stillwater-sc / universal / 30672136635

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(posit): clear -Wsign-conversion in posit_impl convert_ (#1281) (#1286)

Sub-issue of Epic #1279. Four sites in the posit encode/rounding path
(convert_), all value-preserving:

- `exponent = e % (1ull << es)`: the signed scale `e` was already promoted
  to unsigned long long by the modulo (the other operand is 1ull); make it
  explicit. This is correct for a power-of-two modulus -- 2^64 mod 2^es is
  0, so the unsigned reinterpretation of a two's-complement negative e
  yields exactly the low es bits, which is the posit exponent field (and
  matches the arithmetic-shift regime split e >> es).
- `regime <<= es + nrFbits + 1u`, `exponent <<= nrFbits + 1u`,
  `pt_bits <<= pt_len - len`: unsigned shift-amount / left-align distance
  flowing into operator<<=(int); each is >= 1 and small by construction.
  Cast to int explicitly.

Proven behavior-neutral: optimized assembly (g++ -O2) for a TU
instantiating convert_ across posit<8,0>/<16,1>/<32,2>/<64,3> is
byte-for-byte identical before and after -- the casts are explicit forms
of conversions the compiler already performed implicitly.

The dead `convert_to_bb` (inside `#ifdef TBD`, never called) has the same
idiom but does not compile, so it is intentionally left untouched.

Verified posit_impl -Wsign-conversion AND -Wconversion clean on gcc AND
clang; posit conversion/assignment/multiplication regressions PASS on
both. (The posit<64,2..4> division failures are pre-existing on main --
identical counts with the fix reverted -- and out of scope here.)

Relates to #1279
Resolves #1281

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

Coverage Stats

4 of 4 new or added lines in 1 file covered. (100.0%)

36 existing lines in 4 files now uncovered.

41708 of 48936 relevant lines covered (85.23%)

7271799.74 hits per line