Severe performance degradation for tracing under 3.11 #93516
Comments
|
Related issue where a user reported that code with cProfile slowed down 1.6x, and the only thing I could pinpoint was that tracing itself slowed down, not cProfile #93381. |
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
|
I now realize this is probably due to using the Python tracer rather than the C tracer, will try again. |
|
Okay, now with Script used:import coverage
from time import perf_counter
def fib(n):
if n <= 1:
return n
else:
return fib(n-1) + fib(n-2)
t0 = perf_counter()
cov = coverage.Coverage()
cov.start()
fib(35)
cov.stop()
cov.save()
t1 = perf_counter()
print(t1 - t0)Profile on 3.10 branch: took 13.3 secondsFunctions taking more than 1% of CPU:
Profile on main branch: took 23.0 secondsFunctions taking more than 1% of CPU:
|
|
This seems to suggest to me that this could be made much faster via a dedicated C API function: https://github.com/nedbat/coveragepy/blob/master/coverage/ctracer/util.h#L41 #define MyCode_GetCode(co) (PyObject_GetAttrString((PyObject *)(co), "co_code")) |
|
Indeed, on main, if I add Each of these calls to |
|
Ah! |
|
Some performance degradation under tracing is expected, but not as much as reported. From the profiles that @sweeneyde gathered it looks like we are seeing slowdowns in:
The new line number table gives better error messages without taking up too much extra space, but it is more complex and thus slower to parse. Once nedbat/coveragepy#1394 is merged, we can look at the profile again to see if much has changed. |
New profile of fib_cover.py on main branch, with the PyCode_GetCode addition (took 18.1 seconds)
I wonder: would there be any way to retain some specializations during tracing? Some specialized opcodes are certainly incorrect to use during tracing, e.g., |
Well, two times slower than 3.10, as some of @nedbat's benchmarks show, is not acceptable in my opinion. We should try to get this below 20% at the very least. I marked this as a release blocker and this will block any future releases. |
|
The problem with saying that something is unacceptable, is that implies there is an acceptable alternative. What does it mean making this a release blocker? Are you not going to release 3.11 if coverage is >20% slower? I don't think cutoffs, such as 20%, are constructive. We should do as well as we are reasonably able, regardless of whether we achieve some arbitrary number. |
I don't know, but there is always the possibility to revert the changes that made coverage slower (worst case scenario).
I will check with the SC, but I can already tell you that I would not be comfortable releasing 3.11 if is 2 times slower in coverage, unless the SC mandates otherwise.
The problem here is that this tradeoff was not discussed anywhere or advertised and no one had the possibility to pronounce or to make tradeoffs or otherwise, and therefore we cannot just "decide" out of the blue what is and what is not acceptable because the community didn't had the change to object. |
|
The largest contributor to the slowdown is computing line numbers. The root cause of that is PEP 657. The proximate cause is my PR to reduce the memory consumed of PEP 657. I think it would be a lot more helpful if we fixed the issue, rather than making dire pronouncements.
Everything we do is public. I even opened nedbat/coveragepy#1287 back in November, so that @nedbat was aware that there might be slowdowns to coverage.py |
That's not true and I feel is misleading. When line numbers and column offsets were separated, PEP 657 had 0 overhead on computing line numbers, so the root cause is certainly not PEP 657. The root cause was merging line numbers and column offsets so you need to pay for decoding both just to get line numbers.
That's not enough. If that were true we would not need to do PEPs because "what we do is already public". The point of having discussion on mailing lists is precisely so everyone can be aware of the tradeoff and give us their opinions. |
|
In any case, there is no point to discuss in these terms because I am not the one making decisions here. As RM my only authority is to consider the current slowdown a blocker and ask for a bigger discussion to be taken if we cannot fix the issue. Let's see what we can do and then let's discuss how to proceed once we understand our options. |
No, that was the proximate cause. It wouldn't have been necessary if not for PEP 657. Claiming that PEP 657 had zero overhead is not true. The original implementation had a large overhead in terms on memory use. |
Mark, I understand that and we are on the same page. I know why you made that change, and IIRC I even reviewed it. The problem here is that that change (or other changes) had unexpected consequences on timing and we need to make decisions either trying to fix together it if we can or potentially reverting it we cannot. That's all I am saying
The claim was that PEP 657 had zero overhead on the time computing the line numbers, not on memory.My sentence was:
|
|
Marking this as a release or deferred blocker is fine and entirely what I'd expect any RM to want to do. That just means we collect more information and make a decision before we exit beta stages. I don't personally have a problem with coverage being slower, even by 2x. It's more about understanding what all is impacted by this and how so we can explain our measured rationale to users. ie: Look beyond just Q1: What else uses tracing? Q1b: What of those use it at other than development time? These questions are informational and do not block performance work on reducing the impact to tracing. |
|
How bad would it be to store an extra field into code objects that is the "decompressed" line number table, lazily allocated for only those code objects which are actually traced? Could it occupy |
In #93383 (comment) I benchmarked adding an additional field to the code object and there was no slowdown in pyperformance. I do not know how memory consumption is affected. I'm not sure co_extra is the right place for that info considering its for consumers of the PEP 523 API. Maybe I can merge your idea into a simple lazily created PyCodeObject {
...
void *co_tracing_info
}
struct _PyCodeTracingInfo {
// cached code bytes object
PyObject *co_code;
// decompressed line number table
...
