bpo-41756: Introduce PyGen_Send C API (GH-22196)

The new API allows to efficiently send values into native generators
and coroutines avoiding use of StopIteration exceptions to signal 
returns.

ceval loop now uses this method instead of the old "private"
_PyGen_Send C API. This translates to 1.6x increased performance
of 'await' calls in micro-benchmarks.

Aside from CPython core improvements, this new API will also allow 
Cython to generate more efficient code, benefiting high-performance
IO libraries like uvloop.

7 files changed, 148 insertions, 39 deletions

diff --git a/Doc/c-api/gen.rst b/Doc/c-api/gen.rst
index 7441092..e098425 100644
--- a/Doc/c-api/gen.rst
+++ b/Doc/c-api/gen.rst
@@ -15,6 +15,11 @@ than explicitly calling :c:func:`PyGen_New` or :c:func:`PyGen_NewWithQualName`.
    The C structure used for generator objects.
 
 
+.. c:type:: PySendResult
+
+   The enum value used to represent different results of :c:func:`PyGen_Send`.
+
+
 .. c:var:: PyTypeObject PyGen_Type
 
    The type object corresponding to generator objects.
@@ -42,3 +47,13 @@ than explicitly calling :c:func:`PyGen_New` or :c:func:`PyGen_NewWithQualName`.
    with ``__name__`` and ``__qualname__`` set to *name* and *qualname*.
    A reference to *frame* is stolen by this function.  The *frame* argument
    must not be ``NULL``.
+
+.. c:function:: PySendResult PyGen_Send(PyGenObject *gen, PyObject *arg, PyObject **presult)
+
+   Sends the *arg* value into the generator *gen*. Coroutine objects
+   are also allowed to be as the *gen* argument but they need to be
+   explicitly casted to PyGenObject*. Returns:
+
+   - ``PYGEN_RETURN`` if generator returns. Return value is returned via *presult*.
+   - ``PYGEN_NEXT`` if generator yields. Yielded value is returned via *presult*.
+   - ``PYGEN_ERROR`` if generator has raised and exception. *presult* is set to ``NULL``.
diff --git a/Doc/data/refcounts.dat b/Doc/data/refcounts.dat
index 355a4d6..6b1bde3 100644
--- a/Doc/data/refcounts.dat
+++ b/Doc/data/refcounts.dat
@@ -959,6 +959,11 @@ PyGen_NewWithQualName:PyFrameObject*:frame:0:
 PyGen_NewWithQualName:PyObject*:name:0:
 PyGen_NewWithQualName:PyObject*:qualname:0:
 
+PyGen_Send:int:::
+PyGen_Send:PyGenObject*:gen:0:
+PyGen_Send:PyObject*:arg:0:
+PyGen_Send:PyObject**:presult:+1:
+
 PyCoro_CheckExact:int:::
 PyCoro_CheckExact:PyObject*:ob:0:
 
diff --git a/Include/genobject.h b/Include/genobject.h
index a76dc92..7488054 100644
--- a/Include/genobject.h
+++ b/Include/genobject.h
@@ -45,6 +45,21 @@ PyAPI_FUNC(PyObject *) _PyGen_Send(PyGenObject *, PyObject *);
 PyObject *_PyGen_yf(PyGenObject *);
 PyAPI_FUNC(void) _PyGen_Finalize(PyObject *self);
 
+typedef enum {
+    PYGEN_RETURN = 0,
+    PYGEN_ERROR = -1,
+    PYGEN_NEXT = 1,
+} PySendResult;
+
+/* Sends the value into the generator or the coroutine. Returns:
+   - PYGEN_RETURN (0) if generator has returned.
+     'result' parameter is filled with return value
+   - PYGEN_ERROR (-1) if exception was raised.
+     'result' parameter is NULL
+   - PYGEN_NEXT (1) if generator has yielded.
+     'result' parameter is filled with yielded value. */
+PyAPI_FUNC(PySendResult) PyGen_Send(PyGenObject *, PyObject *, PyObject **);
+
 #ifndef Py_LIMITED_API
 typedef struct {
     _PyGenObject_HEAD(cr)
diff --git a/Misc/NEWS.d/next/Core and Builtins/2020-09-12-12-55-45.bpo-41756.1h0tbV.rst b/Misc/NEWS.d/next/Core and Builtins/2020-09-12-12-55-45.bpo-41756.1h0tbV.rst
new file mode 100644
index 0000000..b387cfd
--- /dev/null
+++ b/Misc/NEWS.d/next/Core and Builtins/2020-09-12-12-55-45.bpo-41756.1h0tbV.rst
@@ -0,0 +1,2 @@
+Add PyGen_Send function to allow sending value into generator/coroutine
+without raising StopIteration exception to signal return
diff --git a/Modules/_asynciomodule.c b/Modules/_asynciomodule.c
index 4a1c91e..2151f20 100644
--- a/Modules/_asynciomodule.c
+++ b/Modules/_asynciomodule.c
@@ -2621,6 +2621,20 @@ task_set_error_soon(TaskObj *task, PyObject *et, const char *format, ...)
     Py_RETURN_NONE;
 }
 
