Merge pull request #1324 from undingen/small_impr

use PyObject_Malloc in pyston::DenseMap, smaller ICInvalidator, reduce malloc calls

src/core/from_llvm/DenseMap.h

@@ -50,6 +50,9 @@
 #include "llvm/Support/PointerLikeTypeTraits.h"
 #include "llvm/Support/type_traits.h"
 
+// Pyston change: use the python allocator
+#include "Python.h"
+
 namespace pyston {
 
 // This should only take effect for this header file:
@@ -721,6 +724,10 @@ private:
     Buckets = static_cast<BucketT*>(operator new(sizeof(BucketT) * NumBuckets));
     return true;
   }
+
+  // Pyston change: use the python allocator
+  static void* operator new(size_t count) { return PyObject_Malloc(count); }
+  static void operator delete(void* p) { PyObject_Free(p); }
 };
 
 template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4,

src/core/types.h

@@ -25,6 +25,7 @@
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/iterator_range.h"
+#include "llvm/ADT/SmallPtrSet.h"
 #include "Python.h"
 
 #include "core/common.h"
@@ -288,7 +289,7 @@ struct ParamReceiveSpec {
 class ICInvalidator {
 private:
     int64_t cur_version;
-    std::unordered_set<ICSlotInfo*> dependents;
+    llvm::SmallPtrSet<ICSlotInfo*, 2> dependents;
 
 public:
     ICInvalidator() : cur_version(0) {}

src/runtime/int.cpp

@@ -342,17 +342,17 @@ extern "C" PyObject* PyInt_FromString(const char* s, char** pend, int base) noex
 #ifdef Py_USING_UNICODE
 extern "C" PyObject* PyInt_FromUnicode(Py_UNICODE* s, Py_ssize_t length, int base) noexcept {
     PyObject* result;
-    char* buffer = (char*)malloc(length + 1);
+    char* buffer = (char*)PyMem_MALLOC(length + 1);
 
     if (buffer == NULL)
         return PyErr_NoMemory();
 
     if (PyUnicode_EncodeDecimal(s, length, buffer, NULL)) {
-        free(buffer);
+        PyMem_FREE(buffer);
         return NULL;
     }
     result = PyInt_FromString(buffer, NULL, base);
-    free(buffer);
+    PyMem_FREE(buffer);
     return result;
 }
 #endif

src/runtime/set.cpp

@@ -515,9 +515,6 @@ Box* setUnion(BoxedSet* self, BoxedTuple* args) {
     BoxedSet* rtn = makeNewSet(self->cls, self);
     AUTO_DECREF(rtn);
 
-    for (auto&& p : self->s)
-        _setAdd(rtn, p);
-
     for (auto container : args->pyElements()) {
         AUTO_DECREF(container);
         for (auto elt : container->pyElements()) {

src/runtime/tuple.cpp

@@ -265,7 +265,7 @@ Box* tupleRepr(Box* _t) {
 
     BoxedTuple* t = (BoxedTuple*)_t;
     int n;
-    std::vector<char> chars;
+    llvm::SmallVector<char, 128> chars;
     int status = Py_ReprEnter((PyObject*)t);
     n = t->size();
     if (n == 0) {
@@ -427,7 +427,7 @@ extern "C" Box* tupleNew(Box* _cls, BoxedTuple* args, BoxedDict* kwargs) {
             return r;
         }
 
-        std::vector<Box*> elts;
+        llvm::SmallVector<Box*, 16> elts;
         try {
             for (auto e : elements->pyElements())
                 elts.push_back(e);
@@ -437,10 +437,16 @@ extern "C" Box* tupleNew(Box* _cls, BoxedTuple* args, BoxedDict* kwargs) {
             throw e;
         }
 
-        auto rtn = BoxedTuple::create(elts.size(), cls);
-        memcpy(&rtn->elts[0], &elts[0], elts.size() * sizeof(Box*));
+        if (elts.empty()) {
+            if (cls == tuple_cls)
+                return incref(EmptyTuple);
+            return BoxedTuple::create(0, cls);
+        } else {
+            auto rtn = BoxedTuple::create(elts.size(), cls);
+            memcpy(&rtn->elts[0], &elts[0], elts.size() * sizeof(Box*));
+            return rtn;
+        }
 
-        return rtn;
     } else {
         if (cls == tuple_cls)
             return incref(EmptyTuple);