} |
An early design for the compressed line number table did just this. It isn't just tracing that needs line numbers, tracebacks need them too. I think we should see how much faster we can make tracing with the compressed table first. If that isn't good enough, then we can reconsidered the decompressed table. |
Debuggers and profiling tools, like cProfile. Q1b: What of those use it at other than development time? Hopefully, no one. Tracing has always imposed a considerable overhead. |
@pablogsal I've tried with both (the benchmarks I posted are for Python 3.11 beta 3 and Python 3.11 tip which I compiled locally -- note that I did not compile with PGO just release -- i.e.: |
|
Better formatted table for
|
|
As a random Python developer, I would be unhappy with a 20% slowdown in the speed of coverage, under the assumption that that would make all of my continuous integration runs take 20% longer. :/ |
|
I did a benchmark, and the latest tip of 3.11 (commit c966e08) is looking better, though still slower than 3.10:
|
|
Quite frankly, those 3.11 tip numbers look well within reason to accept. Tracing is a relatively rare operation in terms of all CPU cycles consumed by CPython processes on all computers around the world. It is worth sacrificing some tracing performance there if deemed necessary in order to gain performance where it is actually statistically significant: in deployed python programs. |
|
As a top line conclusion, I agree that with these improvements this no longer needs to block the release. That said, I think it's important to note that coverage performance has a substantial impact on developer experience. For many projects tests-with-coverage is a blocking element of their CI pipeline, so this performance ultimately dictates their iteration cycle time. It cannot happen for 3.11, but I think on a longer timeline it's worth exploring what can be done to improve coverage's performance (which perhaps is another external tool like https://github.com/plasma-umass/slipcover) -- particularly if it's going to be in tension with other optimizations we want to land for the general case! |
|
We still have Mark's PR #94231. Once that's merged I suspect the benchmark numbers for 3.11 tip will drop to below a 30% slowdown over 3.10. |
|
Yup, I just did that. About the same as without the change:
|
* GH-93444: remove redundant fields from basicblock: b_nofallthrough, b_exit, b_return (GH-93445) * netrc: Remove unused "import shlex" (#93311) * gh-92886: Fix test that fails when running with `-O` in `test_imaplib.py` (#93237) * Fix missing word in sys.float_info docstring (GH-93489) * [doc] Correct a grammatical error in a docstring. (GH-93441) * gh-93442: Make C++ version of _Py_CAST work with 0/NULL. (#93500) Add C++ overloads for _Py_CAST_impl() to handle 0/NULL. This will allow C++ extensions that pass 0 or NULL to macros using _Py_CAST() to continue to compile. Without this, you get an error like: invalid ‘static_cast’ from type ‘int’ to type ‘_object*’ The modern way to use a NULL value in C++ is to use nullptr. However, we want to not break extensions that do things the old way. Co-authored-by: serge-sans-paille * gh-93442: Add test for _Py_CAST(nullptr). (gh-93505) * gh-90473: wasmtime does not support absolute symlinks (GH-93490) * gh-89973: Fix re.error in the fnmatch module. (GH-93072) Character ranges with upper bound less that lower bound (e.g. [c-a]) are now interpreted as empty ranges, for compatibility with other glob pattern implementations. Previously it was re.error. * Document LOAD_FAST_CHECK opcode (#93498) * gh-93247: Fix assert function in asyncio locks test (#93248) * gh-90473: WASI requires proper open(2) flags (GH-93529) * GH-92308 What's New: list pending removals in 3.13 and future versions (#92562) * gh-90473: Skip POSIX tests that don't apply to WASI (GH-93536) * asyncio.Barrier docs: Fix typo (#93371) taks -> tasks * gh-83728: Add hmac.new default parameter deprecation (GH-91939) * gh-90473: Make chmod a dummy on WASI, skip chmod tests (GH-93534) WASI does not have the ``chmod(2)`` syscall yet. * Remove action=None kwarg from Barrier docs (GH-93538) * [docs] fix some asyncio.Barrier.wait docs grammar (GH-93552) * gh-93475: Expose FICLONE and FICLONERANGE constants in fcntl (#93478) * gh-89018: Improve documentation of `sqlite3` exceptions (#27645) - Order exceptions as in PEP 249 - Reword descriptions, so they match the current behaviour Co-authored-by: Alex Waygood <[email protected]> * bpo-42658: Use LCMapStringEx in ntpath.normcase to match OS behaviour for case-folding (GH-32010) * Fix contributor name in WhatsNew 3.11 (GH-93556) * Grammar fix to socket error string (GH-93523) * gh-86986: bump min sphinx version to 3.2 (GH-93337) * gh-79096: Protect cookie file created by {LWP,Mozilla}CookieJar.save() (GH-93463) Note: This change is not effective on Microsoft Windows. Cookies can store sensitive information and should therefore be protected against unauthorized third parties. This is also described in issue #79096. The filesystem permissions are currently set to 644, everyone can read the file. This commit changes the permissions to 600, only the creater of the file can read and modify it. This improves security, because it reduces the attack surface. Now the attacker needs control of the user that created the cookie or a ways to circumvent the filesystems permissions. This change is backwards incompatible. Systems that rely on world-readable cookies will breake. However, one could argue that those are misconfigured in the first place. * gh-93162: Add ability to configure QueueHandler/QueueListener together (GH-93269) Also, provide getHandlerByName() and getHandlerNames() APIs. Closes #93162. * gh-57539: Increase calendar test coverage (GH-93468) Co-authored-by: Sean Fleming Co-authored-by: Adam Turner <[email protected]> Co-authored-by: Łukasz Langa <[email protected]> * gh-88831: In docs for asyncio.create_task, explain why strong references to tasks are needed (GH-93258) Co-authored-by: Łukasz Langa <[email protected]> * Shrink the LOAD_METHOD cache by one codeunit. (#93537) * Fix MSVC compiler warnings in ceval.c (#93569) * gh-93162: test_config_queue_handler requires threading (GH-93572) * gh-84461: Emscripten's faccessat() does not accept flags (GHß92353) * gh-92592: Allow logging filters to return a LogRecord. (GH-92591) * Fix `PurePath.relative_to` links in the pathlib documentation. (GH-93268) These are currently broken as they refer to :meth:`Path.relative_to` rather than :meth:`PurePath.relative_to`, and `relative_to` is a method on `PurePath`. * GH-93481: