revisiting the architecture support on Power

[Dan Horák]

Hi,

there were some changes in the toolchain for Power so it's time to
discuss them also on the list. There are multiple components involved,
there are defaults and additional optimized runtimes, new hardware
appearing, etc so the whole is quite complicated, but still should be
kept in sync.

The components are
- the compiler (gcc)
- the C library (glibc)
- the distro-wide default (redhat-rpm-config)

The Fedora architectures are
- ppc64 - big endian, historically supported power4/ppc970 and up,
during the time ppc64p7 subarch was added to provide Power7
(and -O3) optimized libs/binaries (about 25 packages)

- ppc64le - the new little endian variant, we have started with Power7
(and up) because it was the hardware we used in the buildsystem

The current distro defaults are (/usr/lib/rpm/redhat/rpmrc)
optflags: ppc64 %{__global_cflags} -m64
optflags: ppc64p7 %{__global_cflags} -m64 -O3 -mcpu=power7 -mtune=power7
optflags: ppc64le %{__global_cflags} -m64 -mcpu=power7 -mtune=power8

And gcc has its own defaults which can be set during the build.

The Fedora infrastructure is currently using Power8 HW for the main Koji
build system [1] to produce the binary rpms for the distrobution and we
have planned to reuse the previous Power7+ builders for COPR [2] and
Fedora Cloud.

Carlos, Florian, could you please comment on the situation?

[1] http://ppc.koji.fedoraproject.org/
[2] https://copr.fedoraproject.org/


	Thanks

		Dan

[Florian Weimer]

On 01/20/2016 12:16 PM, Dan Horák wrote:

>  The Fedora infrastructure is currently using Power8 HW for the main
> Koji
>  build system [1] to produce the binary rpms for the distrobution and we
>  have planned to reuse the previous Power7+ builders for COPR [2] and
>  Fedora Cloud.
>  
>  Carlos, Florian, could you please comment on the situation? 
Dan,

I think there is a fundamental misunderstanding here.  As long as their
is active upstream support, the glibc team will provide whatever
architectures you request, and make a best-effort attempt at isolating
issues.  (Personally, I lack ppc64 assembly knowledge, so it's rare that
I can fix architecture-specific bugs.)

We declined to fix bug 1295797 because the Fedora POWER wiki page

<https://fedoraproject.org/wiki/Architectures/PowerPC#Supported_Architectu...

clearly says that on ppc64le, the minimum architecture requirement is
POWER8.  If this is not what you want, please update the wiki page, make
sure that both gcc, redhat-rpm-config, and binutils have been configured
in the appropriate way, and rebuild glibc to see if the problematic
instructions are still there.  If you have trouble getting the toolchain
side sorted out, please let me know.  Bug 1295797 appears to be a
toolchain configuration issue because the offending file is not compiled
with -mcpu=power8 or anything like that (assuming that the debugging
information can be trusted, which says that it's not the POWER8 memcpy
which triggers the crash, but the C implementation).

But the key task for you is that you need to decide what you want and
document that.  On the glibc side, we do not have a strong preference
for POWER8.  POWER7 should just work fine for ppc64le with a
properly-configured toolchain.  (Earlier architecture versions may be
more difficult, I have not checked that.)

Florian

[Gustavo Luiz Duarte]

On 01/20/2016 10:13 AM, Florian Weimer wrote:
>  On 01/20/2016 12:16 PM, Dan Horák wrote:
>
>  > The Fedora infrastructure is currently using Power8 HW for the main Koji
>  > build system [1] to produce the binary rpms for the distrobution and we
>  > have planned to reuse the previous Power7+ builders for COPR [2] and
>  > Fedora Cloud.
>  >
>  > Carlos, Florian, could you please comment on the situation?
>
>  Dan,
>
>  I think there is a fundamental misunderstanding here.  As long as their
>  is active upstream support, the glibc team will provide whatever
>  architectures you request, and make a best-effort attempt at isolating
>  issues.  (Personally, I lack ppc64 assembly knowledge, so it's rare that
>  I can fix architecture-specific bugs.)
>
>  We declined to fix bug 1295797 because the Fedora POWER wiki page
>
>  <https://fedoraproject.org/wiki/Architectures/PowerPC#Supported_Architectu...
>
>  clearly says that on ppc64le, the minimum architecture requirement is
>  POWER8.  If this is not what you want, please update the wiki page, make
>  sure that both gcc, redhat-rpm-config, and binutils have been configured
>  in the appropriate way, and rebuild glibc to see if the problematic
>  instructions are still there.  If you have trouble getting the toolchain
>  side sorted out, please let me know.  Bug 1295797 appears to be a
>  toolchain configuration issue because the offending file is not compiled
>  with -mcpu=power8 or anything like that (assuming that the debugging
>  information can be trusted, which says that it's not the POWER8 memcpy
>  which triggers the crash, but the C implementation).
>
>  But the key task for you is that you need to decide what you want and
>  document that.  On the glibc side, we do not have a strong preference
>  for POWER8.  POWER7 should just work fine for ppc64le with a
>  properly-configured toolchain.  (Earlier architecture versions may be
>  more difficult, I have not checked that.) 
I understand that there are benefits to moving up to POWER8, and that includes performance
gain and we will be safer in a completely supported setup (supported by IBM and upstream
communities).

