Merge pull request #866 from undingen/perf_pyxl

from_cpython/Modules/_sre.c

@@ -257,7 +257,9 @@ static void
 data_stack_dealloc(SRE_STATE* state)
 {
     if (state->data_stack) {
-        PyMem_FREE(state->data_stack);
+        // Pyston change: use malloc
+        // PyMem_FREE(state->data_stack);
+        free(state->data_stack);
         state->data_stack = NULL;
     }
     state->data_stack_size = state->data_stack_base = 0;
@@ -273,7 +275,9 @@ data_stack_grow(SRE_STATE* state, Py_ssize_t size)
         void* stack;
         cursize = minsize+minsize/4+1024;
         TRACE(("allocate/grow stack %" PY_FORMAT_SIZE_T "d\n", cursize));
-        stack = PyMem_REALLOC(state->data_stack, cursize);
+        // Pyston change: use malloc
+        // stack = PyMem_REALLOC(state->data_stack, cursize);
+        stack = realloc(state->data_stack, cursize);
         if (!stack) {
             data_stack_dealloc(state);
             return SRE_ERROR_MEMORY;

src/core/types.h

@@ -435,6 +435,7 @@ class Rewriter;
 class RewriterVar;
 class RuntimeIC;
 class CallattrIC;
+class CallattrCapiIC;
 class NonzeroIC;
 class BinopIC;
 

src/runtime/ics.h

@@ -77,6 +77,17 @@ public:
     }
 };
 
+class CallattrCapiIC : public RuntimeIC {
+public:
+    CallattrCapiIC() : RuntimeIC((void*)callattrCapi, 1, 320) {}
+
+    Box* call(Box* obj, BoxedString* attr, CallattrFlags flags, Box* arg0, Box* arg1, Box* arg2, Box** args,
+              const std::vector<BoxedString*>* keyword_names) {
+        return (Box*)call_ptr(obj, attr, flags, arg0, arg1, arg2, args, keyword_names);
+    }
+};
+
+
 class BinopIC : public RuntimeIC {
 public:
     BinopIC() : RuntimeIC((void*)binop, 2, 240) {}

src/runtime/objmodel.cpp

@@ -3574,6 +3574,16 @@ static Box* astInterpretHelper(CLFunction* f, int num_args, BoxedClosure* closur
     return astInterpretFunction(f, num_args, closure, generator, globals, arg1, arg2, arg3, (Box**)args);
 }
 
+static Box* astInterpretHelperCapi(CLFunction* f, int num_args, BoxedClosure* closure, BoxedGenerator* generator,
+                                   Box* globals, Box** _args) noexcept {
+    try {
+        return astInterpretHelper(f, num_args, closure, generator, globals, _args);
+    } catch (ExcInfo e) {
+        setCAPIException(e);
+        return NULL;
+    }
+}
+
 // TODO: is it better to take the func_ptr last (requiring passing all the args), or is it better to put it
 // first (requiring moving all the args)?
 static Box* capiCallCxxHelper(Box* (*func_ptr)(void*, void*, void*, void*, void*), void* a, void* b, void* c, void* d,
@@ -3593,7 +3603,7 @@ Box* callCLFunc(CLFunction* f, CallRewriteArgs* rewrite_args, int num_output_arg
     CompiledFunction* chosen_cf = pickVersion(f, S, num_output_args, oarg1, oarg2, oarg3, oargs);
 
     if (!chosen_cf) {
-        if (rewrite_args && S == CXX) {
+        if (rewrite_args) {
             RewriterVar::SmallVector arg_vec;
 
             rewrite_args->rewriter->addDependenceOn(f->dependent_interp_callsites);
@@ -3620,8 +3630,10 @@ Box* callCLFunc(CLFunction* f, CallRewriteArgs* rewrite_args, int num_output_arg
             if (num_output_args >= 4)
                 arg_array->setAttr(24, rewrite_args->args);
 
-            assert(S == CXX);
-            rewrite_args->out_rtn = rewrite_args->rewriter->call(true, (void*)astInterpretHelper, arg_vec);
+            if (S == CXX)
+                rewrite_args->out_rtn = rewrite_args->rewriter->call(true, (void*)astInterpretHelper, arg_vec);
+            else
+                rewrite_args->out_rtn = rewrite_args->rewriter->call(true, (void*)astInterpretHelperCapi, arg_vec);
             rewrite_args->out_success = true;
         }
 

src/runtime/types.cpp

@@ -239,32 +239,27 @@ Box* BoxedClass::callHasnextIC(Box* obj, bool null_on_nonexistent) {
 }
 
 extern "C" PyObject* PyIter_Next(PyObject* iter) noexcept {
-    if (iter->cls->tp_iternext != slot_tp_iternext) {
-        PyObject* result;
+    Box* result = NULL;
+    if (iter->cls->tp_iternext != slot_tp_iternext)
         result = (*iter->cls->tp_iternext)(iter);
-        if (result == NULL && PyErr_Occurred() && PyErr_ExceptionMatches(PyExc_StopIteration))
-            PyErr_Clear();
-        return result;
-    }
-
-    try {
-        Box* hasnext = iter->hasnextOrNullIC();
-        if (hasnext) {
-            if (hasnext->nonzeroIC())
-                return iter->cls->callNextIC(iter);
-            else
+    else {
+        try {
+            Box* hasnext = iter->hasnextOrNullIC();
+            if (hasnext && !hasnext->nonzeroIC())
                 return NULL;
-        } else {
-            return iter->cls->callNextIC(iter);
-        }
-    } catch (ExcInfo e) {
-        if (!e.matches(StopIteration))
+        } catch (ExcInfo e) {
             setCAPIException(e);
-        return NULL;
+            return NULL;
+        }
+        result = iter->cls->call_nextIC(iter);
     }
+
+    if (result == NULL && PyErr_Occurred() && PyErr_ExceptionMatches(PyExc_StopIteration))
+        PyErr_Clear();
+    return result;
 }
 
-Box* BoxedClass::callNextIC(Box* obj) {
+Box* BoxedClass::call_nextIC(Box* obj) noexcept {
     assert(obj->cls == this);
 
     // This would work, but it would have been better to just call tp_iternext
@@ -272,7 +267,7 @@ Box* BoxedClass::callNextIC(Box* obj) {
 
     auto ic = next_ic.get();
     if (!ic) {
-        ic = new CallattrIC();
+        ic = new CallattrCapiIC();
         next_ic.reset(ic);
     }
 

src/runtime/types.h

@@ -186,10 +186,11 @@ public:
     HCAttrs attrs;
 
     // TODO: these don't actually get deallocated right now
-    std::unique_ptr<CallattrIC> hasnext_ic, next_ic, repr_ic;
+    std::unique_ptr<CallattrCapiIC> next_ic;
+    std::unique_ptr<CallattrIC> hasnext_ic, repr_ic;
     std::unique_ptr<NonzeroIC> nonzero_ic;
     Box* callHasnextIC(Box* obj, bool null_on_nonexistent);
-    Box* callNextIC(Box* obj);
+    Box* call_nextIC(Box* obj) noexcept;
     Box* callReprIC(Box* obj);
     bool callNonzeroIC(Box* obj);