[llvm] [CodeGen][AArch64] Set min jump table entries to 13 for AArch64 targets (PR #71166)

[via llvm-commits]

Wilco1 wrote:

> I think it's actually noise, don't think I see improvements on the V2. Perhaps that's still okay because it is shows improvements on the V1. But now I am actually wondering if the V1 improvement is an actual improvement, or do we get lucky with secondary effects such as different function/loop/block alignments?

Check the instruction count, cache and branch misses. Much of the gain comes from a huge reduction in instruction count, the rest from avoiding cache misses from loading the table entries and indirect branch misses. As indirect branch prediction improves, a lower value may be slightly better in future cores, however I don't believe it ever makes sense to use a table with less than 8 cases.

https://github.com/llvm/llvm-project/pull/71166