gh-144569: Scalar replacement of `BINARY_SLICE` for list, tuple, and unicode #144590

cocolato · 2026-02-08T08:16:33Z

When the tier 2 optimizer can determine the type of the container being sliced, _BINARY_SLICE can be replaced with a specialized micro-op that bypasses the PySliceObject allocation entirely.

This PR implements scalar replacement of BINARY_SLICE by add new tier2 op _BINARY_SLICE_LIST, _BINARY_SLICE_TUPLE, and _BINARY_SLICE_UNICODE.

Issue: Avoid creating temporary objects in BINARY_SLICE #144569

cocolato · 2026-02-08T09:08:37Z

I think the tests failure that occurred is unrelated to the PR.

Fidget-Spinner

Pretty close, thanks!

Lib/test/test_capi/test_opt.py

Python/bytecodes.c

Fidget-Spinner · 2026-02-08T11:21:22Z

You forgot to add the new recorder to BINARY_SLICE macro op

Python/bytecodes.c

Fidget-Spinner

lgtm just one comment and add news please

Fidget-Spinner · 2026-02-08T12:20:23Z

Python/bytecodes.c

+                PySlice_AdjustIndices(PyUnicode_GET_LENGTH(container_o), &istart, &istop, 1);
+                res_o = PyUnicode_Substring(container_o, istart, istop);
+            }
+            DECREF_INPUTS();


We should split this into the POP_TOP POP_TOP form, can you please do that in a follow up PR?

Fidget-Spinner · 2026-02-08T12:27:05Z

Actually I've changed my mind. Can you please open a PR to convert BINARY_SLICE to the POP_TOP form please? Then after we merge that PR we can update this one

Fidget-Spinner

Let's leave the DECREF_INPUTS alone for now and try to remove it in a future PR.

The code has repeated instances that should be factored out to a micro-op

Fidget-Spinner · 2026-02-08T13:11:35Z

Python/bytecodes.c

+            if (_PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(start), &istart) &&
+                _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(stop), &istop)) {
+                PySlice_AdjustIndices(PyList_GET_SIZE(container_o), &istart, &istop, 1);
+                res_o = PyList_GetSlice(container_o, istart, istop);
+            }


This is repeated code, it should be factored out to a _UNPACK_INDICES micro-op that looks like this:

Suggested change

if (_PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(start), &istart) &&

_PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(stop), &istop)) {

PySlice_AdjustIndices(PyList_GET_SIZE(container_o), &istart, &istop, 1);

res_o = PyList_GetSlice(container_o, istart, istop);

}

int err = _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(start), &istart);

ERROR_IF(err == 0);

err = _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(stop), &istop);

ERROR_IF(err = 0);

...

istart = PyStackRef_Wrap(...)

...

Python/optimizer_bytecodes.c

Misc/NEWS.d/next/Core_and_Builtins/2026-02-08-13-14-00.gh-issue-144569.pjlJVe.rst

…e-144569.pjlJVe.rst

Python/bytecodes.c

Fidget-Spinner · 2026-02-08T16:44:48Z

This is pretty close, I'm going to push some changes in, then wait a day to see if Mark has any objections tomorrow.

cocolato · 2026-02-08T16:48:45Z

Thanks for your guidance!

1. We cannot exit after unpacking indices, as the stack contains tagged ints now which may lead to a crash. We must insert a guard for the type before unpacking indices. 2. It is possible for an indice to not fit in a tagged int, in which case we must deopt. 3. Recorded values do not always mean that that's the actual type we will see at runtime. So we must guard on that as well.

Fidget-Spinner · 2026-02-08T17:15:09Z

@cocolato please check out the latest commit and the extended commit message for the bugs fixed. Thanks!

Fidget-Spinner · 2026-02-08T17:25:15Z

add_int.py

def testfunc(n):
    data = [1, 2, 3, 4, 5]
    a, b = 1, 3
    for _ in range(n):
        x = data[a:b]
    return x

testfunc(50000000)

Results using hyperfine:

Summary
  PYTHON_JIT=1 ./python ../add_int.py ran
    1.12 ± 0.02 times faster than PYTHON_JIT=0 ./python ../add_int.py

So we made it >10% faster. Nice!

cocolato · 2026-02-09T03:10:04Z

@cocolato please check out the latest commit and the extended commit message for the bugs fixed. Thanks!

