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

stillwater-sc / universal / 40061163622
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran 24 Oct 2026 04:28AM UTC
Jobs 1
Files 761
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

24 Oct 2026 04:08AM UTC coverage: 88.115% (+0.009%) from 88.106%
40061163622

push

github

web-flow
perf(efloat): x86-64 mulx/adcx/adox kernels in inline assembly, chosen at run time, and the thresholds re-measured (#1818)

* perf(efloat): x86-64 mulx/adcx/adox kernels in inline assembly, chosen at run time, and the thresholds re-measured

mul_1, addmul_1 and submul_1 run a four-limb loop of mulx with adox
chaining the product's high halves through OF and adcx adding the
destination through CF; submul_1 adds the complement of each product limb
on a CF chain that starts set. add_n and sub_n are adc/sbb chains. The
loops step a negative index with lea and exit on jrcxz, which leave the
flags alone, so the two carry chains survive the loop (intrinsics could
not keep them apart, #1814). The low n mod 4 limbs take the portable
loop and pass their carry in.

Gated on GNU inline asm on __x86_64__; the multiply kernels also need
BMI2 and ADX, detected once by cpuid, so no -march flag is required.
MSVC, ARM, POWER, RISC-V and EFLOAT_MPN_NO_ASM builds compile only the
portable loops (renamed *_portable, which the test compares against).

With the faster basecase the thresholds move: Karatsuba 16 -> 24,
squaring 48 -> 64, Mulders' short product 32 -> 64, divide-and-conquer
division 16 -> 32 limbs; Toom-3's stay.

mul_basecase is 0.92-1.09x GMP's mpn_mul_n at 8-24 limbs (was up to
1.24x). At 8192 bits a*b 3.46 -> 2.61 us, b/a 6.02 -> 4.58, sqrt
3.66 -> 2.91, at or below gmpy2's 2.79, 4.69 and 3.39; every row 3-26%
faster at 2048 and 8192 bits. The kernels are exact, so results are
unchanged (mpfr_check: 0 wrong).

limb_kernels gains VerifyLimbKernels: every kernel against its portable
loop at lengths 1-300, all n mod 4, all-ones runs, all-ones multipliers
and in-place operands (ASan/UBSan do not see inside inline asm).

Resolves #1816

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(efloat): history row for #1818

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(efloat): 64-bit loop index in the x86-64 kernels, and thresholds by th... (continued)

60 of 60 new or added lines in 1 file covered. (100.0%)

1 existing line in 1 file now uncovered.

54365 of 61698 relevant lines covered (88.11%)

9181465.9 hits per line

Coverage Regressions

Lines Coverage ∆ File
1
97.04
-0.06% include/sw/universal/number/cfloat/cfloat_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 40061163622.1 24 Oct 2026 04:28AM UTC 761
88.11
GitHub Action Run
Source Files on build 40061163622
  • Tree
  • List 761
  • Changed 2
  • Source Changed 1
  • Coverage Changed 2
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40061163622
  • 3a687cce on github
  • Prev Build on main (#40047512831)
  • Next Build on main (#40131128161)
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