[Feature] P9: Power9 CPU tuned libraries

[bugproxy]

glibc provides minimal support for POWER9 since version 2.23.
 glibc 2.25 is the first version to provide POWER9 optimizations via
 STT_GNU_IFUNC. The optimizations are enabled by default and will be disabled
 when using --disable-multi-arch (notice that multi-arch in glibc context is
 not the same thing as used by Debian).

 However, this feature request is proposing to provide a second glibc build,
 configured with --with-cpu=power9 and co-installed with the default glibc
 build, but under $(LIBDIR)/power9.
 Other distros usually install these under /lib64/power9. Ubuntu may need to
 use a different directory due to supporting Debian's multiarch.

 The support for this is known to work upstream since glibc 2.23.
 So, I think this is just pending changes in the distro build system.

[Frank Heimes (fheimes)]

Due to the fact that this ticket is tagged with 'severity-low' does it mean the 'Power9 CPU tuned libraries' can be included with 18.04 (rather than with 17.10)?

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2017-07-18 15:56 EDT-------
Canonical - no we need this to be in 17.10. The priority has been re-adjusted. Let me know if there is an issue.

[Matthias Klose (doko)]

I'd like to see things done for the toolchain in 17.10 already, so that this kind of stuff can be tested for the LTS. It's another thing if we can do it, not having p9 hardware...

[Adam Conrad (adconrad)]

As a general rule, I'd prefer not to have extra glibc builds. The bug report points out that most of the critical optimizations are handles by IFUNC, is there a solid argument for why we should have two builds as well? Given that the followup request would almost certainly be to ship both by default, we'd end up with a fair amount of filesystem bloat, and it's also a bit of a packaging nightmare (there was great rejoicing from my end when I killed libc6-i686, the last "tuned" build of glibc we had).

For i386 versus i686, there were some very reasonable arguments. The mandatory FPU, the extra registers, cmov, etc, etc. Generic compiler tuninng of i386 versus i686 gave very obvious real world performance increases. I'm dubious that this would be true for p8 versus p9.

[bugproxy (bugproxy)]

------- Comment From <email address hidden> 2017-09-05 09:53 EDT-------
> As a general rule, I'd prefer not to have extra glibc builds.
> ...
> (there was great rejoicing from my end when I
> killed libc6-i686, the last "tuned" build of glibc we had).

If this is a policy adopted by all architectures, there is no need to create an exception for POWER.

However, I kindly ask Canonical to contact us if/when another architecture starts having CPU-tuned libraries.

So, I'm changing the status of this feature request accordingly.

Thanks!