Merge pull request #958 from undingen/perf_new5

Misc perf improvements

CMakeLists.txt

@@ -286,7 +286,7 @@ macro(add_pyston_test testname directory)
 endmacro()
 
 # tests         testname directory arguments
-add_pyston_test(defaults tests --order-by-mtime)
+add_pyston_test(defaults tests --order-by-mtime -t50)
 add_pyston_test(force_llvm tests -a=-n -a=-x -t50)
 if(${CMAKE_BUILD_TYPE} STREQUAL "Release")
   add_pyston_test(max_compilation_tier tests -a=-O -a=-x -t50)

src/codegen/ast_interpreter.cpp

@@ -68,8 +68,7 @@ class ASTInterpreter {
 public:
     ASTInterpreter(CLFunction* clfunc, Box** vregs);
 
-    void initArguments(int nargs, BoxedClosure* closure, BoxedGenerator* generator, Box* arg1, Box* arg2, Box* arg3,
-                       Box** args);
+    void initArguments(BoxedClosure* closure, BoxedGenerator* generator, Box* arg1, Box* arg2, Box* arg3, Box** args);
 
     static Box* execute(ASTInterpreter& interpreter, CFGBlock* start_block = NULL, AST_stmt* start_at = NULL);
     static Box* executeInner(ASTInterpreter& interpreter, CFGBlock* start_block, AST_stmt* start_at);
@@ -251,8 +250,8 @@ ASTInterpreter::ASTInterpreter(CLFunction* clfunc, Box** vregs)
     assert(scope_info);
 }
 
-void ASTInterpreter::initArguments(int nargs, BoxedClosure* _closure, BoxedGenerator* _generator, Box* arg1, Box* arg2,
-                                   Box* arg3, Box** args) {
+void ASTInterpreter::initArguments(BoxedClosure* _closure, BoxedGenerator* _generator, Box* arg1, Box* arg2, Box* arg3,
+                                   Box** args) {
     passed_closure = _closure;
     generator = _generator;
 
@@ -280,8 +279,7 @@ void ASTInterpreter::initArguments(int nargs, BoxedClosure* _closure, BoxedGener
             val = createDict();
         doStore(param_names.kwarg_name, Value(val, 0));
     }
-
-    assert(nargs == i);
+    assert(i == param_names.totalParameters());
 }
 
 void ASTInterpreter::startJITing(CFGBlock* block, int exit_offset) {
@@ -1722,8 +1720,8 @@ static int calculateNumVRegs(CLFunction* clfunc) {
     return cfg->sym_vreg_map.size();
 }
 
-Box* astInterpretFunction(CLFunction* clfunc, int nargs, Box* closure, Box* generator, Box* globals, Box* arg1,
-                          Box* arg2, Box* arg3, Box** args) {
+Box* astInterpretFunction(CLFunction* clfunc, Box* closure, Box* generator, Box* globals, Box* arg1, Box* arg2,
+                          Box* arg3, Box** args) {
     UNAVOIDABLE_STAT_TIMER(t0, "us_timer_in_interpreter");
 
     SourceInfo* source_info = clfunc->source.get();
@@ -1731,12 +1729,8 @@ Box* astInterpretFunction(CLFunction* clfunc, int nargs, Box* closure, Box* gene
     assert((!globals) == source_info->scoping->areGlobalsFromModule());
     bool can_reopt = ENABLE_REOPT && !FORCE_INTERPRETER;
 
-    // If the cfg hasn't been computed yet, just conservatively say that it will be a big function.
-    // It shouldn't matter, since the cfg should only be NULL if this is the first execution of this
-    // function.
-    int num_blocks = source_info->cfg ? source_info->cfg->blocks.size() : 10000;
-    int threshold = num_blocks <= 20 ? (REOPT_THRESHOLD_BASELINE / 3) : REOPT_THRESHOLD_BASELINE;
-    if (unlikely(can_reopt && (FORCE_OPTIMIZE || !ENABLE_INTERPRETER || clfunc->times_interpreted > threshold))) {
+    if (unlikely(can_reopt
+                 && (FORCE_OPTIMIZE || !ENABLE_INTERPRETER || clfunc->times_interpreted > REOPT_THRESHOLD_BASELINE))) {
         clfunc->times_interpreted = 0;
 
         EffortLevel new_effort = EffortLevel::MODERATE;
@@ -1744,7 +1738,7 @@ Box* astInterpretFunction(CLFunction* clfunc, int nargs, Box* closure, Box* gene
             new_effort = EffortLevel::MAXIMAL;
 
         std::vector<ConcreteCompilerType*> arg_types;
-        for (int i = 0; i < nargs; i++) {
+        for (int i = 0; i < clfunc->param_names.totalParameters(); i++) {
             Box* arg = getArg(i, arg1, arg2, arg3, args);
 
             assert(arg || i == clfunc->param_names.kwargsIndex()); // only builtin functions can pass NULL args
@@ -1813,7 +1807,7 @@ Box* astInterpretFunction(CLFunction* clfunc, int nargs, Box* closure, Box* gene
         interpreter.setGlobals(source_info->parent_module);
     }
 
