https://github.com/arsenm commented: I think the code change looks fine, but the AArch64 and RISCV cases look like regressions. Is there an explanation for those? Is the type transform loop excessively widening these now? https://github.com/llvm/llvm-project/pull/96151