[llvm] [VectorCombine] Fold insert(binop(x,y),binop(a,b),idx) --> binop(insert(x,a,idx),insert(x,a,idx)) (PR #124909)

[David Green via llvm-commits]

https://github.com/davemgreen approved this pull request.

AArch64 unfortunately has some odd vector insert costs for lane zero of a integer vector, which it treats as free. It makes sense for float vectors but when we try to remove it for integers it causes perf issues.

I ran some tests and it didn't seem to be an issue though, LGTM.

https://github.com/llvm/llvm-project/pull/124909