gh-119109: improve functools.partial vectorcall with keywords

Misc/NEWS.d/next/Library/2024-10-17-01-12-22.gh-issue-119109.u4hcvb.rst

@@ -0,0 +1 @@
+:func:`functools.partial` calls are now faster when keyword arguments are used.

Modules/_functoolsmodule.c

@@ -367,19 +367,6 @@ partial_descr_get(PyObject *self, PyObject *obj, PyObject *type)
  return PyMethod_New(self, obj);
 }
 
-/* Merging keyword arguments using the vectorcall convention is messy, so
- * if we would need to do that, we stop using vectorcall and fall back
- * to using partial_call() instead. */
-Py_NO_INLINE static PyObject *
-partial_vectorcall_fallback(PyThreadState *tstate, partialobject *pto,
- PyObject *const *args, size_t nargsf,
- PyObject *kwnames)
-{
- pto->vectorcall = NULL;
- Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
- return _PyObject_MakeTpCall(tstate, (PyObject *)pto, args, nargs, kwnames);
-}
-
 static PyObject *
 partial_vectorcall(PyObject *self, PyObject *const *args,
  size_t nargsf, PyObject *kwnames)
@@ -388,10 +375,7 @@ partial_vectorcall(PyObject *self, PyObject *const *args,
  PyThreadState *tstate = _PyThreadState_GET();
  Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
 
- /* pto->kw is mutable, so need to check every time */
- if (PyDict_GET_SIZE(pto->kw)){
- return partial_vectorcall_fallback(tstate, pto, args, nargsf, kwnames);
- }
+ /* Placeholder check */
  Py_ssize_t pto_phcount = pto->phcount;
  if (nargs < pto_phcount){
  PyErr_Format(PyExc_TypeError,
@@ -400,50 +384,143 @@ partial_vectorcall(PyObject *self, PyObject *const *args,
  return NULL;
  }
 
- Py_ssize_t nargskw = nargs;
- if (kwnames != NULL){
- nargskw += PyTuple_GET_SIZE(kwnames);
- }
-
  PyObject **pto_args = _PyTuple_ITEMS(pto->args);
  Py_ssize_t pto_nargs = PyTuple_GET_SIZE(pto->args);
+ Py_ssize_t pto_nkwds = PyDict_GET_SIZE(pto->kw);
+ Py_ssize_t nkwds = kwnames == NULL ? 0 : PyTuple_GET_SIZE(kwnames);
+ Py_ssize_t nargskw = nargs + nkwds;
+
+ /* Special cases */
+ if (!pto_nkwds){
+ /* Fast path if we're called without arguments */
+ if (nargskw == 0){
+ return _PyObject_VectorcallTstate(tstate, pto->fn, pto_args,
+ pto_nargs, NULL);
+ }
 
- /* Fast path if we're called without arguments */
- if (nargskw == 0){
- return _PyObject_VectorcallTstate(tstate, pto->fn,
- pto_args, pto_nargs, NULL);
+ /* Use PY_VECTORCALL_ARGUMENTS_OFFSET to prepend a single
+ * positional argument. */
+ if (pto_nargs == 1 && (nargsf & PY_VECTORCALL_ARGUMENTS_OFFSET)){
+ PyObject **newargs = (PyObject **)args - 1;
+ PyObject *tmp = newargs[0];
+ newargs[0] = pto_args[0];
+ PyObject *ret = _PyObject_VectorcallTstate(tstate, pto->fn, newargs,
+ nargs + 1, kwnames);
+ newargs[0] = tmp;
+ return ret;
+ }
  }
 
- /* Fast path using PY_VECTORCALL_ARGUMENTS_OFFSET to prepend a single
- * positional argument */
- if (pto_nargs == 1 && (nargsf & PY_VECTORCALL_ARGUMENTS_OFFSET)){
- PyObject **newargs = (PyObject **)args - 1;
- PyObject *tmp = newargs[0];
- newargs[0] = pto_args[0];
- PyObject *ret = _PyObject_VectorcallTstate(tstate, pto->fn,
- newargs, nargs + 1, kwnames);
- newargs[0] = tmp;
- return ret;
- }
+ /* Total sizes */
+ Py_ssize_t tot_nargs = pto_nargs + nargs - pto_phcount;
+ Py_ssize_t tot_nkwds = pto_nkwds + nkwds;
+ Py_ssize_t tot_nargskw = tot_nargs + tot_nkwds;
 
- PyObject *small_stack[_PY_FASTCALL_SMALL_STACK];
- PyObject **stack;
+ PyObject *pto_kw_merged = NULL; // pto_kw with duplicates merged (if any)
+ PyObject *tot_kwnames;
 
- Py_ssize_t tot_nargskw = pto_nargs + nargskw - pto_phcount;
- if (tot_nargskw <= (Py_ssize_t)Py_ARRAY_LENGTH(small_stack)){
+ /* Allocate Stack
+ * Note, _PY_FASTCALL_SMALL_STACK is optimal for positional only
+ * This case might have keyword arguments
+ * furthermore, it might use extra stack space for temporary key storage
+ * thus, double small_stack size is used, which is 10 * 8 = 80 bytes */
+ PyObject *small_stack[_PY_FASTCALL_SMALL_STACK * 2];
+ PyObject **tmp_stack, **stack;
+ Py_ssize_t init_stack_size = tot_nargskw;
+ if (pto_nkwds){
+ // If pto_nkwds, allocate additional space for temporary new keys
+ init_stack_size += nkwds;
+ }
+ if (init_stack_size <= (Py_ssize_t)Py_ARRAY_LENGTH(small_stack)){
  stack = small_stack;
  }
  else{
- stack = PyMem_Malloc(tot_nargskw * sizeof(PyObject *));
+ stack = PyMem_Malloc(init_stack_size * sizeof(PyObject *));
  if (stack == NULL){
- PyErr_NoMemory();
- return NULL;
+ return PyErr_NoMemory();
  }
  }
 
