https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88074
--- Comment #24 from rguenther at suse dot de <rguenther at suse dot de> --- On Wed, 21 Nov 2018, joseph at codesourcery dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88074 > > --- Comment #23 from joseph at codesourcery dot com <joseph at codesourcery > dot com> --- > And, yes, at least one extra bit in emin is needed for that sticky > rounding code to work (because the user's source code may have a decimal > constant that is slightly over half the least subnormal, and mpfr_strtofr > needs to return a nonzero value for that code to work, meaning a value > whose exponent is one smaller than the smallest exponent of any subnormal > value that is actually representable in the floating-point format). OK, I see. Adding yet another extra bit, this time to emax, doesn't solve the Fortran norm2_3.f90 regression though.
