gh-148211: decompose _POP_TWO/_POP_CALL(_ONE/_TWO) in JIT (GH-148377)

Author: NekoAsakura

Include/internal/pycore_uop_ids.h

@@ -1643,7 +1643,6 @@ def testfunc(n):
         self.assertIsNotNone(ex)
         uops = get_opnames(ex)
         self.assertNotIn("_BINARY_OP_ADD_INT", uops)
-        self.assertNotIn("_POP_TWO", uops)
         self.assertNotIn("_GUARD_NOS_INT", uops)
         self.assertNotIn("_GUARD_TOS_INT", uops)
 
@@ -1862,7 +1861,6 @@ def testfunc(n):
         self.assertIsNotNone(ex)
         uops = get_opnames(ex)
         self.assertNotIn("_COMPARE_OP", uops)
-        self.assertNotIn("_POP_TWO", uops)
 
     def test_compare_op_int_insert_two_load_const_inline_borrow(self):
         def testfunc(n):
@@ -1930,7 +1928,6 @@ def testfunc(n):
         self.assertIsNotNone(ex)
         uops = get_opnames(ex)
         self.assertNotIn("_CONTAINS_OP", uops)
-        self.assertNotIn("_POP_TWO", uops)
 
     def test_to_bool_bool_contains_op_set(self):
         """

Include/internal/pycore_uop_metadata.h

@@ -381,11 +381,6 @@ dummy_func(
             PyStackRef_CLOSE_SPECIALIZED(value, _PyUnicode_ExactDealloc);
         }
 
-        tier2 op(_POP_TWO, (nos, tos --)) {
-            PyStackRef_CLOSE(tos);
-            PyStackRef_CLOSE(nos);
-        }
-
         op(_POP_TOP_OPARG, (args[oparg] -- )) {
             _PyStackRef_CloseStack(args, oparg);
             DEAD(args);
@@ -5922,27 +5917,6 @@ dummy_func(
             value = PyStackRef_FromPyObjectBorrow(ptr);
         }
 
-        tier2 op(_POP_CALL, (callable, null --)) {
-            (void)null; // Silence compiler warnings about unused variables
-            DEAD(null);
-            PyStackRef_CLOSE(callable);
-        }
-
-        tier2 op(_POP_CALL_ONE, (callable, null, pop --)) {
-            PyStackRef_CLOSE(pop);
-            (void)null; // Silence compiler warnings about unused variables
-            DEAD(null);
-            PyStackRef_CLOSE(callable);
-        }
-
-        tier2 op(_POP_CALL_TWO, (callable, null, pop1, pop2 --)) {
-            PyStackRef_CLOSE(pop2);
-            PyStackRef_CLOSE(pop1);
-            (void)null; // Silence compiler warnings about unused variables
-            DEAD(null);
-            PyStackRef_CLOSE(callable);
-        }
-
         tier2 op(_SHUFFLE_3_LOAD_CONST_INLINE_BORROW, (ptr/4, callable, null, arg -- res, a, c)) {
             res = PyStackRef_FromPyObjectBorrow(ptr);
             a = arg;

Lib/test/test_capi/test_opt.py

@@ -677,6 +677,7 @@ const uint16_t op_without_push[MAX_UOP_ID + 1] = {
     [_LOAD_FAST] = _NOP,
     [_LOAD_FAST_BORROW] = _NOP,
     [_LOAD_SMALL_INT] = _NOP,
+    [_PUSH_NULL] = _NOP,
 };
 
 const bool op_skip[MAX_UOP_ID + 1] = {
@@ -692,13 +693,6 @@ const uint16_t op_without_pop[MAX_UOP_ID + 1] = {
     [_POP_TOP_INT] = _NOP,
     [_POP_TOP_FLOAT] = _NOP,
     [_POP_TOP_UNICODE] = _NOP,
-    [_POP_TWO] = _POP_TOP,
-    [_POP_CALL_TWO] = _POP_CALL_ONE,
-    [_POP_CALL_ONE] = _POP_CALL,
-};
-
-const uint16_t op_without_pop_null[MAX_UOP_ID + 1] = {
-    [_POP_CALL] = _POP_TOP,
 };
 
 
@@ -733,10 +727,10 @@ remove_unneeded_uops(_PyUOpInstruction *buffer, int buffer_size)
             default:
             {
                 // Cancel out pushes and pops, repeatedly. So:
-                //     _LOAD_FAST + _POP_TWO + _LOAD_CONST_INLINE_BORROW + _POP_TOP
+                //     _LOAD_FAST + _POP_TOP + _POP_TOP + _LOAD_CONST_INLINE_BORROW + _POP_TOP
                 // ...becomes:
-                //     _NOP + _POP_TOP + _NOP + _NOP
-                while (op_without_pop[opcode] || op_without_pop_null[opcode]) {
+                //     _NOP + _NOP + _POP_TOP + _NOP + _NOP
+                while (op_without_pop[opcode]) {
                     _PyUOpInstruction *last = &buffer[pc - 1];
                     while (op_skip[last->opcode]) {
                         last--;
@@ -749,14 +743,6 @@ remove_unneeded_uops(_PyUOpInstruction *buffer, int buffer_size)
                             pc = (int)(last - buffer);
                         }
                     }
-                    else if (last->opcode == _PUSH_NULL) {
-                        // Handle _POP_CALL separately.
-                        // This looks for a preceding _PUSH_NULL instruction and
-                        // simplifies to _POP_TOP.
-                        last->opcode = _NOP;
-                        opcode = buffer[pc].opcode = op_without_pop_null[opcode];
-                        assert(opcode);
-                    }
                     else {
                         break;
                     }

Python/bytecodes.c

@@ -1272,7 +1272,10 @@ dummy_func(void) {
                 out = Py_True;
             }
             sym_set_const(res, out);
-            ADD_OP(_POP_CALL_TWO, 0, 0);
+            ADD_OP(_POP_TOP, 0, 0);
+            ADD_OP(_POP_TOP, 0, 0);
+            ADD_OP(_POP_TOP_NOP, 0, 0);
+            ADD_OP(_POP_TOP, 0, 0);
             ADD_OP(_LOAD_CONST_INLINE_BORROW, 0, (uintptr_t)out);
         }
     }