Parallels probably work by decompiling DirectX shader bytecode back into GLSL, which is already optimized for OS X shader compiler, therefore there is less burden on the optimizer.
Another example of poor drivers are various GLES drivers on mobile, mostly trivial optimizations can drastically increase the performance: http://aras-p.info/blog/2010/09/29/glsl-optimizer/ and https://github.com/aras-p/glsl-optimizer
Are you sure? I'd think it isn't passing function calls, but rather direct hardware access. I might be wrong.
Either way, there's clearly a way to make the game run significantly faster on OSX, but Valve isn't doing it for some reason.
They use the platform graphics APIs like any other app.