CI: make-sqlite3 fails on ubuntu-latest

Open Beginner friendly
#2,320 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 11 days

Nobody has claimed this yet.

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
68/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
c, github-actions
Domain
ci-cd, databases

Research direction

Start by reading .github/workflows/sqlite3.yml and the make-sqlite3 job; the issue provides a reproduction in libsql-sqlite3 using testfixture and test/libsql_vector.test. Pin the job to ubuntu-24.04 as described, then confirm the workflow configuration reflects that runner. The issue notes this workflow only runs when libsql-sqlite3/** changes, so the isolated workflow edit will not trigger it.

Written by the indexing model from the issue text.

Description

Current origin/main has a bug.

The make-sqlite3 job ("Makefile CI with Wasm enabled", .github/workflows/sqlite3.yml) fails on current ubuntu-latest runners, which now resolve to Ubuntu 26.04.

One test out of 330,804 fails: vector-1-conversion-f8 in test/libsql_vector.test.

Note:

This is a different job and a different cause from #2315 / #2317, and #2317 does not cover it.

  • Last green run, on ubuntu-24.04: 42522832842
  • Failing run, on ubuntu-26.04 (image 20261111.196.1): 43845107230

Symptom

One element prints as 1.63294 where 1.63295 is expected. In the same test one more value prints as 10.19 instead of 10.1899, and the f32/f64 hex of the converted vectors differs in the last bits.

Cause

The 26.04 runner image installs Ubuntu's amd64v3 packages (the job log shows tcl8.6-dev_..._amd64v3.deb). The gcc of that variant defaults to -march=x86-64-v3, which enables FMA. testfixture is built with plain gcc -g -O2, and gcc's default is -ffp-contract=fast, so this expression in vectorConvertFromF8() (libsql-sqlite3/src/vector.c) is compiled to one fused multiply-add:

dstF32[i] = alpha * src[i] + shift;

A fused multiply-add rounds once instead of twice, so the result differs in the last bit and the test's printed values change. The f8 blob itself is the same either way. Only the values decoded from it differ. Wasm and gcc 15 as such are not the cause.

How to reproduce

On any x86-64 machine with FMA, no container needed:

cd libsql-sqlite3 && mkdir bld && cd bld
CFLAGS="-g -O2 -march=x86-64-v3" ../configure && make testfixture
./testfixture ../test/libsql_vector.test     # vector-1-conversion-f8 fails

With CFLAGS="-g -O2 -march=x86-64-v3 -ffp-contract=off" the test passes.

What I checked:

  • In an ubuntu:26.04 container with APT::Architecture-Variants "amd64v3", gcc 15.2.0 reports -march=x86-64-v3 and defines __FMA__. The stock ubuntu:26.04 gcc reports -march=x86-64.
  • A testfixture built in that container with the default ./configure fails vector-1-conversion-f8, and its "got" output is byte-identical to the one in the CI log.
  • On my machine -march=x86-64-v3 gives the same output, and adding -ffp-contract=off makes the test pass.
  • In an ubuntu:24.04 container (gcc 13.3, -march=x86-64) the test passes.

Not checked: I did not inspect the gcc on the runner itself, only that its apt installs amd64v3 packages. I ran only libsql_vector.test, not the full make test, and without --enable-wasm-runtime-dynamic.

Fix to unblock CI

Pin the job to ubuntu-24.04, the same approach as #2317. It could go into #2317 or be a PR of its own:

--- a/.github/workflows/sqlite3.yml
+++ b/.github/workflows/sqlite3.yml
@@ -19,7 +19,10 @@ concurrency:
 
 jobs:
   make-sqlite3:
-    runs-on: ubuntu-latest
+    # The ubuntu-26.04 image uses amd64v3 packages, whose gcc defaults to
+    # -march=x86-64-v3. The resulting FMA contraction changes the floats
+    # expected by vector-1-conversion-f8 in test/libsql_vector.test.
+    runs-on: ubuntu-24.04
     defaults:
       run:
         working-directory: libsql-sqlite3

Note that this workflow only runs when libsql-sqlite3/** changes, so a PR with only this change will not run it.

Longer-term options

The pin hides the fact that the result of vector_extract() on an f8 vector depends on the compiler's target flags. Anyone who builds libSQL with -march=x86-64-v3, -march=native, or on a distribution whose gcc defaults to it gets the other rounding. Options, for the maintainers to choose:

  1. Build with -ffp-contract=off. That keeps today's output everywhere, but only for builds that use the flag.
  2. Make the code independent of contraction. gcc ignores #pragma STDC FP_CONTRACT, so that means something like storing the product in a volatile temporary before the addition. It costs some speed, and the bundled amalgamations need regenerating.
  3. Make the test tolerant of the last bit (round the printed values, drop the hex comparison).
Dominant language
C
Stars
17.3k
Forks
562
Avg merge
10d 15h
Merged PRs (30d)
1

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from tursodatabase/libsql

All issues in tursodatabase/libsql

Similar issues

More C issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.