Suppress expected deprecation warning in test_pyclbr (GH-93483) * gh-93370: Deprecate sqlite3.version and sqlite3.version_info (#93482) Co-authored-by: Alex Waygood <[email protected]> Co-authored-by: Adam Turner <[email protected]> Co-authored-by: Erlend E. Aasland <[email protected]> * GH-93521: For dataclasses, filter out `__weakref__` slot if present in bases (GH-93535) * gh-93421: Update sqlite3 cursor.rowcount only after SQLITE_DONE (#93526) * gh-93584: Make all install+tests targets depends on all (GH-93589) All install targets use the "all" target as synchronization point to prevent race conditions with PGO builds. PGO builds use recursive make, which can lead to two parallel `./python setup.py build` processes that step on each others toes. "test" targets now correctly compile PGO build in a clean repo. * gh-87961: Remove outdated notes from functions that aren't in the Limited API (GH-93581) * Remove outdated notes from functions that aren't in the Limited API Nowadays everything that *is* in the Limited API has a note added automatically. These notes could mislead people to think that these functions could never be added to the limited API. Remove them. * Also remove forgotten note on tp_vectorcall_offset not being finalized * gh-93180: Update os.copy_file_range() documentation (#93182) * gh-93575: Use correct way to calculate PyUnicode struct sizes (GH-93602) * gh-93575: Use correct way to calculate PyUnicode struct sizes * Add comment to keep test_sys and test_unicode in sync * Fix case code < 256 * gh-90473: Define HOSTRUNNER for WASI (GH-93606) * gh-79096: Fix/improve http cookiejar tests (GH-93614) Fixup of GH-93463: - remove stray print - use proper way to check file mode - add working chmod decorator Co-authored-by: Łukasz Langa <[email protected]> * gh-93616: Fix env changed issue in test_modulefinder (GH-93617) * gh-90494: Reject 6th element of the __reduce__() tuple (GH-93609) copy.copy() and copy.deepcopy() now always raise a TypeError if __reduce__() returns a tuple with length 6 instead of silently ignore the 6th item or produce incorrect result. * Doc: Update references and examples of old, unsupported OSes and uarches (GH-92791) * bpo-45383: Get metaclass from bases in PyType_From* (GH-28748) This checks the bases of of a type created using the FromSpec API to inherit the bases metaclasses. The metaclass's alloc function will be called as is done in `tp_new` for classes created in Python. Co-authored-by: Petr Viktorin <[email protected]> Co-authored-by: Erlend Egeberg Aasland <[email protected]> * Improve logging documentation with example and additional cookbook re… (GH-93644) * gh-90473: disable user site packages on WASI/Emscripten (GH-93633) * gh-90473: Skip get_config_h() tests on WASI (GH-93645) * gh-90549: Fix leak of global named resources using multiprocessing spawn (#30617) Co-authored-by: XD Trol <[email protected]> Co-authored-by: Antoine Pitrou <[email protected]> * gh-92434: Silence compiler warning in Modules/_sqlite/connection.c on 32-bit systems (#93090) * gh-90763: Modernise xx template module initialisation (#93078) Use C APIs such as PyModule_AddType instead of PyModule_AddObject. Also remove incorrect module decrefs if module fails to initialise. * gh-93491: Add support tier detection to configure (GH-93492) Co-authored-by: Adam Turner <[email protected]> Co-authored-by: Steve Dower <[email protected]> Co-authored-by: Erlend Egeberg Aasland <[email protected]> * gh-93466: Document PyType_Spec doesn't accept repeated slot IDs; raise where this was problematic (GH-93471) * gh-93671: Avoid exponential backtracking in deeply nested sequence patterns in match statements (GH-93680) Co-authored-by: Łukasz Langa <[email protected]> * gh-81790: support "UNC" device paths in `ntpath.splitdrive()` (GH-91882) * GH-93621: reorder code in with/async-with exception exit path to reduce the size of the exception table (GH-93622) * gh-93461: Invalidate sys.path_importer_cache entries with relative paths (GH-93653) * gh-91317: Document that Path does not collapse initial `//` (GH-32193) Documentation for `pathlib` says: > Spurious slashes and single dots are collapsed, but double dots ('..') are not, since this would change the meaning of a path in the face of symbolic links: However, it omits that initial double slashes also aren't collapsed. Later, in documentation of `PurePath.drive`, `PurePath.root`, and `PurePath.name` it mentions UNC but: - this abbreviation says nothing to a person who is unaware about existence of UNC (Wikipedia doesn't help either by [giving a disambiguation page](https://en.wikipedia.org/wiki/UNC)) - it shows up only if a person needs to use a specific property or decides to fully learn what the module provides. For context, see the BPO entry. * gh-92886: Fix tests that fail when running with optimizations (`-O`) in `test_zipimport.py` (GH-93236) * gh-92930: _pickle.c: Acquire strong references before calling save() (GH-92931) * gh-84461: Use HOSTRUNNER to run regression tests (GH-93694) Co-authored-by: Brett Cannon <[email protected]> * gh-90473: Skip test_queue when threading is not available (GH-93712) * gh-90153: whatsnew: "z" option in format spec (GH-93624) Add what's new entry for PEP 682 in Python 3.11. * gh-86404: [doc] A make sucpicious false positive. (GH-93710) * Change list to view object (#93661) * gh-84508: tool to generate cjk traditional chinese mappings (gh-93272) * Remove usage of _Py_IDENTIFIER from math module (#93739) * gh-91162: Support splitting of unpacked arbitrary-length tuple over TypeVar and TypeVarTuple parameters (alt) (GH-93412) For example: A[T, *Ts][*tuple[int, ...]] -> A[int, *tuple[int, ...]] A[*Ts, T][*tuple[int, ...]] -> A[*tuple[int, ...], int] * gh-93728: fix memory leak in deepfrozen code objects (GH-93729) * gh-93747: Fix Refleak when handling multiple Py_tp_doc slots (gh-93749) * GH-90699: use statically allocated strings in typeobject.c (gh-93751) * Add more FOR_ITER specialization stats (GH-32151) * gh-89653: PEP 670: Convert PyFunction macros (#93765) Convert PyFunction macros to static inline functions. * Remove ANY_VARARGS() macro from the C API (#93764) The macro was exposed by mistake. * gh-84623: Remove unused imports in stdlib (#93773) * gh-91731: Don't define 'static_assert' in C++11 where is a keyword to avoid UB (GH-93700) * gh-84623: Remove unused imports in tests (#93772) * gh-93353: Fix importlib.resources._tempfile() finalizer (#93377) Fix the importlib.resources.as_file() context manager to remove the temporary file if destroyed late during Python finalization: keep a local reference