Matt Arsenault via llvm-dev
2017-Apr-12 18:05 UTC
[llvm-dev] Constant folding some canonicalize intrinsics
It isn't OK to always fold it since the exact result is platform dependent, but is there some subset of cases where it is OK? I would hope cases like constant 0 are universally 0. Would it be OK to generically constant fold canonicalize intrinsics with an input that is an entirely normal number? By entirely normal I mean not subnormal and not inf/nan. These currently block constant folding some OpenCL builtin library functions, but in practice the constant inputs that appear aren't subnormals that may require flushing. -Matt