Re: [PATCH] gcov: Fix -fprofile-update=atomic

[Sebastian Huber <sebastian.huber@embedded-brains.de>]

On 16.12.22 13:09, Richard Biener wrote:
> On Fri, Dec 16, 2022 at 11:39 AM Sebastian Huber
> <sebastian.huber@embedded-brains.de>  wrote:
>> On 16.12.22 10:47, Richard Biener wrote:
>>>> No, if you select -fprofile-update=atomic, then the updates shall be
>>>> atomic from my point of view. If a fallback is acceptable, then you can
>>>> use -fprofile-update=prefer-atomic. Using the fallback in
>>>> -fprofile-update=atomic is a bug which prevents the use of gcov for
>>>> multi-threaded applications on the lower end targets which do not have
>>>> atomic operations in hardware.
>>> Ah, OK.  So the fallback there is libatomic calls as well then?  Note
>>> not all targets support libatomic, for those the failure mode is likely
>>> a link error (which might be fine, but we eventually might want to
>>> amend documentation to indicate this failure mode).
>> It seems these library calls caused issues in the past:
>>
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77466
>>
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77378
> Hmm, those are testsuite-isms in some way but of course
> users would run into the same issue, needing explicit
> -latomic (where available).  I suppose target specs could
> automatically add that for -fprofile-update=atomic but this
> would need to be specified at link time as well then.

Why can't we provide libatomic for all targets? Is gthr.h not always 
available?

> 
>> One option could be to emit calls to a new libgcov function:
>>
>> __gcov_inc_counter(counter) -> updated value
>>
>> This function could use a __gthread_mutex_t mutex for updates. This ends
>> up probably with quite a bad performance.
> But that's eventually what libatomic will do as well as fallback.

For libatomic, hosts can implement a protect_start() and protect_end() 
function. On RTEMS, this is implemented like this:

#include <machine/_libatomic.h>

static inline UWORD
protect_start (void *ptr)
{
   return _Libatomic_Protect_start (ptr);
}

static inline void
protect_end (void *ptr, UWORD isr_level)
{
   _Libatomic_Protect_end (ptr, isr_level);
}

On single-core systems, this is mapped to interrupt disable/enable. This 
is quite efficient (compared to a mutex).

> 
> I don't have a good idea here.
> 
> Do you have to explicitely link -latomic on RISCV?

For RTEMS, libatomic is always added to the linker command line:

gcc/config/rtems.h:"%{qrtems:--start-group -lrtemsbsp -lrtemscpu 
-latomic -lc -lgcc --end-group}"

For riscv, there seems to be also a linux special case:

gcc/config/riscv/linux.h:  " %{pthread:" LD_AS_NEEDED_OPTION " -latomic 
" LD_NO_AS_NEEDED_OPTION "}"
gcc/config/riscv/linux.h:#define LIB_SPEC GNU_USER_TARGET_LIB_SPEC " 
-latomic "


-- 
embedded brains GmbH
Herr Sebastian HUBER
Dornierstr. 4
82178 Puchheim
Germany
email: sebastian.huber@embedded-brains.de
phone: +49-89-18 94 741 - 16
fax:   +49-89-18 94 741 - 08

Registergericht: Amtsgericht München
Registernummer: HRB 157899
Vertretungsberechtigte Geschäftsführer: Peter Rasmussen, Thomas Dörfler
Unsere Datenschutzerklärung finden Sie hier:
https://embedded-brains.de/datenschutzerklaerung/