On 06/14/2013 05:42 AM, Mattias Ellert wrote:
fre 2013-06-14 klockan 01:14 +0300 skrev Susi Lehtola:
On Thu, 13 Jun 2013 22:49:11 +0200 Mattias Ellert mattias.ellert@fysast.uu.se wrote:
tor 2013-06-13 klockan 11:53 -0700 skrev Toshio Kuratomi:
The FPC discussed this today and added a prohibition to using %{_isa} in BuildRequires to the Guidelines:
Please reconsider this. A specfile without isa in BuildRequires is broken for exactly the same reason a binary rpm without isa in Requires is broken. Not all packages the provide the BR are suitable to fulfil it for the purpose of providing the resources necessary for doing the build.
The difference is that BuildRequires are only relevant on the build system, where the correct architecture will be pulled in by the BuildRequire. Remember, the build environment is prepped separately for each build.
I disagree with this. The BuildRequires are relevant for users wanting to rebuild packages on their own machines. Without isa this is severely broken.
Also, as is noted in the guidelines https://fedoraproject.org/wiki/Packaging:Guidelines#BuildRequires_and_.25.7B... if you use %_isa in the BuildRequires, then e.g. a srpm built on x86_64 won't work on i386.
This "won't work" is an exaggeration. There are a few glitches in some cases, but these are minor compared to the problems you get by not having proper BRs by not using isa.
For binary RPMs the situation is very different - you can't assume anything about the state of the system. %_isa is needed for the case where the system already has, say libfoo(x86-32) installed, and then you install foo(x86-64) that dlopens libfoo. You need the %_isa in the binary rpm requires to make sure that the compatible library gets installed, although the libfoo package already is present on the system.
You can not make assumptions about what packages a user has installed on the system where packages are built.
Exactly. Especially if you take into account that the sane way to build packages in a controlled environment is to use mock.
Users rebuild packages on the same systems as they install binary packages - the same issues arise.
This is definitely not true. No sane person would keep all the needed -devel packages on the final machines where the built binaries will be used. I for one have completely different machines in terms of hardware and installed packages. The build farm ( where mock is used ) has significantly different memory and storage requiements compared to the production machines. And except for python and maybe perl, no build tools ( especially gcc, make & Co.) exist on the production machines.