On Mon, Aug 16, 2004 at 05:22:38PM -0300, bulia byak wrote:
Today it crashes when I try to move a rect in the sample file.
Thanks for the info. *runs it* Ah, I've fixed that at home without committing. *edit* Now I've committed to CVS what I think I did at home. Still haven't tested much, but it doesn't crash.
The current code in CVS preserves the shape of the connector (modulo a form of rotation/scaling), but the calculation of which point on the bounding box to use doesn't take the curve shape into consideration. This is most noticeable a problem when the two attached objects are very close to each other (e.g. overlapping), in which case we'd like the connectors to be able to come from the "sides" of the two objects.
There are a number of ways of modifying a curve to have different end points (e.g. different distance or direction) while retaining the "same" curviness. Different transforms do the right or wrong thing for different cases.
An affine transform would be the first thing that comes to mind. E.g. rotate then scale. When stretching/squashing, it's not clear whether we should scale uniformly or scale just in the direction of the new angle. E.g. suppose that the end points lie on a horizontal line and the user's just dragging left & right. Should we scale just the x coordinates or scale x and y coordinates by the same amount? Just-x can be desirable when the new points are very far from each other, whereas both-equally can be better if the new x distance is less than the old y height of the curve.
The main thing to be careful of when using an affine transform is not to lose information if the new distance is zero. Perhaps use the path's transform= attribute, or make the path invisible if the new distance is zero.
The way I've implemented it is not an affine transform on the path. It's vaguely spring-like, in that the amount that each point of the curve is moved is proportional to how far along the curve that point is rather than what its position is. For bezier segments, I've used a simple approximation to that rather than try to refit a new curve to an exact transformation.
At some point in the future we may allow playing around with different choices for how the curve is transformed, but we can't make a fair comparison until we improve said calculation of where to attach to to take curve shape (end tangents) into consideration.
[Allowing choice of connector shapes would be good in future]
Most interesting for us is getting connectors to choose a good route that tries to avoid crossing other "objects". This requires that the software know what objects should be avoided, and what objects deliberately overlap or come close to the connector. Similarly, it is often desirable that certain objects not overlap each other. State charts are interesting, in that objects A1 & A2 mustn't overlap each other, while objects B1 & B2 are required to overlap (and be contained within) A1, while not overlapping each other.
pjrm.