minor improvements

src/asm_writing/assembler.cpp

@@ -1138,21 +1138,25 @@ void Assembler::skipBytes(int num) {
     addr += num;
 }
 
-ForwardJump::ForwardJump(Assembler& assembler, ConditionCode condition)
+template <int MaxJumpSize>
+ForwardJumpBase<MaxJumpSize>::ForwardJumpBase(Assembler& assembler, ConditionCode condition)
     : assembler(assembler), condition(condition), jmp_inst(assembler.curInstPointer()) {
-    assembler.jmp_cond(JumpDestination::fromStart(assembler.bytesWritten() + max_jump_size), condition);
+    assembler.jmp_cond(JumpDestination::fromStart(assembler.bytesWritten() + MaxJumpSize), condition);
     jmp_end = assembler.curInstPointer();
 }
 
-ForwardJump::~ForwardJump() {
+template <int MaxJumpSize>
+ForwardJumpBase<MaxJumpSize>::~ForwardJumpBase() {
     uint8_t* new_pos = assembler.curInstPointer();
     int offset = new_pos - jmp_inst;
-    RELEASE_ASSERT(offset < max_jump_size, "");
+    RELEASE_ASSERT(offset < MaxJumpSize, "");
     assembler.setCurInstPointer(jmp_inst);
     assembler.jmp_cond(JumpDestination::fromStart(assembler.bytesWritten() + offset), condition);
     while (assembler.curInstPointer() < jmp_end)
         assembler.nop();
     assembler.setCurInstPointer(new_pos);
 }
+template class ForwardJumpBase<128>;
+template class ForwardJumpBase<1048576>;
 }
 }

src/asm_writing/assembler.h

@@ -214,19 +214,22 @@ public:
 // This class helps generating a forward jump with a relative offset.
 // It keeps track of the current assembler offset at construction time and in the destructor patches the
 // generated conditional jump with the correct offset depending on the number of bytes emitted in between.
-class ForwardJump {
+template <int MaxJumpSize=128>
+class ForwardJumpBase {
 private:
-    const int max_jump_size = 1048587;
     Assembler& assembler;
     ConditionCode condition;
     uint8_t* jmp_inst;
     uint8_t* jmp_end;
 
 public:
-    ForwardJump(Assembler& assembler, ConditionCode condition);
-    ~ForwardJump();
+    ForwardJumpBase(Assembler& assembler, ConditionCode condition);
+    ~ForwardJumpBase();
 };
 
+#define ForwardJump ForwardJumpBase<128>
+#define LargeForwardJump ForwardJumpBase<1048576>
+
 uint8_t* initializePatchpoint2(uint8_t* start_addr, uint8_t* slowpath_start, uint8_t* end_addr, StackInfo stack_info,
                                const std::unordered_set<int>& live_outs);
 }

src/codegen/ast_interpreter.cpp

@@ -414,9 +414,11 @@ Box* ASTInterpreter::executeInner(ASTInterpreter& interpreter, CFGBlock* start_b
             interpreter.current_inst = s;
             if (interpreter.jit)
                 interpreter.jit->emitSetCurrentInst(s);
-            Py_XDECREF(v.o);
-            if (v.var)
-                v.var->xdecref();
+            if (v.o) {
+                Py_XDECREF(v.o);
+                if (v.var)
+                    v.var->decref();
+            }
             v = interpreter.visit_stmt(s);
         }
     }

src/codegen/entry.cpp

@@ -529,9 +529,6 @@ extern "C" void Py_Initialize() noexcept {
 }
 
 void teardownCodegen() {
-    if (PROFILE)
-        g.func_addr_registry.dumpPerfMap();
-
     for (int i = 0; i < g.jit_listeners.size(); i++) {
         g.engine->UnregisterJITEventListener(g.jit_listeners[i]);
         delete g.jit_listeners[i];

src/runtime/types.cpp

@@ -4325,6 +4325,9 @@ extern "C" void Py_Finalize() noexcept {
 
     // wait_for_thread_shutdown();
 
+    if (PROFILE)
+        g.func_addr_registry.dumpPerfMap();
+
     call_sys_exitfunc();
     // initialized = 0;
 

tools/annotate.py

@@ -57,6 +57,7 @@ def lookupAsHeapAddr(n):
     lines = []
     while True:
         l = _heap_proc.stdout.readline()
+        assert l, "heapmap subprocess exited? code: %r" % _heap_proc.poll()
         if l == '!!!!\n':
             break
         lines.append(l)
@@ -95,7 +96,7 @@ if __name__ == "__main__":
 
     parser = argparse.ArgumentParser(formatter_class=argparse.RawTextHelpFormatter)
     parser.add_argument("func_name", metavar="FUNC_NAME")
-    parser.add_argument("--collapse-nops", dest="collapse_nops", action="store_true", default=True)
+    parser.add_argument("--collapse-nops", dest="collapse_nops", action="store", default=5, type=int)
     parser.add_argument("--no-collapse-nops", dest="collapse_nops", action="store_false")
     parser.add_argument("--heap-map-args", nargs='+', help="""
 Command to run that will provide heap map information.
@@ -157,7 +158,13 @@ equivalent to '--heap-map-args ./pyston_release -i BENCHMARK'.
                 nops = (nops[0] + count, nops[1] + 1, nops[2], addr)
         else:
             if nops:
-                print str(nops[0]).rjust(8), ("    %s-%s              nop*%d" % (nops[2], nops[3], nops[1])).ljust(70)
+                if int(nops[3], 16) - int(nops[2], 16) + 1 <= args.collapse_nops:
+                    nop_count = nops[0]
+                    for addr in xrange(int(nops[2], 16), int(nops[3], 16) + 1):
+                        print str(nop_count).rjust(8), ("    %s              nop" % ("%x:      90" % addr).ljust(29)).ljust(70)
+                        nop_count = 0
+                else:
+                    print str(nops[0]).rjust(8), ("    %s              nop*%d" % (("%s-%s" % (nops[2], nops[3])).ljust(29), nops[1])).ljust(70)
                 nops = None
             print str(count).rjust(8), l.ljust(70), extra