Why do the coordinates swap when doing matrix multiplication while rotating a rectangle?

[brandxntu]

I am applying the derotation matrix to a rectangle on a page that is rotated 270 degrees, however I am noticing some strange behavior in the output of that rectangle.

In this case, the derotation matrix is Matrix(0, 1, -1, 0, h, 0)
(h corresponding to the height of the page)
For example if the rectangle I am using is [x1 ,y1 ,x2, y2], multiplying the rectangle by the derotation matrix will give me an output of [h-y2, x1, h-y1, x2].

However, when doing the matrix multiplication by hand for each point, the result is:

[x1, y1, 1] * Matrix(0, 1, -1, 0, h, 0) = [h-y1, x1]

[x2, y2, 1] * Matrix(0, 1, -1, 0, h, 0) = [h-y2, x2]

While [h-y2, x1, h-y1, x2] and [h-y1, x1, h-y2, x2] are functionally the same rectangle (albeit with different corners), why is this swap of the x values done? Furthermore, how is this done mathematically in the code? What rule is used to swap these coordinates?

[JorjMcKie]

Your manual multiplication - corner point by corner point - leads to an "infinite" rectangle, see documentation.
The logic takes care that the result rect is defined by the diagonal of the points top-left and bottom-right.
To see the algorithm (which ultimately is written in C), lookup fz_transform_rect in MuPDF source code.