https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79487
--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Dominik Vogt from comment #4) > What happens in Cse1 is that the constant is propagated into the > FLOAT_EXTEND expression, resulting in > > (float_expand:DD (const_double:SD -9223372036854775808)) > > which is eventually simplified using simplify_const_unary_operation. This > has a case dealing with FLOAT_EXTEND: > > case FLOAT_EXTEND: > /* All this does is change the mode, unless changing > mode class. */ > /* Don't perform the operation if flag_signaling_nans is on > and the operand is a signaling NaN. */ > if ((GET_MODE_CLASS (mode) != GET_MODE_CLASS (GET_MODE (op)) > || d.decimal) > && !(HONOR_SNANS (mode) && REAL_VALUE_ISSIGNALING_NAN (d))) > real_convert (&d, mode, &d); > break; > > This test fails, so the following code just strips the FLOAT_EXTEND:DD and > replaces the mode in const_double with the target mode: > > (const_double:DD -9223372036854775808) > > which is wrong. Shouldn't this truncate the constant to the source mode > first before converting it to the target mode? This isn't truncation, but extension (SDmode to DDmode). I presume all SDmode values are representable in DDmode, so I don't see anything wrong on that.
