Repository navigation
New tier 2 counters break some C extensions due to order of field mismatch #117549
Description
Activity
Oh, I think I see what it is. I swapped the order in which the fields
valueandbackoffare declared, and they have a compiler (flag?) that insists that the initializer lists the fields in the same order. @mdboom are you planning to make a PR to fix this? I can do it too (it's raining here :-).Yeah, I'm planning on making a PR. I think that's all it is ... though I think it's not a compiler flag so much as they are compiling C++ which is maybe strict about these things.
Reacted by Guido van Rossum- added a commit that references this issue
on Apr 4, 2024 Apparently, designated initializers are a C++20 feature, i.e. they require a really recent C++ compiler with appropriate configuration. I see this in MSVC with Py3.13a6, so I don't think it's fixed:
C:\...\Python\3.13.0-alpha.6\x64\include\internal\pycore_backoff.h(47): error C7555: use of designated initializers requires at least '/std:c++20' C:\...\Python\3.13.0-alpha.6\x64\include\internal\pycore_backoff.h(47): error C4576: a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntaxThe reason why this appears on user side is probably that Cython needs to include
pycore_frame.hin Py3.11+ because some parts of frames that it needs for its coroutine implementation were moved fromframeobject.hinto internal header files at the time.The discussion that lead to this change is here:
#90992So if I changed pycore_backoff.h to use something like this, the problem will be solved?
diff --git a/Include/internal/pycore_backoff.h b/Include/internal/pycore_backoff.h index decf92bc419..50e88487a52 100644 --- a/Include/internal/pycore_backoff.h +++ b/Include/internal/pycore_backoff.h @@ -44,7 +44,10 @@ make_backoff_counter(uint16_t value, uint16_t backoff) { assert(backoff <= 15); assert(value <= 0xFFF); - return (_Py_BackoffCounter){.backoff = backoff, .value = value}; + _Py_BackoffCounter counter; + counter.value = value; + counter.backoff = backoff; + return counter; } static inline _Py_BackoffCounter
I know there's another place this is used in that header, I'll fix that too. I'll whip up a PR.
So if I changed pycore_backoff.h to use something like this, the problem will be solved?
Probably, yes. Looks good to me. I'm not sure how to reproduce the issue locally, though, because GCC seems to be fairly relaxed about non-standard features whereas MSVC has the tendency to simply not support them, and I don't have MSVC on my side.
Okay, I've asked you to review gh-118580 anyways, since we're so close to the beta1 release date. Hopefully it'll go in over the weekend and somebody else reading this can test with a nightly build or something. Otherwise we can of course revise this after beta1.
The fix has been merged, let's keep this issue open for a bit to wait for confirmation that this fixes things.
Closing now. If you find a problem with this in beta 1, leave a comment here and I'll reopen.
Bug report
Bug description:
@tacaswell reported here:
#177144 appears to have broken building scipy
conformed scipy builds with 63bbe77
CPython versions tested on:
CPython main branch
Operating systems tested on:
Linux
Linked PRs