stillwater-sc / universal / 29657033445

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(efloat): parse() high precision without silent garbage (+ BigBits budget) (#1144)

* fix(efloat): parse() high precision without silent garbage; add BigBits budget

parse() silently returned ~0 once the requested precision approached the d2b
working budget (issue #1141). Root cause: parse capped target_bits at
default_big_bits-8 (2040) and called convert<2048>, but convert's reduction
left-shifts the digit-integer by ~(target + 3*neg_E) bits (neg_E = fractional
places + any negative scientific exponent). For a target anywhere near 2048 that
shift overflows the BigBits integer and convert returns garbage -- so the 2040
cap was itself unsafe, and even a short literal like pi produced ~0 at high
precision.

Two changes:
- Template parse on BigBits (default decimal_to_binary::default_big_bits), so
  parse(s,x) is unchanged and callers can request a larger exact budget, e.g.
  parse<16384>(s,x) to represent ~1000-digit constants. The round-up bigint now
  honors BigBits (was hardcoded to the default).
- Size target_bits overflow-safely from the scanned literal: compute the
  effective decimal exponent and significant-digit count, and cap target so the
  convert shift stays within BigBits. A request beyond the safe ceiling is now
  correctly rounded to that ceiling instead of returning garbage.

Forward declaration in efloat_fwd.hpp carries the BigBits default (it is what
efloat::assign() sees); efloat_fwd.hpp now includes decimal_to_binary.hpp for
default_big_bits.

Verified: default parse of short/long pi at precision 256..3400 now returns the
correct value (was ~0 at >=2040); parse<16384> of a 1000-digit pi matches
efloat_pi<128>() to ~1000 digits, and parse<16384>("0.1")*10 == 1 to ~3300 bits.
New test elastic/efloat/conversion/parse_precision.cpp; all 22 existing efloat
regression tests pass on gcc 13.3.0 and clang 18.1.3.

Resolves #1141

* fix(efloat): budget parse() for positive exponents too (CodeRabbit #1144)

The overflow-safe target ceiling ... (continued)

Coverage Stats

18 of 18 new or added lines in 1 file covered. (100.0%)

10 existing lines in 2 files now uncovered.

49476 of 58413 relevant lines covered (84.7%)

6874568.46 hits per line