Towards the end, when the author tries to replace the inner branch with arithmetic ("Replacing the branch with arithmetic"), one thing I noticed is that the bra

Towards the end, when the author tries to replace the inner branch with arithmetic ("Replacing the branch with arithmetic"), one thing I noticed is that the branchless code actually reduced branches by 5x the number of eliminated branch mispredictions (233k vs 43k). That reads to me that the processor was actually doing a pretty good job predicting the branches for large inputs (toward the end of the sort, when values are close together?), not the lack of CMOV that the author asserts is at fault. edit: that said, I don't think that's the full explanation why that's slower, given that the small inputs actually do okay compared to the overlined (although most of that is recouping a bit of the losses against the Jedi baseline).