Re: [kvm-unit-tests RFC PATCH 1/6] powerpc: add pmu tests

[Chinmay Rath <rathc@linux.ibm.com>]

On 6/10/26 17:03, Thomas Huth wrote:
> On 10/06/2026 11.27, Chinmay Rath wrote:
>>
>> On 6/8/26 19:27, Thomas Huth wrote:
>>> On 08/06/2026 14.45, Thomas Huth wrote:
>>>> On 02/06/2026 08.48, Chinmay Rath wrote:
>>>>> From: Nicholas Piggin <npiggin@gmail.com>
>>>>>
>>>>> Add some initial PMU testing.
>>>>>
>>>>> - PMC5/6 tests
>>>>> - PMAE / PMI test
>>>>> - BHRB basic tests
>>>>>
>>>>> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
>>>>> Signed-off-by: Chinmay Rath <rathc@linux.ibm.com>
>>>>> ---
>>>>>   lib/powerpc/asm/processor.h |   2 +
>>>>>   lib/powerpc/asm/reg.h       |   9 +
>>>>>   lib/powerpc/asm/setup.h     |   1 +
>>>>>   lib/powerpc/setup.c         |  20 ++
>>>>>   powerpc/Makefile.common     |   3 +-
>>>>>   powerpc/pmu.c               | 567 
>>>>> ++++++++++++++++++++++++++++++++++++
>>>>>   powerpc/unittests.cfg       |   3 +
>>>>>   7 files changed, 604 insertions(+), 1 deletion(-)
>>>>>   create mode 100644 powerpc/pmu.c
>>>>
>>>>   Hi Chinmay,
>>>>
>>>> the problem with Clang on Travis [*] still seems to persist:
>>>>
>>>>   https://app.travis-ci.com/github/huth/kvm-unit-tests/jobs/639614142
>>>>
>>>> Could you please have a look?
>>>>
>>>>   Thanks,
>>>>    Thomas
>>>>
>>>>
>>>> [*] This already happened with Nicolas' last version:
>>>>
>>>>   https://www.spinics.net/lists/kvm/msg351218.html
>>>
>>> I managed to get access to a ppc64 machine. The error is:
>>>
>>> /tmp/pmu-eab466.s: Assembler messages:
>>> /tmp/pmu-eab466.s:1649: Error: unrecognized opcode: `ldat'
>>> clang: error: assembler command failed with exit code 1 (use -v to 
>>> see invocation)
>>> make: *** [<builtin>: powerpc/pmu.o] Error 1
>>>
>>>  HTH,
>>>   Thomas
>> Hi Thomas,
>>
>> Thanks for looking into this and providing the exact error message.
>>
>> I was looking into this Travis CI job that you pointed to and noticed 
>> that the clang version being used was 14.
>> I was wondering, would it be possible to use a newer version of clang 
>> for the job since LDAT is a legit PPC insn,
>> that was introduced with Power 9, ISA version 3.0, way back a decade 
>> in 2016 !
>> So I was wondering if using a newer version of clang that recognizes 
>> the instruction would be a better approach.
>
> I can reproduce the very same issue on a ppc64le box with Clang 22:
>
> # clang --version
> clang version 22.1.6 (Fedora 22.1.6-1.fc44)
> Target: ppc64le-redhat-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
> Configuration file: /etc/clang/ppc64le-redhat-linux-gnu-clang.cfg
> # make
> clang -no-integrated-as -std=gnu99 -ffreestanding -O2 -msoft-float 
> -mno-altivec  -I /root/kvm-unit-tests/lib -I 
> /root/kvm-unit-tests/lib/libfdt -I lib -Wa,-mregnames -g -MMD -MP -MF 
> powerpc/.pmu.d -fno-strict-aliasing -fno-common -Wall -Wwrite-strings 
> -Wempty-body -Wuninitialized -Wignored-qualifiers -Wno-missing-braces 
> -Werror  -fomit-frame-pointer -fno-stack-protector    
> -Wno-frame-address   -fno-pic -Wunused-but-set-parameter  
> -Wno-override-init -Wmissing-prototypes -Wstrict-prototypes  
> -mlittle-endian   -c -o powerpc/pmu.o powerpc/pmu.c
> /tmp/pmu-f0c247.s: Assembler messages:
> /tmp/pmu-f0c247.s:1656: Error: unrecognized opcode: `ldat'
> clang: error: assembler command failed with exit code 1 (use -v to see 
> invocation)
> make: *** [<builtin>: powerpc/pmu.o] Error 1
>
> Does compiling with Clang work for you? If so, which version?
>
>  Thomas
>
>
Thanks for sharing this observation Thomas. I did some digging on a 
ppc64 machine and found the following :

The problem isn't with clang. Clang is able to perfectly identify LDAT 
insn. This can be verified by writing a small C program using LDAT 
within a similar asm volatile block and compiling with Clang. Sails smooth.
The problem is, the GNU assembler does not identify LDAT, unless 
"-mpower10" flag is passed to it. Can be verified by writing a small .S 
program making use of LDAT and passing to 'as', the GNU assembler binary.
In your Travis CI log, as I can check, the following config is used :
$ export CONFIG="--arch=ppc64 --endian=little --cc=clang 
--cflags=-no-integrated-as"
"no-integrated-as" defers the assembling job to GNU assembler, and since 
no "-mpower10" flag is passed, the compilation fails.
Using the following flag in the config instead, passes -mpower10 to GNU 
assembler and gets the compilation job done : 
  --cflags="-no-integrated-as -Wa,-mpower10"

Thanks,
Chinmay

PS: I will be away traveling for a while, will get back to this as soon 
as I am back. Thanks for understanding   :)