GH-87849: Simplify stack effect of SEND and specialize it for generators and coroutines. (GH-101788)

6 files changed, 139 insertions, 78 deletions

diff --git a/Python/bytecodes.c b/Python/bytecodes.c
index 2b9f12f..429cd7f 100644
--- a/Python/bytecodes.c
+++ b/Python/bytecodes.c
@@ -680,51 +680,66 @@ dummy_func(
             PREDICT(LOAD_CONST);
         }
 
-        inst(SEND, (receiver, v -- receiver if (!jump), retval)) {
+        family(for_iter, INLINE_CACHE_ENTRIES_FOR_ITER) = {
+            SEND,
+            SEND_GEN,
+        };
+
+        inst(SEND, (unused/1, receiver, v -- receiver, retval)) {
+            #if ENABLE_SPECIALIZATION
+            _PySendCache *cache = (_PySendCache *)next_instr;
+            if (ADAPTIVE_COUNTER_IS_ZERO(cache->counter)) {
+                assert(cframe.use_tracing == 0);
+                next_instr--;
+                _Py_Specialize_Send(receiver, next_instr);
+                DISPATCH_SAME_OPARG();
+            }
+            STAT_INC(SEND, deferred);
+            DECREMENT_ADAPTIVE_COUNTER(cache->counter);
+            #endif  /* ENABLE_SPECIALIZATION */
             assert(frame != &entry_frame);
-            bool jump = false;
-            PySendResult gen_status;
-            if (tstate->c_tracefunc == NULL) {
-                gen_status = PyIter_Send(receiver, v, &retval);
-            } else {
-                if (Py_IsNone(v) && PyIter_Check(receiver)) {
-                    retval = Py_TYPE(receiver)->tp_iternext(receiver);
-                }
-                else {
-                    retval = PyObject_CallMethodOneArg(receiver, &_Py_ID(send), v);
-                }
-                if (retval == NULL) {
-                    if (tstate->c_tracefunc != NULL
-                            && _PyErr_ExceptionMatches(tstate, PyExc_StopIteration))
-                        call_exc_trace(tstate->c_tracefunc, tstate->c_traceobj, tstate, frame);
-                    if (_PyGen_FetchStopIterationValue(&retval) == 0) {
-                        gen_status = PYGEN_RETURN;
-                    }
-                    else {
-                        gen_status = PYGEN_ERROR;
-                    }
+            if (Py_IsNone(v) && PyIter_Check(receiver)) {
+                retval = Py_TYPE(receiver)->tp_iternext(receiver);
+            }
+            else {
+                retval = PyObject_CallMethodOneArg(receiver, &_Py_ID(send), v);
+            }
+            if (retval == NULL) {
+                if (tstate->c_tracefunc != NULL
+                        && _PyErr_ExceptionMatches(tstate, PyExc_StopIteration))
+                    call_exc_trace(tstate->c_tracefunc, tstate->c_traceobj, tstate, frame);
+                if (_PyGen_FetchStopIterationValue(&retval) == 0) {
+                    assert(retval != NULL);
+                    JUMPBY(oparg);
                 }
                 else {
-                    gen_status = PYGEN_NEXT;
+                    assert(retval == NULL);
+                    goto error;
                 }
             }
-            if (gen_status == PYGEN_ERROR) {
-                assert(retval == NULL);
-                goto error;
-            }
-            Py_DECREF(v);
-            if (gen_status == PYGEN_RETURN) {
-                assert(retval != NULL);
-                Py_DECREF(receiver);
-                JUMPBY(oparg);
-                jump = true;
-            }
             else {
-                assert(gen_status == PYGEN_NEXT);
                 assert(retval != NULL);
             }
         }
 
