No, you are just not understanding the tradeoff. They fixed some canvas not being accelerated when they should have been. This means they can be hardware composited, drawn to and need to only exist once as a GPU texture. This is a massive win for performance and power consumption.

Here is what is not efficient, ever: getting a GPU texture into a raster format CPU accessible buffer. At best this requires synchronization and a full memory copy into a new buffer, but it can easily be much much more expensive because GPUs, and mobile GPUs basically never, store textures in raster formats. So now you need a full copy and a bunch of extra SIMD to convert it from the GPU preferred format into raster format for your getImageData call.

And all this for what is essentially a trap API! Chances are if you are calling getImageData regularly, you are doing something wrong and inefficient.