to the os.remove() function. Patch by Victor Stinner. * gh-84461: Fix parallel testing on WebAssembly (GH-93768) * gh-89653: PEP 670: Macros always cast arguments in cpython/ (#93766) Header files in the Include/cpython/ are only included if the Py_LIMITED_API macro is not defined. * gh-93353: Add test.support.late_deletion() (#93774) * gh-93741: Add private C API _PyImport_GetModuleAttrString() (GH-93742) It combines PyImport_ImportModule() and PyObject_GetAttrString() and saves 4-6 lines of code on every use. Add also _PyImport_GetModuleAttr() which takes Python strings as arguments. * gh-79512: Fixed names and __module__ value of weakref classes (GH-93719) Classes ReferenceType, ProxyType and CallableProxyType have now correct atrtributes __module__, __name__ and __qualname__. It makes them (types, not instances) pickleable. * gh-91810: Fix regression with writing an XML declaration with encoding='unicode' (GH-93426) Suppress writing an XML declaration in open files in ElementTree.write() with encoding='unicode' and xml_declaration=None. If file patch is passed to ElementTree.write() with encoding='unicode', always open a new file in UTF-8. * gh-93761: Fix test to avoid simple delay when synchronizing. (GH-93779) * gh-89546: Clean up PyType_FromMetaclass (GH-93686) When changing PyType_FromMetaclass recently (GH-93012, GH-93466, GH-28748) I found a bunch of opportunities to improve the code. Here they are. Fixes: #89546 Automerge-Triggered-By: GH:encukou * gh-91321: Fix compatibility with C++ older than C++11 (#93784) Fix the compatibility of the Python C API with C++ older than C++11. _Py_NULL is only defined as nullptr on C++11 and newer. * GH-93662: Make sure that column offsets are correct in multi-line method calls. (GH-93673) * GH-93516: Store offset of first traceable instruction in code object (GH-93769) * gh-90473: Include stdlib dir in wasmtime PYTHONPATH (GH-93797) * GH-93429: Merge `LOAD_METHOD` back into `LOAD_ATTR` (GH-93430) * gh-93353: regrtest checks for leaked temporary files (#93776) When running tests with -jN, create a temporary directory per process and mark a test as "environment changed" if a test leaks a temporary file or directory. * gh-79579: Improve DML query detection in sqlite3 (#93623) The fix involves using pysqlite_check_remaining_sql(), not only to check for multiple statements, but now also to strip leading comments and whitespace from SQL statements, so we can improve DML query detection. pysqlite_check_remaining_sql() is renamed lstrip_sql(), to more accurately reflect its function, and hardened to handle more SQL comment corner cases. * GH-93678: reduce boilerplate and code repetition in the compiler (GH-93682) * gh-91877: Fix WriteTransport.get_write_buffer_{limits,size} docs (#92338) - Amend docs for WriteTransport.get_write_buffer_limits - Add docs for WriteTransport.get_write_buffer_size * GH-93429: Document `LOAD_METHOD` removal (GH-93803) * Include freelists in allocation total. (GH-93799) * gh-93795: Use test.support TESTFN/unlink in sqlite3 tests (#93796) * Remove LOAD_METHOD stats. (GH-93807) * Rename 'LOAD_METHOD' specialization stat consts to 'ATTR'. (GH-93812) * gh-93353: Fix regrtest for -jN with N >= 2 (GH-93813) * [docs] Fix LOAD_ATTR version changed (GH-93816) * gh-93814: Add infinite test for itertools.chain.from_iterable (GH-93815) fix #93814 Automerge-Triggered-By: GH:rhettinger * gh-93735: Split Docs CI to speed-up the build (GH-93736) * gh-93183: Adjust wording in socket docs (#93832) package => packet Co-authored-by: Victor Norman * gh-93829: In sqlite3, replace Py_BuildValue with faster APIs (#93830) - In Modules/_sqlite/connection.c, use PyLong_FromLong - In Modules/_sqlite/microprotocols.c, use PyTuple_Pack * Add test.support.busy_retry() (#93770) Add busy_retry() and sleeping_retry() functions to test.support. * gh-87260: Update sqlite3 signature docs to reflect actual implementation (#93840) Align the docs for the following methods with the actual implementation: - sqlite3.complete_statement() - sqlite3.Connection.create_function() - sqlite3.Connection.create_aggregate() - sqlite3.Connection.set_progress_handler() * test_thread uses support.sleeping_retry() (#93849) test_thread.test_count() now fails if it takes longer than LONG_TIMEOUT seconds. * Use support.sleeping_retry() and support.busy_retry() (#93848) * Replace time.sleep(0.010) with sleeping_retry() to use an exponential sleep. * support.wait_process(): reuse sleeping_retry(). * _test_eintr: remove unused variables. * Update includes in call.c (GH-93786) * gh-93857: Fix broken audit-event targets in sqlite3 docs (#93859) Corrected targets for the following audit-events: - sqlite3.enable_load_extension => sqlite3.Connection.enable_load_extension - sqlite3.load_extension => sqlite3.Connection.load_extension * GH-93850: Fix test_asyncio exception ignored tracebacks (#93854) * gh-93824: Reenable installation of shell extension on Windows ARM64 (GH-93825) * test_asyncio: run_until() implements exponential sleep (#93866) run_until() of test.test_asyncio.utils now uses an exponential sleep delay (max: 1 second), rather than a fixed delay of 1 ms. Similar design than support.sleeping_retry() wait strategy that applies exponential backoff. * test_asyncore: Optimize capture_server() (#93867) Remove time.sleep(0.01) in test_asyncore capture_server(). The sleep was redundant and inefficient, since the loop starts with select.select() which also implements a sleep (poll for socket data with a timeout). * Tests call sleeping_retry() with SHORT_TIMEOUT (#93870) Tests now call busy_retry() and sleeping_retry() with SHORT_TIMEOUT or LONG_TIMEOUT (of test.support), rather than hardcoded constants. Add also WAIT_ACTIVE_CHILDREN_TIMEOUT constant to _test_multiprocessing. * gh-84461: Document how to install SDKs manually (GH-93844) Co-authored-by: Brett Cannon <[email protected]> * gh-93820: Fix copy() regression in enum.Flag (GH-93876) GH-26658 introduced a regression in copy / pickle protocol for combined `enum.Flag`s. `copy.copy(re.A | re.I)` would fail with `AttributeError: ASCII|IGNORECASE`. `enum.Flag` now has a `__reduce_ex__()` method that reduces flags by combined value, not by combined name. * Call busy_retry() and sleeping_retry() with error=True (#93871) Tests no longer call busy_retry() and sleeping_retry() with error=False: raise an exception if the loop times out. * gh-87347: Add parenthesis around PyXXX_Check() arguments (#92815) * gh-91321: Fix test_cppext for C++03 (#93902) Don't build _testcppext.cpp with -Wzero-as-null-pointer-constant when testing C++03: only use this compiler flag with C++11. * gh-91577: SharedMemory move imports out of methods (#91579) SharedMemory.unlink() uses the unregister() function from resource_tracker. Previously it was imported in the method, but this can fail if the method is called during interpreter shutdown, for example when unlink is part of a __del__() method. Moving the import to the top of the file, means that the unregister() method is available during interpreter shutdown. The register call in SharedMemory.