Re: [PATCH] perf: fix building error in x86_64 when dwarf unwind is on

[Wang Nan <wangnan0@huawei.com>]

On 2015/1/7 16:39, Jiri Olsa wrote:
> On Wed, Jan 07, 2015 at 02:40:04PM +0900, Namhyung Kim wrote:
>> Hi,
>>
>> On Wed, Jan 07, 2015 at 10:53:52AM +0800, Wang Nan wrote:
>>> Ping...
>>
>> Sorry for long delay.
>>
>>>
>>> On 2014/12/29 16:14, Wang Nan wrote:
>>>> On 2014/12/29 15:56, Namhyung Kim wrote:
>>>>> Hi Wang,
>>>>>
>>>>> (Adding Arnaldo and Jiri to CC)
>>>>>
>>>>> On Sat, Dec 27, 2014 at 09:26:11AM +0800, Wang Nan wrote:
>>>>>> When build with 'make ARCH=x86' and dwarf unwind is on, there is a
>>>>>> compiling error:
>>>>>>
>>>>>>    CC       /home/wn/perf/arch/x86/util/unwind-libdw.o
>>>>>>    CC       /home/wn/perf/arch/x86/tests/regs_load.o
>>>>>>  arch/x86/tests/regs_load.S: Assembler messages:
>>>>>>  arch/x86/tests/regs_load.S:65: Error: operand type mismatch for `push'
>>>>>>  arch/x86/tests/regs_load.S:72: Error: operand type mismatch for `pop'
>>>>>>  make[1]: *** [/home/wn/perf/arch/x86/tests/regs_load.o] Error 1
>>>>>>  make[1]: INTERNAL: Exiting with 25 jobserver tokens available; should be 24!
>>>>>>  make: *** [all] Error 2
>>>>>>  ...
>>>>>>
>>>>>> Which is caused by incorrectly undefine macro HAVE_ARCH_X86_64_SUPPORT.
>>>>>> 'config/Makefile.arch' tests __x86_64__ only when 'ARCH=x86_64'. However,
>>>>>> with 'ARCH=x86', the underlying compile may also be x86_64, which causes
>>>>>> mismatching.
>>>>>
>>>>> Hmm.. how did you compile this?  I guess ARCH=x86 requires -m32 flag
>>>>> to the gcc, did you pass it (like via EXTRA_CFLAGS=-m32)?
>>>>>
>>>>> I'm confused by 'underlying compile may also be x86_64' part..
>>>>>
>>>>
>>>> I hit this problem when building perf with Yocto (https://www.yoctoproject.org/), which
>>>> is a famous building system for embeded system.
>>>>
>>>> When building kernel, we can simply use 'make ARCH=x86' and select 'CONFIG_64BIT'
>>>> in menuconfig to get a x86_64 kernel. As a result, there building framework
>>>> like Yocto doesn't ever consider 'x86_64' as a valid ARCH option. See:
>>>> https://git.yoctoproject.org/cgit/cgit.cgi/poky/tree/meta/classes/kernel-arch.bbclass
>>>> Therefore, when building with such framework, it uses a x86_64 compiler and ARCH=x86.
>>
>> Hmm.. okay.  I think that it should also be checked with the new build
>> system for perf, Jiri? ;)
> 
> yea.. however the config/Makefile still stays as it was
> 
> but whatever change you'll make for Makefile.perf I need
> to transform.. but this one does not seem too massive ;-)
> 
>>
>> Anyway, I cleaned up the code like below.. As __LP64__ is defined for
>> x86_64 as well, we can consolidate the __x86_64__ check to the
>> __LP64__ check and get rid of the IS_X86_64 IMHO.
>>
>> Thanks,
>> Namhyung
>>
>>
>> diff --git a/tools/perf/Makefile.perf b/tools/perf/Makefile.perf
>> index 67a03a825b3c..eb3e2f3e14b4 100644
>> --- a/tools/perf/Makefile.perf
>> +++ b/tools/perf/Makefile.perf
>> @@ -462,10 +462,12 @@ BUILTIN_OBJS += $(OUTPUT)builtin-bench.o
>>  # Benchmark modules
>>  BUILTIN_OBJS += $(OUTPUT)bench/sched-messaging.o
>>  BUILTIN_OBJS += $(OUTPUT)bench/sched-pipe.o
>> +ifeq ($(ARCH),x86)
>>  ifeq ($(RAW_ARCH),x86_64)
>>  BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy-x86-64-asm.o
>>  BUILTIN_OBJS += $(OUTPUT)bench/mem-memset-x86-64-asm.o
>>  endif
>> +endif
> 
> isn't the check for x86 superfluous here? the x86_64 check is stronger
> 
> otherwise it looks ok to me.. Wang Nan, could you please check
> if it fixes the issue for you?
> 

Sorry, although Namhyung Kim's patch solves my ARCH=x86 problem, it breaks ARM building:

...
| bench/../../../arch/x86/lib/memcpy_64.S: Assembler messages:
| bench/../../../arch/x86/lib/memcpy_64.S:27: Error: junk at end of line, first unrecognized character is `,'
| bench/../../../arch/x86/lib/memcpy_64.S:29: Error: bad instruction `movq %rdi,%rax'
| bench/../../../arch/x86/lib/memcpy_64.S:30: Error: bad instruction `movq %rdx,%rcx'
| bench/../../../arch/x86/lib/memcpy_64.S:31: Error: bad instruction `shrq $3,%rcx'
| bench/../../../arch/x86/lib/memcpy_64.S:32: Error: bad instruction `andl $7,%edx'
...

It builds x86_64 bench when I corss compiling perf for arm (use ARCH=arm). This is caused by Makefile.perf:

...
BUILTIN_OBJS += $(OUTPUT)bench/sched-pipe.o
ifeq ($(RAW_ARCH),x86_64)
BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy-x86-64-asm.o
BUILTIN_OBJS += $(OUTPUT)bench/mem-memset-x86-64-asm.o
endif
BUILTIN_OBJS += $(OUTPUT)bench/mem-memcpy.o
...

Which rely on RAW_ARCH, but RAW_ARCH is not a cross-compiling-friendly variable now.

> thanks,
> jirka
>