-    interpreter.initArguments(nargs, (BoxedClosure*)closure, (BoxedGenerator*)generator, arg1, arg2, arg3, args);
+    interpreter.initArguments((BoxedClosure*)closure, (BoxedGenerator*)generator, arg1, arg2, arg3, args);
     Box* v = ASTInterpreter::execute(interpreter);
     return v ? v : None;
 }
@@ -1829,7 +1823,7 @@ Box* astInterpretFunctionEval(CLFunction* clfunc, Box* globals, Box* boxedLocals
     }
 
     ASTInterpreter interpreter(clfunc, vregs);
-    interpreter.initArguments(0, NULL, NULL, NULL, NULL, NULL, NULL);
+    interpreter.initArguments(NULL, NULL, NULL, NULL, NULL, NULL);
     interpreter.setBoxedLocals(boxedLocals);
 
     assert(!clfunc->source->scoping->areGlobalsFromModule());
@@ -1860,8 +1854,7 @@ static Box* astInterpretDeoptInner(CLFunction* clfunc, AST_expr* after_expr, AST
     }
 
     ASTInterpreter interpreter(clfunc, vregs);
-
-    assert(clfunc->source->scoping->areGlobalsFromModule());
+    if (source_info->scoping->areGlobalsFromModule())
         interpreter.setGlobals(source_info->parent_module);
 
     for (const auto& p : *frame_state.locals) {
@@ -1873,6 +1866,9 @@ static Box* astInterpretDeoptInner(CLFunction* clfunc, AST_expr* after_expr, AST
             interpreter.setPassedClosure(p.second);
         } else if (name == CREATED_CLOSURE_NAME) {
             interpreter.setCreatedClosure(p.second);
+        } else if (name == PASSED_GLOBALS_NAME) {
+            assert(!source_info->scoping->areGlobalsFromModule());
+            interpreter.setGlobals(p.second);
         } else {
             InternedString interned = clfunc->source->getInternedStrings().get(name);
             interpreter.addSymbol(interned, p.second, false);
@@ -1886,7 +1882,7 @@ static Box* astInterpretDeoptInner(CLFunction* clfunc, AST_expr* after_expr, AST
     while (true) {
         if (enclosing_stmt->type == AST_TYPE::Assign) {
             auto asgn = ast_cast<AST_Assign>(enclosing_stmt);
-            assert(asgn->value == after_expr);
+            RELEASE_ASSERT(asgn->value == after_expr, "%p %p", asgn->value, after_expr);
             assert(asgn->targets.size() == 1);
             assert(asgn->targets[0]->type == AST_TYPE::Name);
             auto name = ast_cast<AST_Name>(asgn->targets[0]);
@@ -1895,6 +1891,7 @@ static Box* astInterpretDeoptInner(CLFunction* clfunc, AST_expr* after_expr, AST
             break;
         } else if (enclosing_stmt->type == AST_TYPE::Expr) {
             auto expr = ast_cast<AST_Expr>(enclosing_stmt);
+            RELEASE_ASSERT(expr->value == after_expr, "%p %p", expr->value, after_expr);
             assert(expr->value == after_expr);
             break;
         } else if (enclosing_stmt->type == AST_TYPE::Invoke) {

src/codegen/ast_interpreter.h

@@ -70,8 +70,8 @@ struct Value {
     Value(Box* o, RewriterVar* var) : o(o), var(var) {}
 };
 
-Box* astInterpretFunction(CLFunction* f, int nargs, Box* closure, Box* generator, Box* globals, Box* arg1, Box* arg2,
-                          Box* arg3, Box** args);
+Box* astInterpretFunction(CLFunction* f, Box* closure, Box* generator, Box* globals, Box* arg1, Box* arg2, Box* arg3,
+                          Box** args);
 Box* astInterpretFunctionEval(CLFunction* cf, Box* globals, Box* boxedLocals);
 Box* astInterpretDeopt(CLFunction* cf, AST_expr* after_expr, AST_stmt* enclosing_stmt, Box* expr_val,
                        FrameStackState frame_state);

src/codegen/irgen/hooks.cpp

@@ -326,7 +326,7 @@ void compileAndRunModule(AST_Module* m, BoxedModule* bm) {
     }
 
