universal2 build improvements

[radarhere]

Two suggested changes when building universal2 wheels.

1. Instead of presuming the length of the string to replace in the following code, use sed. This will allow $py_osx_ver to vary in length.
   

   multibuild/osx_utils.sh

   Lines 513 to 514 in 2e10577

   	if [[ "$whl" == *macosx_${py_osx_ver}_x86_64.whl ]]; then
   	whl_base=$(echo $whl | rev | cut -c 23- | rev)

2. When building for universal2 on x86-64, run the cross build first and then the native build. This will mean that the more natural native variables are in place when the code is finished, for anything that might run afterwards.

[isuruf]

Some shared libraries require a native version of the library/tools for cross compiling. Changing the order would prevent that.

[radarhere]

But the native version is already built last on arm64?

multibuild/osx_utils.sh

Lines 527 to 532 in 2e10577

	function wrap_wheel_builder {
	if [[ "${PLAT:-}" == "universal2" ]]; then
	if [[ "$(uname -m)" == "arm64" ]]; then
	(macos_intel_cross_build_setup && $@)
	rm -rf *-stamp
	(macos_arm64_native_build_setup && $@)

[isuruf]

Ah yes. That should be fixed. Arm64 native CI is untested AFAIK

[radarhere]

I've removed the second change, leaving just the sed change.

[radarhere]

Some shared libraries require a native version of the library/tools for cross compiling. Changing the order would prevent that.

Could you give an example of one such library?

[isuruf]

ncurses

[radarhere]

I attempted to test this with a custom

function build_ncurses {
    fetch_unpack https://invisible-island.net/datafiles/release/ncurses.tar.gz
    (cd ncurses-6.3 \
        && ./configure --prefix=$BUILD_PREFIX \
        && make -j4 \
        && make install)
    touch "ncurses-stamp"
}

but it wouldn't even build on arm64. If you're able to demonstrate, that would be helpful.

[radarhere]

If you aren't able to demonstrate, it would be good if you were able to Approve this PR for my sed change - just to clarify that our discussion is not a blocker.

[isuruf]

I'll have a look at the ordering later this week. Sed change is fine.

[radarhere]

To give more information about my situation, build_libjpeg_turbo is installing cjpeg and djpeg binaries. Because the cross build is run last, the binaries built raise "Bad CPU type in executable" during subsequent testing.

Sure, a workaround for this could be for me to just run build_libjpeg_turbo again after both builds - but running the native build last seemed like the more general solution to this type of problem.

[isuruf]

I think the issue with both ncurses and build_libjpeg_turbo is that the PATH is set incorrectly for the cross compilation. The native prefix should be first in the PATH for cross compilation.