Re: [PATCH] powerpc: Remove backtrace implementation

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 22/02/2021 19:16, Tulio Magno Quites Machado Filho wrote:
> Adhemerval Zanella <adhemerval.zanella@linaro.org> writes:
> 
>> On 12/02/2021 14:21, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>> Does this really work well (enough) on systems which do not build
>>> everything with asynchronous unwind tables?  If I recall correctly, GCC
>>> enabled them by default only quite recently.
>>
>> This should work as well any other architecture since it uses the libgcc_s
>> routines.  Afaik the powerpc implementation was done as optimization, not
>> for correctness.
> 
> I think Florian is correct.
> 
> GCC enabled -fasynchronous-unwind-tables by default on ppc* ELF targets on
> version 8 [1].
> So if you take a Debian 10 ppc64le, build the backtrace example from
> linux-manpages [2] with gcc-7 or with a newer GCC using
> 'gcc-9 -g0 -fno-asynchronous-unwind-tables', this executable won't be able to
> print the correct backtrace, e.g.:
> 
> $ ./testrun.sh ./tst-no-asynchronous-unwind-tables 3
> backtrace() returned 1 addresses
> ./tst-no-asynchronous-unwind-tables(+0xb48) [0x7fff99a00b48]
> 
> 8 addresses were expected in the previous test.
> 
> With that said, I don't think the powerpc files are just an optimization and
> removing them may cause old binaries to stop getting the right backtrace.
> 
> [1] https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=0359465c703a
> [2] https://www.man7.org/linux/man-pages/man3/backtrace.3.html
> 

Afaik this is true for every other architecture that uses libgcc for
backtrace implementation, currently *everything* but powerpc and
microblaze:

This is on aarch64 with gcc 5.4:

$ gcc -g0 -fno-asynchronous-unwind-tables backtrace.c -o backtrace
$ ./backtrace 3
backtrace() returned 1 addresses
./backtrace() [0x4009dc]

And this is sparc64 with gcc-10:

$ gcc -g0 -fno-asynchronous-unwind-tables backtrace.c -o backtrace
$ ./backtrace 3
backtrace() returned 1 addresses
./backtrace(+0xa74) [0x10000000a74]

So for all other architectures we require -fasynchronous-unwind-tables
to get backtrace work and I plan to remove microblaze implementation
so powerpc will be only outlier.

The question is: should we fix all other architectures to provide a
non-libgcc way to get around this issue? What about architectures that
backtrace through frame layout is not really possible or require hackish
ways (-fomit-frame-pointer for some architecture or thumb mode with
gcc and clang objects)? 

I really don't think we should keep this fast-unwinders (sanitizers
has a hard time maintaining for some architectures) so I still think
we should move to use *just* libgcc as backend and document the flags
 required to get a meaningful stacktrace