__init__() can also use this imported resource_tracker. * gh-92547: Amend What's New (#93872) * Fix BINARY_SUBSCR_GETITEM stats (GH-93903) * gh-93847: Fix repr of enum of generic aliases (GH-93885) * gh-93353: regrtest supports checking tmp files with -j2 (#93909) regrtest now also implements checking for leaked temporary files and directories when using -jN for N >= 2. Use tempfile.mkdtemp() to create the temporary directory. Skip this check on WASI. * GH-91389: Fix dis position information for CACHEs (GH-93663) * gh-91985: Ensure in-tree builds override platstdlib_dir in every path calculation (GH-93641) * GH-83658: make multiprocessing.Pool raise an exception if maxtasksperchild is not None or a positive int (GH-93364) Closes #83658. * test_logging: Fix BytesWarning in SysLogHandlerTest (GH-93920) * gh-91404: Revert "bpo-23689: re module, fix memory leak when a match is terminated by a signal or allocation failure (GH-32283) (#93882) Revert "bpo-23689: re module, fix memory leak when a match is terminated by a signal or memory allocation failure (GH-32283)" This reverts commit 6e3eee5. Manual fixups to increase the MAGIC number and to handle conflicts with a couple of changes that landed after that. Thanks for reviews by Ma Lin and Serhiy Storchaka. * gh-89745: Avoid exact match when comparing program_name in test_embed on Windows (GH-93888) * gh-93852: Add test.support.create_unix_domain_name() (#93914) test_asyncio, test_logging, test_socket and test_socketserver now create AF_UNIX domains in the current directory to no longer fail with OSError("AF_UNIX path too long") if the temporary directory (the TMPDIR environment variable) is too long. Modify the following tests to use create_unix_domain_name(): * test_asyncio * test_logging * test_socket * test_socketserver test_asyncio.utils: remove unused time import. * gh-77782: Py_FdIsInteractive() now uses PyConfig.interactive (#93916) * gh-74953: Add _PyTime_FromMicrosecondsClamp() function (#93942) * gh-74953: Fix PyThread_acquire_lock_timed() code recomputing the timeout (#93941) Set timeout, don't create a local variable with the same name. * gh-77782: Deprecate global configuration variable (#93943) Deprecate global configuration variable like Py_IgnoreEnvironmentFlag: the Py_InitializeFromConfig() API should be instead. Fix declaration of Py_GETENV(): use PyAPI_FUNC(), not PyAPI_DATA(). * gh-93911: Specialize `LOAD_ATTR_PROPERTY` (GH-93912) * gh-92888: Fix memoryview bad `__index__` use after free (GH-92946) Co-authored-by: chilaxan <[email protected]> Co-authored-by: Serhiy Storchaka <[email protected]> * GH-89858: Fix test_embed for out-of-tree builds (GH-93465) * gh-92611: Add details on replacements for cgi utility funcs (GH-92792) Per @brettcannon 's [suggestions on the Discourse thread](https://discuss.python.org/t/pep-594-take-2-removing-dead-batteries-from-the-standard-library/13508/51), discussed in #92611 and as a followup to PR #92612 , this PR add additional specific per-function replacement information for the utility functions in the `cgi` module deprecated by PEP 594 (PEP-594). @brettcannon , should this be backported (without the `deprecated-removed` , which I would update it accordingly and re-add in my other PR adding that to the others for 3.11+), or just go in 3.11+? * GH-77403: Fix tests which fail when PYTHONUSERBASE is not normalized (GH-93917) * gh-91387: Strip trailing slash from tarfile longname directories (GH-32423) Co-authored-by: Brett Cannon <[email protected]> * Add jaraco as primary owner of importlib.metadata and importlib.resources. (#93960) * Add jaraco as primary owner of importlib.metadata and importlib.resources. * Align indentation. Co-authored-by: Ezio Melotti <[email protected]> Co-authored-by: Ezio Melotti <[email protected]> * gh-84461: Fix circulare dependency on BUILDPYTHON (GH-93977) * gh-89828: Do not relay the __class__ attribute in GenericAlias (#93754) list[int].__class__ returned type, and isinstance(list[int], type) returned True. It caused numerous problems in code that checks isinstance(x, type). * gh-84461: Fix pydebug Emscripten browser builds (GH-93982) wasm_assets script did not take the ABIFLAG flag of sysconfigdata into account. * gh-93955: Use unbound methods for slot `__getattr__` and `__getattribute__` (GH-93956) * gh-91387: Fix tarfile test on WASI (GH-93984) WASI's rmdir() syscall does not like the trailing slash. * gh-93975: Nicer error reporting in test_venv (GH-93959) - gh-93957: Provide nicer error reporting from subprocesses in test_venv.EnsurePipTest.test_with_pip. - Update changelog This change does three things: 1. Extract a function for trapping output in subprocesses. 2. Emit both stdout and stderr when encountering an error. 3. Apply the change to `ensurepip._uninstall` check. * GH-93990: fix refcounting bug in `add_subclass` in `typeobject.c` (GH-93989) * What's new in 3.10: fix link to issue (#93968) * What's new in 3.10: fix link to issue * What's new in 3.10: fix link to GH issue Co-authored-by: Ezio Melotti <[email protected]> Co-authored-by: Ezio Melotti <[email protected]> * gh-93761: Fix test_logging test_config_queue_handler() race condition (#93952) Fix a race condition in test_config_queue_handler() of test_logging. * gh-74953: Reformat PyThread_acquire_lock_timed() (#93947) Reformat the pthread implementation of PyThread_acquire_lock_timed() using a mutex and a conditioinal variable. * Add goto to avoid multiple indentation levels and exit quickly * Use "while(1)" and make the control flow more obvious. * PEP 7: Add braces around if blocks. * gh-93937, C API: Move PyFrame_GetBack() to Python.h (#93938) Move the follow functions and type from frameobject.h to pyframe.h, so the standard <Python.h> provide frame getter functions: * PyFrame_Check() * PyFrame_GetBack() * PyFrame_GetBuiltins() * PyFrame_GetGenerator() * PyFrame_GetGlobals() * PyFrame_GetLasti() * PyFrame_GetLocals() * PyFrame_Type Remove #include "frameobject.h" from many C files. It's no longer needed. * gh-93991: Use boolean instead of 0/1 for condition check (GH-93992) # gh-93991: Use boolean instead of 0/1 for condition