feature module: Fix handling of multiple conflicts per attribute

[Vogtinator]

• Attributes without match were never actually added to the list
• Only the last conflict actually had an effect, earlier results were discarded

[mattip]

Also see mesonbuild#11307 which will make this module available in upstream meson

[mattip]

@seiko2plus, @rgommers could you take a look?

[rgommers]

This seems reasonable to me, but it's @seiko2plus's code and it needs to be integrated into his feature branch too in order not to lose this fix when we rebase or discard this fork. So I'll leave it to @seiko2plus to review/incorporate.

[Vogtinator]

Ping @seiko2plus

[Vogtinator]

Ping.

[rgommers]

Sorry this stalled @Vogtinator. Since @seiko2plus isn't available, I'll attempt to get this merged. I tested an integration branch with this change included in numpy/numpy#27550, and got one relevant failure on armhf under QEMU:

 $ /usr/bin/python -c 'import sys; del sys.path[0]; import numpy'
Traceback (most recent call last):
  File "/numpy/build-install/usr/lib/python3/dist-packages/numpy/_core/__init__.py", line 23, in <module>
    from . import multiarray
  File "/numpy/build-install/usr/lib/python3/dist-packages/numpy/_core/multiarray.py", line 10, in <module>
    from . import overrides
  File "/numpy/build-install/usr/lib/python3/dist-packages/numpy/_core/overrides.py", line 7, in <module>
    from numpy._core._multiarray_umath import (
ImportError: /numpy/build-install/usr/lib/python3/dist-packages/numpy/_core/_multiarray_umath.cpython-310-arm-linux-gnueabihf.so: undefined symbol: _ZN2np7highway10qsort_simd11QSort_ASIMDIjEEvPT_i

From the description and numpy/numpy#26151 it seems that you were actually trying to fix something for armhf? I don't have local access to armhf, but either way it would be helpful to have a clearer description of how exactly things are broken for you without this PR.

[Vogtinator]

From the description and numpy/numpy#26151 it seems that you were actually trying to fix something for armhf? I don't have local access to armhf, but either way it would be helpful to have a clearer description of how exactly things are broken for you without this PR.

Yes, probably caused by:

  ASIMD.update(args: [
    {'val': '-mfpu=neon-fp-armv8', 'match': '-mfpu=.*'},
    '-march=armv8-a+simd'
  ])

In the build log the issue is visible as:

Test features "ASIMD" : Unsupported due to Compiler fails against the test code of "ASIMD"
Test features "ASIMDHP" : Unsupported due to Implied feature "ASIMD" is not supported
Test features "ASIMDFHM" : Unsupported due to Implied feature "ASIMD" is not supported
Test features "SVE" : Unsupported due to Implied feature "ASIMD" is not supported
Configuring npy_cpu_dispatch_config.h using configuration
Message: 
CPU Optimization Options
  baseline:
    Requested : min
    Enabled   : 
  dispatch:
    Requested : max -xop -fma4
    Enabled   : NEON NEON_FP16 NEON_VFPV4

With this PR, ASIMD should show up as supported by the compiler and get used for runtime dispatch.

[Vogtinator]

Ping

[seiko2plus]

@Vogtinator, sorry for being away for a while. I confirm the bug you reported on armhf and your patch fixes it. Regarding the error @rgommers caught from CI, it's highway/QSort related, likely marked as incompatible platform via VQSORT_COMPILER_COMPATIBLE. As a quick fix, I will apply your patch to bring asimd support for arm7 for the rest of the kernels.

[seiko2plus]

@Vogtinator, LGTM, thank you!

Regarding your note about test failures on armhf under QEMU (mentioned in numpy/numpy#27550): @rgommers, I can confirm that the armhf test now passes in PR numpy/numpy#28635 after fixing the link error."