On the other hand, the fact that POWER7 is able to run little endian has helped Fedora a
lot (we built the whole ppc64le distro solely based on POWER7). Moving to POWER8 as a
requirement now would have a great impact on those people who are in the same situation we
were not long ago, relying on POWER7 hardware to run little endian.

I vote for updating the wiki and keep the POWER7 support on ppc64le for now. We can always
re-evaluate this decision in the future if we like.

[]'s
Gustavo

>  Florian
>  _______________________________________________
>  ppc mailing list
>  ppc(a)lists.fedoraproject.org
>  http://lists.fedoraproject.org/admin/lists/[e-mail]
>

[Steven Munroe]

On Wed, 2016-01-20 at 13:13 +0100, Florian Weimer wrote:
>  On 01/20/2016 12:16 PM, Dan Horák wrote:
>  
>  > The Fedora infrastructure is currently using Power8 HW for the main Koji
>  > build system [1] to produce the binary rpms for the distrobution and we
>  > have planned to reuse the previous Power7+ builders for COPR [2] and
>  > Fedora Cloud.
>  > 
>  > Carlos, Florian, could you please comment on the situation?
>  
>  Dan,
>  
>  I think there is a fundamental misunderstanding here.  As long as their
>  is active upstream support, the glibc team will provide whatever
>  architectures you request, and make a best-effort attempt at isolating
>  issues.  (Personally, I lack ppc64 assembly knowledge, so it's rare that
>  I can fix architecture-specific bugs.)
>  
>  We declined to fix bug 1295797 because the Fedora POWER wiki page
>  
>  <https://fedoraproject.org/wiki/Architectures/PowerPC#Supported_Architectu...
>  
>  clearly says that on ppc64le, the minimum architecture requirement is
>  POWER8.  If this is not what you want, please update the wiki page, make
>  sure that both gcc, redhat-rpm-config, and binutils have been configured
>  in the appropriate way, and rebuild glibc to see if the problematic
>  instructions are still there.  If you have trouble getting the toolchain
>  side sorted out, please let me know.  Bug 1295797 appears to be a
>  toolchain configuration issue because the offending file is not compiled
>  with -mcpu=power8 or anything like that (assuming that the debugging
>  information can be trusted, which says that it's not the POWER8 memcpy
>  which triggers the crash, but the C implementation).
>  
>  But the key task for you is that you need to decide what you want and
>  document that.  On the glibc side, we do not have a strong preference
>  for POWER8.  POWER7 should just work fine for ppc64le with a
>  properly-configured toolchain.  (Earlier architecture versions may be
>  more difficult, I have not checked that.)
>   
Actually POWER7 does not work just fine for ppc64le. Unaligned
load/stores take a big performance hit for micro-code in LE mode. 

POWER7LE require special firmware and was only intended as a lab/disro
bring up platform for LE development. IBM does not support POWER7LE for
production use.

POWER8 was designed from the ground up for LE and is the first
officially supported LE platform for general availability.

