When normalize coordinate, return instead of assert() when divide by 0

core/include/fxcrt/fx_coordinates.h

Index: core/include/fxcrt/fx_coordinates.h
diff --git a/core/include/fxcrt/fx_coordinates.h b/core/include/fxcrt/fx_coordinates.h
index 9a63e2a4ad5946afd582975dc2a84c1e465169f8..6ec3b5e1ebb2fd69fb1ecf05f5508f69fe4a1c79 100644
--- a/core/include/fxcrt/fx_coordinates.h
+++ b/core/include/fxcrt/fx_coordinates.h
@@ -156,7 +156,9 @@ public:
     void		Normalize()
     {
         FX_FLOAT fLen = FXSYS_sqrt(FXT_PSV::x * FXT_PSV::x + FXT_PSV::y * FXT_PSV::y);
-        FXSYS_assert(fLen >= 0.0001f);
+        if (fLen < 0.0001f) {

[Tom Sepez, 2014/08/04 17:09:38]

Isn't .0001f a kinda large value for near-zero? I'd expect you could get a few
more sig digits here.

Do we want to set both x and y to something (like 0.0) when the length is 0?

[Bo Xu, 2014/08/04 18:48:09]

For some fuzz test, the x, and y are set to be really large values like over
1e20. After some matrix transformation, the fLen is computed to -1.#IND0004 or
-1.#INF0004. In these situations, I think we should leave these values there for
later application processing but not change to zero.

I can check with the devs on the 0.0001f, but this number is for the objects
spacing in PDF and 0.00001f is a reasonable "0" for a spacing in the PDF world.

+            return;
+        }
         FXT_PSV::x = ((baseType)FXT_PSV::x) / fLen;
         FXT_PSV::y = ((baseType)FXT_PSV::y) / fLen;
     }