Comment by rav

You could use handwritten asmjs in that case. I did it at work to improve performance on some pixel-wise software rendering we were doing. (We have since then switched to WebGL, but asmjs was a nice stepping stone.) If I recall correctly, it basically boils down to using small fixed size arrays and appending "|0" after each uint32 operation and "+" in front of each float32 operation - so it's something you can do by hand. Of course, one of my colleagues then made a tweak which gave a surprisingly huge slowdown - you should probably declare loudly (with comments etc.) that a certain code block is intended to be asmjs if you do this in a large codebase.