https://github.com/paulwalker-arm commented: There's a lot of noise in the test output. Perhaps pass the three 64/128-bit vectors in separately and use llvm.vector.insert to join them? https://github.com/llvm/llvm-project/pull/192972