[Michael J Wolf]

Gustavo Luiz Duarte <gustavold(a)linux.vnet.ibm.com&gt; wrote on 01/20/2016
08:51:06 AM:

>  From: Gustavo Luiz Duarte <gustavold(a)linux.vnet.ibm.com&gt;
>  To: Florian Weimer <fweimer(a)redhat.com&gt;
>  Cc: ppc(a)lists.fedoraproject.org, Dan Horák <dan(a)danny.cz&gt;, 
carlos(a)redhat.com
>  Date: 01/20/2016 08:57 AM
>  Subject: Re: revisiting the architecture support on Power
>
>
>  On 01/20/2016 10:13 AM, Florian Weimer wrote:
>  > On 01/20/2016 12:16 PM, Dan Horák wrote:
>  >
>  > > The Fedora infrastructure is currently using Power8 HW for the main 
Koji
>  > > build system [1] to produce the binary rpms for the
> distrobution and 
we
>  > > have planned to reuse the previous Power7+ builders for
> COPR [2] and
>  > > Fedora Cloud.
>  > >
>  > > Carlos, Florian, could you please comment on the situation?
>  >
>  > Dan,
>  >
>  > I think there is a fundamental misunderstanding here.  As long as their
>  > is active upstream support, the glibc team will provide whatever
>  > architectures you request, and make a best-effort attempt at isolating
>  > issues.  (Personally, I lack ppc64 assembly knowledge, so it's rare 
that
>  > I can fix architecture-specific bugs.)
>  >
>  > We declined to fix bug 1295797 because the Fedora POWER wiki page
>  >
>  > <https://fedoraproject.org/wiki/Architectures/
>  PowerPC#Supported_Architectures>
>  >
>  > clearly says that on ppc64le, the minimum architecture requirement is
>  > POWER8.  If this is not what you want, please update the wiki page, 
make
>  > sure that both gcc, redhat-rpm-config, and binutils have been
configured
>  > in the appropriate way, and rebuild glibc to see if the
> problematic
>  > instructions are still there.  If you have trouble getting the 
toolchain
>  > side sorted out, please let me know.  Bug 1295797 appears to be
> a
>  > toolchain configuration issue because the offending file is not 
compiled
>  > with -mcpu=power8 or anything like that (assuming that the
> debugging
>  > information can be trusted, which says that it's not the POWER8 memcpy
>  > which triggers the crash, but the C implementation).
>  >
>  > But the key task for you is that you need to decide what you want and
>  > document that.  On the glibc side, we do not have a strong preference
>  > for POWER8.  POWER7 should just work fine for ppc64le with a
>  > properly-configured toolchain.  (Earlier architecture versions may be
>  > more difficult, I have not checked that.)
>
>  I understand that there are benefits to moving up to POWER8, and
>  that includes performance gain and we will be safer in a completely
>  supported setup (supported by IBM and upstream communities).
>
>  On the other hand, the fact that POWER7 is able to run little endian
>  has helped Fedora a lot (we built the whole ppc64le distro solely
>  based on POWER7). Moving to POWER8 as a requirement now would have a
>  great impact on those people who are in the same situation we were
>  not long ago, relying on POWER7 hardware to run little endian.
>
>  I vote for updating the wiki and keep the POWER7 support on ppc64le
>  for now. We can always re-evaluate this decision in the future if we 
like.

The wiki page is correct.  IBM will not be supporting LE on the P7+
architecture.
There are no customers using that setup, it requires a special version of
FW in
order to run as well.  So from an IBM point of view LE is support on Power
8 and
newer systems only.

As far as I know Power7 LE is only in the labs for build machines etc.  I
will work
with the Fedora release and infrastructure folks to see what P8 only will
mean to them.

thanks
Mike Wolf
>  []'s
>  Gustavo
>
>  >
>  > Florian
>  > _______________________________________________
>  > ppc mailing list
>  > ppc(a)lists.fedoraproject.org
>  > http://lists.fedoraproject.org/admin/lists/[e-mail]
>  >
>  _______________________________________________
>  ppc mailing list
>  ppc(a)lists.fedoraproject.org
>  http://lists.fedoraproject.org/admin/lists/[e-mail]

[Gustavo Luiz Duarte]

