gh-106529: Support JUMP_BACKWARD in Tier 2 (uops) (#106543)

During superblock generation, a JUMP_BACKWARD instruction is translated to either a JUMP_TO_TOP micro-op (when the target of the jump is exactly the beginning of the superblock, closing the loop), or a SAVE_IP + EXIT_TRACE pair, when the jump goes elsewhere. The new JUMP_TO_TOP instruction includes a CHECK_EVAL_BREAKER() call, so a closed loop can still be interrupted.
python · Jul 11, 2023 · cabd6e8 · cabd6e8
1 parent 292ac4b
commit cabd6e8
Show file tree

Hide file tree

Showing 5 changed files with 63 additions and 30 deletions.
diff --git a/Lib/test/test_capi/test_misc.py b/Lib/test/test_capi/test_misc.py
@@ -2525,7 +2525,6 @@ def testfunc(n):
                 i += 1
 
         opt = _testinternalcapi.get_uop_optimizer()
-
         with temporary_optimizer(opt):
             testfunc(10)
 
@@ -2541,7 +2540,6 @@ def testfunc(n):
                 i += 1
 
         opt = _testinternalcapi.get_uop_optimizer()
-
         with temporary_optimizer(opt):
             testfunc(10)
 
@@ -2550,6 +2548,20 @@ def testfunc(n):
         uops = {opname for opname, _ in ex}
         self.assertIn("_POP_JUMP_IF_TRUE", uops)
 
+    def test_jump_backward(self):
+        def testfunc(n):
+            i = 0
+            while i < n:
+                i += 1
+        opt = _testinternalcapi.get_uop_optimizer()
+        with temporary_optimizer(opt):
+            testfunc(10)
+
+        ex = get_first_executor(testfunc)
+        self.assertIsNotNone(ex)
+        uops = {opname for opname, _ in ex}
+        self.assertIn("JUMP_TO_TOP", uops)
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/Python/ceval.c b/Python/ceval.c
@@ -2783,6 +2783,13 @@ _PyUopExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject
                 break;
             }
 
+            case JUMP_TO_TOP:
+            {
+                pc = 0;
+                CHECK_EVAL_BREAKER();
+                break;
+            }
+
             case SAVE_IP:
             {
                 frame->prev_instr = ip_offset + oparg;

diff --git a/Python/opcode_metadata.h b/Python/opcode_metadata.h
diff --git a/Python/optimizer.c b/Python/optimizer.c
@@ -372,9 +372,7 @@ translate_bytecode_to_trace(
     _PyUOpInstruction *trace,
     int buffer_size)
 {
-#ifdef Py_DEBUG
     _Py_CODEUNIT *initial_instr = instr;
-#endif
     int trace_length = 0;
     int max_length = buffer_size;
 
@@ -456,6 +454,19 @@ translate_bytecode_to_trace(
                 break;
             }
 
+            case JUMP_BACKWARD:
+            {
+                if (instr + 2 - oparg == initial_instr
+                    && trace_length + 3 <= max_length)
+                {
+                    ADD_TO_TRACE(JUMP_TO_TOP, 0);
+                }
+                else {
+                    DPRINTF(2, "JUMP_BACKWARD not to top ends trace\n");
+                }
+                goto done;
+            }
+
             default:
             {
                 const struct opcode_macro_expansion *expansion = &_PyOpcode_macro_expansion[opcode];

diff --git a/Tools/cases_generator/generate_cases.py b/Tools/cases_generator/generate_cases.py
@@ -1346,6 +1346,7 @@ def add(name: str) -> None:
         add("SAVE_IP")
         add("_POP_JUMP_IF_FALSE")
         add("_POP_JUMP_IF_TRUE")
+        add("JUMP_TO_TOP")
 
         for instr in self.instrs.values():
             if instr.kind == "op" and instr.is_viable_uop():