gh-107137: Add _PyTupleBuilder API to the internal C API

Add _PyTupleBuilder structure and functions:

* _PyTupleBuilder_Init()
* _PyTupleBuilder_Alloc()
* _PyTupleBuilder_Append()
* _PyTupleBuilder_AppendUnsafe()
* _PyTupleBuilder_Finish()
* _PyTupleBuilder_Dealloc()

The builder tracks the size of the tuple and resize it in
_PyTupleBuilder_Finish() if needed. Don't allocate empty tuple.
Allocate an array of 16 objects on the stack to avoid allocating
small tuple. _PyTupleBuilder_Append() overallocates the tuple by 25%
to reduce the number of _PyTuple_Resize() calls.

Do no track the temporary internal tuple by the GC before
_PyTupleBuilder_Finish() creates the final complete and consistent
tuple object.

Use _PyTupleBuilder API in itertools batched_traverse(),
PySequence_Tuple() and initialize_structseq_dict().

Add also helper functions:

* _PyTuple_ResizeNoTrack()
* _PyTuple_NewNoTrack()

Author: vstinner

Include/internal/pycore_tuple.h

@@ -8,8 +8,8 @@ extern "C" {
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
-#include "tupleobject.h"   /* _PyTuple_CAST() */
-
+extern PyObject* _PyTuple_NewNoTrack(Py_ssize_t size);
+extern int _PyTuple_ResizeNoTrack(PyObject **pv, Py_ssize_t newsize);
 
 /* runtime lifecycle */
 
@@ -73,6 +73,142 @@ typedef struct {
     PyTupleObject *it_seq; /* Set to NULL when iterator is exhausted */
 } _PyTupleIterObject;
 
+
+// --- _PyTupleBuilder API ---------------------------------------------------
+
+typedef struct _PyTupleBuilder {
+    PyObject* small_tuple[16];
+    PyObject *tuple;
+    PyObject **items;
+    size_t size;
+    size_t allocated;
+} _PyTupleBuilder;
+
+static inline int
+_PyTupleBuilder_Alloc(_PyTupleBuilder *builder, size_t size)
+{
+    if (size > (size_t)PY_SSIZE_T_MAX) {
+        /* Check for overflow */
+        PyErr_NoMemory();
+        return -1;
+    }
+    if (size <= builder->allocated) {
+        return 0;
+    }
+
+    if (size <= Py_ARRAY_LENGTH(builder->small_tuple)) {
+        assert(builder->tuple == NULL);
+        builder->items = builder->small_tuple;
+        builder->allocated = Py_ARRAY_LENGTH(builder->small_tuple);
+        return 0;
+    }
+
+    assert(size >= 1);
+    if (builder->tuple != NULL) {
+        if (_PyTuple_ResizeNoTrack(&builder->tuple, (Py_ssize_t)size) < 0) {
+            return -1;
+        }
+    }
+    else {
+        builder->tuple = _PyTuple_NewNoTrack((Py_ssize_t)size);
+        if (builder->tuple == NULL) {
+            return -1;
+        }
+
+        if (builder->size > 0) {
+            memcpy(_PyTuple_ITEMS(builder->tuple),
+                   builder->small_tuple,
+                   builder->size * sizeof(builder->small_tuple[0]));
+        }
+    }
+    builder->items = _PyTuple_ITEMS(builder->tuple);
+    builder->allocated = size;
+    return 0;
+}
+
+static inline int
+_PyTupleBuilder_Init(_PyTupleBuilder *builder, Py_ssize_t size)
+{
+    memset(builder, 0, sizeof(*builder));
+
+    int res;
+    if (size > 0) {
+        res = _PyTupleBuilder_Alloc(builder, (size_t)size);
+    }
+    else {
+        res = 0;
+    }
+    return res;
+}
+
+// The tuple builder must have already enough allocated items to store item.
+static inline void
+_PyTupleBuilder_AppendUnsafe(_PyTupleBuilder *builder, PyObject *item)
+{
+    assert(builder->items != NULL);
+    assert(builder->size < builder->allocated);
+    builder->items[builder->size] = item;
+    builder->size++;
+}
+
+static inline int
+_PyTupleBuilder_Append(_PyTupleBuilder *builder, PyObject *item)
+{
+    if (builder->size >= (size_t)PY_SSIZE_T_MAX) {
+        // prevent integer overflow
+        PyErr_NoMemory();
+        return -1;
+    }
+    if (builder->size >= builder->allocated) {
+        size_t allocated = builder->size;
+        allocated += (allocated >> 2);  // Over-allocate by 25%
+        if (_PyTupleBuilder_Alloc(builder, allocated) < 0) {
+            return -1;
+        }
+    }
+    _PyTupleBuilder_AppendUnsafe(builder, item);
+    return 0;
+}
+
+static inline void
+_PyTupleBuilder_Dealloc(_PyTupleBuilder *builder)
+{
+    Py_CLEAR(builder->tuple);
+    builder->items = NULL;
+    builder->size = 0;
+    builder->allocated = 0;
+}
+
+static inline PyObject*
+_PyTupleBuilder_Finish(_PyTupleBuilder *builder)
+{
+    if (builder->size == 0) {
+        _PyTupleBuilder_Dealloc(builder);
+        // return the empty tuple singleton
+        return PyTuple_New(0);
+    }
+
+    if (builder->tuple != NULL) {
+        if (_PyTuple_ResizeNoTrack(&builder->tuple, (Py_ssize_t)builder->size) < 0) {
+            _PyTupleBuilder_Dealloc(builder);
+            return NULL;
+        }
+
+        PyObject *result = builder->tuple;
+        builder->tuple = NULL;
+        // Avoid _PyObject_GC_TRACK() to avoid including pycore_object.h
+        PyObject_GC_Track(result);
+        return result;
+    }
+    else {
+        PyObject *tuple = _PyTuple_FromArraySteal(builder->items,
+                                                  (Py_ssize_t)builder->size);
+        builder->size = 0;
+        return tuple;
+    }
+}
+
+
 #ifdef __cplusplus
 }
 #endif

