`test_strftime_y2k` fails on embedded Linux

[picnixz]

Bug report

Bug description:

This is an issue to track the progress of fixing the JIT. Branches that trigger the embedded Linux tests fail on test_strftime_y2k, e.g. https://github.com/python/cpython/actions/runs/10655985897/job/29534232332?pr=123546. As a side-effect this causes users getting notification that their JIT workflow failed.

Both the C and F specifiers are failing for this specific test:

test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) ... 
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=1, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=1, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=48, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=48, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=70, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=70, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=999, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDate_Fast.test_strftime_y2k) (year=999, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) ... 
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=1, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=1, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=48, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=48, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=70, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=70, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=999, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTime_Fast.test_strftime_y2k) (year=999, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) ... 
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=1, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=1, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=48, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=48, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=70, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=70, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=999, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestDateTimeTZ_Fast.test_strftime_y2k) (year=999, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) ... 
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=1, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=1, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=48, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=48, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=70, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=70, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=99, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=99, specifier='C') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=999, specifier='F') ... FAIL
test_strftime_y2k (test.datetimetester.TestSubclassDateTime_Fast.test_strftime_y2k) (year=999, specifier='C') ... FAIL

A solution is to add this test to the ignored tests or to fix the detection algorithm.

cc @blhsing @serhiy-storchaka

CPython versions tested on:

CPython main branch

Operating systems tested on:

Other

Linked PRs

• gh-123861
• gh-128444
• gh-136022

Related and proposed PR: gh-128106

[serhiy-storchaka]

It seems that only tests for the C implementation fail.

We can try to move the detection from the compile time to the run time, like for Python implementation. But it is a puzzle why it does not work here.

[picnixz]

(Forget my comment that I deleted, I looked at the wrong commit).

I think we can improve the configure.ac detection by including the weird cases. Or maybe the test is broken because of this:

if _time.strftime('%F', (1900, 1, 1, 0, 0, 0, 0, 1, 0)) == '1900-01-01':
    specifiers += 'FC'

(Those are the two failing specifiers)

[brandtbucher]

This particular test failure could be a bit weird, since this CI job involves cross-compiling Python and then running the test suite under emulation. Not sure if it's a "real" issue, or if we should just skip the test (like we do for others).

[picnixz]

since this CI job involves cross-compiling Python and then running the test suite under emulation

Does it mean that the configure tests might not be correct (or inconsistent because they were done using cross-compilation?) If so, we can just outright ignore the test I guess?

[serhiy-storchaka]

If these configure tests are not correct, then other configure tests can be not correct too.

[picnixz]

Ha! I knew I should have looked more in detail the logs:

On aarch64-unknown-linux-gnu (failing tests), we have two different configure:

https://github.com/python/cpython/commit/e565b87d9dafe89bf262e3ab46edb9c9c8618b27/checks/29534232332/logs

2024-09-01T17:33:50.4537429Z checking whether year with century should be normalized for strftime... yes 2024-09-01T17:33:50.4888417Z checking whether C99-specific strftime specifiers are supported... yes

2024-09-01T17:35:47.4696152Z checking whether year with century should be normalized for strftime... yes 2024-09-01T17:35:47.4698164Z checking whether C99-specific strftime specifiers are supported... >> no <<

So it appears we have different ./configure. I don't know which one is actually being used at runtime in the emulated environment. Now, both '%F' and '%C' are C99-specific (https://en.cppreference.com/w/c/chrono/strftime) so the following

if _time.strftime('%F', (1900, 1, 1, 0, 0, 0, 0, 1, 0)) == '1900-01-01':
    specifiers += 'FC'

is perhaps a too broad check. Since specifiers 'F' and 'C' are used, we actually expect to support %F, but this contradicts the fact that "C99-specific strftime specifiers are supported" (at least for the second one which seems the most recent and probably the one being used under the hood). The configure check is:

if (strftime(full_date, sizeof(full_date), "%F", &date) && !strcmp(full_date, "1900-01-01")) {
    return 0;
}

In other words, either strftime(full_date, sizeof(full_date), "%F", &date) fails or full_date != 1900-01-01. So maybe %F does not fail. We can probably debug the tests to check whether it _time.strftime that bugs, or if it is the configure test that bugs. I don't know if it's possible for _time.strftime to support %F (which does not directly call strftime()) but at the same time fail the strftime(full_date, sizeof(full_date), "%F", &date) && !strcmp(full_date, "1900-01-01") test.

