Multipolygon intersection incorrect for rectangles with rather specific collinear points

[thmshnz]

The source files in the attached archive bug.zip demonstrate an intersection bug that is present in version 1.79.0. Given two intersecting rectangles each with collinear points represented as two multi_polygons, for specific choices of those points, boost::geometry::intersection produces an incorrect result.

nonbug.cpp produces the following output where the second part of the output refers to a case with the y-coordinates of the points on the second rectangle being perturbed ever so slightly. As expected, the output is the same in both cases.

Valid (mpoly1): 1
Valid (mpoly2): 1
Valid (mpoly3): 1
Area:           0.09
Bounding box:   ((-5.8, 5), (-5.5, 5.3))

Valid (mpoly1): 1
Valid (mpoly2): 1
Valid (mpoly3): 1
Area:           0.09
Bounding box:   ((-5.8, 5), (-5.5, 5.3))

However, bug.cpp produces the following output where the first part indicates the bug. After perturbing the y-coordinates of the points on the second rectangle, the problem vanishes.

Valid (mpoly1): 1
Valid (mpoly2): 1
Valid (mpoly3): 1
Area:           0.24
Bounding box:   ((-5.8, 5), (-5, 5.3))

Valid (mpoly1): 1
Valid (mpoly2): 1
Valid (mpoly3): 1
Area:           0.09
Bounding box:   ((-5.8, 5), (-5.5, 5.3))

Note that the difference between both files is as follows.

--- nonbug.cpp  2022-07-21 13:00:55.168046300 +0200
+++ bug.cpp     2022-07-25 11:00:54.132520800 +0200
@@ -165,8 +165,8 @@ double r2[][2] = {
     {-5.54, 5.3},
     {-5.53, 5.3},
     {-5.52, 5.3},
-    {-5.51, 5.3},
-    {-5.5, 5.3},
+    {-5.51, 5.299999999999998},
+    {-5.5, 5.299999999999998},
     {-5.49, 5.3},
     {-5.48, 5.3},
     {-5.47, 5.3},

[barendgehrels]

Thanks for the report! I'll have a look next week.

[thmshnz]

Looking at boost versions 1.76 - 1.80, I noted that the bug is present in version 1.79 and the latest 1.80 release. It is not present in versions 1.76 - 1.78.

[barendgehrels]

Sorry, it's a few weeks later already. It's a numerical issue. Using long double fixes this particular example. Boost 1.79 is different from the versions before, because we stopped using rescaling (which caused several other issues). I will look later if this issue can be fixed in a generic way, but that will not be soon.