1 |
On Tue, Jan 21, 2014 at 1:30 PM, Reinis Danne <rei4dan@×××××.com> wrote: |
2 |
|
3 |
> |
4 |
> |
5 |
> Ok, that is different issue. The question is whether we want, |
6 |
> e.g., eselect blas be build time only or also runtime setting. |
7 |
> |
8 |
> If we use provider specific sonames, then the program will use |
9 |
> whatever provider was selected at build time even if different |
10 |
> provider is set afterwards. Thus it is build time only setting, |
11 |
> with the caveat that package has to be rebuilt if selected |
12 |
> provider is removed (preserved-libs should catch this). |
13 |
> |
14 |
> ⬑ This is how it is now in the overlay. |
15 |
> |
16 |
|
17 |
One other advantage of keeping the generic names is also that maintainer |
18 |
would not have to modify the ebuild for packages to link with the selected |
19 |
provider. Most blas dependent packages assume there is /usr/lib/libblas.so. |
20 |
|
21 |
Some applications (I can see armadillo, scamp) actually use the |
22 |
atlas-specific routines clapack_* routines. |
23 |
We could aim at keeping both the library soname and filename as close to |
24 |
upstream, and add an option to the alternative framework to create a |
25 |
ldscript with a generic soname and filename, turned on for blas and cblas |
26 |
at least. |
27 |
|
28 |
As a user, I don't particularly enjoy rebuilding octave, ROOT, R, or the |
29 |
scipy stack. @preserved-rebuild is more a hack than a feature. |
30 |
I don't know how many people actually switch providers, besides a first |
31 |
benchmark test. My guess is probably not many given the burden of |
32 |
re-compiling all the reverse dependencies and the previous fragility of the |
33 |
system. |
34 |
|
35 |
Sebastien |