As for other floating-point rounding builtins that can be optimized
when build with -fno-math-errno, this patch adds support for lrint
and llrint. It currently only optimize for AArch64 backend.
This patch depends on https://reviews.llvm.org/D62017.