On 01/20/2016 03:34 PM, Michael J Wolf wrote:
>  Gustavo Luiz Duarte <gustavold(a)linux.vnet.ibm.com&gt; wrote on
> 01/20/2016
>  08:51:06 AM:
>
>   > From: Gustavo Luiz Duarte <gustavold(a)linux.vnet.ibm.com&gt;
>   > To: Florian Weimer <fweimer(a)redhat.com&gt;
>   > Cc: ppc(a)lists.fedoraproject.org, Dan Horák <dan(a)danny.cz&gt;,
>  carlos(a)redhat.com
>   > Date: 01/20/2016 08:57 AM
>   > Subject: Re: revisiting the architecture support on Power
>   >
>   >
>   > On 01/20/2016 10:13 AM, Florian Weimer wrote:
>   > > On 01/20/2016 12:16 PM, Dan Horák wrote:
>   > >
>   > > > The Fedora infrastructure is currently using Power8 HW for the
>  main Koji
>   > > > build system [1] to produce the binary rpms for the distrobution
>  and we
>   > > > have planned to reuse the previous Power7+ builders for COPR [2] and
>   > > > Fedora Cloud.
>   > > >
>   > > > Carlos, Florian, could you please comment on the situation?
>   > >
>   > > Dan,
>   > >
>   > > I think there is a fundamental misunderstanding here.  As long as their
>   > > is active upstream support, the glibc team will provide whatever
>   > > architectures you request, and make a best-effort attempt at isolating
>   > > issues.  (Personally, I lack ppc64 assembly knowledge, so it's rare
>  that
>   > > I can fix architecture-specific bugs.)
>   > >
>   > > We declined to fix bug 1295797 because the Fedora POWER wiki page
>   > >
>   > > <https://fedoraproject.org/wiki/Architectures/
>   > PowerPC#Supported_Architectures>
>   > >
>   > > clearly says that on ppc64le, the minimum architecture requirement is
>   > > POWER8.  If this is not what you want, please update the wiki page,
>  make
>   > > sure that both gcc, redhat-rpm-config, and binutils have been
>  configured
>   > > in the appropriate way, and rebuild glibc to see if the problematic
>   > > instructions are still there.  If you have trouble getting the
>  toolchain
>   > > side sorted out, please let me know.  Bug 1295797 appears to be a
>   > > toolchain configuration issue because the offending file is not
>  compiled
>   > > with -mcpu=power8 or anything like that (assuming that the debugging
>   > > information can be trusted, which says that it's not the POWER8 memcpy
>   > > which triggers the crash, but the C implementation).
>   > >
>   > > But the key task for you is that you need to decide what you want and
>   > > document that.  On the glibc side, we do not have a strong preference
>   > > for POWER8.  POWER7 should just work fine for ppc64le with a
>   > > properly-configured toolchain.  (Earlier architecture versions may be
>   > > more difficult, I have not checked that.)
>   >
>   > I understand that there are benefits to moving up to POWER8, and
>   > that includes performance gain and we will be safer in a completely
>   > supported setup (supported by IBM and upstream communities).
>   >
>   > On the other hand, the fact that POWER7 is able to run little endian
>   > has helped Fedora a lot (we built the whole ppc64le distro solely
>   > based on POWER7). Moving to POWER8 as a requirement now would have a
>   > great impact on those people who are in the same situation we were
>   > not long ago, relying on POWER7 hardware to run little endian.
>   >
>   > I vote for updating the wiki and keep the POWER7 support on ppc64le
>   > for now. We can always re-evaluate this decision in the future if we
>  like.
>
>  The wiki page is correct.  IBM will not be supporting LE on the P7+
>  architecture.
>  There are no customers using that setup, it requires a special version
>  of FW in
>  order to run as well.  
I didn't think of this detail... because of the special firmware required, very few
people are relying on POWER7 for little endian... then I agree that keeping POWER7 support
makes much less sense. So I change my vote :)

[]'s
Gustavo

>  So from an IBM point of view LE is support on
>  Power 8 and
>  newer systems only.
>
>  As far as I know Power7 LE is only in the labs for build machines etc.
>    I will work
>  with the Fedora release and infrastructure folks to see what P8 only
>  will mean to them.
>
>  thanks
>  Mike Wolf
>   >
>   > []'s
>   > Gustavo
>   >
>   > >
>   > > Florian
>   > > _______________________________________________
>   > > ppc mailing list
>   > > ppc(a)lists.fedoraproject.org
>   > > http://lists.fedoraproject.org/admin/lists/[e-mail]
>   > >
>   > _______________________________________________
>   > ppc mailing list
>   > ppc(a)lists.fedoraproject.org
>   > http://lists.fedoraproject.org/admin/lists/[e-mail]
>
>
>
>  _______________________________________________
>  ppc mailing list
>  ppc(a)lists.fedoraproject.org
>  http://lists.fedoraproject.org/admin/lists/[e-mail]
>

[Carlos O'Donell]

On 01/20/2016 06:16 AM, Dan Horák wrote:
>  The current distro defaults are (/usr/lib/rpm/redhat/rpmrc)
>  optflags: ppc64 %{__global_cflags} -m64
>  optflags: ppc64p7 %{__global_cflags} -m64 -O3 -mcpu=power7 -mtune=power7
>  optflags: ppc64le %{__global_cflags} -m64 -mcpu=power7 -mtune=power8 
The ppc64le case needs to become '-mcpu=power8 -mtune=power8', as Steven
Munroe notes the only official ppc64le support is with POWER8 hardware.

