Merge pull request #1186 from kmod/numpy_perf

One small change for numpy performance

src/codegen/unwinding.cpp

@@ -546,7 +546,8 @@ public:
     }
 
     std::tuple<FrameInfo*, ExcInfo, PythonStackExtractor> pause() {
-        t.end();
+        static StatCounter stat("us_unwind_session");
+        stat.log(t.end());
         return std::make_tuple(std::move(prev_frame_info), std::move(exc_info), std::move(pystack_extractor));
     }
 

src/core/ast.cpp

@@ -2223,7 +2223,7 @@ public:
     }
 };
 
-void flatten(const std::vector<AST_stmt*>& roots, std::vector<AST*>& output, bool expand_scopes) {
+void flatten(const llvm::SmallVector<AST_stmt*, 4>& roots, std::vector<AST*>& output, bool expand_scopes) {
     FlattenVisitor visitor(&output, expand_scopes);
 
     for (int i = 0; i < roots.size(); i++) {

src/core/ast.h

@@ -1432,7 +1432,7 @@ public:
 // Given an AST node, return a vector of the node plus all its descendents.
 // This is useful for analyses that care more about the constituent nodes than the
 // exact tree structure; ex, finding all "global" directives.
-void flatten(const std::vector<AST_stmt*>& roots, std::vector<AST*>& output, bool expand_scopes);
+void flatten(const llvm::SmallVector<AST_stmt*, 4>& roots, std::vector<AST*>& output, bool expand_scopes);
 void flatten(AST_expr* root, std::vector<AST*>& output, bool expand_scopes);
 // Similar to the flatten() function, but filters for a specific type of ast nodes:
 template <class T, class R> void findNodes(const R& roots, std::vector<T*>& output, bool expand_scopes) {

src/core/cfg.h

@@ -53,12 +53,12 @@ public:
     // contains the address of the entry function
     std::pair<CFGBlock*, Box*>(*entry_code)(void* interpeter, CFGBlock* block, Box** vregs);
 
-    std::vector<AST_stmt*> body;
-    std::vector<CFGBlock*> predecessors, successors;
+    llvm::SmallVector<AST_stmt*, 4> body;
+    llvm::SmallVector<CFGBlock*, 2> predecessors, successors;
     int idx; // index in the CFG
     const char* info;
 
-    typedef std::vector<AST_stmt*>::iterator iterator;
+    typedef llvm::SmallVector<AST_stmt*, 4>::iterator iterator;
 
     CFGBlock(CFG* cfg, int idx) : cfg(cfg), code(NULL), entry_code(NULL), idx(idx), info(NULL) {}
 

src/runtime/capi.cpp

@@ -1714,15 +1714,6 @@ template <ExceptionStyle S>
 Box* BoxedCApiFunction::tppCall(Box* _self, CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Box* arg1, Box* arg2,
                                 Box* arg3, Box** args,
                                 const std::vector<BoxedString*>* keyword_names) noexcept(S == CAPI) {
-    if (S == CAPI) {
-        try {
-            return tppCall<CXX>(_self, NULL, argspec, arg1, arg2, arg3, args, keyword_names);
-        } catch (ExcInfo e) {
-            setCAPIException(e);
-            return NULL;
-        }
-    }
-
     STAT_TIMER(t0, "us_timer_boxedcapifunction__call__", 10);
 
     BoxedCApiFunction* self = static_cast<BoxedCApiFunction*>(_self);
@@ -1791,6 +1782,7 @@ Box* BoxedCApiFunction::tppCall(Box* _self, CallRewriteArgs* rewrite_args, ArgPa
                 rewrite_args->out_rtn->setType(RefType::OWNED);
                 rewrite_args->out_success = true;
             }
+            assert(rtn); // This shouldn't throw, otherwise we would need to check the return value
             return rtn;
         }
         // TODO rewrite these cases specially; tp_new_wrapper just slices the args array,
@@ -1861,18 +1853,20 @@ Box* BoxedCApiFunction::tppCall(Box* _self, CallRewriteArgs* rewrite_args, ArgPa
         }
 
         if (rewrite_args) {
-            rewrite_args->rewriter->checkAndThrowCAPIException(rewrite_args->out_rtn);
+            if (S == CXX)
+                rewrite_args->rewriter->checkAndThrowCAPIException(rewrite_args->out_rtn);
             rewrite_args->out_success = true;
         }
 
         if (S == CXX && !rtn)
             throwCAPIException();
-        assert(rtn && "should have set + thrown an exception!");
         return rtn;
     };
 
-    return rearrangeArgumentsAndCall(paramspec, NULL, self->method_def->ml_name, defaults, rewrite_args, argspec, arg1,
-                                     arg2, arg3, args, keyword_names, continuation);
+    return callCXXFromStyle<S>([=]() {
+        return rearrangeArgumentsAndCall(paramspec, NULL, self->method_def->ml_name, defaults, rewrite_args, argspec,
+                                         arg1, arg2, arg3, args, keyword_names, continuation);
+    });
 }
 
