paperchalice wrote: Currently I couldn't fix the test `CodeGen/ARM/fp16.ll`, because `llvm.convert.to.fp16.f64` in this test is incompatible with fast math flags. https://github.com/llvm/llvm-project/pull/146316