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

stillwater-sc / universal / 40168143550
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran 25 Oct 2026 06:30AM UTC
Jobs 1
Files 762
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

25 Oct 2026 06:10AM UTC coverage: 88.244% (+0.02%) from 88.222%
40168143550

push

github

web-flow
perf(efloat): one- and two-limb quotients and roots, differences as 2X - Y, and copies without store-forwarding stalls (#1829)

The mpfr_bands cells #1827 left above 1.5x MPFR (#1828):

- Quotients of operands of one or two limbs at 64 to 128 bits divide in
  fixed-size code, as mpfr_div_1 and mpfr_div_2: Q = floor(A B^M / D) from
  one divq or two steps of sb_div_qr_short, the guard bit from Q's last bit
  or from 2 R >= D, rounded by store_short, the short products' rounding
  tail, now shared and always inlined.
- Roots of one or two limbs: the exact root and remainder by sqrtrem2 or the
  new limb_kernels::sqrtrem4, dc_sqrtrem's level for four limbs in registers
  (10.8 against 13.7 ns); at 96 bits sqrt_approx4, the second limb from a
  double reciprocal of the top root instead of a second division, within
  2^15 units (6069 the worst of 20 million), rounded from where its bits
  under the guard bit are more than 2^16 from a boundary, else the exact
  root. sqrt() and sqrt_to no longer copy their argument to hold the root.
- A difference one bit apart whose leading bit cancels, which the top limbs
  decide, is formed as 2 X - Y in one pass: limb_kernels::twice_sub_n, a + ~b
  + 1 on adcx and + a on adox where the processor has ADX, no shift of Y
  before and none of the difference after; one to four limbs in a fixed loop.
- copy_limbs no longer vectorizes runs of up to 8 limbs: a 32-byte load of
  limbs just stored one at a time stalls until the stores retire, which
  every result's move into its destination paid. efloat's copies and moves
  are written member by member: clang's defaulted ones copied the state, sign
  and exponent as one 16-byte vector, with the same stall.
- Short sums write the result's own limbs from three limbs up, a fixed-length
  copy then a rare trim below; fma's result no longer starts as a copy of its
  longest operand.

mpfr_bands, i7-12700K pinned, best of 3, g++ 15.2, against #1827: every cell
from 64 to 16384 bits wit... (continued)

243 of 243 new or added lines in 4 files covered. (100.0%)

12 existing lines in 2 files now uncovered.

54962 of 62284 relevant lines covered (88.24%)

9014251.45 hits per line

Coverage Regressions

Lines Coverage ∆ File
11
95.68
-0.44% include/sw/universal/number/efloat/efloat_impl.hpp
1
97.04
0.0% include/sw/universal/number/cfloat/cfloat_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 40168143550.1 25 Oct 2026 06:30AM UTC 762
88.24
GitHub Action Run
Source Files on build 40168143550
  • Tree
  • List 762
  • Changed 6
  • Source Changed 5
  • Coverage Changed 6
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40168143550
  • 169f9cab on github
  • Prev Build on main (#40151466553)
  • Next Build on main (#40230749925)
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