stillwater-sc / universal / 29750128328

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(math): add fma, scalbn, logb, ilogb to efloat and ereal (#1171)

* feat(efloat+ereal): add fma, scalbn, logb, ilogb (both oracles)

The last <cmath> functions missing from both adaptive-precision types, added to
each type's numerics.hpp on native arithmetic:

- scalbn(x, n) = x * 2^n -- delegates to ldexp (exact binary-exponent shift).
- logb(x)  = floor(log2(|x|)) as a float; ilogb(x) = same as int with the
  <cmath> special values (ilogb(0)=FP_ILOGB0, ilogb(inf)=INT_MAX,
  ilogb(nan)=FP_ILOGBNAN; logb(0)=-inf, logb(inf)=+inf, logb(nan)=nan).
  * efloat: exact from scale() (the leading-bit exponent).
  * ereal: from the leading component's std::ilogb, corrected by +/-1 against the
    full magnitude (lower components can pull |x| across a power-of-two boundary).
- fma(x, y, z) = x*y + z. Both types multiply exactly at their working precision,
  so no intermediate rounding is introduced; 0*inf yields NaN through the type's
  own multiply, matching IEEE.

Verified against std:: on gcc and clang for both types, including logb binade
boundaries and the ilogb/logb special values. Tests added to the existing
efloat/ereal numerics suites; ereal mathlib.hpp location note updated.
cppcheck-clean.

Resolves #1166
Relates to #582

* test(math): make fma checks non-tautological + drop C compound literals (CodeRabbit #1171)

- fma tests compared fma(a,b,c) against a*b+c -- the same expression, so they
  could not detect intermediate rounding. Now verify exactness against an
  independently built value: (2^30+1)*(2^30-1) = 2^60-1, which is exact in the
  oracle but rounds to 2^60 in double, with the expected 2^60-1 constructed via
  scalbn/subtract (not the multiply). A rounding fma would fail. (efloat + ereal)
- ereal test used C99 compound literals `(const double[]){...}` (a GCC/Clang
  extension, ill-formed ISO C++); replaced by the direct construction above.
- ereal/mathlib.hpp remaining-work tracker no longer lists the now-implemented
  #1165/#1166 <cmat... (continued)

Coverage Stats

43 of 49 new or added lines in 2 files covered. (87.76%)

7 existing lines in 3 files now uncovered.

49845 of 58793 relevant lines covered (84.78%)

7013644.86 hits per line