     UNAVOIDABLE_STAT_TIMER(t0, "us_timer_interpreted_module_toplevel");
-    Box* r = astInterpretFunction(clfunc, 0, NULL, NULL, NULL, NULL, NULL, NULL, NULL);
+    Box* r = astInterpretFunction(clfunc, NULL, NULL, NULL, NULL, NULL, NULL, NULL);
     assert(r == None);
 }
 

src/runtime/objmodel.cpp

@@ -3936,20 +3936,30 @@ static Box* callChosenCF(CompiledFunction* chosen_cf, BoxedClosure* closure, Box
 // This function exists for the rewriter: astInterpretFunction takes 9 args, but the rewriter
 // only supports calling functions with at most 6 since it can currently only pass arguments
 // in registers.
-static Box* astInterpretHelper(CLFunction* f, int num_args, BoxedClosure* closure, BoxedGenerator* generator,
-                               Box* globals, Box** _args) {
+static Box* astInterpretHelper(CLFunction* f, BoxedClosure* closure, BoxedGenerator* generator, Box* globals,
+                               Box** _args) {
     Box* arg1 = _args[0];
     Box* arg2 = _args[1];
     Box* arg3 = _args[2];
     Box* args = _args[3];
 
-    return astInterpretFunction(f, num_args, closure, generator, globals, arg1, arg2, arg3, (Box**)args);
+    return astInterpretFunction(f, 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 {
+static Box* astInterpretHelperCapi(CLFunction* f, BoxedClosure* closure, BoxedGenerator* generator, Box* globals,
+                                   Box** _args) noexcept {
     try {
-        return astInterpretHelper(f, num_args, closure, generator, globals, _args);
+        return astInterpretHelper(f, closure, generator, globals, _args);
+    } catch (ExcInfo e) {
+        setCAPIException(e);
+        return NULL;
+    }
+}
+
+static Box* astInterpretHelper2ArgsCapi(CLFunction* f, BoxedClosure* closure, BoxedGenerator* generator, Box* globals,
+                                        Box* arg1, Box* arg2) noexcept {
+    try {
+        return astInterpretFunction(f, closure, generator, globals, arg1, arg2, NULL, NULL);
     } catch (ExcInfo e) {
         setCAPIException(e);
         return NULL;
@@ -3988,14 +3998,24 @@ Box* callCLFunc(CLFunction* f, CallRewriteArgs* rewrite_args, int num_output_arg
             // TODO this kind of embedded reference needs to be tracked by the GC somehow?
             // Or maybe it's ok, since we've guarded on the function object?
             arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)f, Location::forArg(0)));
-            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)num_output_args, Location::forArg(1)));
-            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)closure, Location::forArg(2)));
-            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)generator, Location::forArg(3)));
-            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)globals, Location::forArg(4)));
+            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)closure, Location::forArg(1)));
+            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)generator, Location::forArg(2)));
+            arg_vec.push_back(rewrite_args->rewriter->loadConst((intptr_t)globals, Location::forArg(3)));
+
+            if (num_output_args <= 2) {
+                if (num_output_args >= 1)
+                    arg_vec.push_back(rewrite_args->arg1);
+                if (num_output_args >= 2)
+                    arg_vec.push_back(rewrite_args->arg2);
 
+                if (S == CXX)
+                    rewrite_args->out_rtn = rewrite_args->rewriter->call(true, (void*)astInterpretFunction, arg_vec);
+                else
+                    rewrite_args->out_rtn
+                        = rewrite_args->rewriter->call(true, (void*)astInterpretHelper2ArgsCapi, arg_vec);
+            } else {
                 // Hacky workaround: the rewriter can only pass arguments in registers, so use this helper function
                 // to unpack some of the additional arguments:
-            // TODO if there's only one arg we could just pass it normally
                 RewriterVar* arg_array = rewrite_args->rewriter->allocate(4);
                 arg_vec.push_back(arg_array);
                 if (num_output_args >= 1)
@@ -4011,19 +4031,20 @@ Box* callCLFunc(CLFunction* f, CallRewriteArgs* rewrite_args, int num_output_arg
                     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;
         }
 
         if (S == CAPI) {
             try {
-                return astInterpretFunction(f, num_output_args, closure, generator, globals, oarg1, oarg2, oarg3,
-                                            oargs);
+                return astInterpretFunction(f, closure, generator, globals, oarg1, oarg2, oarg3, oargs);
             } catch (ExcInfo e) {
                 setCAPIException(e);
                 return NULL;
             }
         } else {
-            return astInterpretFunction(f, num_output_args, closure, generator, globals, oarg1, oarg2, oarg3, oargs);
+            return astInterpretFunction(f, closure, generator, globals, oarg1, oarg2, oarg3, oargs);
         }
     }