perf: pull set construction out of function #1012
Conversation
Signed-off-by: Henry Schreiner <henryfs@princeton.edu>
|
I guess this makes sense but I do wonder if it switching from a set to an in-line tuple might be faster for such a small set. Looking up a global value is slower than looking up a local value, constructing a tuple is very fast, doing x in (a, b, c) does a fast path of checking the Python object ids first, strings are often interned in CPythin, and the O(1) look up of a set might have a real time slower lookup than a tuple of a handful of items. |
|
To get the full fast path, I believe it does have to be fully inlined. If it is, then it doesn’t matter if it’s a set or a tuple, it produces the same byte code. I was assuming this wasn’t all that hot, and this is a common way to do this sort of thing. |
Is might, but I personally would rather we didn't to retain the semantics that there is no preferred order. |
Based on KRRT7#3; codeflash-ai thinks it can speed up this function by 564% (single function benchmark), but I think most of the speed up comes from avoiding constructing a set in every function call, and I think this is more standard of a structure too.