>  And gcc has its own defaults which can be set during the build.
>  
>  The Fedora infrastructure is currently using Power8 HW for the main Koji
>  build system [1] to produce the binary rpms for the distrobution and we
>  have planned to reuse the previous Power7+ builders for COPR [2] and
>  Fedora Cloud. 
This will mean that COPR cannot build POWER8 enabled binaries, and that
would be a shame for users of newer hardware.
 
>  Carlos, Florian, could you please comment on the situation? 
At present glibc in fedora rawhide requires POWER8 hardware.

Why? We build optimized variants of the runtime for POWER8 hardware and
need POWER8 hardware to test the resulting binaries (native build).

This means that glibc for fedora rawhide will be unbuildable in COPR
until COPR is upgraded to POWER8 machines.

However, we can change that, as Florian suggets, it's up to the Fedora
Power community to make a decision as to what is needed to best serve the
IBM hardware users.

We currently have the following support in place (from glibc.spec in rawhide):

# 64-bit BE:
# - Power 620 / 970 ISA (default runtime, compatile with POWER4 and newer)
#       - Provided for the large number of PowerPC G5 users.
#       - IFUNC support provides optimized core routines for POWER6,
#         POWER7, and POWER8 transparently (if not using specific runtimes
#         below)
# - POWER6 (has power6x symlink to power6, enabled via AT_PLATFORM)
#       - Legacy for old systems. Should be deprecated at some point soon.
# - POWER7 (enabled via AT_PLATFORM)
#       - Existing deployments.
# - POWER8 (enabled via AT_PLATFORM)
#       - Latest generation.
# 64-bit LE:
# - POWER8 LE (default)
#       - Latest generation.

Whatever decisions you make, please tell us and we can adjust the library.

Cheers,
Carlos.

[Dan Horák]

On Wed, 20 Jan 2016 13:46:53 -0500
"Carlos O'Donell" <carlos(a)redhat.com&gt; wrote:

>  On 01/20/2016 06:16 AM, Dan Horák wrote:
>  > The current distro defaults are (/usr/lib/rpm/redhat/rpmrc)
>  > optflags: ppc64 %{__global_cflags} -m64
>  > optflags: ppc64p7 %{__global_cflags} -m64 -O3 -mcpu=power7
>  > -mtune=power7 optflags: ppc64le %{__global_cflags} -m64
>  > -mcpu=power7 -mtune=power8
>  
>  The ppc64le case needs to become '-mcpu=power8 -mtune=power8', as
>  Steven Munroe notes the only official ppc64le support is with POWER8
>  hardware.
>  
>  > And gcc has its own defaults which can be set during the build.
>  > 
>  > The Fedora infrastructure is currently using Power8 HW for the main
>  > Koji build system [1] to produce the binary rpms for the
>  > distrobution and we have planned to reuse the previous Power7+
>  > builders for COPR [2] and Fedora Cloud.
>  
>  This will mean that COPR cannot build POWER8 enabled binaries, and
>  that would be a shame for users of newer hardware.
>   
>  > Carlos, Florian, could you please comment on the situation?
>  
>  At present glibc in fedora rawhide requires POWER8 hardware.
>  
>  Why? We build optimized variants of the runtime for POWER8 hardware
>  and need POWER8 hardware to test the resulting binaries (native
>  build).
>  
>  This means that glibc for fedora rawhide will be unbuildable in COPR
>  until COPR is upgraded to POWER8 machines.
>  
>  However, we can change that, as Florian suggets, it's up to the Fedora
>  Power community to make a decision as to what is needed to best serve
>  the IBM hardware users.
>  
>  We currently have the following support in place (from glibc.spec in
>  rawhide):
>  
>  # 64-bit BE:
>  # - Power 620 / 970 ISA (default runtime, compatile with POWER4 and
>  # newer)
>  #       - Provided for the large number of PowerPC G5 users.
>  #       - IFUNC support provides optimized core routines for POWER6,
>  #         POWER7, and POWER8 transparently (if not using specific
>  # runtimes below)
>  # - POWER6 (has power6x symlink to power6, enabled via AT_PLATFORM)
>  #       - Legacy for old systems. Should be deprecated at some point
>  # soon.
>  # - POWER7 (enabled via AT_PLATFORM)
>  #       - Existing deployments.
>  # - POWER8 (enabled via AT_PLATFORM)
>  #       - Latest generation.
>  # 64-bit LE:
>  # - POWER8 LE (default)
>  #       - Latest generation.
>  
>  Whatever decisions you make, please tell us and we can adjust the
>  library. 
sorry for the late reply, but Fedora PPC team would like keep the
support for Power7 for a while until we will able to replace the
remaining older HW with Power8 one. We have already started the required
actions, but as you can understand it will take some time. I personally
expect we will be able to switch F-25 to pure Power8.

