stillwater-sc / universal / 26346900378

85%

master: 84%

push

github

fix(ereal): IEEE 754 special values for multiplication (#966) (#967)

* fix(ereal): fixing the fuzz iterations

* fix(ereal): IEEE 754 special values for multiplication (#966)

ereal multiplication routed straight into expansion_product, which has no
special-value handling: finite * Inf produced Inf - Inf cross terms that
renormalized to NaN, Inf * Inf collapsed to NaN, and any zero operand was
short-circuited to a positive +0 limb -- so 2 * Inf gave NaN, Inf * Inf
gave NaN, and every signed-zero product (e.g. -0 * +0, 3 * -0) lost its
sign.

Add apply_ieee754_mul_special_values(), a sibling to the addition handler,
and apply it in operator*= before the EFT product. The sign of a product is
signbit(a) XOR signbit(b), including for zero and infinite results:

  NaN * x                 = NaN
  Inf * 0 (either order)  = NaN
  Inf * Inf               = Inf, sign = a ^ b
  Inf * finite-nonzero    = Inf, sign = a ^ b
  0 * finite / finite * 0 = 0,   sign = a ^ b

operator*=(double) now delegates to the ereal overload (matching
operator-=(double)) so the special-value table applies uniformly; the free
operator* overloads already constructed an ereal for the scalar, so in-place
`*= scalar` is now consistent with `x = x * scalar`. The signed-zero result
projects correctly to double via the convert_to_ieee754 seeding fixed in
#964.

The IEEE 754 special-value test block in multiplication.cpp (deferred by the
#946 refactor) is now exercised. Verified foundational + fuzz pass on gcc-13
and clang-18; addition and subtraction unchanged.

Resolves #966

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

22 of 22 new or added lines in 1 file covered. (100.0%)

47303 of 56328 relevant lines covered (83.98%)

5272263.54 hits per line