+static inline int
+gen_status_from_result(PyObject **result)
+{
+    if (*result != NULL) {
+        return PYGEN_NEXT;
+    }
+    if (_PyGen_FetchStopIterationValue(result) == 0) {
+        return PYGEN_RETURN;
+    }
+
+    assert(PyErr_Occurred());
+    return PYGEN_ERROR;
+}
+
 static PyObject *
 task_step_impl(TaskObj *task, PyObject *exc)
 {
@@ -2679,26 +2693,29 @@ task_step_impl(TaskObj *task, PyObject *exc)
         return NULL;
     }
 
+    int gen_status = PYGEN_ERROR;
     if (exc == NULL) {
         if (PyGen_CheckExact(coro) || PyCoro_CheckExact(coro)) {
-            result = _PyGen_Send((PyGenObject*)coro, Py_None);
+            gen_status = PyGen_Send((PyGenObject*)coro, Py_None, &result);
         }
         else {
             result = _PyObject_CallMethodIdOneArg(coro, &PyId_send, Py_None);
+            gen_status = gen_status_from_result(&result);
         }
     }
     else {
         result = _PyObject_CallMethodIdOneArg(coro, &PyId_throw, exc);
+        gen_status = gen_status_from_result(&result);
         if (clear_exc) {
             /* We created 'exc' during this call */
             Py_DECREF(exc);
         }
     }
 
-    if (result == NULL) {
+    if (gen_status == PYGEN_RETURN || gen_status == PYGEN_ERROR) {
         PyObject *et, *ev, *tb;
 
-        if (_PyGen_FetchStopIterationValue(&o) == 0) {
+        if (result != NULL) {
             /* The error is StopIteration and that means that
                the underlying coroutine has resolved */
 
@@ -2709,10 +2726,10 @@ task_step_impl(TaskObj *task, PyObject *exc)
                 res = future_cancel((FutureObj*)task, task->task_cancel_msg);
             }
             else {
-                res = future_set_result((FutureObj*)task, o);
+                res = future_set_result((FutureObj*)task, result);
             }
 
-            Py_DECREF(o);
+            Py_DECREF(result);
 
             if (res == NULL) {
                 return NULL;
diff --git a/Objects/genobject.c b/Objects/genobject.c
index 809838a..24aca98 100644
--- a/Objects/genobject.c
+++ b/Objects/genobject.c
@@ -137,7 +137,7 @@ gen_dealloc(PyGenObject *gen)
 }
 
 static PyObject *
-gen_send_ex(PyGenObject *gen, PyObject *arg, int exc, int closing)
+gen_send_ex(PyGenObject *gen, PyObject *arg, int exc, int closing, int *is_return_value)
 {
     PyThreadState *tstate = _PyThreadState_GET();
     PyFrameObject *f = gen->gi_frame;
@@ -170,6 +170,10 @@ gen_send_ex(PyGenObject *gen, PyObject *arg, int exc, int closing)
                 PyErr_SetNone(PyExc_StopAsyncIteration);
             }
             else {
+                if (is_return_value != NULL) {
+                    *is_return_value = 1;
+                    Py_RETURN_NONE;
+                }
                 PyErr_SetNone(PyExc_StopIteration);
             }
         }
@@ -230,18 +234,33 @@ gen_send_ex(PyGenObject *gen, PyObject *arg, int exc, int closing)
             /* Delay exception instantiation if we can */
             if (PyAsyncGen_CheckExact(gen)) {
                 PyErr_SetNone(PyExc_StopAsyncIteration);
+                Py_CLEAR(result);
             }
             else if (arg) {
-                /* Set exception if not called by gen_iternext() */
-                PyErr_SetNone(PyExc_StopIteration);
+                if (is_return_value != NULL) {
+                    *is_return_value = 1;
+                }
+                else {
+                    /* Set exception if not called by gen_iternext() */
+                    PyErr_SetNone(PyExc_StopIteration);
+                    Py_CLEAR(result);
+                }
+            }
+            else {
+                Py_CLEAR(result);
             }
         }
         else {
             /* Async generators cannot return anything but None */
             assert(!PyAsyncGen_CheckExact(gen));
-            _PyGen_SetStopIterationValue(result);
+            if (is_return_value != NULL) {
+                *is_return_value = 1;
+            }
+            else {
+                _PyGen_SetStopIterationValue(result);
+                Py_CLEAR(result);
+            }
         }
-        Py_CLEAR(result);
     }
     else if (!result && PyErr_ExceptionMatches(PyExc_StopIteration)) {
         const char *msg = "generator raised StopIteration";
@@ -264,7 +283,7 @@ gen_send_ex(PyGenObject *gen, PyObject *arg, int exc, int closing)
         _PyErr_FormatFromCause(PyExc_RuntimeError, "%s", msg);
     }
 
