python prototyping

Adjust the logics to cache python frames

Author: r1viollet

app/python-tests/Dockerfile

@@ -0,0 +1,2 @@
+FROM python:slim
+

include/austin_symbol_lookup.hpp

@@ -16,10 +16,15 @@ namespace ddprof {
 class AustinSymbolLookup {
 public:
   SymbolIdx_t get_or_insert(austin_frame_t *frame, SymbolTable &symbol_table);
-
+  void clear() {
+    for (auto el : _austin_symbols) {
+      _free_list.push_back(el);
+    }
+    _austin_symbols.clear();
+  }
 private:
-  using FrameKeyMap = std::unordered_map<uintptr_t, SymbolIdx_t>;
-  FrameKeyMap _frame_key_map;
+  std::vector<SymbolIdx_t> _austin_symbols;
+  std::vector<SymbolIdx_t> _free_list;
 };
 
 } // namespace ddprof

include/unwind_state.hpp

@@ -43,7 +43,6 @@ struct UnwindState {
         stack(nullptr), stack_sz(0), current_ip(0), austin_handle(nullptr) {
     output.clear();
     output.locs.reserve(DD_MAX_STACK_DEPTH);
-
   }
 
   ddprof::DwflHdr dwfl_hdr;

src/austin_symbol_lookup.cc

@@ -16,20 +16,21 @@ namespace ddprof {
 
 
 SymbolIdx_t AustinSymbolLookup::get_or_insert(austin_frame_t * frame, SymbolTable &symbol_table) {
-  auto iter = _frame_key_map.find(frame->key);
-
-  if (iter != _frame_key_map.end()) {
-    return iter->second;
+  SymbolIdx_t idx = -1;
+  // reuse elements to avoid growing the table
+  if (!_free_list.empty()) {
+    auto el = _free_list.back();
+    _free_list.pop_back();
+    _austin_symbols.push_back(el);
+    idx = el;
+  }
+  if (idx == -1) {
+    idx = symbol_table.size();
+    symbol_table.push_back(Symbol());
   }
-
-  SymbolIdx_t index = symbol_table.size();
   std::string symname = std::string(frame->scope);
-
-  symbol_table.push_back(Symbol(symname, symname, frame->line, std::string(frame->filename)));
-
-  _frame_key_map.emplace(frame->key, index);
-
-  return index;
+  symbol_table[idx] = Symbol(symname, symname, frame->line, std::string(frame->filename));
+  return idx;
 }
 
 } // namespace ddprof

src/unwind_dwfl.cc

@@ -105,6 +105,9 @@ namespace ddprof {
 DDRes unwind_init_dwfl(UnwindState *us) {
   // Create or get the dwfl object associated to cache
   us->_dwfl_wrapper = &(us->dwfl_hdr.get_or_insert(us->pid));
+  // clear at every iteration
+  us->symbol_hdr._austin_symbol_lookup.clear();
+
   if (!us->_dwfl_wrapper->_attached) {
     // we need to add at least one module to figure out the architecture (to
     // create the unwinding backend)
@@ -404,7 +407,6 @@ static DDRes add_python_frame(UnwindState *us, SymbolIdx_t symbol_idx,
                               ElfAddress_t pc, Dwfl_Frame *dwfl_frame) {
   SymbolHdr &unwind_symbol_hdr = us->symbol_hdr;
   SymbolTable &symbol_table = unwind_symbol_hdr._symbol_table;
-
   std::string symname = symbol_table.at(symbol_idx)._symname;
   if (us->austin_handle &&
       (symname.find("PyEval_EvalFrameDefault") != std::string::npos ||
@@ -413,11 +415,11 @@ static DDRes add_python_frame(UnwindState *us, SymbolIdx_t symbol_idx,
         unwind_symbol_hdr._austin_symbol_lookup;
 
     // The register we are interested in is RSI, but it doesn't seem to be
-    // available. So we loop over the first 64 registers and stop if we find
+    // available. So we loop over the available registers and stop if we find
     // a register value that resolves correctly to a Python frame.
-    uint64_t val;
-    for (int i = 0; i < (int)sizeof(*dwfl_frame->regs_set) * 8; i++) {
-      if (__libdwfl_frame_reg_get(dwfl_frame, i, &val)) {
+    uint64_t val = 0;
+    for (int i = 0; i < PERF_REGS_COUNT; i++) {
+      if (__libdwfl_frame_reg_get(dwfl_frame, i, &val) && val) {
         austin_frame_t *frame =
             austin_read_frame(us->austin_handle, (void *)val);
         if (frame) {