Comments on check-abi

[Andreas Jaeger <aj at suse dot de>]

I discussed at the GCCSummit with Benjamin Kosnik check-abi and
promised to send my ideas.  Sorry, for beeing late ;-(

Here're the problems I see with the current way:
- Unify names: We should unify the names of the directories and find
  the right one in case that somebody gives a different name to
  configure.  Also the i586-linux and i686-linux ABIs should be the
  same.  I'll send a patch implementing this in a separate email.

- Create baseline automatically if not exists: We should create the
  baseline automatically if it does not exist so that we encourage
  users to send the results in.

- We should always run the testsuite and consider a failure in
  check-abi a hard error.

- Add option to run with multilibs: The problem is that some compilers
  are multilib enabled, e.g. the MIPS compiler can support three
  multilibs.  For others the multilibs are other archs, e.g. the
  32-bit x86-64 multilib should have the same ABI as the i686
  multilib.

  Currently check-ABI is only run for the default ABI, the test is not
  multi-lib aware.

  I run on x86_64-linux-gnu the testsuite with
   RUNTESTFLAGS="--target_board 'unix{-m32,}'" 
  in the environment to test both 32-bit x86 and 64-bit x86-64 code
  generation.  This gives 32/ subdirectories when building and the
  testsuite handles this fine.

  My suggestion is to add MULTISUBDIR ("32/" for AMD64) to baseline_file in
  aclocal.m4 and then create the special ABI and let us test it

What do others think?

Andreas
-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj