adrian-prantl wrote: So I reverted back to the state where I convert everything to int32_t. This way I don't have to touch the UINT32_MAX sentinels in this patch, which should be done separately and very carefully. https://github.com/llvm/llvm-project/pull/83501