check * gh-84461: Fix Emscripten umask and permission issues (GH-94002) - Emscripten's default umask is too strict, see emscripten-core/emscripten#17269 - getuid/getgid and geteuid/getegid are stubs that always return 0 (root). Disable effective uid/gid syscalls and fix tests that use chmod() current user. - Cannot drop X bit from directory. * gh-84461: Skip test_unwritable_directory again on Emscripten (GH-94007) GH-93992 removed geteuid() and enabled the test again on Emscripten. * gh-93925: Improve clarity of sqlite3 commit/rollback, and close docs (#93926) Co-authored-by: CAM Gerlach <[email protected]> * gh-61162: Clarify sqlite3 connection context manager docs (GH-93890) Explicitly note that transactions are only closed if there is an open transation at `__exit__`, and that transactions are not implicitly opened during `__enter__`. Co-authored-by: CAM Gerlach <[email protected]> Co-authored-by: Stanley <[email protected]> Automerge-Triggered-By: GH:erlend-aasland * gh-79009: sqlite3.iterdump now correctly handles tables with autoincrement (#9621) Co-authored-by: Erlend E. Aasland <[email protected]> * gh-84461: Silence some compiler warnings on WASM (GH-93978) * GH-93897: Store frame size in code object and de-opt if insufficient space on thread frame stack. (GH-93908) * GH-93516: Speedup line number checks when tracing. (GH-93763) * Use a lookup table to reduce overhead of getting line numbers during tracing. * gh-90539: doc: Expand on what should not go into CFLAGS, LDFLAGS (#92754) * gh-87347: Add parenthesis around macro arguments (#93915) Add unit test on Py_MEMBER_SIZE() and some other macros. * gh-93937: PyOS_StdioReadline() uses PyConfig.legacy_windows_stdio (#94024) On Windows, PyOS_StdioReadline() now gets PyConfig.legacy_windows_stdio from _PyOS_ReadlineTState, rather than using the deprecated global Py_LegacyWindowsStdioFlag variable. Fix also a compiler warning in Py_SetStandardStreamEncoding(). * GH-93249: relax overly strict assertion on bounds->ar_start (GH-93961) * gh-94021: Address unreachable code warning in specialize code (GH-94022) * GH-93678: refactor compiler so that optimizer does not need the assembler and compiler structs (GH-93842) * gh-93839: Move Lib/ctypes/test/ to Lib/test/test_ctypes/ (#94041) * Move Lib/ctypes/test/ to Lib/test/test_ctypes/ * Remove Lib/test/test_ctypes.py * Update imports and build system. * gh-93839: Move Lib/unttest/test/ to Lib/test/test_unittest/ (#94043) * Move Lib/unittest/test/ to Lib/test/test_unittest/ * Remove Lib/test/test_unittest.py * Replace unittest.test with test.test_unittest * Remove unittest.load_tests() * Rewrite unittest __init__.py and __main__.py * Update build system, CODEOWNERS, and wasm_assets.py * GH-91432: Specialize FOR_ITER (GH-91713) * Adds FOR_ITER_LIST and FOR_ITER_RANGE specializations. * Adds _PyLong_AssignValue() internal function to avoid temporary boxing of ints. * gh-94028: Clear and reset sqlite3 statements properly in cursor iternext (GH-94042) * gh-94052: Don't re-run failed tests with --python option (#94054) * gh-93839: Use load_package_tests() for testmock (GH-94055) Fixes failing tests on WebAssembly platforms. Automerge-Triggered-By: GH:tiran * gh-54781: Move Lib/lib2to3/tests/ to Lib/test/test_lib2to3/ (#94049) * Move Lib/lib2to3/tests/ to Lib/test/test_lib2to3/. * Remove Lib/test/test_lib2to3.py. * Update imports. * all_project_files(): use different paths and sort files to make the tests more reproducible. * Update references to tests. * gh-74953: _PyThread_cond_after() uses _PyTime_t (#94056) pthread _PyThread_cond_after() implementation now uses the _PyTime_t type to handle properly overflow: clamp to the maximum value. Remove MICROSECONDS_TO_TIMESPEC() function. * GH-93841: Allow stats to be turned on and off, cleared and dumped at runtime. (GH-93843) * gh-86986: Drop compatibility support for Sphinx 2 (GH-93737) * Revert "bpo-42843: Keep Sphinx 1.8 and Sphinx 2 compatibility (GH-24282)" This reverts commit 5c1f15b * Revert "bpo-42579: Make workaround for various versions of Sphinx more robust (GH-23662)" This reverts commit b63a620. * gh-94068: Remove HVSOCKET_CONTAINER_PASSTHRU constant because it has been removed from Windows (GH-94069) Fixes #94068 Automerge-Triggered-By: GH:zware * Closes gh-94038: Update Release Schedule in README.rst from PEP 664 to PEP 693 (GH-94046) * gh-93851: Fix all broken links in Doc/ (GH-93853) * gh-93675: Fix typos in `Doc/` (GH-93676) Closes #93675 * Minor optimization for Fractions.limit_denominator (GH-93730) When we construct the upper and lower candidates in limit_denominator, the numerator and denominator are already relatively prime (and the denominator positive) by construction, so there's no need to go through the usual normalisation in the constructor. This saves a couple of potentially expensive gcd calls. Suggested by Michael Scott Asato Cuthbert in GH-93477. * gh-93240: clarify wording in IO tutorial (GH-93276) Co-authored-by: Adam Turner <[email protected]> * Tutorial: specify match cases don't fall through (GH-93615) * gh-93021: Fix __text_signature__ for __get__ (GH-93023) Because of the way wrap_descr_get is written, the second argument to __get__ methods implemented through the wrapper is always optional. * gh-82927: Update files related to HTML entities. (GH-92504) * DOC: correct bytesarray -> bytearray in comments (GH-92410) * gh-87389: Fix an open redirection vulnerability in http.server. (#93879) Fix an open redirection vulnerability in the `http.server` module when an URI path starts with `//` that could produce a 301 Location header with a misleading target. Vulnerability discovered, and logic fix proposed, by Hamza Avvan (@hamzaavvan). Test and comments authored by Gregory P. Smith [Google]. * gh-89336: Remove configparser APIs that were deprecated for 3.12 (#92503) https://github.com/python/cpython/issue/89336: Remove configparser 3.12 deprecations. Co-authored-by: Hugo van Kemenade <[email protected]> * bpo-30535: [doc] state that sys.meta_path is not empty by default (GH-94098) Co-authored-by: Windson yang <[email protected]> * gh-88123: Implement new Enum __contains__ (GH-93298) Co-authored-by: Ethan Furman <[email protected]> * Stats: Add summary of top instructions for misses and deferred specialization. (GH-94072) * gh-74696: Do not change the current working directory in shutil.make_archive() if possible (GH-93160) It is no longer changed when create a zip or tar archive. It is still changed for custom archivers registered with shutil.register_archive_format() if root_dir is not None. Co-authored-by: Éric <[email