We do not recommend the users to anything else then Power8, that's why
the "supported" state for Power8 in the wiki.

64-bit BE - sounds good

64-bit LE - please add back the Power7 support, and keep the Power8
optimized one, so users can take advantage of their hw


	Thanks

		Dan

[Florian Weimer]

On 01/29/2016 03:55 PM, Dan Horák wrote:

>  64-bit LE - please add back the Power7 support, and keep the Power8
>  optimized one, so users can take advantage of their hw 
As far as I can tell, upstream GCC does not support POWER7 with ppc64le,
or the firmware-based emulation is broken.  Based on my review, we are
not using any special flags to request POWER8 in gcc or glibc, and yet
the object code produced by GCC does not work on your POWER7 machines
(or maybe only some of them?).

Or put differently, based on IBM's comments in this thread, I don't see
any reasonable way to address bug 1295797.

Florian

[Dan Horák]

On Fri, 29 Jan 2016 15:59:07 +0100
Florian Weimer <fweimer(a)redhat.com&gt; wrote:

>  On 01/29/2016 03:55 PM, Dan Horák wrote:
>  
>  > 64-bit LE - please add back the Power7 support, and keep the Power8
>  > optimized one, so users can take advantage of their hw
>  
>  As far as I can tell, upstream GCC does not support POWER7 with
>  ppc64le, or the firmware-based emulation is broken.  Based on my
>  review, we are not using any special flags to request POWER8 in gcc
>  or glibc, and yet the object code produced by GCC does not work on
>  your POWER7 machines (or maybe only some of them?).
>  
>  Or put differently, based on IBM's comments in this thread, I don't
>  see any reasonable way to address bug 1295797. 
might be related to the change gcc made in their spec file, moving the
defaults from
with-cpu-32=power7 --with-tune-32=power8 --with-cpu-64=power7
--with-tune-64=power8
to
http://pkgs.fedoraproject.org/cgit/rpms/gcc.git/tree/gcc.spec?h=f23#n1009
and
http://pkgs.fedoraproject.org/cgit/rpms/gcc.git/tree/gcc.spec#n1002


		Dan

[Carlos O'Donell]

On 01/29/2016 09:55 AM, Dan Horák wrote:
>  sorry for the late reply, but Fedora PPC team would like keep the
>  support for Power7 for a while until we will able to replace the
>  remaining older HW with Power8 one. We have already started the required
>  actions, but as you can understand it will take some time. I personally
>  expect we will be able to switch F-25 to pure Power8. 
Certainly.

What I will do is a little bit different.

How does the following patch look to you?

diff --git a/glibc.spec b/glibc.spec
index 8fc1617..5910b14 100644
--- a/glibc.spec
+++ b/glibc.spec
@@ -885,11 +885,21 @@ build nosegneg -mno-tls-direct-seg-refs
 )
 %endif
 
+#
+# The Fedora POWER community has requested we continue to support doing
+# POWER8 builds on POWER7 hardware.
+# The request is here by Dan Horak:
+#
https://lists.fedoraproject.org/archives/list/[e-mail]...
+#
+# To make the eventual transition to POWER8 easier we change the -mcpu
+# flags and the --with-cpu flag to power7 for the POWER8 build. That means
+# we still only use POWER7 optimized code, but tune for POWER8.
+#
 %if %{buildpower8}
 (
-  AddOns="$AddOns --with-cpu=power8"
-  GCC="$GCC -mcpu=power8 -mtune=power8"
-  GXX="$GXX -mcpu=power8 -mtune=power8"
+  AddOns="$AddOns --with-cpu=power7"
+  GCC="$GCC -mcpu=power7 -mtune=power8"
+  GXX="$GXX -mcpu=power7 -mtune=power8"
   build power8
 )
 %endif
---

This does not address the 64-bit LE POWER8 issues as pointed out by Florian.

