In order to give a better account of how focus and master interact, and
how each operation affects focus, we reimplement the StackSet type as a
two level nested 'Zipper'. To quote Oleg:
A Zipper is essentially an `updateable' and yet pure functional
cursor into a data structure. Zipper is also a delimited
continuation reified as a data structure.
That is, we use the Zipper as a cursor which encodes the window which is
in focus. Thus our data structure tracks focus correctly by
construction! We then get simple, obvious semantics for e.g. insert, in
terms of how it affects focus/master. Our transient-messes-with-focus
bug evaporates. 'swap' becomes trivial.
By moving focus directly into the stackset, we can toss some QC
properties about focus handling: it is simply impossible now for focus
to go wrong. As a benefit, we get a dozen new QC properties for free,
governing how master and focus operate.
The encoding of focus in the data type also simplifies the focus
handling in Operations: several operations affecting focus are now
simply wrappers over StackSet.
For the full story, please read the StackSet module, and the QC
properties.
Finally, we save ~40 lines with the simplified logic in Operations.hs
For more info, see the blog post on the implementation,
http://cgi.cse.unsw.edu.au/~dons/blog/2007/05/17#xmonad_part1b_zipper
When focus was separated from the stack order on each workspace, we
forgot to update the Arbitrary instance to set random focus. As spotted
by David R, this then invalidates 4 of our QC properties. In particular,
the property involving where focus goes after a random transient
(annoying behaviour) appeared to be correct, but wasn't, due to
inadequate coverage.
This patch sets focus to a random window on each workspace. As a result,
we now catch the focus/raise/delete issue people have been complaining
about.
Lesson: make sure your QuickCheck generators are doing what you think
they are.
Using Typeables as the only constraint on layout messages is a bit
scary, as a user can send arbitrary values to layoutMsg, whether they
make sense or not: there's basically no type feedback on the values you
supply to layoutMsg.
Folloing Simon Marlow's dynamically extensible exceptions paper, we use
an existential type, and a Message type class, to constrain valid
arguments to layoutMsg to be valid members of Message.
That is, a user writes some data type for messages their layout
algorithm accepts:
data MyLayoutEvent = Zoom
| Explode
| Flaming3DGlassEffect
deriving (Typeable)
and they then add this to the set of valid message types:
instance Message MyLayoutEvent
Done. We also reimplement the dynamic type check while we're here, to
just directly use 'cast', rather than expose a raw fromDynamic/toDyn.
With this, I'm much happier about out dynamically extensible layout
event subsystem.
