1 |
Hello, |
2 |
|
3 |
Currently in-source builds differ a lot from out-of-source builds. Most |
4 |
notably: |
5 |
|
6 |
- BUILD_DIR points to source dir rather than build dir, |
7 |
|
8 |
- build directories are not overriden, |
9 |
|
10 |
- PYTHONPATH is not set. |
11 |
|
12 |
|
13 |
With the net outcome being that: |
14 |
|
15 |
- in-source builds are theoretically safer for weird build system hacks, |
16 |
|
17 |
- in-source build ebuilds are harder to write, |
18 |
|
19 |
- switching between one and the other requires a fair amount of changes. |
20 |
|
21 |
|
22 |
If we consider then that enabling in-source build should be considered |
23 |
a quick and possibly temporary hack to get the package to build, points |
24 |
2/3 seem to really outweigh the benefit of 1. |
25 |
|
26 |
For that reason, I'd like to 'unify' the behavior between the two. |
27 |
The patches which I will send in reply to this mail will: |
28 |
|
29 |
|
30 |
1) make BUILD_DIR point to the 'build' subdirectory in in-source builds, |
31 |
|
32 |
2) override build locations the same way in both kinds of builds, |
33 |
|
34 |
3) set PYTHONPATH in in-source builds, |
35 |
|
36 |
4) [optional] run *_all() phases in implementation-specific sources |
37 |
copy rather than ${S}. |
38 |
|
39 |
|
40 |
This is an incompatible change but seems to be the least troublesome |
41 |
possibility. Most importantly, any ebuild using BUILD_DIR |
42 |
in python_*_all() phase will become at least partially broken. |
43 |
|
44 |
Hopefully, this ain't as bad as it seems. Those ebuilds which used |
45 |
BUILD_DIR there to append to PYTHONPATH will still work. While |
46 |
the appending will no longer be correct, the newly-introduced |
47 |
PYTHONPATH will take precedence and solve the issue for now. |
48 |
|
49 |
There is also the potential of packages being broken because of |
50 |
--build-* overrides. But the most common case of relying on standard |
51 |
locations would probably still work because the new paths will mostly be |
52 |
absolute-path variants of the default locations. |
53 |
|
54 |
In any case, I'm testing this a bit now and will let you know what gets |
55 |
broken. |
56 |
|
57 |
|
58 |
Rationale: |
59 |
|
60 |
The code appends '/build' subdirectory to BUILD_DIR in distutils-r1 |
61 |
(python-r1's BUILD_DIR still points to the sources). This seemed like |
62 |
the best possible way of achieving the goal since: |
63 |
|
64 |
1) BUILD_DIR is not used commonly in in-source build ebuilds, |
65 |
|
66 |
2) using ${BUILD_DIR}/lib with BUILD_DIR pointing to source root is |
67 |
likely to collide with package sources, |
68 |
|
69 |
3) using ${BUILD_DIR}/build/lib conditionally would cause |
70 |
incompatibility with out-of-source builds, |
71 |
|
72 |
4) using ${BUILD_DIR}/build/lib unconditionally would break a fair |
73 |
number of out-of-source build ebuilds using '${BUILD_DIR}/lib'. |
74 |
|
75 |
Therefore, appending the 'build' subdir to BUILD_DIR seems like |
76 |
the least evil necessary to fix the past mistake. |
77 |
|
78 |
-- |
79 |
Best regards, |
80 |
Michał Górny |