Use tagged ints for faster iteration

[markshannon]

Iteration over tuples and short lists is quite inefficient as we need to create an iterator object, only to have to destroy it again moments later. Not only that, fetching values from iterators involves additional indirection compared to fetching them from sequences.

Instead we can push a pair of values to the stack. For common sequences, like tuple, list, strings, some ranges and a few others, we push the sequence and the integer index (initially 0) to the stack. For other iterables, we push the iterator and NULL.

GET_ITER will have the signature:
iterable -- iter, index_or_null
FOR_ITER now has the signature:
iter, index_or_null -- iter, index_or_null, next.

What makes this efficient is tagged integers. By using tagged integers, no objects need to be created.

Examples

GET_ITER

[ <tuple at ...> ] -> [ <tuple at ...>, 0 ]

[ <file at ...> ] -> [ <file iterator at ...>, NULL ]

FOR_ITER

[ <tuple at ...>, 0 ] -> [ <tuple at ...>, 1, item0 ]

[ <file iterator at ...>, NULL ] -> [ <file iterator at ...>, NULL, line ]

Linked PRs

• GH-132554: "Virtual" iterators #132555
• GH-132554: Add stats for GET_ITER #132592

[markshannon]

Faster CPython discussion

Tagged ints issue

[mdboom]

As discussed in our meeting, I'm going to take a first crack at this.

[mdboom]

As discussed in our meeting, I'm going to take a first crack at this.

Actually, realizing @markshannon already started that, I'm instead going to follow-on and do GET_ITER specializations.