Modules/itertoolsmodule.c

@@ -4,7 +4,7 @@
 #include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "pycore_typeobject.h"    // _PyType_GetModuleState()
 #include "pycore_object.h"        // _PyObject_GC_TRACK()
-#include "pycore_tuple.h"         // _PyTuple_ITEMS()
+#include "pycore_tuple.h"         // _PyTupleBuilder
 #include "structmember.h"         // PyMemberDef
 #include <stddef.h>               // offsetof()
 
@@ -193,47 +193,42 @@ batched_traverse(batchedobject *bo, visitproc visit, void *arg)
 static PyObject *
 batched_next(batchedobject *bo)
 {
-    Py_ssize_t i;
-    Py_ssize_t n = bo->batch_size;
     PyObject *it = bo->it;
-    PyObject *item;
-    PyObject *result;
-
     if (it == NULL) {
         return NULL;
     }
-    result = PyTuple_New(n);
-    if (result == NULL) {
+
+    _PyTupleBuilder builder;
+    Py_ssize_t n = bo->batch_size;
+    if (_PyTupleBuilder_Init(&builder, n) < 0) {
         return NULL;
     }
+
     iternextfunc iternext = *Py_TYPE(it)->tp_iternext;
-    PyObject **items = _PyTuple_ITEMS(result);
-    for (i=0 ; i < n ; i++) {
-        item = iternext(it);
+    for (Py_ssize_t i=0 ; i < n; i++) {
+        PyObject *item = iternext(it);
         if (item == NULL) {
-            goto null_item;
+            if (PyErr_Occurred()) {
+                if (!PyErr_ExceptionMatches(PyExc_StopIteration)) {
+                    /* Input raised an exception other than StopIteration */
+                    goto error;
+                }
+                PyErr_Clear();
+                // StopIteration was raised
+            }
+            if (i == 0) {
+                goto error;
+            }
+            break;
         }
-        items[i] = item;
+        _PyTupleBuilder_AppendUnsafe(&builder, item);
     }
-    return result;
+    return _PyTupleBuilder_Finish(&builder);
 
