Re: git: b0b4d9e1eae8 - main - */*: Switch to NumPy 2.x

On 5/11/26 12:59, Vladimir Druzenko wrote:
> 11.05.2026 22:35, Daniel Engberg пишет:
>> On 2026-05-11 21:19, Max Brazhnikov wrote:
>>> On Mon, 11 May 2026 14:42:57 -0400 Charlie Li wrote:
>>>> Vladimir Druzenko wrote:
>>>>> 06.05.2026 21:14, Vladimir Druzenko пишет:
>>>>>> 06.05.2026 20:29, Max Brazhnikov пишет:
>>>>>>> On Wed, 6 May 2026 01:04:14 +0300 Vladimir Druzenko wrote:
>>>>>>>> Hello!
>>>>>>>>
>>>>>>>> 05.05.2026 22:33, Max Brazhnikov пишет:
>>>>>>>>> The branch main has been updated by makc:
>>>>>>>>>
>>>>>>>>> URL: https://cgit.FreeBSD.org/ports/commit/?
>>>>>>>>> id=b0b4d9e1eae890d0de1591e20fb37358439a2ff6
>>>>>>>>>
>>>>>>>>> commit b0b4d9e1eae890d0de1591e20fb37358439a2ff6
>>>>>>>>> Author:     Max Brazhnikov <makc@FreeBSD.org>
>>>>>>>>> AuthorDate: 2026-05-05 19:27:15 +0000
>>>>>>>>> Commit:     Max Brazhnikov <makc@FreeBSD.org>
>>>>>>>>> CommitDate: 2026-05-05 19:29:37 +0000
>>>>>>>>>
>>>>>>>>>        */*: Switch to NumPy 2.x
>>>>>>>>>              PR:             294328
>>>>>>>>>        Exp-run:        antoine
>>>>>>>>> ---
>>>>>>>>>     archivers/py-bitshuffle/Makefile            | 4 ++--
>>>>>>>>>     archivers/py-blosc2/Makefile                | 4 ++--
>>>>>>>>>     astro/astrometry/Makefile                   | 4 ++--
>>>>>>>> By default NumPy 2.x builds with X86_V2, configure log:
>>>>>>>> CPU Optimization Options
>>>>>>>>      baseline:
>>>>>>>>        Requested : min+detect
>>>>>>>>        Enabled   : X86_V2
>>>>>>>>      dispatch:
>>>>>>>>        Requested : max
>>>>>>>>        Enabled   : X86_V3 X86_V4 AVX512_ICL
>>>>>>>>
>>>>>>>> So numpy-config fails with SIGILL.
>>>>>>>> Valgrid prints:
>>>>>>>> RuntimeError: NumPy was built with baseline optimizations:
>>>>>>>> (X86_V2) but your machine doesn't support:
>>>>>>>> (X86_V2).
>>>>>>>>
>>>>>>>> Patch:
>>>>>>>> diff --git a/math/py-numpy/Makefile b/math/py-numpy/Makefile
>>>>>>>> index 9500c87da243..961f1e03ff86 100644
>>>>>>>> --- a/math/py-numpy/Makefile
>>>>>>>> +++ b/math/py-numpy/Makefile
>>>>>>>> @@ -1,6 +1,6 @@
>>>>>>>>     PORTNAME=      numpy
>>>>>>>>     DISTVERSION=   2.4.4
>>>>>>>> -PORTREVISION=  2
>>>>>>>> +PORTREVISION=  3
>>>>>>>>     PORTEPOCH=     1
>>>>>>>>     CATEGORIES=    math python
>>>>>>>>     MASTER_SITES=  PYPI \
>>>>>>>> @@ -24,7 +24,9 @@ TEST_DEPENDS=
>>>>>>>> ${PYTHON_PKGNAMEPREFIX}hypothesis>=5.3.0:devel/py-hypothesis@${PY_
>>>>>>>>     USES=          compiler:c++17-lang cpe fortran pkgconfig
>>>>>>>> python:3.11+
>>>>>>>> shebangfix
>>>>>>>>     USE_PYTHON=    autoplist concurrent cython pep517 pytest
>>>>>>>>
>>>>>>>> -PEP517_BUILD_CONFIG_SETTING= -Csetup-args=-Dallow-noblas=false
>>>>>>>> +PEP517_BUILD_CONFIG_SETTING= -Csetup-args=-Dallow-noblas=false \
>>>>>>>> + -Csetup-args=-Dcpu-baseline="none" \
>>>>>>>> +  -Csetup-args=-Dcpu-baseline-detect="disabled"
>>>>>>> Defining only -Dcpu-baseline="none" should be enough. Could you
>>>>>>> check
>>>>>>> whether
>>>>>>> it resolves the problem for you?
>>>>>>>
>>>>>>> Max
>>>>>> You are right, -Dcpu-baseline="none" is enough.
>>>>>> The -Csetup-args=-Dcpu-baseline-detect="disabled" option causes -
>>>>>> march=* to be ignored when baseline detection occurs (if the user
>>>>>> added CPUTYPE=* to the make.conf).
>>>>>> So in order to respect -march this option should not be added.
>>>>> Hello!
>>>>> Can I commit this change?
>>>>>
>>>> Not yet. Still needs some further evaluation.
>>> -Dcpu-baseline="none" is safe to commit and should be committed if we
>>> claim support for older CPU.
>>>
>>> Cheers,
>>> Max
>> What's probably better to do is to define none only if CPUTYPE is not
>> defined (set) and otherwise omit -Dcpu-baseline as the default cpu-
>> baseline-detect setting should take care of it otherwise according to
>> https://numpy.org/devdocs/reference/simd/build-options.html
>>
>> Best regards,
>> Daniel
> 
> Not quite right: min = X86_V2, and baseline-detect adds any additional
> instructions detected to min.
> So it's "X86_V2+detect", not just "detect". For CPUs that don't support
> X86_V2, this results in a SIGILL.
> Universal solution would be "none+detect".
> 

From what I read, the code generated was using new instructions in ways
expected to work before the code that is intended to detect and report
lack of compatibility. This contributed to why it was crashing in a
bunch of environments, even ones not tied to FreeBSD: it never got so
far as the code that is supposed to report such issues.

-- 
===
Mark Millard
marklmi at yahoo.com