bpo-46407: Optimizing power-of-2 operations

[thatbirdguythatuknownot]

https://bugs.python.org/issue46407

[sweeneyde]

Come to think of it, this is kind of two separate orthogonal changes, right? It may be easier to benchmark, review and get parts merged if this was two PRS, one to implement the 1 << x and 2**x changes, and the other to separate l_divmod from l_mod

[thatbirdguythatuknownot]

Come to think of it, this is kind of two separate orthogonal changes, right? It may be easier to benchmark, review and get parts merged if this was two PRS, one to implement the 1 << x and 2**x changes, and the other to separate l_divmod from l_mod

How should I separate it? Should I get rid of the l_divmod/l_mod changes in this PR, change the title, then open a new PR with the l_divmod/l_mod changes?

[sweeneyde]

Come to think of it, this is kind of two separate orthogonal changes, right? It may be easier to benchmark, review and get parts merged if this was two PRS, one to implement the 1 << x and 2**x changes, and the other to separate l_divmod from l_mod

How should I separate it? Should I get rid of the l_divmod/l_mod changes in this PR, change the title, then open a new PR with the l_divmod/l_mod changes?

Sounds good to me

[thatbirdguythatuknownot]

It seems like this PR has lost all activity.

[mdickinson]

I'm unassigning myself from review. I think @tim-one would be better placed to decide whether this is an optimisation worth making, and if so review the code.

[thatbirdguythatuknownot]

PR bump.

[thatbirdguythatuknownot]

PR bump 2.

[thatbirdguythatuknownot]

PR bump 3.

[thatbirdguythatuknownot]

PR has lost all activity.

[tim-one]

I appreciate the effort, but I'm inclined to reject this change. The increase in code complexity isn't justified by the relatively small (~6%) speedup the bpo report showed for a best-case isolated microbenchmark (x**10000000 after x=2). There are few cases the patch can help with, but testing for those cases slightly slows down all other cases (tests-and-branches to skip the new code), and also blows space in I cache for code that won't be executed most times.

long_pow() in particular is already a long-winded function without this, where existing optimizations have major payoffs that apply in many cases. The generality and high bang-for-the-buck nature of those justify their code complexity.

The other change - introducing remainder-only functions - was much more attractive on these grounds (no runtime tests needed - we pick which we want in each context at compile-time, and they're always faster when they're picked).

[thatbirdguythatuknownot]

I see.