http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21718
--- Comment #29 from Joseph S. Myers <jsm28 at gcc dot gnu.org> --- GCC supports lots of different floating-point formats, not all IEEE, including variants such as "no subnormals" or "no infinities". Using round to zero/stickiness allows the existing GCC code that knows about the peculiarities of these formats to do the final rounding for them, instead of needing to replicate that logic specially for string conversions in order to use mpfr_subnormalize. Simple use of round-to-nearest without setting exponent limits and mpfr_subnormalize would indeed result in problems with double rounding. d1c is slightly above half the least subnormal (d1a is slightly below, d1b is exactly half the least subnormal), so should give the least subnormal as result, which is what I get with current glibc. Note that if GCC is running with an MPFR version before 3.1.1p2, it's possible MPFR bugs will cause incorrect results (I don't know if the bug in question affects the way GCC uses MPFR). (On 32-bit i686-pc-linux-gnu you'd have issues with excess precision meaning constants are first interpreted as long double then later rounded to their semantic type, hence the FLT_EVAL_METHOD conditionals.)
