Skip to content

gh-131798: JIT: Propagate the result in _BINARY_OP_SUBSCR_TUPLE_INT #133003

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Fidget-Spinner merged 4 commits into from
Apr 26, 2025
Merged

gh-131798: JIT: Propagate the result in _BINARY_OP_SUBSCR_TUPLE_INT #133003

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c3cd21c
Make _BINARY_OP_SUBSCR_TUPLE_INT smarter
tomasr8 Apr 24, 2025
c7b7b48
Add news entry
tomasr8 Apr 26, 2025
dddd56f
Fix a case when the tuple length is not known
tomasr8 Apr 26, 2025
00b00cb
Keep the same names for stack variables
tomasr8 Apr 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions Lib/test/test_capi/test_opt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1923,6 +1923,23 @@ def testfunc(n):
self.assertNotIn("_GUARD_TOS_INT", uops)
self.assertIn("_CALL_LEN", uops)

def test_binary_op_subscr_tuple_int(self):
def testfunc(n):
x = 0
for _ in range(n):
y = (1, 2)
if y[0] == 1: # _COMPARE_OP_INT + _GUARD_IS_TRUE_POP are removed
x += 1
return x

res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
self.assertEqual(res, TIER2_THRESHOLD)
self.assertIsNotNone(ex)
uops = get_opnames(ex)
self.assertIn("_BINARY_OP_SUBSCR_TUPLE_INT", uops)
self.assertNotIn("_COMPARE_OP_INT", uops)
self.assertNotIn("_GUARD_IS_TRUE_POP", uops)


def global_identity(x):
return x
Expand Down
2 changes: 2 additions & 0 deletions Misc/NEWS.d/next/Core_and_Builtins/2025-04-26-13-57-13.gh-issue-131798.Gt8CGE.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,2 @@
Propagate the return type of ``_BINARY_OP_SUBSCR_TUPLE_INT`` in JIT. Patch
by Tomas Roun
14 changes: 14 additions & 0 deletions Python/optimizer_bytecodes.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -370,6 +370,20 @@ dummy_func(void) {
res = sym_new_type(ctx, &PyUnicode_Type);
}

op(_BINARY_OP_SUBSCR_TUPLE_INT, (left, right -- res)) {
assert(sym_matches_type(left, &PyTuple_Type));
if (sym_is_const(ctx, right)) {
assert(PyLong_CheckExact(sym_get_const(ctx, right)));
long index = PyLong_AsLong(sym_get_const(ctx, right));
assert(index >= 0);
assert(index < sym_tuple_length(left));
Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure these asserts are actually needed. The instruction will deopt when this is not true so perhaps we can remove them?

Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Interesting, we are actually hitting the assertion:

Python/optimizer_cases.c.h:630: optimize_uops: Assertion `index < sym_tuple_length(left)' failed.

I guess I need to move the DEOPT_IF checks into a separate guard for it to work?

Copy link
Copy Markdown
Member

@Fidget-Spinner Fidget-Spinner Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Check out the code for sym_tuple_length. It can return -1 if the length is not known. So for example if you propagate a PyTuple_Type, but not the length, it will fail.

You need to check that the length is not -1.

Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Got it, thanks! I somehow didn't realize you can know something is a tuple but not know its length..

res = sym_tuple_getitem(ctx, left, index);
Copy link
Copy Markdown
Member

@Fidget-Spinner Fidget-Spinner Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a bit unfortunate. I would wish we could automatically evaluate this, but it seems tuples are a special category so we can't.

#132733

Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 Apr 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I haven't fully digested your PR yet, but maybe there is some way to extend it to support tuples as well?

}
else {
res = sym_new_not_null(ctx);
}
Comment on lines +375 to +391
Copy link
Copy Markdown
Member

@brandtbucher brandtbucher May 6, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@tomasr8, this can be improved to handle abstract tuples that may not be constant (and it cleans things up, since it includes a length check):

Suggested change
if (sym_is_const(ctx, sub_st)) {
assert(PyLong_CheckExact(sym_get_const(ctx, sub_st)));
long index = PyLong_AsLong(sym_get_const(ctx, sub_st));
assert(index >= 0);
int tuple_length = sym_tuple_length(tuple_st);
if (tuple_length == -1) {
// Unknown length
res = sym_new_not_null(ctx);
}
else {
assert(index < tuple_length);
res = sym_tuple_getitem(ctx, tuple_st, index);
}
}
else {
res = sym_new_not_null(ctx);
}
long index = PyLong_AsLong(sym_get_const(ctx,
sub_st));
// ...PyLong_AsLong can fail, need to check for error here...
res = sym_tuple_getitem(ctx, tuple_st, index)

Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 May 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this can be improved to handle abstract tuples that may not be constant

@brandtbucher, I think this already handles abstract tuples? Assuming that abstract means those that have JIT_SYM_TUPLE_TAG rather than JIT_SYM_KNOWN_VALUE_TAG? For tuples that just have JIT_SYM_KNOWN_CLASS_TAG we don't know the length so there's nothing we can do I believe.

But yeah, we can simplify the code since _Py_uop_sym_tuple_getitem can handle unknown length. The only difference is that it sets the type to unknown rather than non null as we do here. Is that a problem? Shouldn't _Py_uop_sym_tuple_getitem always return at least JIT_SYM_NON_NULL_TAG?

Copy link
Copy Markdown
Member

@brandtbucher brandtbucher May 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Your current code is gated on sym_is_const, which abstract tuples fail. The rest of your understanding is correct, though.

Is that a problem? Shouldn't _Py_uop_sym_tuple_getitem always return at least JIT_SYM_NON_NULL_TAG?

Sounds like you found your next PR. :)

Copy link
Copy Markdown
Member

@brandtbucher brandtbucher May 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, I completely misunderstood, sorry. I thought you were checking whether the tuple was const, not the index. Disregard!

Copy link
Copy Markdown
Member Author

@tomasr8 tomasr8 May 8, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

no worries :) I think we can still simplify the code a bit though!

}

op(_TO_BOOL, (value -- res)) {
int already_bool = optimize_to_bool(this_instr, ctx, value, &res);
if (!already_bool) {
Expand Down
16 changes: 15 additions & 1 deletion Python/optimizer_cases.c.h
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

Loading