stillwater-sc / universal / 26001208466

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(hfloat): implement parse() for decimal string input (#849) (#859)

* feat(hfloat): implement parse() for decimal string input (Resolves #849)

hfloat::parse was a stub that returned false unconditionally. This
implements it via the route established in Phase B2/C/D of #835:
  decimal_to_binary::convert -> distill<1> -> double -> convert_ieee754

That gives bit-exact correctly-rounded conversion for the hfloat
configurations whose precision fits in a double (24 fbits for
hfloat_short, 56 fbits for hfloat_long). Wider hfloat<28,*> sees the
same precision loss the existing operator=(double) path already has;
upgrading that to a wider distill<N> is a separate enhancement.

The hfloat encoding (truncation rounding, overflow saturation to
maxpos/maxneg, no NaN/Inf) is preserved by routing through the
existing convert_ieee754 entry point.

Special-value handling:
- nan / inf / infinity tokens REJECTED (return false). hfloat has no
  NaN or Inf encoding; silently mapping these would be wrong.
- Empty string, "+", "-", ".", "e10" REJECTED (no-digit guard).
- Trailing junk REJECTED (1.5abc, 1.5.0, 1e, 1ea).

operator>> hygiene:
- Extraction guard: `if (!(istr >> txt)) return istr;`
- setstate(std::ios::failbit) on parse failure (in addition to the
  existing cerr diagnostic).

Tests (static/float/hfloat/conversion/string_parse.cpp):
- hfloat_short and hfloat_long: canonical decimals match constructor
- nan / inf / infinity tokens (all 11 spellings) rejected
- Overflow saturation: parse("1e100") -> maxpos, parse("-1e100") -> maxneg
- Malformed input rejection (11 cases)
- operator>> failbit on bad token + on EOF

Verified on gcc (build_ci) and clang (build_ci_clang): hfloat_string_parse
PASS, hfloat_api regression PASS.

Resolves #849

* fix(hfloat): unify mantissa->hex path, address user + CodeRabbit round-1 review on PR #859

Three issues addressed:

1. The original PR description claimed "hfloat_long (56 fbits) -- fully
   exact (well within double's 53)... (continued)

Coverage Stats

85 of 88 new or added lines in 1 file covered. (96.59%)

19 existing lines in 4 files now uncovered.

46682 of 55519 relevant lines covered (84.08%)

6423600.74 hits per line