Currently CanvasRenderingContext2D keeps its own DrawTarget which it draws into and copies the whole thing in a TextureClient after each frame to share it with the Compositor. We should do the following things: * draw directly into a TextureClient * share this TextureClient with the compositor without copying it at the end of the frame * use a new TextureClient for each frame and only copy from the previous TextureClient if the first drawing operation does not overwrite all of the canvas' pixels Eventually use a pool to reduce the overhead of creating and destroying textures each frame, but we already pay that cost so we can track the pool as an orthogonal improvement (could be done in bug 1008221).