-
-
Notifications
You must be signed in to change notification settings - Fork 30.4k
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.
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
gh-115999: Add free-threaded specialization for COMPARE_OP #126410
base: main
Are you sure you want to change the base?
Conversation
…E_OP specialization in general.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM! Please add a short explanation of why the specialized instructions are thread-safe to either the pull request or the commit (when it's merged).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The approach looks sound, but the test will need moving and should not rely on specific sequences of instructions.
@@ -1335,6 +1335,32 @@ def test_call_specialize(self): | |||
got = self.get_disassembly(co, adaptive=True) | |||
self.do_disassembly_compare(got, call_quicken) | |||
|
|||
@cpython_only |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can we move this test to a more appropriate location, like test_opcache
?
Also, can we avoid hardcoding sequences of instructions into the tests and use a behavioral approach: testing that the behavior is the same with and without specialization.
We should already have tests for specialization of COMPARE_OP
somewhere.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Behavioural tests specifically for specialization sound like a great idea regardless of whether we're free-threaded or not. They should still make sure that actual specialization happened, and not just exercise some code. As far as I can tell we don't have any of those right now. I assume that was on purpose.
I agree we should already have tests for specialization of COMPARE_OP, but we don't. That's why I added the new test, similar to the existing test for specialization of BINARY_OP, which Matt adapted for free-threaded specialization in #123926. That test existed here, in test_dis. I assume they were added here on purpose. Should all those tests be moved to a more appropriate location?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Please take a look: #126498,
even I am not sure Mark intended these kinds of tests.
(If I add COMPARE_OP tests, it is properly failed.)
When you're done making the requested changes, leave the comment: |
@@ -2393,12 +2394,12 @@ _Py_Specialize_CompareOp(_PyStackRef lhs_st, _PyStackRef rhs_st, _Py_CODEUNIT *i | |||
goto failure; | |||
} | |||
if (PyFloat_CheckExact(lhs)) { | |||
instr->op.code = COMPARE_OP_FLOAT; | |||
specialized_op = COMPARE_OP_FLOAT; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Now you can use helpers to write it, like specialize(instr, COMPARE_OR_FLOAT)
which looks clearer.
(You need to merge main to get changes from 9ce4fa0)
goto success; | ||
} | ||
} | ||
SPECIALIZATION_FAIL(COMPARE_OP, compare_op_fail_kind(lhs, rhs)); | ||
failure: | ||
STAT_INC(COMPARE_OP, failure); | ||
instr->op.code = COMPARE_OP; | ||
SET_OPCODE_OR_RETURN(instr, COMPARE_OP); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The same applies to that piece of code, you can remove both of failure and success blocks, now you just need to do unspecialize(instr, compare_op_fail_kind(lhs, rhs))
. And in fact there's no more a SET_OPCODE_OR_RETURN
macro :)
Also don't forget to remove the #ifdef Py_STATS
directive near the compare_op_fail_kind
.
Add free-threaded specialization for COMPARE_OP, and tests for COMPARE_OP specialization in general. This relies on thread-local bytecode and atomic operations to update the counters and bytecode, which are single operations.
--disable-gil
builds #115999