Re: atomic update of profile counters (issue7000044)

[Rong Xu <xur at google dot com>]

Hi,

Here is a new patch. The only difference is to declare
__atomic_fetch_add as weak. This is
needed for targets without sync/atomic builtin support. The patch
contains a call to the builtin regardless of the new options
-fprofile-gen-atomic. This results in a unsat in these targets even
for regular profile-gen built.

With this new patch, if the user uses -fprofile-gen-atomic in these
target, the generated code will seg fault.

We think a better solution is to emit the builtin call only in these
targets with the support, and give warning for non-supported target.
But I did not find any target hook for this. Does anyone know how to
do this?

Thanks,

-Rong


On Fri, Dec 28, 2012 at 11:35 AM, Xinliang David Li <davidxl@google.com> wrote:
> It would be great if this can make into gcc4.8. The patch has close to
> 0 impact on code stability.
>
> David
>
> On Fri, Dec 28, 2012 at 11:32 AM, Rong Xu <xur@google.com> wrote:
>> Hi Honza,
>>
>> In the other thread of discussion (similar patch in google-4_7
>> branch), you said you were thinking if to let this patch into trunk in
>> stage 3. Can you give some update?
>>
>> Thanks,
>>
>> -Rong
>>
>> On Fri, Dec 21, 2012 at 10:37 AM, Rong Xu <xur@google.com> wrote:
>>> On Fri, Dec 21, 2012 at 1:25 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
>>>>> Hi,
>>>>>
>>>>> This patch adds support of atomic update of profiles counters. The goal is to improve
>>>>> the poor counter values for highly thread programs.
>>>>>
>>>>> The atomic update is under a new option -fprofile-gen-atomic=<N>
>>>>> N=0: default, no atomic update
>>>>> N=1: atomic update edge counters.
>>>>> N=2: atomic update some of value profile counters (currently indirect-call and one value profile).
>>>>> N=3: both edge counter and the above value profile counters.
>>>>> Other value: fall back to the default.
>>>>>
>>>>> This patch is a simple porting of the version in google-4_7 branch. It uses __atomic_fetch_add
>>>>> based on Andrew Pinski's suggestion. Note I did not apply to all the value profiles as
>>>>> the indirect-call profile is the most relevant one here.
>>>>>
>>>>> Test with bootstrap.
>>>>>
>>>>> Comments and suggestions are welcomed.
>>>>>
>>>>> Thanks,
>>>>>
>>>>> -Rong
>>>>>
>>>>>
>>>>> 2012-12-20  Rong Xu  <xur@google.com>
>>>>>
>>>>>       * libgcc/libgcov.c (__gcov_one_value_profiler_body_atomic): New
>>>>>         function. Atomic update profile counters.
>>>>>       (__gcov_one_value_profiler_atomic): Ditto.
>>>>>       (__gcov_indirect_call_profiler_atomic): Ditto.
>>>>>       * gcc/gcov-io.h: Macros for atomic update.
>>>>>       * gcc/common.opt: New option.
>>>>>       * gcc/tree-profile.c (gimple_init_edge_profiler): Atomic
>>>>>         update profile counters.
>>>>>       (gimple_gen_edge_profiler): Ditto.
>>>>
>>>> The patch looks resonable.  Eventually we probably should provide rest of the value counters
>>>> in thread safe manner.  What happens on targets not having atomic operations?
>>>
>>> From http://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html#_005f_005fsync-Builtins,
>>> it says:
>>>       "If a particular operation cannot be implemented on the target
>>> processor, a warning is generated and a call an external function is
>>> generated. "
>>>
>>> So I think there will be a warning and eventually a link error of unsat.
>>>
>>> Thanks,
>>>
>>> -Rong
>>>
>>>
>>>>
>>>> Honza

Attachment: patch2.diff
Description: Binary data