Re: add dbgcnt and opt-info support for devirtualization

[Xinliang David Li <davidxl at google dot com>]

On Fri, May 16, 2014 at 9:03 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
>> Hi, debugging runtime bugs due to devirtualization can be hard for
>> very large C++ programs with complicated class hierarchy. This patch
>> adds the support to report this high level transformation via
>> -fopt-info (not hidden inside dump file) and the ability the do binary
>> search with cutoff.
>>
>> Ok for trunk after build and test?
>
> Seems resonable to me.
>>
>> thanks,
>>
>> David
>
>> Index: ChangeLog
>> ===================================================================
>> --- ChangeLog (revision 210479)
>> +++ ChangeLog (working copy)
>> @@ -1,3 +1,18 @@
>> +2014-05-15  Xinliang David Li  <davidxl@google.com>
>> +
>> +     * cgraphunit.c (walk_polymorphic_call_targets): Add
>> +     dbgcnt and fopt-info support.
>> +     2014-05-15  Xinliang David Li  <davidxl@google.com>
>> +
>> +             * cgraphunit.c (walk_polymorphic_call_targets): Add
>> +             dbgcnt and fopt-info support.
>> +             * ipa-prop.c (ipa_make_edge_direct_to_target): Ditto.
>> +             * ipa-devirt.c (ipa_devirt): Ditto.
>> +             * ipa.c (walk_polymorphic_call_targets): Ditto.
>> +             * gimple-fold.c (fold_gimple_assign): Ditto.
>> +             (gimple_fold_call): Ditto.
>> +             * dbgcnt.def: New counter.
>> +
>>  2014-05-15  Martin Jambor  <mjambor@suse.cz>
>>
>>       PR ipa/61085
>> Index: ipa-prop.c
>> ===================================================================
>> --- ipa-prop.c        (revision 210479)
>> +++ ipa-prop.c        (working copy)
>> @@ -59,6 +59,7 @@ along with GCC; see the file COPYING3.
>>  #include "ipa-utils.h"
>>  #include "stringpool.h"
>>  #include "tree-ssanames.h"
>> +#include "dbgcnt.h"
>>
>>  /* Intermediate information about a parameter that is only useful during the
>>     run of ipa_analyze_node and is not kept afterwards.  */
>> @@ -2494,6 +2495,13 @@ ipa_make_edge_direct_to_target (struct c
>>           fprintf (dump_file, "ipa-prop: Discovered direct call to non-function"
>>                               " in %s/%i, making it unreachable.\n",
>>                    ie->caller->name (), ie->caller->order);
>> +          else if (dump_enabled_p ())
>> +         {
>> +           location_t loc = gimple_location (ie->call_stmt);
>> +           dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc,
>> +                            "Discovered direct call to non-function in %s, "
>> +                            "making it unreachable\n", ie->caller->name ());
>
> Perhaps "turning it to __builtin_unreachable call" and similarly in the other cases
> we introduce __builtin_unreachable? I think that could be easier for user to work
> out.

Ok.

>
> What king of problems in devirtualizatoin you are seeing?

I have been chasing a runtime failure of a very large test built with
gcc-4_9. The bad code either calls a pure function or turn a virtual
call into __builtin_unreachable (incomplete target set). The indirect
info shows the otr type to be !maybe_derived_type, and the outer-type
gets cleared during inline update. I isolated a small test case -- but
the good news is that gcc-4_9 @head already fixed the problem.

I will  check in the test case to trunk later.

thanks,

David

>
>
> Honza