protected]> Co-authored-by: Łukasz Langa <[email protected]> * gh-94101 Disallow instantiation of SSLSession objects (GH-94102) Fixes #94101 Automerge-Triggered-By: GH:tiran * Fix typo in _io.TextIOWrapper Clinic input (#94037) Co-authored-by: Łukasz Langa <[email protected]> * gh-93951: In test_bdb.StateTestCase.test_skip, avoid including auxiliary importers. (GH-93962) Co-authored-by: Brett Cannon <[email protected]> * gh-91172: Create a workflow for verifying bundled pip and setuptools (GH-31885) Co-authored-by: Hugo van Kemenade <[email protected]> Co-authored-by: Adam Turner <[email protected]> * gh-94114: Remove obsolete reference to python.org mirrors (GH-94115) * gh-94114 * gh-84623: Remove unused imports (#94132) * gh-54781: Move Lib/tkinter/test/test_ttk/ to Lib/test/test_ttk/ (#94070) * Move Lib/tkinter/test/test_tkinter/ to Lib/test/test_tkinter/. * Move Lib/tkinter/test/test_ttk/ to Lib/test/test_ttk/. * Add Lib/test/test_ttk/__init__.py based on test_ttk_guionly.py. * Add Lib/test/test_tkinter/__init__.py * Remove old Lib/test/test_tk.py. * Remove old Lib/test/test_ttk_guionly.py. * Add __main__ sub-modules. * Update imports and update references to rename files. * gh-84623: Move imports in doctests (#94133) Move imports in doctests to prevent false alarms in pyflakes. * Add ABI dump Makefile target (#94136) * gh-84623: Remove unused imports in idlelib (#94143) Remove commented code in test_debugger_r.py. Co-authored-by: Terry Jan Reedy <[email protected]> * gh-85308: argparse: Use filesystem encoding for arguments file (GH-93277) * Closes gh-94152: Update pyvideo.org URL (GH-94075) The URL is now https://pyvideo.org, which uses HTTPS and avoids a redirect. * gh-91456: [Enum] Deprecate default auto() behavior with mixed value types (GH-91457) When used with plain Enum, auto() returns the last numeric value assigned, skipping any incompatible member values (such as strings); starting in 3.13 the default auto() for plain Enums will require all the values to be of compatible types, and will return a new value that is 1 higher than any existing value. Co-authored-by: Ethan Furman <[email protected]> * gh-84461: Fix test_sqlite for Emscripten/WASI (#94125) * gh-86404: [doc] Fix missing backtick and double target name. (#94120) * gh-89121: Keep the number of pending SQLite statements to a minimum (#30379) Make sure statements that have run to completion or errored are reset and cleared off the cursor for all paths in execute() and executemany(). * GH-91742: Fix pdb crash after jump (GH-94171) * [Enum] fix typo (GH-94158) * gh-92858: Improve error message for some suites with syntax error before ':' (#92894) * gh-93771: Clarify how deepfreeze.py is run (#94150) * gh-91219: Add an index_pages default list and parameter to SimpleHTTPRequestHandler (GH-31985) * Add an index_pages default list to SimpleHTTPRequestHandler and an optional constructor parameter that allows the default indexes pages list to be overridden. This makes it easy to set a new index page name without having to override send_head. * [Enum] Remove automatic docstring generation (GH-94188) * Add ABI dump script (#94135) * Add more tests for throwing into yield from (GH-94097) * gh-94169: Remove deprecated io.OpenWrapper (#94170) Remove io.OpenWrapper and _pyio.OpenWrapper, deprecated in Python 3.10: just use :func:`open` instead. The open() (io.open()) function is a built-in function. Since Python 3.10, _pyio.open() is also a static method. * gh-94199: Remove ssl.RAND_pseudo_bytes() function (#94202) Remove the ssl.RAND_pseudo_bytes() function, deprecated in Python 3.6: use os.urandom() or ssl.RAND_bytes() instead. * gh-94196: Remove gzip.GzipFile.filename attribute (#94197) gzip: Remove the filename attribute of gzip.GzipFile, deprecated since Python 2.6, use the name attribute instead. In write mode, the filename attribute added '.gz' file extension if it was not present. * gh-93692: remove "build finished successfully" message from setup.py (#93693) The message was only emitted when the build succeeded _and_ there were missing modules. * gh-84461: Fix ctypes and test_ctypes on Emscripten (#94142) - c_longlong and c_longdouble need experimental WASM bigint. - Skip tests that need threading - Define ``CTYPES_MAX_ARGCOUNT`` for Emscripten. libffi-emscripten 2022-06-23 supports up to 1000 args. * gh-94205: Ensures all required DLLs are copied on Windows for underpth tests (GH-94206) * gh-84461: Build Emscripten with WASM BigInt support (#94219) * gh-94172: urllib.request avoids deprecated check_hostname (#94193) The urllib.request no longer uses the deprecated check_hostname parameter of the http.client module. Add private http.client._create_https_context() helper to http.client, used by urllib.request. Remove the now redundant check on check_hostname and verify_mode in http.client: the SSLContext.check_hostname setter already implements the check. * IDLE: replace if statement with expression (#94228) * Docs: Remove `Provides [...]` from `multiprocessing.shared_memory` description (#92761) * gh-93382: Sync up `co_code` changes with 3.11 (GH-94227) Sync up co_code changes with 3.11 commit 852b4d4. * gh-94217: Skip import tests when _testcapi is a builtin (GH-94218) * gh-85308: Add argparse tests for reading non-ASCII arguments from file (GH-94160) * bpo-46642: Explicitly disallow subclassing of instaces of TypeVar, ParamSpec, etc (GH-31148) The existing test covering this case passed only incidentally. We explicitly disallow doing this and add a proper error message. Co-authored-by: Serhiy Storchaka <[email protected]> * bpo-26253: Add compressionlevel to tarfile stream (GH-2962) `tarfile` already accepts a compressionlevel argument for creating files. This patch adds the same for stream-based tarfile usage. The default is 9, the value that was previously hard-coded. * gh-70441: Fix test_tarfile on systems w/o bz2 (gh-2962) (#94258) * gh-94199: Remove ssl.match_hostname() function (#94224) * gh-94207: Fix struct module leak (GH-94239) Make _struct.Struct a GC type This fixes a memory leak in the _struct module, where as soon as a Struct object is stored in the cache, there's a cycle from the _struct module to the cache to Struct objects to the Struct type back to the module. If _struct.Struct is not gc-tracked, that cycle is never collected. This PR makes _struct.Struct GC-tracked, and adds a regression test. * gh-94245: Test pickling and copying of typing.Tuple[()] (GH-94259) * gh-77560: Report possible errors in restoring builtins at finalization (GH-94255) Seems in the past the copy of builtins was not made in some scenarios, and