Currently, TorchDynamo calls PyFrame_FastToLocalsWithError (docs).

Internally in cpython, this converts between the "frame fast variables" (a flat C array of PyObject*) and frame->f_locals (a python dictionary).

frame->f_locals starts as NULL, and is only populated by this function.

This works, but is higher overhead than it needs to be.

We should rewrite the generated guards to use the "fast variables" data structure directly.

cc @ezyang @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @msaroufim @zou3519 @aakhundov @soumith @wconstab @ngimel @yanboliang @Xia-Weiwen @desertfire