stillwater-sc / universal / 29703104434

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(efloat): implement long double conversion (was yielding 0) (#1161)

convert_ieee754 only builds mantissa limbs for sizeof(Real) of 4 or 8; the else
branch was a no-op static_assert(true), so a wide long double (sizeof 16 on
x86-64) pushed no limbs and normalized to 0. Both the constructor/assignment
(input) and operator long double() (output) were affected -- input silently
produced 0, output capped at double precision.

Fix (both directions, portable, exact):
- operator=(long double): frexp into m * 2^e with m in [0.5,1) (inside double's
  range), decompose m into an exact sum of doubles via the working double path,
  then apply e with setexponent(). Preserves the full significand AND the
  extended exponent range.
- operator long double(): the dual -- bring |x| into [1,2), extract as an exact
  sum of doubles, scale by the true exponent with ldexp.
- Platforms where long double == double take a direct convert_ieee754(double)
  path (if constexpr).

Verified round-trip-exact across normal values, beyond-double precision
(1 + 2^-60 survives), the extended exponent range (2^1100, 2^-1100, 2^16000),
and special values (+/-0, +/-inf, NaN), on gcc and clang. New regression test
elastic/efloat/conversion/long_double.cpp (wide-only checks gated on
sizeof(long double) > sizeof(double)).

This also removes the need for the nexttoward long double -> double workaround
in math/next.hpp (#1120); left as a follow-up.

Resolves #1160

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Coverage Stats

39 of 40 new or added lines in 1 file covered. (97.5%)

3 existing lines in 1 file now uncovered.

49760 of 58697 relevant lines covered (84.77%)

7033938.62 hits per line