the error was silenced. Write it now to stderr, so we have a chance to see it. * gh-90016: Reword sqlite3 adapter/converter docs (#93095) Also add adapters and converter recipes. Co-authored-by: CAM Gerlach <[email protected]> Co-authored-by: Alex Waygood <[email protected] * bpo-39971: Change examples to be runnable (GH-32172) * gh-70474: [doc] fix wording of GET_ANEXT doc (GH-94048) * gh-93259: Validate arg to ``Distribution.from_name``. (GH-94270) Syncs with importlib_metadata 4.12.0. Co-authored-by: Irit Katriel <[email protected]> Co-authored-by: Ulises Ojeda <[email protected]> Co-authored-by: jackh-ncl <[email protected]> Co-authored-by: Mark Dickinson <[email protected]> Co-authored-by: Colin Delahunty <[email protected]> Co-authored-by: Neil Schemenauer <[email protected]> Co-authored-by: Christian Heimes <[email protected]> Co-authored-by: Dennis Sweeney <[email protected]> Co-authored-by: Cyker Way <[email protected]> Co-authored-by: Hugo van Kemenade <[email protected]> Co-authored-by: Omer Katz <[email protected]> Co-authored-by: Stanley <[email protected]> Co-authored-by: Thomas Grainger <[email protected]> Co-authored-by: Illia Volochii <[email protected]> Co-authored-by: Erlend Egeberg Aasland <[email protected]> Co-authored-by: Alex Waygood <[email protected]> Co-authored-by: AN Long <[email protected]> Co-authored-by: Samodya Abeysiriwardane <[email protected]> Co-authored-by: Evorage <[email protected]> Co-authored-by: Davide Rizzo <[email protected]> Co-authored-by: Pascal Wittmann <[email protected]> Co-authored-by: Vinay Sajip <[email protected]> Co-authored-by: Adam Turner <[email protected]> Co-authored-by: Łukasz Langa <[email protected]> Co-authored-by: Andreas Grommek <[email protected]> Co-authored-by: Mark Shannon <[email protected]> Co-authored-by: Ken Jin <[email protected]> Co-authored-by: Adrian Garcia Badaracco <[email protected]> Co-authored-by: jacksonriley <[email protected]> Co-authored-by: Kalyan <[email protected]> Co-authored-by: Bluenix <[email protected]> Co-authored-by: Petr Viktorin <[email protected]> Co-authored-by: CAM Gerlach <[email protected]> Co-authored-by: Sebastian Berg <[email protected]> Co-authored-by: Leo Trol <[email protected]> Co-authored-by: XD Trol <[email protected]> Co-authored-by: Antoine Pitrou <[email protected]> Co-authored-by: neonene <[email protected]> Co-authored-by: Steve Dower <[email protected]> Co-authored-by: Pablo Galindo Salgado <[email protected]> Co-authored-by: Barney Gale <[email protected]> Co-authored-by: Oleg Iarygin <[email protected]> Co-authored-by: Brett Cannon <[email protected]> Co-authored-by: John Belmonte <[email protected]> Co-authored-by: Julien Palard <[email protected]> Co-authored-by: Pamela Fox <[email protected]> Co-authored-by: Dong-hee Na <[email protected]> Co-authored-by: Kumar Aditya <[email protected]> Co-authored-by: Victor Stinner <[email protected]> Co-authored-by: Sanket Shanbhag <[email protected]> Co-authored-by: Jeong YunWon <[email protected]> Co-authored-by: Steve Dower <[email protected]> Co-authored-by: samtygier <[email protected]> Co-authored-by: Ken Jin <[email protected]> Co-authored-by: Brandt Bucher <[email protected]> Co-authored-by: Gregory P. Smith <[email protected]> Co-authored-by: chilaxan <[email protected]> Co-authored-by: Serhiy Storchaka <[email protected]> Co-authored-by: Chris Fernald <[email protected]> Co-authored-by: Jason R. Coombs <[email protected]> Co-authored-by: Ezio Melotti <[email protected]> Co-authored-by: Lei Zhang <[email protected]> Co-authored-by: Erlend Egeberg Aasland <[email protected]> Co-authored-by: itssme <[email protected]> Co-authored-by: Matthias Köppe <[email protected]> Co-authored-by: MilanJuhas <[email protected]> Co-authored-by: luzpaz <[email protected]> Co-authored-by: paulreece <[email protected]> Co-authored-by: max <[email protected]> Co-authored-by: Jelle Zijlstra <[email protected]> Co-authored-by: Thomas A Caswell <[email protected]> Co-authored-by: Windson yang <[email protected]> Co-authored-by: Carl Bordum Hansen <[email protected]> Co-authored-by: Ethan Furman <[email protected]> Co-authored-by: Éric <[email protected]> Co-authored-by: chgnrdv <[email protected]> Co-authored-by: fikotta <[email protected]> Co-authored-by: partev <[email protected]> Co-authored-by: Terry Jan Reedy <[email protected]> Co-authored-by: Inada Naoki <[email protected]> Co-authored-by: Oscar R <[email protected]> Co-authored-by: wookie184 <[email protected]> Co-authored-by: Guido van Rossum <[email protected]> Co-authored-by: Myron Walker <[email protected]> Co-authored-by: Sam Ezeh <[email protected]> Co-authored-by: Ken Jin <[email protected]> Co-authored-by: Gregory Beauregard <[email protected]> Co-authored-by: Yaron de Leeuw <[email protected]> Co-authored-by: Mark Dickinson <[email protected]>
The ``assert`` is broken on big endian platforms and not present in the main branch. Drop it. Correct version would be ``_PyOpcode_Deopt[_Py_OPCODE(...)]`` instead of ``_PyOpcode_Deopt[...]``.
|
I now agree with Gregory and Alex that the performance regression is within acceptable levels. IMO we can leave this issue open but it shouldn't block the 3.11 release. |
|
The current sources seem to be significantly faster than b3. Can the line number checks be eliminated when frame->f_trace_lines is false? This wouldn't help with coverage but would help debuggers that set f_trace_lines to false when line events aren't needed. |
|
I just reran the performance for
|
|
Hmm, I may be confusing myself but I'm seeing tip take about 50% of the time of b3 when running one of our test files. I'm testing with a C trace function that doesn't do anything except return 0. Looking at the diffs since b3, I think performance partially depends on the number of code objects. |
|
On the off chance that some native flame graphs could be of help here. These are taken from running @nedbat's benchmarks for the bm_sudoku with 3.10
bm_sudoku with 3.11
|
|
Thanks for the report @P403n1x87! Unfortunately, the results you show are not very useful (is mostly what we are already getting by running But thanks a lot for the help 🙏, I am sure that if we need to measure the interaction between the benchmark python code and the interpreter we can use your tool 👍 |
iritkatriel
added
the
interpreter-core
Bug report
Coverage.py is seeing a significant increase in overhead for tracing code in 3.11 compared to 3.10: nedbat/coveragepy#1287
As an example:
(This is the output of lab/benchmark.py.)
Your environment
Linked PRs
The text was updated successfully, but these errors were encountered: