simpleperf: Remove stack and reg fields in samples after unwinding

perf_data_converter doesn't support stack and regs fields.
This is to create recording file parseable by perf_data_converter.

Bug: 273341791
Test: run simpleperf_unit_test
Change-Id: Ie05c683eb2f3eec3e08e482b82cbc0be4abe834a

6 files changed, 39 insertions, 3 deletions

diff --git a/simpleperf/cmd_record.cpp b/simpleperf/cmd_record.cpp
index 974be878..912d3e77 100644
--- a/simpleperf/cmd_record.cpp
+++ b/simpleperf/cmd_record.cpp
@@ -1371,6 +1371,12 @@ bool RecordCommand::SetEventSelectionFlags() {
 
 bool RecordCommand::CreateAndInitRecordFile() {
   EventAttrIds attrs = event_selection_set_.GetEventAttrWithId();
+  bool remove_regs_and_stacks = unwind_dwarf_callchain_ && !post_unwind_;
+  if (remove_regs_and_stacks) {
+    for (auto& attr : attrs) {
+      ReplaceRegAndStackWithCallChain(attr.attr);
+    }
+  }
   record_file_writer_ = CreateRecordFile(record_filename_, attrs);
   if (record_file_writer_ == nullptr) {
     return false;
@@ -1841,6 +1847,15 @@ bool RecordCommand::PostUnwindRecords() {
   if (!reader) {
     return false;
   }
+  // Write new event attrs without regs and stacks fields.
+  EventAttrIds attrs = reader->AttrSection();
+  for (auto& attr : attrs) {
+    ReplaceRegAndStackWithCallChain(attr.attr);
+  }
+  if (!record_file_writer_->WriteAttrSection(attrs)) {
+    return false;
+  }
+
   sample_record_count_ = 0;
   auto callback = [this](std::unique_ptr<Record> record) {
     return SaveRecordAfterUnwinding(record.get());
diff --git a/simpleperf/cmd_record_test.cpp b/simpleperf/cmd_record_test.cpp
index 3882a32c..395bfd19 100644
--- a/simpleperf/cmd_record_test.cpp
+++ b/simpleperf/cmd_record_test.cpp
@@ -257,7 +257,16 @@ TEST(record_cmd, dwarf_callchain_sampling) {
   ASSERT_TRUE(RunRecordCmd({"-p", pid, "--call-graph", "dwarf"}));
   ASSERT_TRUE(RunRecordCmd({"-p", pid, "--call-graph", "dwarf,16384"}));
   ASSERT_FALSE(RunRecordCmd({"-p", pid, "--call-graph", "dwarf,65536"}));
-  ASSERT_TRUE(RunRecordCmd({"-p", pid, "-g"}));
+  TemporaryFile tmpfile;
+  ASSERT_TRUE(RunRecordCmd({"-p", pid, "-g"}, tmpfile.path));
+  auto reader = RecordFileReader::CreateInstance(tmpfile.path);
+  ASSERT_TRUE(reader);
+  const EventAttrIds& attrs = reader->AttrSection();
+  ASSERT_GT(attrs.size(), 0);
+  // Check that reg and stack fields are removed after unwinding.
+  for (const auto& attr : attrs) {
+    ASSERT_EQ(attr.attr.sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER), 0);
+  }
 }
 
 TEST(record_cmd, system_wide_dwarf_callchain_sampling) {
diff --git a/simpleperf/event_attr.cpp b/simpleperf/event_attr.cpp
index 4aaaff10..b8acc341 100644
--- a/simpleperf/event_attr.cpp
+++ b/simpleperf/event_attr.cpp
@@ -254,4 +254,11 @@ std::string GetEventNameByAttr(const perf_event_attr& attr) {
   return name;
 }
 
+void ReplaceRegAndStackWithCallChain(perf_event_attr& attr) {
+  attr.sample_type &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER);
+  attr.exclude_callchain_user = 0;
+  attr.sample_regs_user = 0;
+  attr.sample_stack_user = 0;
+}
+
 }  // namespace simpleperf
diff --git a/simpleperf/event_attr.h b/simpleperf/event_attr.h
index a4b989ae..5521f328 100644
--- a/simpleperf/event_attr.h
+++ b/simpleperf/event_attr.h
@@ -48,6 +48,7 @@ bool IsCpuSupported(const perf_event_attr& attr);
 // Return event name with modifier if the event is found, otherwise return "unknown".
 // This function is slow for using linear search, so only used when reporting.
 std::string GetEventNameByAttr(const perf_event_attr& attr);
+void ReplaceRegAndStackWithCallChain(perf_event_attr& attr);
 
 inline bool operator==(const perf_event_attr& attr1, const perf_event_attr& attr2) {
   return memcmp(&attr1, &attr2, sizeof(perf_event_attr)) == 0;
diff --git a/simpleperf/record.cpp b/simpleperf/record.cpp
index 40d3287a..c364b2fc 100644
--- a/simpleperf/record.cpp
+++ b/simpleperf/record.cpp
@@ -713,8 +713,10 @@ SampleRecord::SampleRecord(const perf_event_attr& attr, uint64_t id, uint64_t ip
 void SampleRecord::ReplaceRegAndStackWithCallChain(const std::vector<uint64_t>& ips) {
   uint32_t size_added_in_callchain = sizeof(uint64_t) * (ips.size() + 1);
   uint32_t size_reduced_in_reg_stack =
-      regs_user_data.reg_nr * sizeof(uint64_t) + stack_user_data.size + sizeof(uint64_t);
+      (regs_user_data.reg_nr + 1) * sizeof(uint64_t) + stack_user_data.size + sizeof(uint64_t) * 2;
+
   uint32_t new_size = size() + size_added_in_callchain - size_reduced_in_reg_stack;
+  sample_type &= ~(PERF_SAMPLE_STACK_USER | PERF_SAMPLE_REGS_USER);
   BuildBinaryWithNewCallChain(new_size, ips);
 }
 
diff --git a/simpleperf/record_test.cpp b/simpleperf/record_test.cpp
index ddecc9bd..8d9d151a 100644
--- a/simpleperf/record_test.cpp
+++ b/simpleperf/record_test.cpp
@@ -104,11 +104,13 @@ TEST_F(RecordTest, SampleRecord_exclude_kernel_callchain) {
 }
 
 TEST_F(RecordTest, SampleRecord_ReplaceRegAndStackWithCallChain) {
-  event_attr.sample_type |= PERF_SAMPLE_CALLCHAIN | PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER;
+  event_attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
   SampleRecord expected(event_attr, 0, 1, 2, 3, 4, 5, 6, {}, {1, PERF_CONTEXT_USER, 2, 3, 4, 5}, {},
                         0);
   for (size_t stack_size : {8, 1024}) {
+    event_attr.sample_type |= PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER;
     SampleRecord r(event_attr, 0, 1, 2, 3, 4, 5, 6, {}, {1}, std::vector<char>(stack_size), 10);
+    event_attr.sample_type &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER);
     r.ReplaceRegAndStackWithCallChain({2, 3, 4, 5});
     CheckRecordMatchBinary(r);
     CheckRecordEqual(r, expected);