Merge pull request #650 from kmod/perf7

Switch BoxedMethodDescriptor to tpp_call

microbenchmarks/itertools_chain_bench.py

@@ -7,7 +7,8 @@ def f(c):
         pass
 
 l = []
+r = range(500)
 for i in xrange(100):
-    l.append(itertools.chain(*[range(500) for j in xrange(500)]))
+    l.append(itertools.chain(*[r for j in r]))
 c = itertools.chain(*l)
 f(c)

src/core/stats.cpp

@@ -25,6 +25,7 @@ namespace pyston {
 #if STAT_TIMERS
 
 __thread StatTimer* StatTimer::stack;
+__thread uint64_t* StatTimer::counter_override;
 
 StatTimer* StatTimer::swapStack(StatTimer* s) {
     uint64_t at_time = getCPUTicks();

src/core/stats.h

@@ -129,10 +129,22 @@ private:
     int avoidability;
     bool reset_avoidability;
 
+    static __thread uint64_t* counter_override;
+
 public:
     StatTimer(uint64_t* counter, int avoidability, bool reset_avoidability = false)
         : _statcounter(counter), avoidability(avoidability), reset_avoidability(reset_avoidability) {}
 
+    static void overrideCounter(uint64_t* new_counter) {
+        assert(!counter_override);
+        counter_override = new_counter;
+    }
+
+    static void finishOverride() {
+        assert(counter_override);
+        counter_override = NULL;
+    }
+
     void pushNonTopLevel() {
 #ifndef NDEBUG
         _start_time = 0;
@@ -187,7 +199,10 @@ private:
         assert(at_time > _start_time);
 
         uint64_t _duration = at_time - _start_time;
-        Stats::log(_statcounter, _duration);
+        if (counter_override)
+            Stats::log(counter_override, _duration);
+        else
+            Stats::log(_statcounter, _duration);
 
         _start_time = 0;
     }

src/runtime/cxx_unwind.cpp

@@ -561,6 +561,9 @@ static inline void unwind_loop(ExcInfo* exc_data) {
             // we're transfering control to a non-cleanup landing pad.
             // i.e. a catch block.  thus ends our unwind session.
             endPythonUnwindSession(unwind_session);
+#if STAT_TIMERS
+            pyston::StatTimer::finishOverride();
+#endif
         }
         static_assert(THREADING_USE_GIL, "have to make the unwind session usage in this file thread safe!");
         // there is a python unwinding implementation detail leaked
@@ -660,6 +663,10 @@ extern "C" void __cxa_end_catch() {
 #define EXCINFO_TYPE_INFO _ZTIN6pyston7ExcInfoE
 extern "C" std::type_info EXCINFO_TYPE_INFO;
 
+#if STAT_TIMERS
+static uint64_t* unwinding_stattimer = pyston::Stats::getStatCounter("us_timer_unwinding");
+#endif
+
 extern "C" void __cxa_throw(void* exc_obj, std::type_info* tinfo, void (*dtor)(void*)) {
     assert(!pyston::in_cleanup_code);
     assert(exc_obj);
@@ -671,6 +678,9 @@ extern "C" void __cxa_throw(void* exc_obj, std::type_info* tinfo, void (*dtor)(v
     pyston::ExcInfo* exc_data = (pyston::ExcInfo*)exc_obj;
     checkExcInfo(exc_data);
 
+#if STAT_TIMERS
+    pyston::StatTimer::overrideCounter(unwinding_stattimer);
+#endif
     // let unwinding.cpp know we've started unwinding
     pyston::throwingException(pyston::getActivePythonUnwindSession());
     pyston::unwind(exc_data);

src/runtime/types.h

@@ -927,8 +927,8 @@ public:
 
     static Box* __get__(BoxedMethodDescriptor* self, Box* inst, Box* owner);
     static Box* __call__(BoxedMethodDescriptor* self, Box* obj, BoxedTuple* varargs, Box** _args);
-    static Box* callInternal(BoxedFunctionBase* f, CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Box* arg1,
-                             Box* arg2, Box* arg3, Box** args, const std::vector<BoxedString*>* keyword_names);
+    static Box* tppCall(Box* _self, CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Box* arg1, Box* arg2, Box* arg3,
+                        Box** args, const std::vector<BoxedString*>* keyword_names);
     static void gcHandler(GCVisitor* v, Box* _o);
 };