+        inst(SEND_GEN, (unused/1, receiver, v -- receiver)) {
+            assert(cframe.use_tracing == 0);
+            PyGenObject *gen = (PyGenObject *)receiver;
+            DEOPT_IF(Py_TYPE(gen) != &PyGen_Type &&
+                     Py_TYPE(gen) != &PyCoro_Type, SEND);
+            DEOPT_IF(gen->gi_frame_state >= FRAME_EXECUTING, SEND);
+            STAT_INC(SEND, hit);
+            _PyInterpreterFrame *gen_frame = (_PyInterpreterFrame *)gen->gi_iframe;
+            frame->yield_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;
+            JUMPBY(INLINE_CACHE_ENTRIES_SEND + oparg);
+            DISPATCH_INLINED(gen_frame);
+        }
+
         inst(YIELD_VALUE, (retval -- unused)) {
             // NOTE: It's important that YIELD_VALUE never raises an exception!
             // The compiler treats any exception raised here as a failed close()
@@ -796,12 +811,13 @@ dummy_func(
             }
         }
 
-        inst(CLEANUP_THROW, (sub_iter, last_sent_val, exc_value -- value)) {
+        inst(CLEANUP_THROW, (sub_iter, last_sent_val, exc_value -- none, value)) {
             assert(throwflag);
             assert(exc_value && PyExceptionInstance_Check(exc_value));
             if (PyErr_GivenExceptionMatches(exc_value, PyExc_StopIteration)) {
                 value = Py_NewRef(((PyStopIterationObject *)exc_value)->value);
                 DECREF_INPUTS();
+                none = Py_NewRef(Py_None);
             }
             else {
                 _PyErr_SetRaisedException(tstate, Py_NewRef(exc_value));
diff --git a/Python/compile.c b/Python/compile.c
index a3c915c..b49eda3 100644
--- a/Python/compile.c
+++ b/Python/compile.c
@@ -1789,6 +1789,8 @@ compiler_add_yield_from(struct compiler *c, location loc, int await)
     ADDOP(c, loc, CLEANUP_THROW);
 
     USE_LABEL(c, exit);
+    ADDOP_I(c, loc, SWAP, 2);
+    ADDOP(c, loc, POP_TOP);
     return SUCCESS;
 }
 
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index a224d4e..093ebff 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -882,57 +882,69 @@
         }
 
         TARGET(SEND) {
+            PREDICTED(SEND);
             PyObject *v = PEEK(1);
             PyObject *receiver = PEEK(2);
             PyObject *retval;
+            #if ENABLE_SPECIALIZATION
+            _PySendCache *cache = (_PySendCache *)next_instr;
+            if (ADAPTIVE_COUNTER_IS_ZERO(cache->counter)) {
+                assert(cframe.use_tracing == 0);
+                next_instr--;
+                _Py_Specialize_Send(receiver, next_instr);
+                DISPATCH_SAME_OPARG();
+            }
+            STAT_INC(SEND, deferred);
+            DECREMENT_ADAPTIVE_COUNTER(cache->counter);
+            #endif  /* ENABLE_SPECIALIZATION */
             assert(frame != &entry_frame);
-            bool jump = false;
-            PySendResult gen_status;
-            if (tstate->c_tracefunc == NULL) {
-                gen_status = PyIter_Send(receiver, v, &retval);
-            } else {
-                if (Py_IsNone(v) && PyIter_Check(receiver)) {
-                    retval = Py_TYPE(receiver)->tp_iternext(receiver);
-                }
-                else {
-                    retval = PyObject_CallMethodOneArg(receiver, &_Py_ID(send), v);
-                }
-                if (retval == NULL) {
-                    if (tstate->c_tracefunc != NULL
-                            && _PyErr_ExceptionMatches(tstate, PyExc_StopIteration))
-                        call_exc_trace(tstate->c_tracefunc, tstate->c_traceobj, tstate, frame);
-                    if (_PyGen_FetchStopIterationValue(&retval) == 0) {
-                        gen_status = PYGEN_RETURN;
-                    }
-                    else {
-                        gen_status = PYGEN_ERROR;
-                    }
+            if (Py_IsNone(v) && PyIter_Check(receiver)) {
+                retval = Py_TYPE(receiver)->tp_iternext(receiver);
+            }
+            else {
+                retval = PyObject_CallMethodOneArg(receiver, &_Py_ID(send), v);
+            }
+            if (retval == NULL) {
+                if (tstate->c_tracefunc != NULL
+                        && _PyErr_ExceptionMatches(tstate, PyExc_StopIteration))
+                    call_exc_trace(tstate->c_tracefunc, tstate->c_traceobj, tstate, frame);
+                if (_PyGen_FetchStopIterationValue(&retval) == 0) {
+                    assert(retval != NULL);
+                    JUMPBY(oparg);
                 }
                 else {
-                    gen_status = PYGEN_NEXT;
+                    assert(retval == NULL);
+                    goto error;
                 }
             }
-            if (gen_status == PYGEN_ERROR) {
-                assert(retval == NULL);
-                goto error;
-            }
-            Py_DECREF(v);
-            if (gen_status == PYGEN_RETURN) {
-                assert(retval != NULL);
-                Py_DECREF(receiver);
-                JUMPBY(oparg);
-                jump = true;
-            }
             else {
-                assert(gen_status == PYGEN_NEXT);
                 assert(retval != NULL);
             }
-            STACK_SHRINK(1);
-            STACK_GROW(((!jump) ? 1 : 0));
             POKE(1, retval);
+            JUMPBY(1);
             DISPATCH();
         }
 
