[Scwm-devs] random thoughts and ideas

[Scott L. <sl...@cs...>]

Hi all.  [Sorry this message isn't formatted for 80 columns]

I've had about all the testing I can take at the moment, so I'm taking
a short break to familarize myself with the C code and start thinking
about plans and enhancements to make.  I should get back to testing
again shortly.  Anyway, here are some random thoughts, ideas, comments,
and questions for feedback.

I just noticed the stuff about window properties that Maciej was adding.  I'm
not sure that this is a good approach.  The way I see it, this basically
replaces a lot of specific functions that do clear things with a general
dispatching function that needs to have lots of magic constants/sysmbols/etc.
defined for it to work.  I don't really see much gain here.  The main gain
I see here is a logical place to attach new properties to windows.  With an
appropriate naming scheme, I think the individual functions capture most
of the information here.  Maybe what we need here is just to have the hooks
for the specific events call a generic property change hook after they run.
This seems to get most of the benefits (although I'm not even convinced this
is necessary) without resorting to processing every kind of hook on every
kind of event.  I think the scwm-options stuff is more like the kind of
interface we want.  What we do need is some way to add information to the
window thats needs to be used by some modules.  Thoughts on this?  I'm
pretty fuzzy on what the problem even is that we are trying to solve here.
I see a need for extensions to the information for each window, but I'm
not sure this gets us anywhere.

Anyway, that was actually just a side note.  The main thing I've been
thinking about is how to do the event/decoration rewrite correctly.  I've
grouped them together here since in my mind they are related and need to
be thought about at the same time (because events can be bound to decoration
elements).  Probably need to implemented in the order event rewrite then
decoration rewrite.  I've looked over the stuff in doc/dev to gather up
the ideas people had been thinking about.  In general, I think the event
proposal in doc/dev/events.gjb is pretty good, although I have a few comments
and thoughts.  First off, I think the idea of making events (make-key-event
and friends) might a good idea IF we can wrap this for the user such
that they can specify things more easily.  It would be pretty irritating to
have to make an event before you can bind a key in a keymap which then needs
to be installed before it has any effect.  I'm not sure this is a clear win
though.  Added this layer allows us to construct a little C structure to
represent that type of event which could allow us to avoid reparsing the
arguments if the event is reused.  However, many events are global and thus
only used once.  It's also not clear this translation is expensive.  I also
think the repesentation of the event map is unlikely to use these structures
since the event map should probably be represented as a hash.  Then having
these lying around in memory is just wasting space.  Also, if we use these
make-*-event routines, they should have a different name.  As is, they sound
like they are making an event that you could send to an app whereas they are
being used to define slots to which event handlers can be bound.

  I like your ideas for event map objects in general.  I agree that they should
be attached to X windows (button, titles, frames, client windows, etc.).
I'm not sure I agree with your commentary in make-event-map about relying on the
geometry-based chaining of X11.  Currently, we use the XShape extension to handle
our squashed title bars.  Fluxbox has a similar display element for their tabbed
windows which looks similar on screen.  However, instead of the shape extension
they use a seperate window which has the root window as parent.  This avoids the
shape extension but means that they have to explicitly remember to move both
windows whenever they move one.  Which is better depends on the efficiency of the
xshape extension and presumably on the area to which the shape extension is
applied.  I don't know enough to say whether this is a better way to do things,
but relying on X to propagate the events precludes doing things this way.  I'm not
sure if X even gives the option of not propagating the event upwards though.
I think allowing multiple parent event maps is confusing, error prone and doesn't
buy us much.  Hmm... just thought how that could be used to provide a nice keymap
mixin ability though.  I think we can get this same mixin type ability just
through the normal parent mechanism if we seperate the concept of event maps
and installed event maps (I think we might need this anyway to do the grabs
properly).  Then the installed event map could have the parent pointer and the
event map itself has no parent.  Oh, btw, the add-event-binding! primitive should
be able to take a keymap as the way to process a key-binding so that we can get
prefix key capability.  There should be some mechanism for having the sub event
map time out back to the normal key map for people who want it.  Anyway, back
to the parenting thing.  A mixin would just correspond to inserting another keymap
as the child of the current keymap.  Not sure what the primitives would look like
for all of this.  Just spouting ideas to get feedback right now.  For the event
callbacks, I don't see why different kinds of events couldn't have different
callback types.  Makes things less orthogonal but saves the callbacks accessing
some mysterious global variable.  I don't understand the comment you made about
the Common Lisp Object System method combination.  I don't understand how CLOS
is applicable, but maybe I'm just being dense.  I also think we want to have
some semantic classes to which event maps can be attached so you could say
things like (attach-event-map 'maximize-button max-button-event-map).  Semantic
classes would also be useful for supporting the various wm hints.

