[cairo] create_surface_for_cg_context behavior in Cairo 1.9 vs. 1.12

[cu]

The test case is quite simple:

// .. set up source_surf as an image surface of the same size as the
UIView, draw into it...

CGContextRef cgref = UIGraphicsGetCurrentContext();
qsurf = cairo_quartz_surface_create_for_cg_context(cgref,
source_surf_width, source_surf_height);
qpat = cairo_pattern_create_for_surface(source_surf);
qcr = cairo_create(qsurf);
cairo_set_source(qcr, qpat);
cairo_paint(qcr);

Under 1.9.8 the resulting cairo surface backed by CGContext is measured
in "points", covers the entire view and seems to have native iOS scaling
applied.
Under 1.12 it is in pixels with no scaling applied, and covers upper
left quadrant of the view.

I tried to manually add scaling to CGContext surface while using cairo
1.12 by setting scaling transform and increasing surface dimensions
times 2. The result was not the same as when using cairo 1.9.8. Image
appeared poorly scaled (not very sharp, looked like the usual pixman
upscaling) and rendering was very slow. It would appear that 1.9.8
somehow is using native up-scaling of iOS which is faster and creates
better quality images on "retina" display. We are using this
functionality to move main image  to screen, where both performance and
quality of upscaling are crucial.


>
> Quite a lot of changes happened between 1.12 and 1.9.8.
> Could you provide a small testcase to help keeping track of this
> behavior (and to bisect the change).
>
> The API reference says:
> "The CGContext is assumed to be in the standard Cairo coordinate space
> (that is, with the origin at the upper left and the Y axis increasing
> downward). If the CGContext is in the Quartz coordinate space (with
> the origin at the bottom left), then it should be flipped before this
> function is called."
> so I would expect cairo-quartz to respect the original transform
> (although I guess bad things might happen if there are fallbacks
> and/or the surface is used as a source).
>
>