joker-eph wrote: > change should really only affect the performance of very large splat constants. So, I'm not sure how this should be tested. In this case we tag the PR with [NFC] (no functional change) to convey this. https://github.com/llvm/llvm-project/pull/115229