Add CHECK_PERIOIDIC after every call, then skip it for Python call specializations

Author: markshannon

Include/internal/pycore_opcode_metadata.h

@@ -106,6 +106,7 @@ def inner():
             ('SET_FUNCTION_ATTRIBUTE', 8, 2),
             ('PUSH_NULL', None, 1),
             ('CALL', 0, 2),                     # (lambda: x)()
+            ('CHECK_PERIODIC', 0, 2),
             ('LOAD_CONST', 2, 2),               # 2
             ('BINARY_OP', 6, 2),                # %
             ('RETURN_VALUE', None, 2)

Include/internal/pycore_uop_metadata.h

@@ -2545,6 +2545,13 @@ dummy_func(
             pushed_frame->localsplus[0] = owner;
             DEAD(owner);
             new_frame = PyStackRef_Wrap(pushed_frame);
+            /* Can't use _SAVE_RETURN_OFFSET as there is no follwoing CHECK_PERIODIC to skip */
+            #if TIER_ONE
+            frame->return_offset = (uint16_t)(next_instr - this_instr);
+            #endif
+            #if TIER_TWO
+            frame->return_offset = oparg;
+            #endif
         }
 
         macro(LOAD_ATTR_PROPERTY) =
@@ -2553,7 +2560,6 @@ dummy_func(
             _GUARD_TYPE_VERSION +
             unused/2 +
             _LOAD_ATTR_PROPERTY_FRAME +
-            _SAVE_RETURN_OFFSET +
             _PUSH_FRAME;
 
         inst(LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN, (unused/1, type_version/2, func_version/2, getattribute/4, owner -- unused)) {
@@ -4352,6 +4358,7 @@ dummy_func(
         macro(CALL_LIST_APPEND) =
             unused/1 +
             unused/2 +
+            CHECK_PERIODIC + // Do this first to avoid deopting in the middle of the qinstruction
             _GUARD_CALLABLE_LIST_APPEND +
             _GUARD_NOS_NOT_NULL +
             _GUARD_NOS_LIST +
@@ -5203,8 +5210,10 @@ dummy_func(
         }
 
         op(_SAVE_RETURN_OFFSET, (--)) {
+            // Skips following CHEKC_PERIODIC
             #if TIER_ONE
-            frame->return_offset = (uint16_t)(next_instr - this_instr);
+            assert(next_instr->op.code == CHECK_PERIODIC);
+            frame->return_offset = (uint16_t)(next_instr - this_instr)+1;
             #endif
             #if TIER_TWO
             frame->return_offset = oparg;

Lib/test/test_compiler_assemble.py

@@ -463,6 +463,7 @@ codegen_call_exit_with_nones(compiler *c, location loc)
     ADDOP_LOAD_CONST(c, loc, Py_None);
     ADDOP_LOAD_CONST(c, loc, Py_None);
     ADDOP_I(c, loc, CALL, 3);
+    ADDOP(c, loc, CHECK_PERIODIC);
     return SUCCESS;
 }
 
@@ -1504,6 +1505,7 @@ codegen_function(compiler *c, stmt_ty s, int is_async)
             ADDOP(c, loc, PUSH_NULL);
             ADDOP_I(c, loc, CALL, 0);
         }
+        ADDOP(c, loc, CHECK_PERIODIC);
     }
 
     RETURN_IF_ERROR(codegen_apply_decorators(c, decos));
@@ -1777,6 +1779,8 @@ codegen_typealias(compiler *c, stmt_ty s)
         RETURN_IF_ERROR(ret);
         ADDOP(c, loc, PUSH_NULL);
         ADDOP_I(c, loc, CALL, 0);
+        ADDOP(c, loc, CHECK_PERIODIC);
+
     }
     RETURN_IF_ERROR(codegen_nameop(c, loc, name, Store));
     return SUCCESS;
@@ -2965,6 +2969,7 @@ codegen_assert(compiler *c, stmt_ty s)
     if (s->v.Assert.msg) {
         VISIT(c, expr, s->v.Assert.msg);
         ADDOP_I(c, LOC(s), CALL, 0);
+        ADDOP(c, LOC(s), CHECK_PERIODIC);
     }
     ADDOP_I(c, LOC(s->v.Assert.test), RAISE_VARARGS, 1);
 
@@ -4135,6 +4140,7 @@ codegen_joined_str(compiler *c, expr_ty e)
             ADDOP_I(c, loc, LIST_APPEND, 1);
         }
         ADDOP_I(c, loc, CALL, 1);
+        ADDOP(c, loc, CHECK_PERIODIC);
     }
     else {
         VISIT_SEQ(c, expr, e->v.JoinedStr.values);
@@ -4871,6 +4877,7 @@ codegen_comprehension(compiler *c, expr_ty e, int type,
 
     VISIT(c, expr, outermost->iter);
     ADDOP_I(c, loc, CALL, 0);
+    ADDOP(c, loc, CHECK_PERIODIC);
 
     if (is_async_comprehension && type != COMP_GENEXP) {
         ADDOP_I(c, loc, GET_AWAITABLE, 0);
@@ -5012,6 +5019,7 @@ codegen_async_with_inner(compiler *c, stmt_ty s, int pos)
     ADDOP_I(c, loc, SWAP, 3);
     ADDOP_I(c, loc, LOAD_SPECIAL, SPECIAL___AENTER__);
     ADDOP_I(c, loc, CALL, 0);
+    ADDOP(c, loc, CHECK_PERIODIC);
     ADDOP_I(c, loc, GET_AWAITABLE, 1);
     ADDOP_LOAD_CONST(c, loc, Py_None);
     ADD_YIELD_FROM(c, loc, 1);
@@ -5121,6 +5129,7 @@ codegen_with_inner(compiler *c, stmt_ty s, int pos)
     ADDOP_I(c, loc, SWAP, 3);
     ADDOP_I(c, loc, LOAD_SPECIAL, SPECIAL___ENTER__);
     ADDOP_I(c, loc, CALL, 0);
+    ADDOP(c, loc, CHECK_PERIODIC);
     ADDOP_JUMP(c, loc, SETUP_WITH, final);
 
     /* SETUP_WITH pushes a finally block. */

Python/bytecodes.c

@@ -769,7 +769,7 @@ translate_bytecode_to_trace(
                                 oparg = orig_oparg & 0xF;
                                 break;
                             case OPARG_SAVE_RETURN_OFFSET:  // op=_SAVE_RETURN_OFFSET; oparg=return_offset
-                                oparg = offset;
+                                oparg = offset + 1;
                                 assert(uop == _SAVE_RETURN_OFFSET);
                                 break;
                             case OPARG_REPLACED: