Mark replaced uops in expansion table

Author: markshannon

Include/internal/pycore_opcode_metadata.h

@@ -176,6 +176,7 @@ def f():
         with temporary_optimizer(opt):
             f()
         exe = get_first_executor(f)
+        self.assertIsNotNone(exe)
         self.assertTrue(exe.is_valid())
         _testinternalcapi.invalidate_executors(f.__code__)
         self.assertFalse(exe.is_valid())
@@ -196,7 +197,7 @@ def testfunc(x):
         self.assertIsNotNone(ex)
         uops = {opname for opname, _, _ in ex}
         self.assertIn("_SET_IP", uops)
-        self.assertIn("LOAD_FAST", uops)
+        self.assertIn("_LOAD_FAST", uops)
 
     def test_extended_arg(self):
         "Check EXTENDED_ARG handling in superblock creation"

Lib/test/test_capi/test_opt.py

@@ -633,14 +633,6 @@ translate_bytecode_to_trace(
                                         oparg += extras;
                                     }
                                 }
-                                if (_PyUOp_Replacements[uop]) {
-                                    uop = _PyUOp_Replacements[uop];
-                                    if (uop == _FOR_ITER_TIER_TWO) {
-                                        target += 1 + INLINE_CACHE_ENTRIES_FOR_ITER + oparg + 1;
-                                        assert(_PyCode_CODE(code)[target-1].op.code == END_FOR ||
-                                               _PyCode_CODE(code)[target-1].op.code == INSTRUMENTED_END_FOR);
-                                    }
-                                }
                                 break;
                             case OPARG_CACHE_1:
                                 operand = read_u16(&instr[offset].cache);
@@ -661,7 +653,15 @@ translate_bytecode_to_trace(
                                 oparg = offset;
                                 assert(uop == _SAVE_RETURN_OFFSET);
                                 break;
-
+                            case OPARG_REPLACED:
+                                uop = _PyUOp_Replacements[uop];
+                                assert(uop != 0);
+                                if (uop == _FOR_ITER_TIER_TWO) {
+                                    target += 1 + INLINE_CACHE_ENTRIES_FOR_ITER + oparg + 1;
+                                    assert(_PyCode_CODE(code)[target-1].op.code == END_FOR ||
+                                            _PyCode_CODE(code)[target-1].op.code == INSTRUMENTED_END_FOR);
+                                }
+                                break;
                             default:
                                 fprintf(stderr,
                                         "opcode=%d, oparg=%d; nuops=%d, i=%d; size=%d, offset=%d\n",

Python/optimizer.c

@@ -35,6 +35,8 @@
     "OPARG_TOP": 5,
     "OPARG_BOTTOM": 6,
     "OPARG_SAVE_RETURN_OFFSET": 7,
+    # Skip 8 as the other powers of 2 are sizes
+    "OPARG_REPLACED": 9
 }
 
 
@@ -244,6 +246,8 @@ def generate_expansion_table(
                     # Skip specializations
                     if "specializing" in part.annotations:
                         continue
+                    if "replaced" in part.annotations:
+                        size = OPARG_SIZES["OPARG_REPLACED"]
                     expansions.append((part.name, size, offset if size else 0))
                 offset += part.size
         expansions_table[inst.name] = expansions
@@ -267,9 +271,11 @@ def is_viable_expansion(inst: Instruction) -> bool:
     "An instruction can be expanded if all its parts are viable for tier 2"
     for part in inst.parts:
         if isinstance(part, Uop):
-            # Skip specializations
+            # Skip specializing and replaced uops
             if "specializing" in part.annotations:
                 continue
+            if "replaced" in part.annotations:
+                continue
             if part.properties.tier_one_only or not part.is_viable():
                 return False
     return True