From: Andi Kleen <a...@linux.intel.com>

With perf script --itrace=cr we can synthesize calls and returns out of
a PT log. However both calls and returns are marked with the same event,
called branches. This makes it difficult to read and post process,
because calls and returns are somewhat diffferent.

Create a separate return event and mark the returns as return.

Cc: adrian.hun...@intel.com
v2:
Add extra filter for returns.
Signed-off-by: Andi Kleen <a...@linux.intel.com>
---
 tools/perf/util/intel-pt.c | 61 ++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 54 insertions(+), 7 deletions(-)

diff --git a/tools/perf/util/intel-pt.c b/tools/perf/util/intel-pt.c
index 137196990012..c72b9074e86e 100644
--- a/tools/perf/util/intel-pt.c
+++ b/tools/perf/util/intel-pt.c
@@ -82,9 +82,13 @@ struct intel_pt {
        u64 instructions_id;
 
        bool sample_branches;
+       bool sample_returns;
        u32 branches_filter;
        u64 branches_sample_type;
+       u64 returns_sample_type;
+       u32 returns_filter;
        u64 branches_id;
+       u64 returns_id;
 
        bool sample_transactions;
        u64 transactions_sample_type;
@@ -960,7 +964,9 @@ static int intel_pt_inject_event(union perf_event *event,
        return perf_event__synthesize_sample(event, type, 0, sample, swapped);
 }
 
-static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
+static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq,
+                                       bool is_return,
+                                       u32 filter)
 {
        int ret;
        struct intel_pt *pt = ptq->pt;
@@ -971,7 +977,7 @@ static int intel_pt_synth_branch_sample(struct 
intel_pt_queue *ptq)
                struct branch_entry     entries;
        } dummy_bs;
 
-       if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
+       if (filter && !(filter & ptq->flags))
                return 0;
 
        if (pt->synth_opts.initial_skip &&
@@ -990,8 +996,13 @@ static int intel_pt_synth_branch_sample(struct 
intel_pt_queue *ptq)
        sample.pid = ptq->pid;
        sample.tid = ptq->tid;
        sample.addr = ptq->state->to_ip;
-       sample.id = ptq->pt->branches_id;
-       sample.stream_id = ptq->pt->branches_id;
+       if (is_return) {
+               sample.id = ptq->pt->returns_id;
+               sample.stream_id = ptq->pt->returns_id;
+       } else {
+               sample.id = ptq->pt->branches_id;
+               sample.stream_id = ptq->pt->branches_id;
+       }
        sample.period = 1;
        sample.cpu = ptq->cpu;
        sample.flags = ptq->flags;
@@ -1014,6 +1025,8 @@ static int intel_pt_synth_branch_sample(struct 
intel_pt_queue *ptq)
 
        if (pt->synth_opts.inject) {
                ret = intel_pt_inject_event(event, &sample,
+                                           is_return ?
+                                           pt->returns_sample_type :
                                            pt->branches_sample_type,
                                            pt->synth_needs_swap);
                if (ret)
@@ -1241,7 +1254,15 @@ static int intel_pt_sample(struct intel_pt_queue *ptq)
                thread_stack__set_trace_nr(ptq->thread, state->trace_nr);
 
        if (pt->sample_branches) {
-               err = intel_pt_synth_branch_sample(ptq);
+               err = intel_pt_synth_branch_sample(ptq, false,
+                                                  pt->branches_filter);
+               if (err)
+                       return err;
+       }
+
+       if (pt->sample_returns) {
+               err = intel_pt_synth_branch_sample(ptq, true,
+                                                  pt->returns_filter);
                if (err)
                        return err;
        }
@@ -1956,7 +1977,33 @@ static int intel_pt_synth_events(struct intel_pt *pt,
                }
                pt->sample_branches = true;
                pt->branches_sample_type = attr.sample_type;
-               pt->branches_id = id;
+               pt->branches_id = id++;
+       }
+       if (pt->synth_opts.returns) {
+               attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
+               attr.sample_period = 1;
+               attr.sample_type |= PERF_SAMPLE_ADDR;
+               attr.sample_type &= ~(u64)PERF_SAMPLE_CALLCHAIN;
+               attr.sample_type &= ~(u64)PERF_SAMPLE_BRANCH_STACK;
+               pr_debug("Synthesizing 'return' event with id %" PRIu64 " 
sample type %#" PRIx64 "\n",
+                        id, (u64)attr.sample_type);
+               err = intel_pt_synth_event(session, &attr, id);
+               if (err) {
+                       pr_err("%s: failed to synthesize 'return' event type\n",
+                              __func__);
+                       return err;
+               }
+               pt->sample_returns = true;
+               pt->returns_sample_type = attr.sample_type;
+               pt->returns_id = id;
+               evlist__for_each(evlist, evsel) {
+                       if (evsel->id && evsel->id[0] == pt->returns_id) {
+                               if (evsel->name)
+                                       zfree(&evsel->name);
+                               evsel->name = strdup("return");
+                               break;
+                       }
+               }
        }
 
        pt->synth_needs_swap = evsel->needs_swap;
@@ -2155,7 +2202,7 @@ int intel_pt_process_auxtrace_info(union perf_event 
*event,
                pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
                                       PERF_IP_FLAG_TRACE_END;
        if (pt->synth_opts.returns)
-               pt->branches_filter |= PERF_IP_FLAG_RETURN |
+               pt->returns_filter |= PERF_IP_FLAG_RETURN |
                                       PERF_IP_FLAG_TRACE_BEGIN;
 
        if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
-- 
2.5.5

Reply via email to