-    if (!result ||  _PyFrameHasCompleted(f)) {
+    if ((is_return_value && *is_return_value) || !result ||  _PyFrameHasCompleted(f)) {
         /* generator can't be rerun, so release the frame */
         /* first clean reference cycle through stored exception traceback */
         _PyErr_ClearExcState(&gen->gi_exc_state);
@@ -283,7 +302,19 @@ return next yielded value or raise StopIteration.");
 PyObject *
 _PyGen_Send(PyGenObject *gen, PyObject *arg)
 {
-    return gen_send_ex(gen, arg, 0, 0);
+    return gen_send_ex(gen, arg, 0, 0, NULL);
+}
+
+PySendResult
+PyGen_Send(PyGenObject *gen, PyObject *arg, PyObject **result)
+{
+    assert(result != NULL);
+
+    int is_return_value = 0;
+    if ((*result = gen_send_ex(gen, arg, 0, 0, &is_return_value)) == NULL) {
+        return PYGEN_ERROR;
+    }
+    return is_return_value ? PYGEN_RETURN : PYGEN_NEXT;
 }
 
 PyDoc_STRVAR(close_doc,
@@ -365,7 +396,7 @@ gen_close(PyGenObject *gen, PyObject *args)
     }
     if (err == 0)
         PyErr_SetNone(PyExc_GeneratorExit);
-    retval = gen_send_ex(gen, Py_None, 1, 1);
+    retval = gen_send_ex(gen, Py_None, 1, 1, NULL);
     if (retval) {
         const char *msg = "generator ignored GeneratorExit";
         if (PyCoro_CheckExact(gen)) {
@@ -413,7 +444,7 @@ _gen_throw(PyGenObject *gen, int close_on_genexit,
             gen->gi_frame->f_state = state;
             Py_DECREF(yf);
             if (err < 0)
-                return gen_send_ex(gen, Py_None, 1, 0);
+                return gen_send_ex(gen, Py_None, 1, 0, NULL);
             goto throw_here;
         }
         if (PyGen_CheckExact(yf) || PyCoro_CheckExact(yf)) {
@@ -465,10 +496,10 @@ _gen_throw(PyGenObject *gen, int close_on_genexit,
             assert(gen->gi_frame->f_lasti >= 0);
             gen->gi_frame->f_lasti += sizeof(_Py_CODEUNIT);
             if (_PyGen_FetchStopIterationValue(&val) == 0) {
-                ret = gen_send_ex(gen, val, 0, 0);
+                ret = gen_send_ex(gen, val, 0, 0, NULL);
                 Py_DECREF(val);
             } else {
-                ret = gen_send_ex(gen, Py_None, 1, 0);
+                ret = gen_send_ex(gen, Py_None, 1, 0, NULL);
             }
         }
         return ret;
@@ -522,7 +553,7 @@ throw_here:
     }
 
     PyErr_Restore(typ, val, tb);
-    return gen_send_ex(gen, Py_None, 1, 0);
+    return gen_send_ex(gen, Py_None, 1, 0, NULL);
 
 failed_throw:
     /* Didn't use our arguments, so restore their original refcounts */
@@ -551,7 +582,7 @@ gen_throw(PyGenObject *gen, PyObject *args)
 static PyObject *
 gen_iternext(PyGenObject *gen)
 {
-    return gen_send_ex(gen, NULL, 0, 0);
+    return gen_send_ex(gen, NULL, 0, 0, NULL);
 }
 
