Hi Sergey,
I guess my primary goal in maintaining this code is to try to find the
simplest and most directed fix unless a rewrite is in order - i.e.
simply normalizing incoming empty clips so that we end up in a stable
"clip is empty" state would be my preference, but perhaps the proposed
fix isn't so complicated that I need to worry about that.
I've looked deeper into the various Rectangle methods and found lots of
little issues with the way that the various bounds methods on
Rectangle/Rectangle2D are implemented that complicate all of this. I
think a larger rewrite of the clip maintenance code is likely in order
at some point. In particular, while your changes have shifted from one
set of problematic bounds-setting code to another and may have solved
some test case we are currently looking at, I'm not convinced that other
bugs, or perhaps even just as many different bugs, are not waiting to
bite us with the new code. :(
In the meantime, with respect to the new method you've added, I'd like
to see the following changes:
- The method has nothing really to do with "normalizing a matrix" and is
probably misled by the fact that the "matrix" variable in the calling
method should never have been called "matrix" in the first place since
it is really just a list of coordinates. I'd suggest a name better
suited to its purpose, like "fixRectangleOrientation" or
"checkEmptyRectangle".
- For performance - rather than call signum 4 times, why not use more
direct tests like:
if ((clip.getWidth() > 0) != (m[2] - m[0] > 0))
// i.e. we only care if their "non-emptiness" is different
or:
if (clip.getWidth() * (m[2] - m[0]) < 0)
// i.e. we only care if their non-zero size is flipped
- I still think that we don't owe anybody any need to preserve the size
of the empty clip so the adjustments for cases that fail the above tests
could simply be "m[2] = m[0]" and thus we get an empty result. Why did
you feel that we need to maintain the sizes of empty rectangles? Was it
just JCK test failures? Does flipping the results keep us from failing
those tests?
---------------------
Eventually I'd like to see the following simplifications to our SG2D
clip handling:
- we only keep a shape around so we can return a shape and we spend a
lot of time maintaining it. I'd like to see the shape be only
optionally kept if there is a non-rectangle handed in. Otherwise we
just maintain the clipRegion.
- we use Area to intersect non-rectangular shapes even though it is rare
for anyone to call getClip() and even rarer that they do anything with
it except to use setClip() on it to restore the clip. I'd rather see
keeping an array of the incoming shapes, return a special object from
getClip() that just holds that array, and then on setClip() simply
restore the region - thus avoiding ever having to use Area. Only if
they actually iterate the path of the returned getClip() would it
dynamically use Area on the fly to produce that path iteration.
- Avoid use of methods on Rectangle[2D] and write our own code that
ensures that we perform the correct flipping and integer normalization.
Most of that would happen in the Region bounds code if we do the above
anyway and that code is much simpler than some of the corresponding
methods on Rectangle[2D], plus we can fix bugs in there without worrying
about possibly changing the behavior for other apps that may have
started to rely on what we see as "bugs".
...jim
On 12/21/2012 3:21 AM, Sergey Bylokhov wrote:
21.12.2012 6:05, Sergey Bylokhov пишет:
Hi, Jim.
21.12.2012 4:59, Jim Graham wrote:
The Object.equals() method is not intended to compare geometries.
While Area.equals() attempts to perform geometric comparison I think
that was a bad idea in retrospect for many reasons:
- In practice you can only really compare within a tolerance due to
the many ways for computations to every so slightly shift the results
- There is no decent way to satisfy the equals/hashcode contract with
such a comparison so Area is forever broken with respect to that
contract.
- It's a complicated test that can only serve to convince developers
to invite performance drains into their code by making them think it
is a reasonable thing to compare.
For testing we should probably use Shape.contains() to verify answers
for rectangle arguments and if we want to perform verification on
arbitrary shape arguments then we'd need to write our own fuzzy
comparator.
But why fix is incorrect? It just use the same code as sffd for the
correct shapes and have additional code for incorrect. Only
conditionchanged from width/height should be positive, to
width/height should be the same direction as incoming rectangle.
There isn't even any guarantee that we will hand them back a rectangle
from getClip().
One of the jck tests expect this, but I think that this is the wrong
test.
There is nothing in the documentation that guarantees a specific type
returned from the method. If a JCK test was written to expect a
Rectangle from the method then that is a separate problem that we'll
have to address at some point. For now I'm only suggesting modifying
the process for negative-empty rectangles. Does that trigger the JCK
failure? If so, we are still satisfying the documented spec and so
we will simply have to confince JCK to allow this exception.
Problem in jck come from the fact that Rectangle differs from
Rectangle2d. Note that everything is ok if graphics is not transformed
or just translated. But if we scale graphics we get:
1 getClip() return non equivalent object.
2 getClipBounds for Rectangle return it as is, but for Rectangle2D it
is call getBounds, which return an empty bounds or wrong bounds if
rectangle2d was normalized by sffd.
3 Some tests expects class equivalence between set/get clip. I'll
prepare CR for them soon.
Also there is the difference in the real clip, if we use type for the
empty Rectangle to Rectangle2D the real clip became correct.
typo: Also there is the difference in the real clip, if we change type
from the empty Rectangle to the empty Rectangle2D, the real clip will
become correct.
All these issues are carried out in the testcase.
After all, if you end up in the drop through case then we hand the
paths to Area to do intersections and Area will do all sorts of
surgery on the geometry.
But I sure this area will be equal to the user's clip?
For all practical purposes it should be, but not only will the Area
code encapsulate the answer in a different class (even if the result
is a rectangle), it may reverse the order of the segments, it may
represent the area as two separate rectangles that abut each other in
a way that they are equivalent to one larger rectangle, but the Area
code failed to notice that optimization, etc. All it will guarantee
is that it returns a shape that describes the same interior, it will
make no guarantees about minimalism or association of segments...
...jim
--
Best regards, Sergey.
--
Best regards, Sergey.