+        TARGET(SEND_GEN) {
+            PyObject *v = PEEK(1);
+            PyObject *receiver = PEEK(2);
+            assert(cframe.use_tracing == 0);
+            PyGenObject *gen = (PyGenObject *)receiver;
+            DEOPT_IF(Py_TYPE(gen) != &PyGen_Type &&
+                     Py_TYPE(gen) != &PyCoro_Type, SEND);
+            DEOPT_IF(gen->gi_frame_state >= FRAME_EXECUTING, SEND);
+            STAT_INC(SEND, hit);
+            _PyInterpreterFrame *gen_frame = (_PyInterpreterFrame *)gen->gi_iframe;
+            frame->yield_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;
+            JUMPBY(INLINE_CACHE_ENTRIES_SEND + oparg);
+            DISPATCH_INLINED(gen_frame);
+        }
+
         TARGET(YIELD_VALUE) {
             PyObject *retval = PEEK(1);
             // NOTE: It's important that YIELD_VALUE never raises an exception!
@@ -1026,6 +1038,7 @@
             PyObject *exc_value = PEEK(1);
             PyObject *last_sent_val = PEEK(2);
             PyObject *sub_iter = PEEK(3);
+            PyObject *none;
             PyObject *value;
             assert(throwflag);
             assert(exc_value && PyExceptionInstance_Check(exc_value));
@@ -1034,13 +1047,15 @@
                 Py_DECREF(sub_iter);
                 Py_DECREF(last_sent_val);
                 Py_DECREF(exc_value);
+                none = Py_NewRef(Py_None);
             }
             else {
                 _PyErr_SetRaisedException(tstate, Py_NewRef(exc_value));
                 goto exception_unwind;
             }
-            STACK_SHRINK(2);
+            STACK_SHRINK(1);
             POKE(1, value);
+            POKE(2, none);
             DISPATCH();
         }
 
