[gentoo-dev] [PATCH 3/3] distutils-r1.eclass: fix src_configure to handle flag-o-matic correctly

[Eli Schwartz <eschwartz93@gmail.com>]

Use of python_configure_all is a bit broken, because distutils-r1 is a
bit broken. It has the intriguing value-claim that *FLAGS shall be
modified with local -x as part of run-phase, which means that all
modifications are reset as soon as any given phase exits. Including
sub-phases. If you make changes in python_configure or
python_configure_all, as you are expected to do instead of defining
src_configure and then running the eclass phases manually, your changes
inherently get erased before you actually *get* to running builds (i.e.
reading the variables).

For an example of how this works, consider dev-python/scipy. It builds
with filter-lto, for the standard reasons. However, filter-lto gets run
inside python_configure_all, so it gets erased and ignored.

Note: it "worked" anyways prior to reworking meson.eclass to pass
-Db_lto based on `is-flagq flto`. This is because filter-lto removed it
from LDFLAGS, and since distutils-r1 doesn't localize that variable, the
changes stuck around. As a result:

- the previous fix to scipy caused scipy to be compiled with LTO, but
  not linked with LTO
- meson.eclass changes caused scipy to be built with -Db_lto=true, which
  affects both compiling and linking

It's absolutely unsafe to have flag-o-matic be ignored like this, and it
is fascinating to end up with LTO restored into compile flags while
still being filtered from LDFLAGS... simply as a side effect of
distutils-r1 not modifying the latter.

- this patch causes scipy to be built with -Db_lto=false

Consequences of the change make no difference to standard dev-python/
packages, but affect packages that use both distutils-r1 and other
packaging infrastructure:

- global variables for tools are exported. This is supposed to be a
  safe, albeit unnecessary for better-than-setuptools build systems,
  option.

- the "debug" option applies the DEBUG preprocessor macro to more than
  just python code. This was already the case for python projects that
  built non-CPython API C/C++ code and then linked them with thin python
  wrappers, so projects with python bindings shouldn't be any different.
  Also, it is a "debug" use flag so it's pretty silly if it only toggles
  debug on python bindings but not the rest of the package, so just...
  deal with it I guess?

- any project that uses cython, now ignores the Modern C Porting flag
  "incompatible-pointer-types". This is terrible, because code which
  should trigger that error is terrible. But it's also terrible for
  python projects with mixed Cython and handwritten C, to squelch that
  error just because one file happens to use Cython. Yet, we do it
  because we don't have a way to make extremely specific files built
  with different CFLAGS compared to the rest of the project. There's no
  actual reason to treat handwritten C python modules different from
  non-distutils phases.

Signed-off-by: Eli Schwartz <eschwartz93@gmail.com>
---
 eclass/distutils-r1.eclass | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/eclass/distutils-r1.eclass b/eclass/distutils-r1.eclass
index 35825d4c3aa6..873421bbcee8 100644
--- a/eclass/distutils-r1.eclass
+++ b/eclass/distutils-r1.eclass
@@ -1813,16 +1813,15 @@ distutils-r1_run_phase() {
 	fi
 
 	# Set up build environment, bug #513664.
-	local -x AR=${AR} CC=${CC} CPP=${CPP} CXX=${CXX}
 	tc-export AR CC CPP CXX
 
 	if [[ ${DISTUTILS_EXT} ]]; then
 		if [[ ${BDEPEND} == *dev-python/cython* ]] ; then
 			# Workaround for https://github.com/cython/cython/issues/2747 (bug #918983)
-			local -x CFLAGS="${CFLAGS} $(test-flags-CC -Wno-error=incompatible-pointer-types)"
+			append-cflags $(test-flags-CC -Wno-error=incompatible-pointer-types)
 		fi
 
-		local -x CPPFLAGS="${CPPFLAGS} $(usex debug '-UNDEBUG' '-DNDEBUG')"
+		append-cppflags $(usex debug '-UNDEBUG' '-DNDEBUG')
 		# always generate .c files from .pyx files to ensure we get latest
 		# bug fixes from Cython (this works only when setup.py is using
 		# cythonize() but it's better than nothing)
-- 
2.43.0