Re: [PATCH perf/core] perf intel-pt: Fix clang build failure in intel_pt_synth_pebs_sample

[Ian Rogers]

On Thu, May 14, 2020 at 3:04 PM Gustavo A. R. Silva
<gustavoars@xxxxxxxxxx> wrote:
>
> On Thu, May 14, 2020 at 12:06:48PM -0700, Ian Rogers wrote:
> > On Thu, May 14, 2020 at 8:01 AM Gustavo A. R. Silva
> > <gustavoars@xxxxxxxxxx> wrote:
> > >
> > > On Thu, May 14, 2020 at 10:10:30AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > Em Wed, May 13, 2020 at 06:47:38PM -0500, Gustavo A. R. Silva escreveu:
> > > > > Fix the following build failure generated with command
> > > > > $ make CC=clang HOSTCC=clang -C tools/ perf:
> > > > >
> > > > > util/intel-pt.c:1802:24: error: field 'br_stack' with variable sized type 'struct branch_stack' not at the end of a struct or class is a GNU extension [-Werror,-Wgnu-variable-sized-type-not-at-end]
> > > > >                         struct branch_stack br_stack;
> > > > >                                             ^
> > > > > 1 error generated.
> > > > >
> > > > > Fix this by reordering the members of struct br.
> > > >
> > > > Yeah, I noticed that as far back as with ubuntu 16.04's clang:
> > > >
> > > > clang version 3.8.0-2ubuntu4 (tags/RELEASE_380/final)
> > > >
> > > > util/intel-pt.c:1802:24: error: field 'br_stack' with variable sized type 'struct branch_stack' not at the end of a struct or class is a GNU
> > > >       extension [-Werror,-Wgnu-variable-sized-type-not-at-end]
> > > >                         struct branch_stack br_stack;
> > > >                                             ^
> > > > 1 error generated.
> > > >
> > > >
> > > > Will fold this with the bug introducing the problem to avoid bisection
> > > > problems.
> > > >
> > >
> > > I agree. Also, the commit hash of the "Fixes" tag only applies to the
> > > perf/core branch and, I guess that might create confusion.
> >
> >
> > So while this fixes the warning I believe it breaks the intent of the code.
> >
> > tools/perf/util/branch.h:
> > struct branch_stack {
> >        u64                     nr;
> >        u64                     hw_idx;
> >        struct branch_entry     entries[];
> > };
> >
> > tools/perf/util/intel-pt.c:
> >                struct {
> >                        struct branch_stack br_stack;
> >                        struct branch_entry entries[LBRS_MAX];
> >                } br;
> >
> > The array in br is trying to extend branch_stack's entries array. You
> > might have to do something like:
> >
> > alignas(alignof(branch_stack)) char storage[sizeof(branch_stack) +
> > sizeof(branch_entry) * LBRS_MAX];
> > struct branch_stack *br = &storage;
> >
> > malloc/free may be nicer on the eyeballs.
> >
>
> Yep, I'd go for zalloc/free. There are a couple of places where dynamic
> memory is being allocated for struct branch_stack:
>
> tools/perf/util/cs-etm.c-256-   if (etm->synth_opts.last_branch) {
> tools/perf/util/cs-etm.c:257:           size_t sz = sizeof(struct branch_stack);
> tools/perf/util/cs-etm.c-258-
> tools/perf/util/cs-etm.c-259-           sz += etm->synth_opts.last_branch_sz *
> tools/perf/util/cs-etm.c-260-                 sizeof(struct branch_entry);
> tools/perf/util/cs-etm.c-261-           tidq->last_branch = zalloc(sz);
>
> tools/perf/util/thread-stack.c-148-     if (br_stack_sz) {
> tools/perf/util/thread-stack.c:149:             size_t sz = sizeof(struct branch_stack);
> tools/perf/util/thread-stack.c-150-
> tools/perf/util/thread-stack.c-151-             sz += br_stack_sz * sizeof(struct branch_entry);
> tools/perf/util/thread-stack.c-152-             ts->br_stack_rb = zalloc(sz);
>
> there is even function intel_pt_alloc_br_stack().
>
> Just out of curiosity, why the need of such a hack in this case (the
> on-stack extension of branch_stack's entries array)?

My guess would be that the lbr size is an architectural constant and
so avoiding malloc/free in what could be a hot loop was desirable.
As this is part of a larger patch set, is this the only place this
problem has been encountered? Perhaps a macro could perform the
complicated stack allocation I suggested. It may be nice to save
cycles if code this pattern is widespread and the code hot.

Thanks,
Ian

> Thanks
> --
> Gustavo
>