On Thu, Oct 23, 2025 at 08:50:22AM +0200, Stefan Schulze Frielinghaus via Gcc wrote: [...] > My understanding is that for > -fexcess-precision=16 it is sound having targetm.c.excess_precision > (EXCESS_PRECISION_TYPE_FLOAT16) returning > FLT_EVAL_METHOD_PROMOTE_TO_FLOAT---which would align with the default > implementation---or FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 but I might be > wrong here.
To clarify: this statement stems from the fact that with and without -fexcess-precision=16, FLT_EVAL_METHOD equals zero.