- Py_ssize_t tot_nargs;
+ /* Copy keywords to stack */
+ if (!pto_nkwds){
+ tot_kwnames = kwnames;
+ if (nkwds){
+ /* if !pto_nkwds & nkwds, then simply append kw */
+ memcpy(stack + tot_nargs, args + nargs, nkwds * sizeof(PyObject*));
+ }
+ }
+ else{
+ /* stack is now [<positionals>, <pto_kwds>, <kwds>, <kwds_keys>]
+ * Will resize later to [<positionals>, <merged_kwds>] */
+ PyObject *key, *val;
+
+ /* Merge kw to pto_kw or add to tail (if not duplicate) */
+ Py_ssize_t n_tail = 0;
+ for (Py_ssize_t i = 0; i < nkwds; ++i){
+ key = PyTuple_GET_ITEM(kwnames, i);
+ val = args[nargs + i];
+ if (PyDict_Contains(pto->kw, key)){
+ if (pto_kw_merged == NULL){
+ pto_kw_merged = PyDict_Copy(pto->kw);
+ if (pto_kw_merged == NULL){
+ goto error;
+ }
+ }
+ if (PyDict_SetItem(pto_kw_merged, key, val) < 0){
+ Py_DECREF(pto_kw_merged);
+ goto error;
+ }
+ }
+ else{
+ /* Copy keyword tail to stack */
+ stack[tot_nargs + pto_nkwds + n_tail] = val;
+ stack[tot_nargskw + n_tail] = key;
+ n_tail++;
+ }
+ }
+ Py_ssize_t n_merges = nkwds - n_tail;
+
+ /* Create total kwnames */
+ tot_kwnames = PyTuple_New(tot_nkwds - n_merges);
+ if (tot_kwnames == NULL){
+ Py_XDECREF(pto_kw_merged);
+ goto error;
+ }
+ for (Py_ssize_t i = 0; i < n_tail; ++i){
+ key = Py_NewRef(stack[tot_nargskw + i]);
+ PyTuple_SET_ITEM(tot_kwnames, pto_nkwds + i, key);
+ }
+
+ /* Copy pto_keywords with overlapping call keywords merged
+ * Note, tail is already coppied. */
+ Py_ssize_t pos = 0, i = 0;
+ while (PyDict_Next(n_merges ? pto_kw_merged : pto->kw, &pos, &key, &val)){
+ assert(i < pto_nkwds);
+ PyTuple_SET_ITEM(tot_kwnames, i, Py_NewRef(key));
+ stack[tot_nargs + i] = val;
+ i++;
+ }
+ assert(i == pto_nkwds);
+ Py_XDECREF(pto_kw_merged);
+
+ /* Resize Stack if the removing overallocation saves some noticable memory
+ * NOTE: This whole block can be removed without breaking anything */
+ Py_ssize_t noveralloc = n_merges + nkwds;
+ if (stack != small_stack && noveralloc > 6 && noveralloc > init_stack_size / 10){
+ tmp_stack = PyMem_Realloc(stack, (tot_nargskw - n_merges) * sizeof(PyObject *));
+ if (tmp_stack == NULL){
+ Py_DECREF(tot_kwnames);
+ if (stack != small_stack){
+ PyMem_Free(stack);
+ }
+ return PyErr_NoMemory();
+ }
+ stack = tmp_stack;
+ }
+ }
+
+ /* Copy Positionals to stack */
  if (pto_phcount){
- tot_nargs = pto_nargs + nargs - pto_phcount;
  Py_ssize_t j = 0; // New args index
  for (Py_ssize_t i = 0; i < pto_nargs; i++){
  if (pto_args[i] == pto->placeholder){
@@ -455,22 +532,31 @@ partial_vectorcall(PyObject *self, PyObject *const *args,
  }
  }
  assert(j == pto_phcount);
- if (nargskw > pto_phcount){
- memcpy(stack + pto_nargs, args + j, (nargskw - j) * sizeof(PyObject*));
+ /* Add remaining args from new_args */
+ if (nargs > pto_phcount){
+ memcpy(stack + pto_nargs, args + j, (nargs - j) * sizeof(PyObject*));
  }
  }
  else{
- tot_nargs = pto_nargs + nargs;
- /* Copy to new stack, using borrowed references */
  memcpy(stack, pto_args, pto_nargs * sizeof(PyObject*));
- memcpy(stack + pto_nargs, args, nargskw * sizeof(PyObject*));
+ memcpy(stack + pto_nargs, args, nargs * sizeof(PyObject*));
  }
- PyObject *ret = _PyObject_VectorcallTstate(tstate, pto->fn,
- stack, tot_nargs, kwnames);
+
+ PyObject *ret = _PyObject_VectorcallTstate(tstate, pto->fn, stack,
+ tot_nargs, tot_kwnames);
  if (stack != small_stack){
  PyMem_Free(stack);
  }
+ if (pto_nkwds){
+ Py_DECREF(tot_kwnames);
+ }
  return ret;
+
+ error:
+ if (stack != small_stack){
+ PyMem_Free(stack);
+ }
+ return NULL;
 }
 
 /* Set pto->vectorcall depending on the parameters of the partial object */