Division on Zero

[ShiftZ]

slembcke
Hi,

cpSpaceHash.c 512:

Code: Select all

cpFloat dt_dx = 1.0f/cpfabs(b.x - a.x), dt_dy = 1.0f/cpfabs(b.y - a.y);

Getting devision on zero.
Do you have any technical agreement about division on zero. Is it legal?

[slembcke]

Looks like you are trying to do a zero length segment query. I think the correct behavior in that case would be to ignore the query altogether. With the way segment queries work, it would be impossible to detect a hit with any shape.

[ShiftZ]

Looks like you are trying to do a zero length segment query.

No,
b.x - a.x = 0.
Its just a segment parallel to y - axis. and

b.y - a.y = 0,
parallel to x - axis.

Zero length when both are equal to zero.

[slembcke]

Oh, I see what you are getting at. Handling horizonal and vertical lines is handled later in that function by this:

Code: Select all

	// fix NANs in horizontal directions
	cpFloat next_h = (temp_h ? temp_h*dt_dx : dt_dx);
	cpFloat next_v = (temp_v ? temp_v*dt_dy : dt_dy);

Well, division of a finite floating point number by zero gives you infinity. So yes, it's fine. In fact, Chipmunk uses this in a number of places to deal with infinite masses and such.

[ShiftZ]

I noticed that the division by zero greatly stall iPhone processor. I had a bug once, an unexplainable lags on iPhone when player is dead. Problem was that maxSpeed of player was zero and it was divisor in some places. So i switched on floating point exceptions to catch such situations and catched problem described above.

[slembcke]

Hrm. That's strange, I've never heard of that before. Division by zero to get infinity is the standard floating point behavior. It shouldn't be producing any floating point exceptions, but I guess I'm not totally surprised that it's handled in a way that produces a stall though. Floating division generally already stalls the CPU pipeline on most CPUs.

[ShiftZ]

It shouldn't be producing any floating point exceptions

As far as i know cpu can call a signal processor to trap a signal. So its your decision does it generate C++ fp exception or not.
If you mean hardware exception, it will be generated on all kind of proccessors i know. Not sure about stall, but armv6 lagging hardly.
Maybe there is a way to control it, i dont know, ARM support suggests to avoid division by zero at all, and offers some ways to trap it.

[slembcke]

Strange. Floating point divide by zero is not supposed to be considered exceptional. It seems a bit odd that ARM decided to make it optionally exceptional.

I made the following change in trunk:

Code: Select all

	cpFloat dx = cpfabs(b.x - a.x), dy = cpfabs(b.y - a.y);
	cpFloat dt_dx = (dx ? 1.0f/dx : INFINITY), dt_dy = (dy ? 1.0f/dy : INFINITY);

[ShiftZ]

Another catch in cpShape.c 292:

legal arguments values:

n = {1.0000000, -0.00016872486 }
a = (506.35046, 504.56598};
b = {506.35031 505.58569 }

Code: Select all

cpFloat an = cpvdot(a, n);
cpFloat bn = cpvdot(b, n);

an = -506.43561
bn = -506.43561

Code: Select all

cpFloat t = (d - an)/(bn - an);  

Div by zero exception;

[slembcke]

Wrapped an if around it. Looks like I have some other unneeded code in there too. I'l have to look into that at some point.