diff --git a/Python/opcode_metadata.h b/Python/opcode_metadata.h
index db1dfd3..d622eb1 100644
--- a/Python/opcode_metadata.h
+++ b/Python/opcode_metadata.h
@@ -104,6 +104,8 @@ _PyOpcode_num_popped(int opcode, int oparg, bool jump) {
             return 1;
         case SEND:
             return 2;
+        case SEND_GEN:
+            return 2;
         case YIELD_VALUE:
             return 1;
         case POP_EXCEPT:
@@ -453,7 +455,9 @@ _PyOpcode_num_pushed(int opcode, int oparg, bool jump) {
         case GET_AWAITABLE:
             return 1;
         case SEND:
-            return ((!jump) ? 1 : 0) + 1;
+            return 2;
+        case SEND_GEN:
+            return 1;
         case YIELD_VALUE:
             return 1;
         case POP_EXCEPT:
@@ -465,7 +469,7 @@ _PyOpcode_num_pushed(int opcode, int oparg, bool jump) {
         case END_ASYNC_FOR:
             return 0;
         case CLEANUP_THROW:
-            return 1;
+            return 2;
         case LOAD_ASSERTION_ERROR:
             return 1;
         case LOAD_BUILD_CLASS:
@@ -763,7 +767,8 @@ const struct opcode_metadata _PyOpcode_opcode_metadata[256] = {
     [GET_AITER] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IX },
     [GET_ANEXT] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IX },
     [GET_AWAITABLE] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IB },
-    [SEND] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IB },
+    [SEND] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IBC },
+    [SEND_GEN] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IBC },
     [YIELD_VALUE] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IX },
     [POP_EXCEPT] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IX },
     [RERAISE] = { DIR_NONE, DIR_NONE, DIR_NONE, true, INSTR_FMT_IB },
diff --git a/Python/opcode_targets.h b/Python/opcode_targets.h
index eceb246..301ec6e 100644
--- a/Python/opcode_targets.h
+++ b/Python/opcode_targets.h
@@ -165,7 +165,7 @@ static void *opcode_targets[256] = {
     &&TARGET_SET_UPDATE,
     &&TARGET_DICT_MERGE,
     &&TARGET_DICT_UPDATE,
-    &&_unknown_opcode,
+    &&TARGET_SEND_GEN,
     &&_unknown_opcode,
     &&_unknown_opcode,
     &&_unknown_opcode,
diff --git a/Python/specialize.c b/Python/specialize.c
index 908ad6d..4ede312 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -128,6 +128,7 @@ print_spec_stats(FILE *out, OpcodeStats *stats)
     fprintf(out, "opcode[%d].specializable : 1\n", BINARY_SLICE);
     fprintf(out, "opcode[%d].specializable : 1\n", COMPARE_OP);
     fprintf(out, "opcode[%d].specializable : 1\n", STORE_SLICE);
+    fprintf(out, "opcode[%d].specializable : 1\n", SEND);
     for (int i = 0; i < 256; i++) {
         if (_PyOpcode_Caches[i]) {
             fprintf(out, "opcode[%d].specializable : 1\n", i);
@@ -1084,7 +1085,7 @@ PyObject *descr, DescriptorClassification kind)
             if (dict) {
                 SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_ATTR_NOT_MANAGED_DICT);
                 return 0;
-            }  
+            }
             assert(owner_cls->tp_dictoffset > 0);
             assert(owner_cls->tp_dictoffset <= INT16_MAX);
             _py_set_opcode(instr, LOAD_ATTR_METHOD_LAZY_DICT);
@@ -2183,3 +2184,25 @@ success:
     STAT_INC(FOR_ITER, success);
     cache->counter = adaptive_counter_cooldown();
 }
+
+void
+_Py_Specialize_Send(PyObject *receiver, _Py_CODEUNIT *instr)
+{
+    assert(ENABLE_SPECIALIZATION);
+    assert(_PyOpcode_Caches[SEND] == INLINE_CACHE_ENTRIES_SEND);
+    _PySendCache *cache = (_PySendCache *)(instr + 1);
+    PyTypeObject *tp = Py_TYPE(receiver);
+    if (tp == &PyGen_Type || tp == &PyCoro_Type) {
+        _py_set_opcode(instr, SEND_GEN);
+        goto success;
+    }
+    SPECIALIZATION_FAIL(SEND,
+                        _PySpecialization_ClassifyIterator(receiver));
+    STAT_INC(SEND, failure);
+    _py_set_opcode(instr, SEND);
+    cache->counter = adaptive_counter_backoff(cache->counter);
+    return;
+success:
+    STAT_INC(SEND, success);
+    cache->counter = adaptive_counter_cooldown();
+}