A (very) slight speed improvement for iterating over bytes (#21705)
My mentee @xvxvxvxvxv noticed that iterating over array.array is
slightly faster than iterating over bytes. Looking at the source I
observed that arrayiter_next() calls `getitem(ao, it->index++)` wheras
striter_next() uses the idiom (paraphrased)
item = PyLong_FromLong(seq->ob_sval[it->it_index]);
if (item != NULL)
++it->it_next;
return item;
I'm not 100% sure but I think that the second version has fewer
opportunity for the CPU to overlap the `index++` operation with the
rest of the code (which in both cases involves a call). So here I am
optimistically incrementing the index -- if the PyLong_FromLong() call
fails, this will leave the iterator pointing at the next byte, but
honestly I doubt that anyone would seriously consider resuming use of
the iterator after that kind of failure (it would have to be a
MemoryError). And the author of arrayiter_next() made the same
consideration (or never ever gave it a thought :-).
With this, a loop like
for _ in b: pass
is now slightly *faster* than the same thing over an equivalent array,
rather than slightly *slower* (in both cases a few percent).
diff --git a/Objects/bytesobject.c b/Objects/bytesobject.c
index 3a922d3..836a736 100644
--- a/Objects/bytesobject.c
+++ b/Objects/bytesobject.c
@@ -3139,7 +3139,6 @@
striter_next(striterobject *it)
{
PyBytesObject *seq;
- PyObject *item;
assert(it != NULL);
seq = it->it_seq;
@@ -3148,11 +3147,8 @@
assert(PyBytes_Check(seq));
if (it->it_index < PyBytes_GET_SIZE(seq)) {
- item = PyLong_FromLong(
- (unsigned char)seq->ob_sval[it->it_index]);
- if (item != NULL)
- ++it->it_index;
- return item;
+ return PyLong_FromLong(
+ (unsigned char)seq->ob_sval[it->it_index++]);
}
it->it_seq = NULL;