwards. All undoable operations call setUndoParams() to create a new bead. The list of beads does not branch; all undoable operations (except the Undo and Redo commands themselves) delete any beads following the newly created bead. I did not invent this model of unlimited undo. I first came across it in the documentation for Apple’s Yellow Box classes.

Leo’s load process

Leo reads local files twice. The first load discovers the settings to be used in the second load. This ensures that proper settings are available during the second load.

Ctors init settings “early”, before calling the ctors for subsidiary objects. This ensures that proper settings are in effect for the subsidiary ctors.

After creating all subsidiary objects, c.__init__ simply calls c.finishCreate! Creating Commands objects is now completely self-contained. In particular, c.__init__ now creates the fully-inited gui frame. This is a revolution in Leo’s startup code! c.__init__ no longer needs a frame argument, a surprisingly important conceptual simplification.

The old code inited Leo windows in several places (c.__init__, the end of g.app.newLeoCommanderAndFrame and the end of g.openWithFileName) and in several phases, (g.app.newCommanderAndFrame, g.openWithFileName and c.finishCreate.)

The new code has c.__init__ do all the work, in one place, and in one phase. This is supremely important for future maintainers. The old code was difficult for me to understand yesterday, even after a full week of study. The new code is a simple as could possibly be imagined. This is a gigantic step forward for Leo.