LGTM! Thanks again for the fix, I learned a lot from it.

Results using hyperfine:

This is my first time learning about this tool, and it looks great. In the past, when I needed to run benchmarks locally, timeit was somewhat inaccurate, and pyperformance was a bit too heavy.

cocolato · 2026-02-09T03:20:16Z

At the same time, I've noticed that when set PYTHON_OPT_DEBUG=5 is enabled, DUMP_UOP currently only outputs the first ADDED uop. I'll fix this in other PR.

markshannon

Thanks for doing this.
This looks sound, but there are some inefficiencies.

Beware the size of uops. They will be converted to machine code, so keeping code size down is important for performance.

markshannon · 2026-02-09T16:21:08Z

Lib/test/test_capi/test_opt.py

        self.assertIn("_UNPACK_SEQUENCE_TWO_TUPLE", uops)
        self.assertNotIn("_GUARD_TOS_TUPLE", uops)

+    def test_binary_slice_list(self):


Can you add tests that ensure that it deopts correctly.
Something like:

def f(seq): return f[0:1] for _ range(TIER2_THRESHOLD): f([1, 2]) f((1,2))

markshannon · 2026-02-09T16:25:22Z

Python/optimizer_bytecodes.c

-            type == &PyList_Type ||
-            type == &PyTuple_Type)
-        {
+        PyTypeObject *type = sym_get_probable_type(container);


Trying getting the known type as well and only adding the type guard if the type is not known

markshannon · 2026-02-09T16:56:21Z

Python/bytecodes.c

+                ERROR_NO_POP();
+            }
+            PySlice_AdjustIndices(len, &istart, &istop, 1);
+            DEOPT_IF(!PyStackRef_CanTagInt(istart));


You can replace these with asserts.
Once the indices have been adjusted, they cannot be outside the range of tagged integers. We cannot allocate objects that large.

markshannon · 2026-02-09T17:40:26Z

Python/bytecodes.c

        }

-        macro(BINARY_SLICE) = _SPECIALIZE_BINARY_SLICE + _BINARY_SLICE;
+        tier2 op(_UNPACK_INDICES, (container, start, stop -- container, sta, sto)) {


This is quite bulky for jitted code. Could you factor out as much as possible as a helper function?
So that this uop would look like:

tier2 op(_UNPACK_INDICES, (container, start, stop -- container, sta, sto)) { Py_ssize_t istart, istop; int err = _PyEval_UnpackIndices(container, &istart, &istop); if (err == 0) { ERROR_NO_POP(); } PyStackRef_CLOSE(stop); PyStackRef_CLOSE(start); sta = PyStackRef_TagInt(istart); sto = PyStackRef_TagInt(istop);

markshannon · 2026-02-09T17:46:53Z

Python/optimizer_bytecodes.c

        }
    }

+    op(_UNPACK_INDICES, (container, start, stop -- container, sta, sto)) {


start and stop may well be constants. Slices like [:n] and [n:] are quite common, so this can be optimized, popping the None and replacing it with tagged(0) for start, or tagged(MAX) for the end.

If both values are constant we can eliminate the _UNPACK_INDICES. If only one is constant, then we can replace _UNPACK_INDICES with a non-negative compact int guard and tagging.

markshannon · 2026-02-09T17:49:14Z

Python/optimizer_bytecodes.c


+    op(_UNPACK_INDICES, (container, start, stop -- container, sta, sto)) {
+        (void)container;
+        sta = sym_new_compact_int(ctx);


Suggested change

sta = sym_new_compact_int(ctx);

sta = sym_new_type(ctx, &PyLong_Type);

Compact ints may not be as large as indices.

markshannon · 2026-02-09T17:49:20Z

Python/optimizer_bytecodes.c

+    op(_UNPACK_INDICES, (container, start, stop -- container, sta, sto)) {
+        (void)container;
+        sta = sym_new_compact_int(ctx);
+        sto = sym_new_compact_int(ctx);


bedevere-app · 2026-02-09T17:51:55Z

A Python core developer has requested some changes be made to your pull request before we can consider merging it. If you could please address their requests along with any other requests in other reviews from core developers that would be appreciated.

Once you have made the requested changes, please leave a comment on this pull request containing the phrase I have made the requested changes; please review again. I will then notify any core developers who have left a review that you're ready for them to take another look at this pull request.

