cor3ntin wrote: My feeling is that it's not worth the complexity. Getting an accurate benchmak would be difficult, it would only manifest in extremely templated code I wonder if @AaronBallman has opinion. https://github.com/llvm/llvm-project/pull/89494