On Tue, Feb 26, 2019 at 01:33:30PM -0500, Steven Rostedt wrote: > On Tue, 26 Feb 2019 08:19:56 -0800 > Ira Weiny <ira.weiny@xxxxxxxxx> wrote: > > > > > Instead of: > > > > > > + create_mad_addr_info(wr, qp_info, > > > + &__entry->dlid, &__entry->sl, > > > + &__entry->pkey, &__entry->rqpn, > > > + &__entry->rqkey); > > > > > > You do: > > > > > > create_mad_addr_info(wr, qp_info, __entry); > > > > > > And then in that function you assign the values directly to the structure. > > > > Is there any reason this is preferred over what I did? I'm not seeing the > > benefit? > > > > Matters if gcc optimizes it into a static inline (which you may not > want, because that goes directly into the code that calls the trace > points, for each event in that class. Thus, I believe it is a true > function call. In which, it is passing 7 arguments, on x86, that > requires storing on the stack, and takes time to set up and tear down. > > Passing the function pointer directly, makes it only use 3 arguments > (all in registers) and is much faster, especially, if this is a fast > path). > > Do a disassemble of both approaches to see what I mean. Right... However, this results in quite the chicken and the egg problem because calling create_mad_addr_info() within the definition of trace_event_raw_ib_mad_send_template results in errors that I'm declaring the structure in the parameter list... :-/ I've tried moving create_mad_addr_info() around a bit and a couple of things with void pointers but the best I can get is that I'm dereferencing an incomplete type because the output of the preprocessor always seems to put at least 1 call of create_mad_addr_info() prior to the struct definition. Is there an example of this somewhere? This code is not that performance critical but I am curious at this point. Ira > > -- Steve
