Re: [PATCH v4 1/2] perf parse: Refactor struct perf_evsel_config_term

[Mathieu Poirier]

On Thu, Jan 09, 2020 at 01:08:06PM +0800, Leo Yan wrote:
> Hi Mathieu,
> 
> On Wed, Jan 08, 2020 at 10:58:31AM -0700, Mathieu Poirier wrote:
> 
> [...]
> 
> > > diff --git a/tools/perf/util/evsel_config.h b/tools/perf/util/evsel_config.h
> > > index 1f8d2fe0b66e..4e5b3ebf09cf 100644
> > > --- a/tools/perf/util/evsel_config.h
> > > +++ b/tools/perf/util/evsel_config.h
> > > @@ -33,21 +33,8 @@ struct perf_evsel_config_term {
> > >         struct list_head      list;
> > >         enum evsel_term_type  type;
> > >         union {
> > > -               u64           period;
> > > -               u64           freq;
> > > -               bool          time;
> > > -               char          *callgraph;
> > > -               char          *drv_cfg;
> > > -               u64           stack_user;
> > > -               int           max_stack;
> > > -               bool          inherit;
> > > -               bool          overwrite;
> > > -               char          *branch;
> > > -               unsigned long max_events;
> > > -               bool          percore;
> > > -               bool          aux_output;
> > > -               u32           aux_sample_size;
> > > -               u64           cfg_chg;
> > > +               u64           num;
> > > +               char          *str;
> > 
> > That is a lot more than just dealing with the "char *" members.  Given
> > the pervasiveness of the changes I would have been happy to leave
> > other members alone for the time being.
> 
> I think actually you are suggesting like below which add general
> members and also keep the old members.  If so, I prefer to add two
> general members 'num' and 'str'.

If we are to deal with all flields of the union, I think it should be as below:

        union {
                bool            cfg_bool;
                int             cfg_int;
                unsigned long   cfg_ulong;
                u32             cfg_u32;
                char            *cfg_str;
        } val;

But just dealing with the "char *" as below would also be fine with me:

        union {
                u64           period;
                u64           freq;
                bool          time;
                u64           stack_user;
                int           max_stack;
                bool          inherit;
                bool          overwrite;
                unsigned long max_events;
                bool          percore;
                bool          aux_output;
                u32           aux_sample_size;
                u64           cfg_chg;
                u64           num;
                char          *str;
        } val;

> 
> struct perf_evsel_config_term {
>         struct list_head      list;
>         enum evsel_term_type  type;
>         union {
>                 u64           period;
>                 u64           freq;
>                 bool          time;
>                 char          *callgraph;
>                 char          *drv_cfg;
>                 u64           stack_user;
>                 int           max_stack;
>                 bool          inherit;
>                 bool          overwrite;
>                 char          *branch;
>                 unsigned long max_events;
>                 bool          percore;
>                 bool          aux_output;
>                 u32           aux_sample_size;
>                 u64           cfg_chg;
> +               u64           num;
> +               char          *str;
>         } val;
>         bool weak;
> };
> 
> > I will let Jiri make the
> > final call but if we are to proceed this way I think we should have a
> > member per type to avoid casting issues.
> 
> Yeah, let's see what's Jiri thinking.
> 
> Just note, with this change, I don't see any casting warning or errors
> when built perf on arm64/arm32.

At this time you may not, but they will happen and it will be very hard to
debug.

> 
> Thanks,
> Leo Yan