https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121570
--- Comment #11 from Francois-Xavier Coudert <fxcoudert at gcc dot gnu.org> --- The IEEE intrinsics where we currently save and restore the FP state are: - IEEE_LOGB_RINT - IEEE_REM - IEEE_NEXT_AFTER The intrinsics where we currently don't save and restore the FP state, because I somehow convinced myself it was not necessary: - IEEE_IS_NAN (calling the GCC __builtin_isnan) - IEEE_IS_FINITE (calling the GCC __builtin_isfinite) - IEEE_IS_UNORDERED (calling the GCC __builtin_isunordered) - IEEE_SIGNBIT (__builtin_signbit) - IEEE_IS_NORMAL (__builtin_isnormal) - IEEE_IS_NEGATIVE (__builtin_isnegative) - IEEE_SCALB (__builtin_scalbn) - IEEE_COPY_SIGN (__builtin_copysign) - IEEE_CLASS (__builtin_fpclassify, __builtin_issignaling, __builtin_signbit) - IEEE_VALUE - IEEE_FMA (__builtin_fma) - IEEE_MIN_NUM, IEEE_MAX_NUM, and their _MAG variants (__builtin_isgreater, __builtin_isless, __builtin_fmax, __builtin_fmin) - all IEEE_QUIET_ and IEEE_SIGNALING_ comparison intrinsics (calling __builtin_iseqsig, __builtin_isgreaterequal, __builtin_isgreater, __builtin_islessequal, __builtin_isless) All that code is in trans-intrinsic.cc, in the functions gfc_conv_ieee_*. See which of these functions call gfc_save_fp_state and gfc_restore_fp_state, and which don't. Of course, I would love to have help in reviewing these past decisions!
