On Thu, 31 Mar 2022 at 23:46, Segher Boessenkool
[off-list ref] wrote:

On Thu, Mar 31, 2022 at 12:19:52PM -0300, Murilo Opsfelder Araújo wrote:

quoted

My understanding is that the default cpu type for -mcpu=powerpc64 can
change.

Different subtargets (Linux, AIX, Darwin, the various BSDs, bare ELF,
etc.) have different default CPUs.  It also can be set at configure time
for most subtargets.

Linux can be built with compilers not targetting *-linux*, so it would
be best to specify a specific CPU always.

quoted

quoted

I did a little test using my buildroot compiler which has
with-cpu=power10. I used the presence of PCREL relocations as evidence
that it was build for power10.

$ powerpc64le-buildroot-linux-gnu-gcc -mcpu=power10 -c test.c
$ readelf -r test.o |grep -c PCREL
24

It respected the -mcpu=power10 you provided.

Of course.

quoted

And that's my concern.  We're relying on the compiler default cpu type.

That is not the compiler default.  It is the default from who built the
compiler.  It can vary wildly and unpredictably.

The actual compiler default will not change so easily at all, basically
only when its subtarget drops support for an older CPU.

quoted

If gcc defaults -mcpu=powerpc64le to power10, you're going to have
the same problem again.

That will not happen before power10 is the minimum supported CPU.
Anything else is madness.

Murilo, does Segher's explanation address your concerns?

I believe the patch I sent fixes the problem that you're worried
about. It should be compatible into the future too.

Cheers,

Joel

quoted

It happens that the power8 default cpu type
is compatible to your system by coincidence.

No, power8 is (and always was) the minimum supported CPU type for
powerpc64le-linux.

quoted

In gcc/config/rs6000/rs6000-cpus.def, they are set to different processors:

    254 RS6000_CPU ("powerpc64", PROCESSOR_POWERPC64, MASK_PPC_GFXOPT |
    MASK_POWERPC64)
    255 RS6000_CPU ("powerpc64le", PROCESSOR_POWER8, MASK_POWERPC64 |
    ISA_2_7_MASKS_SERVER

Those can and will change though, over time.  But -mcpu=powerpc64 (etc.)
always will mean what the current documentation says it does:
     '-mcpu=powerpc', '-mcpu=powerpc64', and '-mcpu=powerpc64le' specify
     pure 32-bit PowerPC (either endian), 64-bit big endian PowerPC and
     64-bit little endian PowerPC architecture machine types, with an
     appropriate, generic processor model assumed for scheduling
     purposes.

quoted

My suggestion was to explicitly set -mcpu=power8 instead of
-mcpu=powerpc64le.

That is implied anyway, it is the minimum supported for
powerpc64le-linux.  Using -mcpu=powerpc64le might schedule better for
newer CPUs, in the future (but the code will always work on all still
supported CPUs).


Segher