 /*
@@ -1051,13 +1082,13 @@ coro_wrapper_dealloc(PyCoroWrapper *cw)
 static PyObject *
 coro_wrapper_iternext(PyCoroWrapper *cw)
 {
-    return gen_send_ex((PyGenObject *)cw->cw_coroutine, NULL, 0, 0);
+    return gen_send_ex((PyGenObject *)cw->cw_coroutine, NULL, 0, 0, NULL);
 }
 
 static PyObject *
 coro_wrapper_send(PyCoroWrapper *cw, PyObject *arg)
 {
-    return gen_send_ex((PyGenObject *)cw->cw_coroutine, arg, 0, 0);
+    return gen_send_ex((PyGenObject *)cw->cw_coroutine, arg, 0, 0, NULL);
 }
 
 static PyObject *
@@ -1570,7 +1601,7 @@ async_gen_asend_send(PyAsyncGenASend *o, PyObject *arg)
     }
 
     o->ags_gen->ag_running_async = 1;
-    result = gen_send_ex((PyGenObject*)o->ags_gen, arg, 0, 0);
+    result = gen_send_ex((PyGenObject*)o->ags_gen, arg, 0, 0, NULL);
     result = async_gen_unwrap_value(o->ags_gen, result);
 
     if (result == NULL) {
@@ -1926,7 +1957,7 @@ async_gen_athrow_send(PyAsyncGenAThrow *o, PyObject *arg)
 
     assert(o->agt_state == AWAITABLE_STATE_ITER);
 
-    retval = gen_send_ex((PyGenObject *)gen, arg, 0, 0);
+    retval = gen_send_ex((PyGenObject *)gen, arg, 0, 0, NULL);
     if (o->agt_args) {
         return async_gen_unwrap_value(o->agt_gen, retval);
     } else {
diff --git a/Python/ceval.c b/Python/ceval.c
index f747faa..3de372f 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -2223,29 +2223,53 @@ main_loop:
         case TARGET(YIELD_FROM): {
             PyObject *v = POP();
             PyObject *receiver = TOP();
-            int err;
-            if (PyGen_CheckExact(receiver) || PyCoro_CheckExact(receiver)) {
-                retval = _PyGen_Send((PyGenObject *)receiver, v);
+            int is_gen_or_coro = PyGen_CheckExact(receiver) || PyCoro_CheckExact(receiver);
+            int gen_status;
+            if (tstate->c_tracefunc == NULL && is_gen_or_coro) {
+                gen_status = PyGen_Send((PyGenObject *)receiver, v, &retval);
             } else {
-                _Py_IDENTIFIER(send);
-                if (v == Py_None)
-                    retval = Py_TYPE(receiver)->tp_iternext(receiver);
-                else
-                    retval = _PyObject_CallMethodIdOneArg(receiver, &PyId_send, v);
+                if (is_gen_or_coro) {
+                    retval = _PyGen_Send((PyGenObject *)receiver, v);
+                }
+                else {
+                    _Py_IDENTIFIER(send);
+                    if (v == Py_None) {
+                        retval = Py_TYPE(receiver)->tp_iternext(receiver);
+                    }
+                    else {
+                        retval = _PyObject_CallMethodIdOneArg(receiver, &PyId_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, f);
+                    if (_PyGen_FetchStopIterationValue(&retval) == 0) {
+                        gen_status = PYGEN_RETURN;
+                    }
+                    else {
+                        gen_status = PYGEN_ERROR;
+                    }
+                }
+                else {
+                    gen_status = PYGEN_NEXT;
+                }
             }
             Py_DECREF(v);
-            if (retval == NULL) {
-                PyObject *val;
-                if (tstate->c_tracefunc != NULL
-                        && _PyErr_ExceptionMatches(tstate, PyExc_StopIteration))
-                    call_exc_trace(tstate->c_tracefunc, tstate->c_traceobj, tstate, f);
-                err = _PyGen_FetchStopIterationValue(&val);
-                if (err < 0)
-                    goto error;
+            if (gen_status == PYGEN_ERROR) {
+                assert (retval == NULL);
+                goto error;
+            }
+            if (gen_status == PYGEN_RETURN) {
+                assert (retval != NULL);
+
                 Py_DECREF(receiver);
-                SET_TOP(val);
+                SET_TOP(retval);
+                retval = NULL;
                 DISPATCH();
             }
+            assert (gen_status == PYGEN_NEXT);
             /* receiver remains on stack, retval is value to be yielded */
             /* and repeat... */
             assert(f->f_lasti >= (int)sizeof(_Py_CODEUNIT));