tacaswell left a comment

Python.h defines several visibility macros, which only work if defined before any stdlib header is included.This is needed for things like pybind11's strdup to get declared.

Not a pass, but not the fault of the test.TST,BUG: Move pytest import to just before its use.Hopefully this fixes the doc builds.TST: Cygwin: Mark more fork failures as xfail.Not a pass, but not matplotlib's fault.

This involved a good bit of debugging of the build process, especiallyfor the bundled FreeType.  Eventually removing the debugging flags letit just work.There are still fork failures, but I've done everything I know to do.CYGWIN=detect_bloda may reveal more, but may not.parent48e5934author DWesl <22566757+DWesl@users.noreply.github.com> 1651882104 -0400committer DWesl <22566757+DWesl@users.noreply.github.com> 1670678538 -0500parent48e5934author DWesl <22566757+DWesl@users.noreply.github.com> 1651882104 -0400committer DWesl <22566757+DWesl@users.noreply.github.com> 1670678484 -0500CI: Add a Cygwin run to GHA CI.Let's see if I translated the workflow even close to correctly.CI: Pin Cygwin setuptools<60 and NumPy !=1.21setuptools>60 breaks Pillow install, and may cause problems for NumPy.NumPy 1.21 has intermittent segfaults on GHA runners, so let's just skip that.CI: Pin Cygwin setuptools<60 and avoid problem upgradessetuptools>=60 runs intopython-pillow/Pillow#6216 andpypa/setuptools#3304.NumPy==1.21.* has intermittent segfaults due to SIMD attempts on unaligned memory.kiwisolver, numpy, and pillow have problems installing, so just verify that they are present, don't try to upgrade them.CI: Upgrade Cygwin pipCI: Install g++ and print info on Python.hHopefully one of these will convince NumPy to compile.CI: Install libjpeg-devel for Pillow installLet's see if it keeps setuptools old enough for this to work.CI: BLD: Use system freetype and qhullGiven that I'm checking a build problem at the moment, this should allow that to show up faster.This should probably be reverted before merging.CI: Specify shell for Cygwin pytest step.CI: Rebase Matplotlib DLLs before running pytestTry to avoid "BlockingIOError: Resource Unavailable"with a fork() error on stderr.CI: Compile against bundled freetype.See if this fixes the various test errors.It will take a few tries to get freetype to actually compile, I think.CI: Install make for libfreetype buildLet's see if this lets the build run to completion.BLD: Run autoconf before installing bundled freetypeLet's see if this fixes the install problems.CI: Try to get Cygwin build of FreeType working.It doesn't fail the same way locally.CI: Revert to actually-configured python alternatives.BLD: Fix autoconf call on Cygwin.BLD: Fix directory for Cygwin autoreconf.CI: Fix directory for Cygwin autoconf in FreeType.BLD: Fix autoreconf command line.BLD: Run only autoconf on Cygwin, not autoreconf.libtoolize still fails; let's see if this at least gets through configure now.BLD: Run verbose libtoolize on bundled FreeTypeLet's see if this produces useful information.BLD: Copy the libtool files and ensure write permissionsCopy instead of symlinking, and make sure the current user has write permissions for the target files.BLD: Add print statements to update on Autotools progressLet's see if I this narrows down where the error happens.BLD: Add permissions for everybody to fix libtoolizeIt's the only thing I can think of that would make the copy fail.CI: Try to clean up error reporting.Errors importing matplotlib are distinct from test failures; pip's logs of compile failures have a lot of extra stuff not related to the actual compile failureCI: Skip broken FreeType build on CygwinI'll have to update many tests, but that should be doable.CI: Add FreeType and QHull to Cygwin CII'm going the hard way through the testing process.BLD: Use auto-typed result from dlsymTry to avoid the Cygwin-vs-Linux dlsym signature incompatibility.(Cygwin returns FARPROC aka `long long int (*)()`, Linux returns void *)CI: Add Cygwin dependencies and try to rebase moreLet's hope this sidesteps the BlockingIOErrors in subprocess.runCI: Split Cygwin rebase stepOne step for finding the files.a separate step for rebasing, using only rebase.CI: Only rebase files in /usr/ and /usr/local/GHA gives permissions to write to /usr/local, so pip puts the new files there.CI: Add premissions to the image comparisonsSee if this lets the artifact upload step work.CI: Use the bundled FreeType on Cygwin againCI: Try to change rebase to avoid fork failuresBLD: Fix definition of FT_CONFIG_CONFIG_H on CygwinThe default definition doesn't actually define everything assumed to be defined by that header.It was changed a few versions before, but they seem to have only checked whether it worked a few places.BLD: Add autoconf and libtoolize back inI'd forgotten configure crashed.BLD: Continue on libtoolize failureCI: Try to get more information from makeCI: Uses Cygwin FreeType package again.CI: Try doubled backslashes for result directory.Let's see if this allows upload-artifact to succeed.BUG: Rebase DLLs in ediable installI skipped those earlier, which doesn't do anything useful.CI, BUG: Install git on Cygwin runner.Some of the tests use setuptools_scm, which needs git.CI, BUG: Run actions/checkout before Cygwin installOtherwise git panics about someone else owning the repository midway through cloning.CI, BUG: Mark the repo safe againGit is apparently picky about how I specify Windows paths.CI, DBG: Print Windows permissions for image dirInspired byhttps://social.technet.microsoft.com/wiki/contents/articles/31139.powershell-how-to-get-folder-permissions.aspxApparently Windows "dir" doesn't have an option to output permissions, that's a separate utility.CI, DBG: Print more permissions for result_imagesTry printing permissions of the images that fail to upload.  