• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

stillwater-sc / universal / 40888380477
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran 29 Oct 2026 02:27PM UTC
Jobs 1
Files 768
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

29 Oct 2026 02:07PM UTC coverage: 88.251%. Remained the same
40888380477

push

github

web-flow
fix(verification): the random-operand verifiers round their double reference once (#1862)

VerifyBinaryOperatorThroughRandoms, in both the generic and posit test
suites, computed its reference as TestType(da op db): one rounding to
double, a second to TestType. For a type with 27 <= p <= 52 the first can
land exactly on a TestType midpoint the exact result is not on, and the
second then breaks that false tie the wrong way. cfloat<48,8> addition
failed on 1.2e-5 of random pairs, every time with the cfloat result right
and the reference 1 ulp off; it surfaced as a one-off cfloat_fff_addition
failure on macOS (#1859).

correctly_rounded_reference.hpp computes the sign of exact - reference
exactly: TwoSum for add and sub, TwoProduct and the division residual for
mul and div, with the operands scaled to [1,2) so no error term underflows
when the result is subnormal. When the reference is exactly the midpoint of
its two TestType neighbours and that sign is nonzero, it takes the
neighbour on that side. Types with p >= 53 (c64, dd, qd, posit<64,x>) keep
the converted double.

False failures over 1M random pairs for each of the 8 operators, old
reference against new:
  cfloat<48,8> fff/ttf/fft  211 / 215 / 223, now 0
  cfloat<56,11> tff         6234, now 0
  posit<40,2>               5, now 0
  posit<48,2> / <48,3>      2515 / 1650, now 0
  c16, c32, c40, c64, bfloat16, posit<24..32>: 0 before and after
and 0 at 4M on gcc and clang.

That retires the "double rounding errors in the test bench" workarounds in
the cfloat subsuper suites: c48 randoms run in addition, multiplication and
both divisions, and c32 in the divisions. c64 stays off there, now with
the real reason: its max-exponent binade lies beyond double, so the
reference overflows to inf where c64 is finite (#1863).

Also adds verification to the PR-title scope allow-list: the verifiers
serve every number system, so a fix there had no scope of its own.

Resolves #1861

Co-Authored-By: Claude Opus 5.5 <... (continued)

47 of 50 new or added lines in 2 files covered. (94.0%)

3 existing lines in 2 files now uncovered.

55484 of 62871 relevant lines covered (88.25%)

8777523.72 hits per line

Uncovered Changes

Lines Coverage ∆ File
2
95.24
include/sw/universal/verification/correctly_rounded_reference.hpp
1
33.68
1.28% include/sw/universal/verification/test_suite_randoms.hpp

Coverage Regressions

Lines Coverage ∆ File
2
82.83
-0.3% include/sw/universal/number/qd/qd_impl.hpp
1
97.04
-0.06% include/sw/universal/number/cfloat/cfloat_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 40888380477.1 29 Oct 2026 02:27PM UTC 768
88.25
GitHub Action Run
Source Files on build 40888380477
  • Tree
  • List 768
  • Changed 3
  • Source Changed 1
  • Coverage Changed 3
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40888380477
  • 18e61196 on github
  • Prev Build on main (#40830633095)
  • Next Build on main (#40901049687)
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc