vectorcall: fix bindings & use for call0 and call_method0 #1287
Conversation
davidhewitt
force-pushed
the
vectorcall
branch
2 times, most recently
from
84f6834 to
c85c25b
Compare
davidhewitt
force-pushed
the
vectorcall
branch
from
c85c25b to
423c575
Compare
|
Windows failing to link on abi3 - presumably it's us doing something wrong; I'm asking about it at https://bugs.python.org/issue42415 |
davidhewitt
force-pushed
the
vectorcall
branch
2 times, most recently
from
2ef2771 to
9df2e15
Compare
|
Rather than wait for that bug to be resolved (which I guess might not happen until Python 3.9.1) I've put a TODO in the code and just enabling the optimization for the unlimited API for now. |
davidhewitt
force-pushed
the
vectorcall
branch
from
9df2e15 to
e48399c
Compare
|
|
||
| #[cfg(all(Py_3_8, not(PyPy)))] | ||
| #[inline(always)] | ||
| pub unsafe fn _PyObject_FastCallTstate( |
There was a problem hiding this comment.
Where is this function? (I'm looking at https://github.com/python/cpython/blob/v3.8.6/Include/cpython/abstract.h ).
There was a problem hiding this comment.
I'm looking at the same file but the version in master (so 3.10-dev). I figured it was best for these inline functions to be as up-to-date as possible.
| // skipped _PyObject_VectorcallMethodId | ||
| // skipped _PyObject_CallMethodIdNoArgs | ||
| // skipped _PyObject_CallMethodIdOneArg | ||
|
|
||
| // skipped _PyObject_HasLen |
There was a problem hiding this comment.
Just curious, why are they skipped?
There was a problem hiding this comment.
I didn't consider them relevant to my PR; I could add them but I'd also quite like to keep as few unlimited API definitions as possible (as they're more likely for us to be work to maintain).
See also #1289 which talks a bit about what I'm thinking.
| @@ -0,0 +1,3 @@ | |||
| pub mod abstract_; | |||
There was a problem hiding this comment.
👍 for creating this new module. Maybe we can reduce cfg using this.
There was a problem hiding this comment.
Yes, I think for sure (again see #1289).
davidhewitt
force-pushed
the
vectorcall
branch
2 times, most recently
from
6d5c5cf to
4d14ac8
Compare
davidhewitt
force-pushed
the
vectorcall
branch
from
4d14ac8 to
9e99f01
Compare
davidhewitt
force-pushed
the
vectorcall
branch
from
9e99f01 to
87bacf1
Compare
|
Thanks! |
This PR fixes the broken vectorcall ffi bindings, and proves that the new ones work by using them in
.call0and.call_method0implementations (where possible).This yields quite a tidy ~20% performance speedup on those functions, because it no longer creates an empty
PyTupleandPyDicteach time:before:
after:
I'd love to follow this PR up with some ways to use vectorcall for functions with arguments too. Will leave that for #684 instead of pushing this further right now.
Closes #1283