gh-106581: Split CALL_PY_EXACT_ARGS into uops (#107760)

* Split `CALL_PY_EXACT_ARGS` into uops

This is only the first step for doing `CALL` in Tier 2.
The next step involves tracing into the called code object and back.
After that we'll have to do the remaining `CALL` specialization.
Finally we'll have to deal with `KW_NAMES`.

Note: this moves setting `frame->return_offset` directly in front of
`DISPATCH_INLINED()`, to make it easier to move it into `_PUSH_FRAME`.

Author: gvanrossum

Include/internal/pycore_opcode_metadata.h

@@ -2618,6 +2618,23 @@ def testfunc(it):
             with self.assertRaises(StopIteration):
                 next(it)
 
+    def test_call_py_exact_args(self):
+        def testfunc(n):
+            def dummy(x):
+                return x+1
+            for i in range(n):
+                dummy(i)
+
+        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("_PUSH_FRAME", uops)
+
+
 
 if __name__ == "__main__":
     unittest.main()

Lib/test/test_capi/test_misc.py

@@ -956,13 +956,13 @@ dummy_func(
             {
                 PyGenObject *gen = (PyGenObject *)receiver;
                 _PyInterpreterFrame *gen_frame = (_PyInterpreterFrame *)gen->gi_iframe;
-                frame->return_offset = oparg;
                 STACK_SHRINK(1);
                 _PyFrame_StackPush(gen_frame, v);
                 gen->gi_frame_state = FRAME_EXECUTING;
                 gen->gi_exc_state.previous_item = tstate->exc_info;
                 tstate->exc_info = &gen->gi_exc_state;
                 SKIP_OVER(INLINE_CACHE_ENTRIES_SEND);
+                frame->return_offset = oparg;
                 DISPATCH_INLINED(gen_frame);
             }
             if (Py_IsNone(v) && PyIter_Check(receiver)) {
@@ -995,13 +995,13 @@ dummy_func(
             DEOPT_IF(gen->gi_frame_state >= FRAME_EXECUTING, SEND);
             STAT_INC(SEND, hit);
             _PyInterpreterFrame *gen_frame = (_PyInterpreterFrame *)gen->gi_iframe;
-            frame->return_offset = oparg;
             STACK_SHRINK(1);
             _PyFrame_StackPush(gen_frame, v);
             gen->gi_frame_state = FRAME_EXECUTING;
             gen->gi_exc_state.previous_item = tstate->exc_info;
             tstate->exc_info = &gen->gi_exc_state;
             SKIP_OVER(INLINE_CACHE_ENTRIES_SEND);
+            frame->return_offset = oparg;
             DISPATCH_INLINED(gen_frame);
         }
 
@@ -2587,14 +2587,14 @@ dummy_func(
             DEOPT_IF(gen->gi_frame_state >= FRAME_EXECUTING, FOR_ITER);
             STAT_INC(FOR_ITER, hit);
             _PyInterpreterFrame *gen_frame = (_PyInterpreterFrame *)gen->gi_iframe;
-            frame->return_offset = oparg;
             _PyFrame_StackPush(gen_frame, Py_None);
             gen->gi_frame_state = FRAME_EXECUTING;
             gen->gi_exc_state.previous_item = tstate->exc_info;
             tstate->exc_info = &gen->gi_exc_state;
             SKIP_OVER(INLINE_CACHE_ENTRIES_FOR_ITER);
             assert(next_instr[oparg].op.code == END_FOR ||
                    next_instr[oparg].op.code == INSTRUMENTED_END_FOR);
+            frame->return_offset = oparg;
             DISPATCH_INLINED(gen_frame);
         }
 
@@ -2949,32 +2949,72 @@ dummy_func(
             GO_TO_INSTRUCTION(CALL_PY_EXACT_ARGS);
         }
 
