The code in render/picture.c assumes that if the source drawable
doesn't have alpha then the source doesn't have alpha, but that
isn't true for RepeatNone - sampling off the edge of the picture
will introduce alpha.
Will attach a patch fixing the problem. It doesn't try to perfectly
capture every case where the optimization *is* possible, though
it does catch the simple common case where there is no transform
on the source picture.
Created attachment 15706 [details] [review]
Fixed in 1.5 branch and master, thanks!