stillwater-sc / universal / 34785019544

86%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(dfloat): parse counts significant digits, not digit characters (#1489)

dfloat::assign(), which parse() and operator>> use, kept the first ndigits digit
characters of the input as the significand and got the scale wrong three ways:

- Leading zeros used up the precision. decimal32 "0.0001234567" read as 1.23e-8,
  "0.000001234567" as 1e-12, and "0000000123" as 0.
- A digit past the precision after the decimal point still moved the exponent,
  although it was never stored. decimal32 "1.23456789" read as 0.01234567, and
  "12345678.9" as 1234567. decimal64 "3.14159265358979323846" read as
  3.141592653589793e-5.
- The exponent digits accumulated in an int with no bound, so an exponent of ten or
  more digits was signed overflow (UBSan: "999999999 * 10 cannot be represented in
  type 'int'").

The digit loop now keeps value = sig * 10^(decimal_exponent - frac_digits), where
frac_digits counts only the fraction digits appended to sig. Leading zeros take no
precision, and after the point they only set the scale. Digits past the precision
are dropped (truncation, as dfloat rounds everywhere else, #1487): a dropped integer
digit scales the value up, a dropped fraction digit changes nothing. The exponent
saturates at 1000000, beyond every dfloat's range, so such inputs give inf or 0.

New test static/float/dfloat/conversion/parse_digits.cpp (dfloat_parse_digits):
- Random digit strings of 1 to 3*ndigits digits, each written six ways: plain
  positional, with leading integer zeros, with trailing fraction zeros, with the
  point shifted against an explicit exponent, scientific, and integer significand
  with an exponent. Every form must parse to the value truncated to ndigits,
  computed on the digit strings, and parse() must accept it and agree with
  assign(). This runs for dfloat<4>, decimal32 and decimal64 in BID and DPD, and
  decimal128 and dfloat<10> at higher levels.
- The positional round trip of 1.234567e-k for k = 0 to 20, which #1485 proposed.
-... (continued)

Coverage Stats

10 of 10 new or added lines in 1 file covered. (100.0%)

18 existing lines in 2 files now uncovered.

43591 of 50868 relevant lines covered (85.69%)

7486097.6 hits per line