Cheers,
Carlos.

[Steve Munroe]

"Carlos O'Donell" <carlos(a)redhat.com&gt; wrote on 01/29/2016 09:26:36
AM:

>  From: "Carlos O'Donell" <carlos(a)redhat.com
> >  
>  To: Dan Horák <dan(a)danny.cz
> >  
>  Cc:
> ppc(a)lists.fedoraproject.org, fweimer(a)redhat.com, Steve Munroe/
>  Rochester/IBM@IBMUS
>  Date: 01/29/2016 09:26 AM
>  Subject: Re: revisiting the architecture support on Power
>
>  On 01/29/2016 09:55 AM, Dan Horák wrote:
>  > sorry for the late reply, but Fedora PPC team would like keep the
>  > support for Power7 for a while until we will able to replace the
>  > remaining older HW with Power8 one. We have already started the 
required
>  > actions, but as you can understand it will take some time. I
> personally
>  > expect we will be able to switch F-25 to pure Power8.
>
>  Certainly.
>  
I would really like to understand why we have to support P7LE for another
release.

Especial as I know that Mike Wolf is working on getting the new P8 hardware
for COPR.

>  What I will do is a little bit different.
>
>  How does the following patch look to you?
>
>  diff --git a/glibc.spec b/glibc.spec
>  index 8fc1617..5910b14 100644
>  --- a/glibc.spec
>  +++ b/glibc.spec
>  @@ -885,11 +885,21 @@ build nosegneg -mno-tls-direct-seg-refs
>   )
>   %endif
>
>  +#
>  +# The Fedora POWER community has requested we continue to support doing
>  +# POWER8 builds on POWER7 hardware.
>  +# The request is here by Dan Horak:
>  +# https://lists.fedoraproject.org/archives/list/
>  ppc(a)lists.fedoraproject.org/thread/NAF5I5Z7ZZG7E6IHUY5CGP7UY6GF2JQM/
>  +#Steven J. Munroe 
Linux on Power Toolchain Architect
IBM Corporation, Linux Technology Center

>  +# To make the eventual transition to POWER8 easier we change the
> -mcpu
>  +# flags and the --with-cpu flag to power7 for the POWER8 build. That 
means
>  +# we still only use POWER7 optimized code, but tune for POWER8.
>  +#
>   %if %{buildpower8}
>   (
>  -  AddOns="$AddOns --with-cpu=power8"
>  -  GCC="$GCC -mcpu=power8 -mtune=power8"
>  -  GXX="$GXX -mcpu=power8 -mtune=power8"
>  +  AddOns="$AddOns --with-cpu=power7"
>  +  GCC="$GCC -mcpu=power7 -mtune=power8"
>  +  GXX="$GXX -mcpu=power7 -mtune=power8"
>     build power8
>   )
>   %endif
>  ---
>
>  This does not address the 64-bit LE POWER8 issues as pointed out by 
Florian.
>  
Yes the P7+ is not a supported platform for PowerPC Little
Endian. Period!

And the longer we keep is zombie alive the more problems (bugz in GCC and
GLIBC) will accumulate.

Steven J. Munroe
Linux on Power Toolchain Architect
IBM Corporation, Linux Technology Center


> Cheers,
> Carlos.
>

[Al Dunsmuir]

On Friday, January 29, 2016, 9:55:10 AM, Dan Dan Horák wrote:

