I shouldn't have removed this one.

maybe it is cheaper to test if output is `+Inf` or subnormal/zero?

GitHub won't let me comment on collapsed unedited lines, but it is currently assumed that `FE_INEXACT` will be raised if required when `DyadicFloat::operator T()` returns, by these intermediate calculations: https://github.com/llvm/llvm-project/blob/44fb4248d6c63630a3e3f398c1edb086c52a32e6/libc/src/__support/FPUtil/dyadic_float.h#L168-L169 https://github.com/llvm/llvm-project/blob/44fb4248d6c63630a3e3f398c1edb086c52a32e6/libc/src/__support/FPUtil/dyadic_float.h#L176-L177

I think in this case `errno` should also be set to `EDOM` if required.

Should raise `FE_INEXACT` if nonzero bits were shifted out of the mantissa.

Should raise `FE_INEXACT` if non-zero bits were shifted out of the mantissa.

- Should set `ERANGE` if required. - Should raise `FE_UNDERFLOW` and `FE_INEXACT`. - Should use `result_sign` and handle FE_UPWARD and FE_DOWNWARD appropriately.

- Should set `ERANGE` if required. - Should raise `FE_OVERFLOW` and `FE_INEXACT`. - Should use `result_sign` and handle `FE_UPWARD` and `FE_DOWNWARD` appropriately.

Nit: should probably use `using OutType = ...;` instead in C++.

Nit: should probably use `using OutType = ...;` instead in C++.

Not sure why I copy-pasted this.

The `& OutFPBits::SIG_MASK` part is redundant, `FPBits::create_value` does it already.

I think you only need `OutFPBits::FRACTION_LEN + 1` iterations, then the rounding logic could be simpler.

Waiting for #95159 to be merged for this.