stillwater-sc / universal / 24863021025

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(posit): enable constexpr integer construction (Phase 1 of #713) (#714)

* feat(posit): enable constexpr integer construction (Phase 1 of #713)

Adds true constexpr integer-to-posit conversion via uint64_t shift/mask,
bypassing the previous double-cast route through convert_ieee754 (which is
not constexpr because std::frexp is not constexpr until C++26).

Changes:
- promote check_inward_projection_range to constexpr (pure integer fn)
- add convert_unsigned_integer / convert_signed_integer that build the
  posit encoding directly in a uint64_t accumulator with round-to-nearest-
  even on discarded fraction bits; saturate to maxpos/maxneg via the
  existing check_inward_projection_range
- reroute the integer operator= overloads (signed char ... unsigned long
  long) to the new functions; for nbits > 64 they fall back to the
  existing IEEE-754 path
- negation for signed inputs is performed as two's complement on a
  uint64_t (mask + ~x + 1), avoiding the non-constexpr blockbinary
  arithmetic chain
- add static_assert-style smoke tests in static/tapered/posit/api/api.cpp
  that verify constexpr-constructed posits match the runtime encoding
  across nbits in {8, 16, 32, 64} including saturation and INT_MIN

Side effect: integer-to-posit for wide inputs (e.g., int64 into
posit<64,3>) is now bit-exact rather than going through a lossy
double-cast intermediate. The existing regression suite is robust to
this because VerifyIntegerConversion only enumerates posit-representable
integers.

Phases 2 and 3 (constexpr float/double construction, blockbinary
constexpr promotion) are tracked separately.

Resolves #713

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

* fix(posit): char signedness dispatch, independent test reference (CodeRabbit)

Two issues from CodeRabbit on PR #714:

1. operator=(char) hardcoded an unsigned cast, so on signed-char platforms
   (the common case) char(-3) was reinterpreted as 253 and produced the
   wrong pos... (continued)

Coverage Stats

67 of 79 new or added lines in 1 file covered. (84.81%)

2 existing lines in 1 file now uncovered.

44908 of 53215 relevant lines covered (84.39%)

6512805.23 hits per line