>  On Wed, 20 Jan 2016 13:46:53 -0500
>  "Carlos O'Donell" <carlos(a)redhat.com&gt; wrote: 
> > On 01/20/2016 06:16 AM, Dan Horák wrote:
> > > The current distro defaults are (/usr/lib/rpm/redhat/rpmrc)
> > > optflags: ppc64 %{__global_cflags} -m64
> > > optflags: ppc64p7 %{__global_cflags} -m64 -O3 -mcpu=power7
> > > -mtune=power7 optflags: ppc64le %{__global_cflags} -m64
> > > -mcpu=power7 -mtune=power8
> > 
> > The ppc64le case needs to become '-mcpu=power8 -mtune=power8', as
> > Steven Munroe notes the only official ppc64le support is with POWER8
> > hardware.
> > 
> > > And gcc has its own defaults which can be set during the build.
> > > 
> > > The Fedora infrastructure is currently using Power8 HW for the main
> > > Koji build system [1] to produce the binary rpms for the
> > > distrobution and we have planned to reuse the previous Power7+
> > > builders for COPR [2] and Fedora Cloud.
> > 
> > This will mean that COPR cannot build POWER8 enabled binaries, and
> > that would be a shame for users of newer hardware.
> >  
> > > Carlos, Florian, could you please comment on the situation?
> > 
> > At present glibc in fedora rawhide requires POWER8 hardware.
> > 
> > Why? We build optimized variants of the runtime for POWER8 hardware
> > and need POWER8 hardware to test the resulting binaries (native
> > build).
> > 
> > This means that glibc for fedora rawhide will be unbuildable in COPR
> > until COPR is upgraded to POWER8 machines.
> > 
> > However, we can change that, as Florian suggets, it's up to the Fedora
> > Power community to make a decision as to what is needed to best serve
> > the IBM hardware users.
> > 
> > We currently have the following support in place (from glibc.spec in
> > rawhide):
> > 
> > # 64-bit BE:
> > # - Power 620 / 970 ISA (default runtime, compatile with POWER4 and
> > # newer)
> > #       - Provided for the large number of PowerPC G5 users.
> > #       - IFUNC support provides optimized core routines for POWER6,
> > #         POWER7, and POWER8 transparently (if not using specific
> > # runtimes below)
> > # - POWER6 (has power6x symlink to power6, enabled via AT_PLATFORM)
> > #       - Legacy for old systems. Should be deprecated at some point
> > # soon.
> > # - POWER7 (enabled via AT_PLATFORM)
> > #       - Existing deployments.
> > # - POWER8 (enabled via AT_PLATFORM)
> > #       - Latest generation.
> > # 64-bit LE:
> > # - POWER8 LE (default)
> > #       - Latest generation.
> > 
> > Whatever decisions you make, please tell us and we can adjust the
> > library. 
>  sorry for the late reply, but Fedora PPC team would like keep the
>  support for Power7 for a while until we will able to replace the
>  remaining older HW with Power8 one. We have already started the required
>  actions, but as you can understand it will take some time. I personally
>  expect we will be able to switch F-25 to pure Power8. 
>  We do not recommend the users to anything else then Power8,
> that's why
>  the "supported" state for Power8 in the wiki. 
>  64-bit BE - sounds good 
>  64-bit LE - please add back the Power7 support, and keep the Power8
>  optimized one, so users can take advantage of their hw 

>          Thanks 
>                  Dan 
Dan,

I'm  only  a  Fedora  PPC  BE user, and see this discussion as largely
being about PPC LE (so of interest, but not directly relevant to me).

I  am  just bringing IBM Power4 p630 and a P5 550 boxes up (related to
work  on  AIX,  but  will  also  be  used with Fedora), in addition to
working on Mac G5. I did some work on the latter this summer, which is
going  to  have to iterate again (the main Parted developer wanted the
Mac  APM disk format handled with a hidden map partition vs my simpler
addition of new attribute flags).

To  clarify,  the  support for Fedora PPC-64 BE is remaining the same,
correct?

Thanks in advance,
Al

[Carlos O'Donell]

On 01/30/2016 06:56 AM, Al Dunsmuir wrote:
>  To  clarify,  the  support for Fedora PPC-64 BE is remaining the
> same,
>  correct? 
The support for Fedora PPC-64 BE is being extended to include optimized
POWER7 and POWER8 support. Existing support is not being removed.

Cheers,
Carlos.

[Al Dunsmuir]

On Saturday, January 30, 2016, 12:17:12 PM, Carlos O'Donell wrote:
>  On 01/30/2016 06:56 AM, Al Dunsmuir wrote:
> > To  clarify,  the  support for Fedora PPC-64 BE is remaining the same,
> > correct? 
>  The support for Fedora PPC-64 BE is being extended to include
> optimized
>  POWER7 and POWER8 support. Existing support is not being removed. 
Carlos,

Win-win for everyone - that is most excellent news!