https://github.com/davemgreen commented: Would this produce double-rounding errors? The 64bit code is careful to use FCVTXN to round to odd, avoiding the problems with introducing two rounding steps. https://github.com/llvm/llvm-project/pull/179044