Yes, I agree as well.

I agree that this a strict improvement.

It sounds like this is a profitability issue and therefore the LAA users should handle this? Maybe the users would also want to be more general (I think a loop with an iteration count of 2 is probably not worth vectorizing).

Thanks for the reduced testcase!

Thanks! LGTM

Thanks, this all makes sense to me. However, I think there is still an issue with the default
AArch64 target.

Hi Gerolf,

Looks like I spoke too soon (new comment inlined). I think there are some issues with targets without a machine model.

Ah, I just noticed that a header file is missing (AArch64MachineCombinerPattern.h).

Just some minor issues that I've noticed (comments inlined). Otherwise looks good.

It seems to me that the branch mispredict cost for the case where the values of X are random would outweigh the benefits of this transformation for your alternative code sequence, even on OoO cores.

Hi Gerolf,