[PATCH v2 0/5] math: Test libmvec functions with mixed vector lanes

Adhemerval Zanella adhemerval.zanella@linaro.org
Fri Nov 13 12:32:04 GMT 2026


The libmvec tests copy each input to every lane, so they miss bugs in
the code that merges the fast path with the special-case lanes, such as
BZ 34922, BZ 34923 and BZ 34940.  This series runs each input under
several lane layouts: the input in every lane, and the input in the
first or last lane with a control value in the others.

Changes from v1:
  * Add +/-1 to the control values (patch 3).
  * Also check the control lanes (patch 4).
  * Run the x86 sincos tests through the layouts and remove
    TEST_VEC_LOOP (patch 5).

Checked on x86_64-linux-gnu and aarch64-linux-gnu, with SVE vector
lengths of 128, 512 and 2048 bits under QEMU.  The x86_64 AVX512 tests
were built but not run, and the x32 wrappers were only compiled.

Adhemerval Zanella (5):
  math: Test libmvec functions with mixed vector lanes
  math: Test vector pow with a qNaN base and zero exponent
  math: Add +/-1 to the libmvec lane control values
  math: Check the control lanes in the libmvec tests
  math: Run the x86 vector sincos tests through the lane layouts

 math/libm-test-driver.c                       | 130 ++++++--
 math/libm-test-pow.inc                        |   8 +-
 math/test-math-vector.h                       | 289 ++++++++++++++----
 .../aarch64/fpu/test-double-sve-wrappers.c    |  23 +-
 sysdeps/aarch64/fpu/test-float-sve-wrappers.c |  23 +-
 .../aarch64/fpu/test-vpcs-vector-wrapper.h    |  32 +-
 sysdeps/x86/fpu/test-math-vector-sincos.h     | 136 +++++----
 .../fpu/test-float-vlen8-avx2-wrappers.c      |   3 -
 8 files changed, 471 insertions(+), 173 deletions(-)

-- 
2.53.0



More information about the Libc-alpha mailing list