1. The "fancy" operations like the blending modes (I believe these were originally from Photoshop and are now part of PDF, among other things) are definitely doable on GPU, but many of them require shaders and cannot be done in the fixed-function blend pipeline.
2. There are a bunch of optimizations that are still very useful on GPU, including reordering to improve batching of draw calls, and techniques to reduce overdraw when an operation will be completely painted over. Webrender uses hierarchical z-buffer to GPU-accelerate these.
3. There are advanced things like linear-sRGB colorspaces that are very useful, well supported on modern GPU hardware.
4. CoreGraphics on MacOS is pretty slow. On Windows, I've had varying results with Direct2D (I know we chatted some in a previous thread). I think it depends a lot on the quality of drivers.
4. Incremental repaint (as opposed to painting the surface in full) is still useful, if for no other reason than to reduce energy use. It's difficult on some platforms (Mac), well supported on others (Windows), and optional/driver dependent on others (Linux/Android).
5. Clipping is still where much of the performance dies. Sure, you _can_ specify everything in terms of bitmap blending, but do you want to? That's an awful lot of gmem bandwidth you're churning.
And yes, text is fantastically complicated.
Long story short, I don't see any good one-size-fits-all solution here.