stillwater-sc / universal / 26587083869

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(cfloat): native trunc/floor/ceil/round (correct for >53-bit significands) (#1028)

trunc/round/floor/ceil computed `std::X(double(x))`, which is wrong for cfloats
with more than 53 significand bits: any integer above 2^53 loses its low bits in
the double conversion, so e.g. floor(2^60+8) did not return 2^60+8 (#1026).

Now implemented directly on the cfloat encoding:
- trunc: clear the fraction bits below 2^0 per the unbiased scale (scale>=fbits ->
  integral; subnormal or scale<0 -> signed zero). Pure bit op, no arithmetic.
- floor/ceil: trunc, then -/+ 1 for the negative/positive non-integral case.
- round: ties away from zero, decided from the 2^-1 fraction bit for |x|>=1 and
  from 2|x|>=1 for |x|<1 -- never constructing 0.5, which some low-range cfloats
  (e.g. cfloat<8,2>) cannot represent.

Test (static/float/cfloat/math/truncate.cpp): added exhaustive trunc/round
coverage (was floor/ceil only), a wide-precision verifier on cfloat<128,15>
(2^60+8 +/- exact fractions, both signs, vs cfloat-constructed references -- no
double), and fixed the result accumulation (`=` -> `+=`).

Verified gcc + clang: exhaustive cfloat<8,2>/half/bfloat_t match the
representable std reference; wide cfloat<128,15> exact; cfloat_fractional (fmod)
still passes.

Resolves #1026

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

Coverage Stats

33 of 33 new or added lines in 1 file covered. (100.0%)

2 existing lines in 1 file now uncovered.

46923 of 55857 relevant lines covered (84.01%)

5780980.48 hits per line