-        inst(CALL_PY_EXACT_ARGS, (unused/1, func_version/2, callable, self_or_null, args[oparg] -- unused)) {
-            ASSERT_KWNAMES_IS_NULL();
+        op(_CHECK_PEP_523, (--)) {
             DEOPT_IF(tstate->interp->eval_frame, CALL);
-            int argcount = oparg;
-            if (self_or_null != NULL) {
-                args--;
-                argcount++;
-            }
+        }
+
+        op(_CHECK_FUNCTION_EXACT_ARGS, (func_version/2, callable, self_or_null, unused[oparg] -- callable, self_or_null, unused[oparg])) {
+            ASSERT_KWNAMES_IS_NULL();
             DEOPT_IF(!PyFunction_Check(callable), CALL);
             PyFunctionObject *func = (PyFunctionObject *)callable;
             DEOPT_IF(func->func_version != func_version, CALL);
             PyCodeObject *code = (PyCodeObject *)func->func_code;
-            DEOPT_IF(code->co_argcount != argcount, CALL);
+            DEOPT_IF(code->co_argcount != oparg + (self_or_null != NULL), CALL);
+        }
+
+        op(_CHECK_STACK_SPACE, (callable, unused, unused[oparg] -- callable, unused, unused[oparg])) {
+            PyFunctionObject *func = (PyFunctionObject *)callable;
+            PyCodeObject *code = (PyCodeObject *)func->func_code;
             DEOPT_IF(!_PyThreadState_HasStackSpace(tstate, code->co_framesize), CALL);
+        }
+
+        op(_INIT_CALL_PY_EXACT_ARGS, (callable, self_or_null, args[oparg] -- new_frame: _PyInterpreterFrame*)) {
+            int argcount = oparg;
+            if (self_or_null != NULL) {
+                args--;
+                argcount++;
+            }
             STAT_INC(CALL, hit);
-            _PyInterpreterFrame *new_frame = _PyFrame_PushUnchecked(tstate, func, argcount);
+            PyFunctionObject *func = (PyFunctionObject *)callable;
+            new_frame = _PyFrame_PushUnchecked(tstate, func, argcount);
             for (int i = 0; i < argcount; i++) {
                 new_frame->localsplus[i] = args[i];
             }
-            // Manipulate stack directly since we leave using DISPATCH_INLINED().
-            STACK_SHRINK(oparg + 2);
-            SKIP_OVER(INLINE_CACHE_ENTRIES_CALL);
+        }
+
+        // The 'unused' output effect represents the return value
+        // (which will be pushed when the frame returns).
+        // It is needed so CALL_PY_EXACT_ARGS matches its family.
+        op(_PUSH_FRAME, (new_frame: _PyInterpreterFrame* -- unused)) {
+            // Write it out explicitly because it's subtly different.
+            // Eventually this should be the only occurrence of this code.
             frame->return_offset = 0;
-            DISPATCH_INLINED(new_frame);
+            assert(tstate->interp->eval_frame == NULL);
+            _PyFrame_SetStackPointer(frame, stack_pointer);
+            new_frame->previous = frame;
+            CALL_STAT_INC(inlined_py_calls);
+            #if TIER_ONE
+            frame = cframe.current_frame = new_frame;
+            goto start_frame;
+            #endif
+            #if TIER_TWO
+            frame = tstate->cframe->current_frame = new_frame;
+            ERROR_IF(_Py_EnterRecursivePy(tstate), exit_unwind);
+            stack_pointer = _PyFrame_GetStackPointer(frame);
+            ip_offset = (_Py_CODEUNIT *)_PyFrame_GetCode(frame)->co_code_adaptive;
+            #endif
         }
 
+        macro(CALL_PY_EXACT_ARGS) =
+            unused/1 + // Skip over the counter
+            _CHECK_PEP_523 +
+            _CHECK_FUNCTION_EXACT_ARGS +
+            _CHECK_STACK_SPACE +
+            _INIT_CALL_PY_EXACT_ARGS +
+            SAVE_IP +  // Tier 2 only; special-cased oparg
+            SAVE_CURRENT_IP +  // Sets frame->prev_instr
+            _PUSH_FRAME;
+
         inst(CALL_PY_WITH_DEFAULTS, (unused/1, func_version/2, callable, self_or_null, args[oparg] -- unused)) {
             ASSERT_KWNAMES_IS_NULL();
             DEOPT_IF(tstate->interp->eval_frame, CALL);
@@ -3735,6 +3775,16 @@ dummy_func(
             frame->prev_instr = ip_offset + oparg;
         }
 
+        op(SAVE_CURRENT_IP, (--)) {
+            #if TIER_ONE
+            frame->prev_instr = next_instr - 1;
+            #endif
+            #if TIER_TWO
+            // Relies on a preceding SAVE_IP
+            frame->prev_instr--;
+            #endif
+        }
+
         op(EXIT_TRACE, (--)) {
             frame->prev_instr--;  // Back up to just before destination
             _PyFrame_SetStackPointer(frame, stack_pointer);

Python/abstract_interp_cases.c.h

@@ -602,11 +602,6 @@ int _Py_CheckRecursiveCallPy(
     return 0;
 }
 
-static inline int _Py_EnterRecursivePy(PyThreadState *tstate) {
-    return (tstate->py_recursion_remaining-- <= 0) &&
-        _Py_CheckRecursiveCallPy(tstate);
-}
-
 
 static inline void _Py_LeaveRecursiveCallPy(PyThreadState *tstate)  {
     tstate->py_recursion_remaining++;
@@ -770,6 +765,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
 #endif
         {
 
+#define TIER_ONE 1
 #include "generated_cases.c.h"
 
     /* INSTRUMENTED_LINE has to be here, rather than in bytecodes.c,

Python/bytecodes.c

@@ -364,3 +364,8 @@ static const convertion_func_ptr CONVERSION_FUNCTIONS[4] = {
 #else
 #define _Py_atomic_load_relaxed_int32(ATOMIC_VAL) _Py_atomic_load_relaxed(ATOMIC_VAL)
 #endif
+
+static inline int _Py_EnterRecursivePy(PyThreadState *tstate) {
+    return (tstate->py_recursion_remaining-- <= 0) &&
+        _Py_CheckRecursiveCallPy(tstate);
+}

Python/ceval.c

@@ -81,6 +81,7 @@ _PyUopExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject
         OBJECT_STAT_INC(optimization_uops_executed);
         switch (opcode) {
 
+#define TIER_TWO 2
 #include "executor_cases.c.h"
 
             default:
@@ -106,6 +107,7 @@ _PyUopExecute(_PyExecutorObject *executor, _PyInterpreterFrame *frame, PyObject
 pop_2_error:
     STACK_SHRINK(1);
 pop_1_error:
+pop_1_exit_unwind:
     STACK_SHRINK(1);
 error:
     // On ERROR_IF we return NULL as the frame.