bpo-44525: Specialize simple Python calls. #29033
Conversation
…onsume the argument references regardless of whether they succeed or fail.
| else { | ||
| SPECIALIZATION_FAIL(CALL_FUNCTION, call_fail_kind(callable)); | ||
| fail = 1; | ||
| fail = -1; |
There was a problem hiding this comment.
I set specialize_c_call to return 1 to indicate failure. You may have to change the return codes in specialize_c_call for consistency. Sorry!
There was a problem hiding this comment.
We should probably add an enum {SUCCESS, FAILURE} for clarity.
I wouldn't worry too much about it for now, we do much worse elsewhere.
compile.c uses zero for success and non-zero for failure in some places, and the other way around in other places 😞
markshannon
changed the title
markshannon
added
the
🔨 test-with-buildbots
|
🤖 New build scheduled with the buildbot fleet by @markshannon for commit 6f1bfab 🤖 If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again. |
bedevere-bot
removed
the
🔨 test-with-buildbots
|
It would be interesting to see the combined effects of Some extra food for thought: I wonder if we can remove the branch and inline python calling code from the generic |
| } | ||
| int deflen = cache1->defaults_len; | ||
| for (int i = 0; i < deflen; i++) { | ||
| PyObject *def = PyTuple_GET_ITEM(func->func_defaults, cache1->defaults_start+i); |
There was a problem hiding this comment.
I wonder if the compiler is smart enough to realize that func->func_defaults and cache1->defaults_start doesn't change between iterations? I'd say probably not. Perhaps we should micro-optimize and move this out of the loop?
There was a problem hiding this comment.
The common case is that deflen == 0. In which case moving it out of loop could make it worse.
Indeed, I was wondering what this PR + #26934 would yield, instead of just this PR alone. Nonetheless, this should pay for the specialization alone. |
|
|
Great! Thank you for going through the trouble to benchmark. Other than the odd |
|
The usual buildbot failures. |
* main: (263 commits) bpo-45521: Fix a bug in the obmalloc radix tree code. (pythonGH-29051) bpo-45522: Allow to disable freelists on build time (pythonGH-29056) bpo-34451: Document prompt and output toggle feature in html tutorial (pythonGH-27105) bpo-44019: Add operator.call() to __all__ for the operator module (pythonGH-29110) bpo-45315: PyType_FromSpec: Copy spec->name and have the type own the memory for its name (pythonGH-29103) bpo-44220: Export PyStructSequence_UnnamedField in the limited API (pythonGH-26331) bpo-44174: [Enum] add reference to name mangling (pythonGH-29116) bpo-45548: add some missing entries to `Modules/Setup` (pythonGH-29115) bpo-35673: Add a public alias for namespace package __loader__ attribute (python#29049) bpo-45192: Fix a bug that infers the type of an os.PathLike[bytes] object as str (pythonGH-28323) bpo-45527: Don't count cache hits, just misses. (pythonGH-29092) bpo-45320: Remove long-deprecated inspect methods (pythonGH-28618) bpo-41374: Remove obsolete exclusion of netinet/tcp.h on Cygwin (pythonGH-21649) bpo-45532: Replace 'default' with 'main' as default in sys.version (pythonGH-29100) bpo-45464: [doc] Explain that subclassing multiple exceptions is fragile (pythonGH-29094) Cleanup a couple of comments left on PR 28775 post-merge. (pythonGH-29079) bpo-45536: Check OpenSSL APIs in configure (pythonGH-29088) Add PEPs 593 & 647 to list of PEPs at top of typing docs (pythonGH-29097) Add a comment about how to fix bogus test_host_resolution_bad_address failures (python#29085) bpo-44525: Specialize simple Python calls. (pythonGH-29033) ...
Complement to #26934
About 1% speedup
https://bugs.python.org/issue44525