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

stillwater-sc / universal / 38992615351
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran 17 Oct 2026 02:43AM UTC
Jobs 1
Files 724
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

17 Oct 2026 01:51AM UTC coverage: 86.666% (-0.1%) from 86.765%
38992615351

push

github

web-flow
fix(unum1): a scalar unum op gives the tile that holds the exact result (#1705)

Scalar unum assignment and + - * / rounded twice. Each op was
`*this = to_double() OP rhs.to_double()`, so the double rounded to
nearest, the encoder then truncated onto the lattice, and the ubit
ignored which side the exact value was on. An operand's own ubit was
ignored too: an inexact unum counted as its exact lower endpoint.
Overflow clamped the exponent field and built a wrong value.

  1 + 2^-100 in unum<3,4>   was exact 1, now the open tile above 1
  U(1e40), maxpos ~= 6.8e38 was a tile near 6.25e38, now (maxpos, inf)

A scalar unum now names the tile that holds the exact result. Each end
of the result is computed in double. An error-free transform (TwoSum,
TwoProduct, the division or sqrt residual) gives the sign of the
double's rounding, and each end maps to the tile just inside it. When
both ends land in one tile, that tile is the result. When they do not
(an inexact operand whose result spans several tiles), the result is
NaN, and ubound is the type that holds that interval. For * and / no
operand tile crosses zero, so the result is a sign times a magnitude
interval, which settles the open-zero and infinite ends.

lattice.hpp now holds unum_lattice, the error-free transforms and the
tile mapping, all constexpr (bit extraction instead of std::ilogb,
std::ldexp, std::floor), so the scalar operators stay constant-
evaluable. ubound uses the same lattice in place of its private copy.
The integer and long double assignments undo the rounding of their cast
to double: U(2^60 + 1) is the open tile above 2^60. sqrt is directed in
the same way.

Limits:
- (-inf, -maxpos) has the same bit pattern as NaN, so a result below
  -maxpos is NaN.
- Configurations whose lattice does not fit in double (esizesize > 3 or
  fsizesize > 5) keep the old path, bit for bit.
- unum_lattice saturates its binade limits, so unum<8,2> still compiles.

elastic/unum/scalar_tiles.cpp checks thre... (continued)

203 of 208 new or added lines in 3 files covered. (97.6%)

71 existing lines in 2 files now uncovered.

49424 of 57028 relevant lines covered (86.67%)

11249875.4 hits per line

Uncovered Changes

Lines Coverage ∆ File
3
96.63
include/sw/universal/number/unum/lattice.hpp
2
76.74
-16.8% include/sw/universal/number/unum/unum_impl.hpp

Coverage Regressions

Lines Coverage ∆ File
63
76.74
-16.8% include/sw/universal/number/unum/unum_impl.hpp
8
96.37
-0.52% include/sw/universal/number/cfloat/cfloat_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 38992615351.1 17 Oct 2026 02:43AM UTC 724
86.67
GitHub Action Run
Source Files on build 38992615351
  • Tree
  • List 724
  • Changed 4
  • Source Changed 3
  • Coverage Changed 4
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #38992615351
  • 0153d4a4 on github
  • Prev Build on main (#38983239869)
  • Next Build on main (#38995555605)
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