Some thoughts on decorations.  I think we need to think general here.  This means
almost abritrary drawing primitives for drawing various effects (like pixmaps,
gradients, colors, text, you name it) that should eventually live in modules that
get dynamically loaded.  This should allow for things like using a very simple
style for small footprint machines (although, I don't think we are going to do
to well on these anyway due to requiring guile) and using a native widget set
like gtk to blend with the apps better.  This could even allow gtk widgets on
GNOME apps and qt widgets on KDE apps.  I've been trying to keep two constraints
in mind when thinking about this: 1) it should be possible to emulate any other
window manager's look, 2) simple stuff should be simple to specify.  These two
are sort of at odds since 1 implies more features which implies more options
while 2 implies fewer options to have to specify.  I think this problem can be
handled with utility routines to use common styles and reasonable defaults as
appropriate.  Another major issue is how to specify the layout to use.  This
is actually even more complex than it initially appears because of new features
that are starting to appear in other window managers (assuming we stick to
constraint 1).  There are 3 features that I have seen that require special
thought.  The first is tabbed windows ala PWM, Fluxbox, and others.  The basic
idea is to group windows together into one frame and make a bunch of tab buttons
to select which one to display.  This looks cool and I'd like to support it.
The second, which is even more tricky, is subdivided displays and frames.  The
basic idea here is to stick multiple windows in one frame and provide a bar
inbetween them to control the relative size of the windows.  This is usually
a recursive binary partition but other grids are also possible.  See the Ion
window manager for a pretty example of this (http://www.students.tut.fi/~tuomov/ion/)
[what is it with those wacky Fins, they come up with the most strange and
interesting stuff, must be the winter].  I really want to support this as it
leads to a very nice interface for driving the window manager mostly from the
keyboard (the way I tend to do things).  The third feature is a bit more
exotic and not as useful but I'd like to support it if possible anyway.  This
feature is the workspace concept of amiwm (http://www.plig.org/xwinman/amiwm.html,
see the second screen shot).  The basic idea is several desktops that can be
manipulating somewhat like windows.  Sort of a nested window sort of thing.
This is probably more difficult to support and since I don't think it is as
useful, probably the right thing to do is not support at first in the rewrite
but try to leave room for it.  The common thread to these features is that they
remove the 1:1 mapping between client windows and frames/decorations.  I think
the decoration rewrite should reflect this.

Here's basically what I'm thinking at the moment.

Main concepts:

Container - an element in the window manager in which decorations can be arranged
  and client windows can be displayed (roughly equivalent to a frame [although may
  contain several client windows and the a client window can jump from container
  to container {although no 2 at the same time}]).  There would be a n:m mapping
  from client windows to associated containers.  Each client window would have a
  current container of NULL if not currently being displayed in any container.
  Containers can also hold other containers.  Event maps can be attached to containers.
  Each window will have at least a small container always associated with it which
  corresponds to the parent frame currently used.  [Not sure about this part:
  This can also have another frame to support decorations that jump with the window
  from container to container.  For Fvwm style decorations, this would be the only
  container needed.]
Layout - a layout specifies the layout of components within a container and enforces
  any constraints that need to be maintained.  Basically, the layout is reponsible
  for figuring out the location and size of all the widgets in a container.  Examples
  would be a grid layout, horizontal box, vertical box, etc. similar to tables, hboxs,
  vboxs, and the like in gtk.  Layouts can be nested.
Decoration - A primitive section of the screen to which an event map can be attached.
  Basically, these will be X windows.  This will be the smallest thing to which an event
  map can be attached.  Each decoration will have a list of drawing primitives to use
  to draw it.  Decorations would have some sort of shaped flag.
Drawing primitive - The basic primitives which can be used to draw stuff on the
  decorations.  This would include things like relief button, text, pixmap, relief
  pattern, gradient, solid color, line, use gtk button, and so on.  These need to
  be extensible by modules to allow for using widget libraries and such without making
  everyone suffer.  Probably need an object to hold lists of these.

Decorations, layout, and containers without client windows can share their
specification across multiple instantiations.  Probably sort of like the
menu item/menu item in menu distinction.

All of these settings should be changeable on the fly.  There will be styling
routines to help make specification for common styles.  There will be styling
routines for specifying things like fvwm style, wm2 style, etc.

So here's how I see the different cases above being implemented.

fvwm style - one container for each window, each button becomes a decoration, the
  border parts and corners become decorations, the title bar becomes a container
  holding the title, buttons.  The button faces, title faces, and border faces become
  drawing primitive lists.  Icons/menus are containers that don't have a client area.
pwm style [tabbed windows] - one container for each group of windows.  container is
  similar to fvwm stuff except has additional tab section.  Tab buttons bound to
  switching the client window associated with the container's client area.  Other
  bindings split/join groups of windows by creating/destroying containers dynamically.
  Note: this happens even when no client windows are being created.
ion style [subdivided screen] - one container for entire screen.  Container has several
  client areas within it along with decorations for the grab bars to resize the subparts.
  Windows get put into this container when captured and get a new client area created
  for them along with new decoration for the grab bar to resize them.  Note: this means
  need to be able to add decorations to a container on the fly.  Mixed ion/pwm/fvwm style
  can also be supported through similar means.

Other interesting things you could do with this: make a mode that when you bump one window
into another they become part of the same frame and move/iconify/etc together.  This gets
many of the same benefits as the constraint stuff.

Implementation notes:

The drawing primitives are mostly already there in the form of faces.
Many more things need to be dynamic because less info is known up front.  This means more
  dynamic data structures.
Need to be careful in specification of all this to make easy things easy.  This
  means common layout things need to be easy.

Well, those are my thoughts so far.  Let me know what you think, where there are holes,
unimplementable stuff, overly general stuff, etc.  I'm sure there are still major problems
with this and it is still pretty wishy-washy right now but I thought some early feedback
would be helpful.  I see this as probably an immediately post-1.0 thing.  This is going
to shake up the interface a fair amount so this might require a rather major bump of
version number.

Which reminds me, we need to think about a convention for version numbers.  Specifically,
I think something in the version number needs to correspond to changes in the API.  We
need to figure out a correspondence between API changes and revision number bumps so
people can tell what they are in for when upgrading (of course, we should try to keep
backwards compatibility too).  We also need a policy on backwards compatibility since
it'll be too unwieldy to support everything forever.  A lot of the current functions in
the API have pretty arbitrary names.  We should pick a naming scheme and try to stick
to it and change all the old names to new names.  There is the question of whether we
should do this pre-1.0 or not.  The argument against is things work as is and have been
tested this way.  The argument for is that the API should be somewhat stabilized with the
1.0 release and therefore this should happen before the 1.0 release.

Well I better get some sleep now.  Let me know what you'll think.

- Scott