- null_item:
-    if (PyErr_Occurred()) {
-        if (!PyErr_ExceptionMatches(PyExc_StopIteration)) {
-            /* Input raised an exception other than StopIteration */
-            Py_CLEAR(bo->it);
-            Py_DECREF(result);
-            return NULL;
-        }
-        PyErr_Clear();
-    }
-    if (i == 0) {
-        Py_CLEAR(bo->it);
-        Py_DECREF(result);
-        return NULL;
-    }
-    _PyTuple_Resize(&result, i);
-    return result;
+error:
+    _PyTupleBuilder_Dealloc(&builder);
+    Py_CLEAR(bo->it);
+    return NULL;
 }
 
 static PyType_Slot batched_slots[] = {

Objects/abstract.c

@@ -2074,11 +2074,6 @@ PySequence_DelSlice(PyObject *s, Py_ssize_t i1, Py_ssize_t i2)
 PyObject *
 PySequence_Tuple(PyObject *v)
 {
-    PyObject *it;  /* iter(v) */
-    Py_ssize_t n;             /* guess for result tuple size */
-    PyObject *result = NULL;
-    Py_ssize_t j;
-
     if (v == NULL) {
         return null_error();
     }
@@ -2091,66 +2086,53 @@ PySequence_Tuple(PyObject *v)
            a copy, so there's no need for exactness below. */
         return Py_NewRef(v);
     }
-    if (PyList_CheckExact(v))
+    if (PyList_CheckExact(v)) {
         return PyList_AsTuple(v);
+    }
 
-    /* Get iterator. */
-    it = PyObject_GetIter(v);
-    if (it == NULL)
+    _PyTupleBuilder builder;
+    if (_PyTupleBuilder_Init(&builder, 0) < 0) {
         return NULL;
+    }
+
+    /* Get iterator. */
+    PyObject *it = PyObject_GetIter(v);  // iter(v)
+    if (it == NULL) {
+        goto Fail;
+    }
 
     /* Guess result size and allocate space. */
-    n = PyObject_LengthHint(v, 10);
-    if (n == -1)
+    Py_ssize_t n = PyObject_LengthHint(v, 10);  // Guess for result tuple size
+    if (n == -1) {
         goto Fail;
-    result = PyTuple_New(n);
-    if (result == NULL)
+    }
+    if (_PyTupleBuilder_Alloc(&builder, n) < 0) {
         goto Fail;
+    }
 
     /* Fill the tuple. */
+    Py_ssize_t j;
     for (j = 0; ; ++j) {
         PyObject *item = PyIter_Next(it);
         if (item == NULL) {
-            if (PyErr_Occurred())
+            if (PyErr_Occurred()) {
                 goto Fail;
+            }
             break;
         }
-        if (j >= n) {
-            size_t newn = (size_t)n;
-            /* The over-allocation strategy can grow a bit faster
-               than for lists because unlike lists the
-               over-allocation isn't permanent -- we reclaim
-               the excess before the end of this routine.
-               So, grow by ten and then add 25%.
-            */
-            newn += 10u;
-            newn += newn >> 2;
-            if (newn > PY_SSIZE_T_MAX) {
-                /* Check for overflow */
-                PyErr_NoMemory();
-                Py_DECREF(item);
-                goto Fail;
-            }
-            n = (Py_ssize_t)newn;
-            if (_PyTuple_Resize(&result, n) != 0) {
-                Py_DECREF(item);
-                goto Fail;
-            }
+
+        if (_PyTupleBuilder_Append(&builder, item) < 0) {
+            Py_DECREF(item);
+            goto Fail;
         }
-        PyTuple_SET_ITEM(result, j, item);
     }
 
-    /* Cut tuple back if guess was too large. */
-    if (j < n &&
-        _PyTuple_Resize(&result, j) != 0)
-        goto Fail;
-
     Py_DECREF(it);
-    return result;
+    return _PyTupleBuilder_Finish(&builder);
 
 Fail:
-    Py_XDECREF(result);
-    Py_DECREF(it);
+    _PyTupleBuilder_Dealloc(&builder);
+    Py_XDECREF(it);
     return NULL;
 }