Scalar replacement of BINARY_SLICE for list, tuple, and unicode

6ada5ed

cocolato requested review from Fidget-Spinner, markshannon, savannahostrowski and tomasr8 as code owners February 8, 2026 08:16

bedevere-app bot added the awaiting review label Feb 8, 2026

bedevere-app bot mentioned this pull request Feb 8, 2026

Avoid creating temporary objects in BINARY_SLICE #144569

Open

Fidget-Spinner reviewed Feb 8, 2026

View reviewed changes

Lib/test/test_capi/test_opt.py Outdated Show resolved Hide resolved

Python/bytecodes.c Outdated Show resolved Hide resolved

add _RECORD_3OS_TYPE

2b5e88c

cocolato commented Feb 8, 2026

View reviewed changes

Python/bytecodes.c Show resolved Hide resolved

Fidget-Spinner reviewed Feb 8, 2026

View reviewed changes

📜🤖 Added by blurb_it.

3faaac9

Fidget-Spinner reviewed Feb 8, 2026

View reviewed changes

Misc/NEWS.d/next/Core_and_Builtins/2026-02-08-13-14-00.gh-issue-144569.pjlJVe.rst Outdated Show resolved Hide resolved

Fidget-Spinner and others added 2 commits February 8, 2026 13:23

Update Misc/NEWS.d/next/Core_and_Builtins/2026-02-08-13-14-00.gh-issu…

6290916

…e-144569.pjlJVe.rst

add _UNPACK_INDICES

3dd3f3a

cocolato commented Feb 8, 2026

View reviewed changes

Python/bytecodes.c Show resolved Hide resolved

move AdjustIndices to _UNPACK_INDICES

b47e3fc

markshannon requested changes Feb 9, 2026

View reviewed changes

bedevere-app bot removed the awaiting review label Feb 9, 2026

bedevere-app bot added the awaiting changes label Feb 9, 2026

-            if (_PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(start), &istart) &&
-                _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(stop), &istop)) {
-                PySlice_AdjustIndices(PyList_GET_SIZE(container_o), &istart, &istop, 1);
-                res_o = PyList_GetSlice(container_o, istart, istop);
-            }
+    int err = _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(start), &istart);
+    ERROR_IF(err == 0);
+    err = _PyEval_SliceIndex(PyStackRef_AsPyObjectBorrow(stop), &istop);
+    ERROR_IF(err = 0);
+    ...
+    istart = PyStackRef_Wrap(...)
+    ...

	sta = sym_new_compact_int(ctx);
	sta = sym_new_type(ctx, &PyLong_Type);

Uh oh!

gh-144569: Scalar replacement of BINARY_SLICE for list, tuple, and unicode #144590

Are you sure you want to change the base?

gh-144569: Scalar replacement of BINARY_SLICE for list, tuple, and unicode #144590

Uh oh!

Conversation

cocolato commented Feb 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cocolato commented Feb 8, 2026

Uh oh!

Fidget-Spinner left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Fidget-Spinner commented Feb 8, 2026

Uh oh!

Uh oh!

Fidget-Spinner left a comment

Choose a reason for hiding this comment

Uh oh!

Fidget-Spinner Feb 8, 2026

Choose a reason for hiding this comment

Uh oh!

Fidget-Spinner commented Feb 8, 2026

Uh oh!

Fidget-Spinner left a comment

Choose a reason for hiding this comment

Uh oh!

Fidget-Spinner Feb 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Fidget-Spinner commented Feb 8, 2026

Uh oh!

cocolato commented Feb 8, 2026

Uh oh!

Fidget-Spinner commented Feb 8, 2026

Uh oh!

Fidget-Spinner commented Feb 8, 2026

Uh oh!

cocolato commented Feb 9, 2026

Uh oh!

cocolato commented Feb 9, 2026

Uh oh!

markshannon left a comment

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

markshannon Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

bedevere-app bot commented Feb 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

gh-144569: Scalar replacement of `BINARY_SLICE` for list, tuple, and unicode #144590

gh-144569: Scalar replacement of `BINARY_SLICE` for list, tuple, and unicode #144590

cocolato commented Feb 8, 2026 •

edited

Loading

Fidget-Spinner Feb 8, 2026 •

edited

Loading

markshannon Feb 9, 2026 •

edited

Loading