[libcxx-commits] [libcxx] [libc++] Simplify the implementation of std::sort a bit (PR #104902)

[Daniel Kutenin via libcxx-commits]

danlark1 wrote:

We can but I would recommend enabling previous behavior first and then we can refine the threshold. Also in practice, microbenchmarks can lie as the memory is hot and even for low hundred elements we can see improvements because we removed branches.

For example, when the first version with the branch free partitioning was rolled out at Google, we saw 10-15% improvement fleetwide even though microbenchmarks were saying much more.

Unfortunately, we missed the regression from the last year by accident. :(

https://github.com/llvm/llvm-project/pull/104902