 /* extension modules might be compiled with GC support so these

src/runtime/types.cpp

@@ -1162,6 +1162,9 @@ static Box* typeCallInner(CallRewriteArgs* rewrite_args, ArgPassSpec argspec, Bo
         class InitHelper {
         public:
             static Box* call(STOLEN(Box*) made, BoxedClass* cls, Box* args, Box* kwargs) noexcept(S == CAPI) {
+                if (S == CAPI && !made)
+                    return NULL;
+
                 if (!isSubclass(made->cls, cls))
                     return made;
 

test/integration/numpy_test.py

@@ -38,6 +38,9 @@ def print_progress_header(text):
 ENV_NAME = "numpy_test_env_" + os.path.basename(sys.executable)
 DEPENDENCIES = ["nose==1.3.7"]
 
+import platform
+USE_CUSTOM_PATCHES = (platform.python_implementation() == "Pyston")
+
 if not os.path.exists(ENV_NAME) or os.stat(sys.executable).st_mtime > os.stat(ENV_NAME + "/bin/python").st_mtime:
     print "Creating virtualenv to install testing dependencies..."
     VIRTUALENV_SCRIPT = os.path.dirname(__file__) + "/../lib/virtualenv/virtualenv.py"
@@ -69,9 +72,10 @@ if not os.path.exists(CYTHON_DIR):
     subprocess.check_call(["wget", url], cwd=SRC_DIR)
     subprocess.check_call(["tar", "-zxf", "Cython-0.22.tar.gz"], cwd=SRC_DIR)
 
-    PATCH_FILE = os.path.abspath(os.path.join(os.path.dirname(__file__), "Cython-0.22.patch"))
-    subprocess.check_call(["patch", "-p1", "--input=" + PATCH_FILE], cwd=CYTHON_DIR)
-    print ">>> Applied Cython patch"
+    if USE_CUSTOM_PATCHES:
+        PATCH_FILE = os.path.abspath(os.path.join(os.path.dirname(__file__), "Cython-0.22.patch"))
+        subprocess.check_call(["patch", "-p1", "--input=" + PATCH_FILE], cwd=CYTHON_DIR)
+        print ">>> Applied Cython patch"
 
 
     try:
@@ -93,8 +97,9 @@ else:
 
 PATCH_FILE = os.path.abspath(os.path.join(os.path.dirname(__file__), "numpy_patch.patch"))
 
-print_progress_header("Patching NumPy...")
-subprocess.check_call(["patch", "-p1", "--input=" + PATCH_FILE], cwd=NUMPY_DIR)
+if USE_CUSTOM_PATCHES:
+    print_progress_header("Patching NumPy...")
+    subprocess.check_call(["patch", "-p1", "--input=" + PATCH_FILE], cwd=NUMPY_DIR)
 
 try:
     env = os.environ
@@ -107,7 +112,16 @@ try:
     print_progress_header("Installing NumPy...")
     subprocess.check_call([PYTHON_EXE, "setup.py", "install"], cwd=NUMPY_DIR, env=env)
 except:
-    subprocess.check_call(["rm", "-rf", NUMPY_DIR])
+    if USE_CUSTOM_PATCHES:
+        print_progress_header("Unpatching NumPy...")
+        cmd = ["patch", "-p1", "--forward", "-i", NUMPY_PATCH_FILE, "-R", "-d", NUMPY_DIR]
+        subprocess.check_output(cmd, stderr=subprocess.STDOUT)
+
+    # TODO: I'm not sure we need to do this:
+    subprocess.check_call(["rm", "-rf", NUMPY_DIR + "/build"])
+    subprocess.check_call(["rm", "-rf", NUMPY_DIR + "/dist"])
+
+    raise
 
 # From Wikipedia
 script = """
@@ -176,9 +190,10 @@ print_progress_header("Running NumPy test suite...")
 # when all the crashes are fixed.
 # subprocess.check_call([PYTHON_EXE, "-c", numpy_test], cwd=CYTHON_DIR)
 
-print_progress_header("Unpatching NumPy...")
-cmd = ["patch", "-p1", "--forward", "-i", NUMPY_PATCH_FILE, "-R", "-d", NUMPY_DIR]
-subprocess.check_output(cmd, stderr=subprocess.STDOUT)
+if USE_CUSTOM_PATCHES:
+    print_progress_header("Unpatching NumPy...")
+    cmd = ["patch", "-p1", "--forward", "-i", NUMPY_PATCH_FILE, "-R", "-d", NUMPY_DIR]
+    subprocess.check_output(cmd, stderr=subprocess.STDOUT)
 
 print
 print "PASSED"