New test failure on i686 in weekly run against develop HEAD

[mattip]

Starting here in the logs there are many warnings in a failing test on the weekly i686 run. I am not sure what is actually failing. There are at least two warnings about missing {, } around if clauses, and many warnings about using uninitialized values. @martin-frbg is this something known?

[martin-frbg]

there will have been a phase of test failures as I merged all the open PRs related to NAN handling in CSCAL/ZSCAL, but nothing affecting POTRS. I don't think any of the warnings is new (or related to the test failure you are seeing). do you still get this with the develop branch as of last friday ?

[mattip]

Last week's test with 02267d86f passed. I will try to bisect locally to find the problematic commit.

[martin-frbg]

There simply hasn't been any commit (besides the CSCAL/ZSCAL one that replaces the assembly kernel with the generic C one) since June 5 that could affect x86. Maybe something changed in your CI setup ?

[mattip]

Bisecting points to OpenBLAS 6954845d8d3 as the "bad" commit. This only adds tests, so I guess the i686 problem was there all along and only exposed now? I used the older 2025.06.08-1 image used last week, but it also fails on latest docker image from https://quay.io/repository/pypa/manylinux2014_i686?tab=tags&tag=latest. The quickest way to reproduce from this repo is

cd OpenBLAS
git checkout <commit>
git clean -xfdd
cd ..
docker run --rm -e BUILD_PREFIX=/usr/local -e PLAT=i686 -e INTERFACE64= -e NIGHTLY=1 -e PYTHON_VERSION=3.9 -e MB_ML_VER=2014 -e MB_ML_LIBC=manylinux -v $PWD:/io quay.io/pypa/manylinux2014_i686:2025.06.08-1 /io/tools/docker_build_wrap.sh

where tools/docker_build_wrap.sh ends up doing

CFLAGS=' -fvisibility=protected -Wno-uninitialized'
make BUFFERSIZE=20 DYNAMIC_ARCH=1 QUIET_MAKE=1 \
            USE_OPENMP=0 NUM_THREADS=64 \
            DYNAMIC_LIST='PRESCOTT NEHALEM SANDYBRIDGE HASWELL' \
            BINARY=32  \
            OBJCONV=/io/objconv/objconv SYMBOLPREFIX=scipy_ LIBNAMEPREFIX=scipy_ FIXED_LIBNAME=1 \
            TARGET=PRESCOTT

[martin-frbg]

Hmm. I'd think either your bisect is spurious or the HEAD picked up by CI happened to be between introduction of the new test and the switching of CSCAL kernels for all targets to make that test pass ??

[mattip]

Sorry, it was spurious (there were test failures, but not the test_potrs.c one). The first commit where test_potrs.c fails is OpenBLAS e12132abd4b43. The changes are needed to make the zscal tests pass, but they make the test_potrs.c test fail.

[martin-frbg]

I cannot reproduce this, and more importantly I do not see how this could happen unless your compiler is doing something very weird.
The patch introduces code that should perform exactly as before when called internally by OpenBLAS, and enable NAN propagation (matching other implementations) only when SCAL is called directly by user code. Furthermore, it is using the same generic C code as several other platforms, as I saw little value in adding all the extra conditionals to the old handcoded SSE that waa not significantly faster than what a modern compiler would achieve

[mattip]

I cannot reproduce this

OK, thanks for trying. Could it be due to the manylinux2014 standard compiler gcc 10.2.1?

[martin-frbg]

seems unlikely (gcc 10 must still be younger than the code in question) but I can try switching our manylinux-based CI job to BINARY=32. not sure I'll want to have this block the release of 0.3.30 now that I finally have the time and energy to do it, three months later than planned

[mattip]

Fixed on latest develop HEAD. Thanks @martin-frbg for tracking this down even though i686 is not very popular.