Cygwin: can't build Numpy 1.19.1 with GCC < 8.4

[embray]

Since #29766 I cannot build Numpy 1.19.1 on Cygwin, specifically with AVX enabled, due to this issue in Numpy (actually caused by a bug in GCC < 8.4): numpy/numpy#16290

I've been meaning to work on this issue anyways, so I'll see if I can find a clean workaround. If not, it also works to pass some CFLAGS to the Numpy build as noted on the upstream issues.

https://cygwin.com/packages/summary/gcc-core.html

Component: porting: Cygwin

Keywords: cygwin numpy

Issue created by migration from https://trac.sagemath.org/ticket/30392

[mkoeppe]

comment:2

Yes, this is important to fix. Unfortunately the CI infrastructure does not catch this because we cannot control the processor variant that will run the job. For example, https://github.com/sagemath/sage/runs/969113336?check_suite_focus=true https://github.com/sagemath/sage/runs/1002630008 has built numpy successfully.

With which compiler are you observing this failure?

[mkoeppe]

comment:3

The discussion at numpy/numpy#14787 seems to suggest that nothing needs to be fixed with current compilers. Do you still experience the build problem?

[embray]

comment:4

It is a problem with GCC<8.4.

I started on a patch to Numpy the week before last but hit a snag before I was able to finish it. Then I went on vacation :P

An ugly but acceptable workaround is also to set

CFLAGS=-ffixed-xmm16 -ffixed-xmm17 -ffixed-xmm18 -ffixed-xmm19 -ffixed-xmm20 -ffixed-xmm21 -ffixed-xmm22 -ffixed-xmm23 -ffixed-xmm24 -ffixed-xmm25 -ffixed-xmm26 -ffixed-xmm27 -ffixed-xmm28 -ffixed-xmm29 -ffixed-xmm30 -ffixed-xmm31

when compiling Numpy.

[mkoeppe]

comment:5

Current cygwin has gcc 10.2.

I would suggest to limit support of a from-source build of Sage on Cygwin to the latest Cygwin version and encourage users to update Cygwin.

(Does cygwin even have a supported way to install an older version of the distribution?)

[mkoeppe]

comment:10

We can close this one as outdated

[embray]

comment:11

I believe so, yes. Most Cygwin installs are not on GCC<8.4 by default, and I have also fixed the issue in Numpy.