gh-132876: workaround broken ldexp() on Windows 10 #133135
Conversation
Co-authored-by: Tim Peters <tim.peters@gmail.com>
|
Some benchmarking in: #132876 (comment) |
There was a problem hiding this comment.
Since this is a potentially breaking change on Windows, it really needs a NEWS entry, yes? Like
math.ldexp() on Windows failed to round subnormal results before Window 11, and Microsoft probably won't fix it for Windows 10. So Python works around the error on Windows now. This may change results (which were formerly erroneous) in rare cases.
And maybe link to the issue report?
|
It has a NEWS entry, though it could include a bit more detail (such as the ranges of values that may be incorrect). And that entry will already be linked to the bug. So this change LGTM |
Misc/NEWS.d/next/Library/2025-04-29-11-48-46.gh-issue-132876.lyTQGZ.rst
Outdated
Show resolved
Hide resolved
Misc/NEWS.d/next/Library/2025-04-29-11-48-46.gh-issue-132876.lyTQGZ.rst
Outdated
Show resolved
Hide resolved
Misc/NEWS.d/next/Library/2025-04-29-11-48-46.gh-issue-132876.lyTQGZ.rst
Outdated
Show resolved
Hide resolved
Misc/NEWS.d/next/Library/2025-04-29-11-48-46.gh-issue-132876.lyTQGZ.rst
Outdated
Show resolved
Hide resolved
Misc/NEWS.d/next/Library/2025-04-29-11-48-46.gh-issue-132876.lyTQGZ.rst
Outdated
Show resolved
Hide resolved
|
@picnixz, I just added test unconditionally (though, require IEEE 754). |
|
@skirpichev, do you need someone else to merge this? I believe all comments have been addressed. |
Yeah, someone with commit bit :-) |
|
Ah - I didn't realize you don't have that! I'll do it now 😄 , |
|
Thanks @skirpichev for the PR, and @tim-one for merging it 🌮🎉.. I'm working now to backport this PR to: 3.13, 3.14. |
…3135) * pythongh-132876: workaround broken ldexp() on Windows 10 ldexp() fails to round subnormal results before Windows 11, so hide their bug. (cherry picked from commit cf8941c) Co-authored-by: Sergey B Kirpichev <skirpichev@gmail.com> Co-authored-by: Tim Peters <tim.peters@gmail.com>
|
Sorry, @skirpichev and @tim-one, I could not cleanly backport this to |
|
GH-134684 is a backport of this pull request to the 3.14 branch. |
|
FYI, I'm ignoring the auto-generated backport messages. Partly because I don't understand how that works 😦 . |
|
The witch failed to do her job. I'll prepare manual backport for 3.13. |
|
GH-134685 is a backport of this pull request to the 3.13 branch. |
…#134684) gh-132876: workaround broken ldexp() on Windows 10 (GH-133135) * gh-132876: workaround broken ldexp() on Windows 10 ldexp() fails to round subnormal results before Windows 11, so hide their bug. (cherry picked from commit cf8941c) Co-authored-by: Sergey B Kirpichev <skirpichev@gmail.com> Co-authored-by: Tim Peters <tim.peters@gmail.com>
|
Thanks for merging this Tim! FYI, I never merge something after 8 PM or when I'm not on Linux, just to be sure I can quickly revert if I made a mistake. |
|
FYI, I'm retired and "8pm" doesn't mean much of anything to me 😉. If I had to stay awake for another 12 hours to repair problems, that would have been OK. This PR was ready to merge for weeks already, passed exhaustive reviews by multiple capable reviewers, and is a straightforward, localized change. While a chance of unforeseen disaster is never absent, it looked as safe as a PR gets 😄 |
ldexp()on Windows doesn't round subnormal results before Windows 11, but should. Python's :func:math.ldexpwrapper now does round them.