The directory seems to deny only write and delete permissions, which shouldn't cause problems.CI, DBG: Give everyone all permissions to result img dirMaybe this will let upload-artifact succeed.CI: Use dash for /bin/sh on CygwinThis seems to help libtoolize finish without crashing.CI: Try to compile bundled freetype on Cygwin againThis should get farther.CI: FIX: Specify proper shell when setting shell.CI, FIX: Spell shell options correctly to set /bin/shCI, FIX: Use dash to set dash as /bin/sh on CygwinBLD: Specify /bin/dash explicitly for Cygwin FreeType buildTODO: Remember to revert the configure change.CI: Keep bash around when making dash /bin/shCI: Make /bin/sh a symlink to /bin/dashCI: Set Cygwin tempdir to /tmpCI: Explicitly install m4 on Cygwin.CI, DBG: Clarify what /bin/sh is around def change.I want to see how they change before and after I change it from /bin/bash to /bin/dashCI, FIX: Use bash to change /bin/sh to /bin/dashCI, FIX: Clean up syntax to set sh to dash.Also avoid duplicate work for a failure.BLD: Reduce verbosity of FreeType setup on CygwinUse autoreconf instead of autoconf --verboseShould help me figure out where the new failure is.BLD: Stop running autoreconf on CygwinIt doesn't seem to help muchCI, DBG: Only run upload-artifact debuggers if tests ranDon't try to examine images if there hasn't been anything to create them.CI, FIX: Use single quotes for condition.Double quotes cause problemsCI: Allow Cygwin CI to run on patch-2BLD: Revert most changes to setupext.pyStill unconditionally rebuild ./configure on Cygwin, to try to update build system.BLD: Revert the rest of the changes to setupext.pyFreeType seems to build on unmodified matplotlib locally, so this shouldn't be needed.CI: Install matplotlib in Cygwin CIHopefully this pulls in the last dependencies and keeps  the FreeType build from failing.CI: Specify MAKEFLAGS, not MAKEhttps://www.gnu.org/software/make/manual/make.html#Options_002fRecursionsuggests specifying MAKEFLAGS to ensue options are passed to make, not MAKE.  This might help with the weird "directory not found" errors.CI, BUG: Specify --coverage in CPPFLAGS and LDFLAGSSpecifying in CPPFLAGS only causes link to fail.LDFLAGS=-lgcov should also work, but this should be more general.STY: Remove trainling whitespaceGet pre-commit to stop sending me emails about trailing whitespace.CI,BUG: Make sure coverage library gets linkedThere may be a problem in the configtest procedure.CI: Cut separate build stepIt's failing in the pip install step, and I want to be sure that's not because there's already a build tree.CI: Cygwin: Don't add coverage tracing to matplotlibIt causes pip install to fail.CI: Install importlib_resourcesApparently not listed in the requirements*.txtBUG: Fix spellingCI: Cygwin: Drop LaTeX packagesThe PGF tests mostly fail.CI: Set ffmpeg path so Cygwin tests don't use Windows executableCI: Append to mplrc, don't overwrite.Let's see if this fixes the key lookup error on importCI: Cygwin: Create new mplrc file instead of overwriting oldAppending creates load errors in forked tests; replacing the line should fix that, and creating a new version should fix the remaining problems.CI,FIX: Cygwin: Create mpl config directoriesCI: Cygwin: Don't install WX, do install ipythonWX imports keep failing.One test uses IPython and fails, let's make sure that's not because IPython isn't installed.CI: Cygwin: Uninstall wxPythonIt shouldn't be installed, so this should be fast.CI: Cygwin: Try to get remaining tests working.Rebase executables in /bin as well as those in /usr/binRemove default ACLs on /tmpCI,FIX: Fix syntax and style of previous commit//bin is not POSIX-compliant and doesn't work on Cygwin.Delete trailing whitespace.CI: Cygwin: Set username to root and uid to 0This should let checks for "can I access all files regardless of permissions" work the way Linux expects.CI,FIX: Cygwin: Try to fix sed expression for euid=0Sed expression#2, character 0: error: no previous regex.Not really useful, sed.CI: Cygwin: Stop changing username, just change uidCI: Stop rebasing binaries in /binThey are already covered in /usr/bin, and probably the important information is in the file, not a database somewhere (I mean, it's also in a database, but you can turn that off).CI: Cygwin: Upload code coverage dataCI: Cygwin: Update run condition for upstream repoI changed it so I'd get quicker feedback on tests some time ago and forgot to change it back.Update .github/workflows/tests.ymlCo-authored-by: Elliott Sales de Andrade <quantum.analyst@gmail.com>Update .github/workflows/tests.ymlCo-authored-by: Elliott Sales de Andrade <quantum.analyst@gmail.com>CI: Cygwin: Incorporate suggestions from review.CI: Incorporate suggestions from review.CI: Cygwin: Use dash for /bin/sh instead of bash.Faster, and doesn't fail the FreeType build.CI: Test python 3.8 and 3.9 on CygwinCo-authored-by: Thomas A Caswell <tcaswell@gmail.com>CI: Cygwin: Correct matplotlibrc locationApparently it doesn't check ~/.config/matplotlibDOC,CI: Explain why CI runs tests with gid set to 0.The other option is to look into cygutils-extra's cygdrop.I have no idea how to use that, so I'm sticking with this.

I think I did this right, we will find out.

QuLogic left a comment

It might be good to make all the "try fork; xfail on BlockingIOError if running on Cygwin" code look like each other, or all use a single helper function somewhere.Co-authored-by: Elliott Sales de Andrade <quantum.analyst@gmail.com>

This way I have fewer places to get things wrong and fewer places to remember to fix things.Ideally, at least.

…ent.universal_newlines is apparently the back-compatibility name;also text describes the biggest difference in the return types.

Also add a few reference links and types.