---

I'm tired so what I wrote here might be wrong but I'll try to investigate tomorrow.

[blhsing]

2024-09-08T16:49:27.5943515Z checking whether year with century should be normalized for strftime... yes 2024-09-08T16:49:27.6286869Z checking whether C99-specific strftime specifiers are supported... yes

2024-09-08T16:51:33.7240722Z checking whether year with century should be normalized for strftime... yes 2024-09-08T16:51:33.7242911Z checking whether C99-specific strftime specifiers are supported... >> no <<

Ahh how could I have not noticed the second configure in the logs! Was scratching my head over why the detection of C99 support got a yes (in the first configure) but the compiled code ended up not handling C99. It makes total sense now.

The real problem is simply that autoconf does not run the test for the AC_RUN_IFELSE macro at all when cross-compiling, but directly performs action-if-cross-compiling as specified, which I thought should be made a no (to be pessimistic) when it should really be a yes because we should make sure to handle %F and %C ourselves when we are unsure if the target platform can handle them.

[picnixz]

Ahh how could I have not noticed the second configure in the logs!

Well, isn't it trivial to spot it among 11985 lines 😄? Actually, I found the duplicate configure in the logs by a CTRL+F and... saw that I had two results.

[blhsing]

Ahh how could I have not noticed the second configure in the logs!

Well, isn't it trivial to spot it among 11985 lines 😄? Actually, I found the duplicate configure in the logs by a CTRL+F and... saw that I had two results.

I used Ctrl-F too, but the browser limits the search to only what's loaded in the memory, when the output of the second configure is hidden because it is loaded on demand by an AJAX call only if I scroll to it. I should've used the "Search Logs" text box in the page to perform the search for the keyword on the server side instead.

Thanks again!

[picnixz]

I used Ctrl-F too, but the browser limits the search to only what's loaded in the memory, when the output of the second configure is hidden because it is loaded on demand by an AJAX call only if I scroll to it. I should've used the "Search Logs" text box in the page to perform the search for the keyword on the server side instead.

Actually, the trick is to "view raw logs". You'll get the entire logs without buffering (easier to read IMO).

[picnixz]

FTR: https://github.com/python/cpython/pull/124466 was merged as a temporary workaround so we need to keep the issue opened even though the job looks successful (it's just that we skip it).

[picnixz]

Closing this one in favor of gh-128104 which has a better description and an alternative PR.

[serhiy-storchaka]

Why not test this at runtime instead of compile time, like in the Python implementation?

[serhiy-storchaka]

We still need to fix a check for Py_NORMALIZE_CENTURY which does not work in cross-compiler. It is only possible at run time.

If we require C99 compliant strftime (it is not on Linux for %C), we need to make the following changes:

• Remove the runtime check from the Python implementation.
• Update the documentation (it still refers to C89). Document all C99 codes, including the E and O modifiers.

[erlend-aasland]

it is not on Linux for %C

%C works swell for me on Debian and Ubuntu. Please specify where it does not work instead of putting out such a broad claim. We also explicitly test the %C specifier in our test suite. If it did not work on Linux, CI would be broken.

[serhiy-storchaka]

>>> time.strftime("%C", (99, 1, 1, 0, 0, 0, 0, 1, 0))
'0'

In C99 it should be 2 digits.

[encukou]

So, we need a test for normalize_century (ideally at run-time, to avoid cross-compilation issues), but in CPython we can assume %C and %F are available, right?

[serhiy-storchaka]

They are in C99. https://en.cppreference.com/w/c/chrono/strftime

It seems that the main reason of adding checks for them (see #122272) is that they are not documented in Python, and therefore their support is not guaranteed. If we guarantee their support, we should also document them and add tests. We can make this a new feature in 3.15 or even in 3.14, but I have doubts about 3.13.

[encukou]

As I see it, they're guaranteed in CPython (via the C99 build requirement), but not in Python the language.

[serhiy-storchaka]

This is an interesting point of view. Then we can remove the check in the C code (which is CPython specific), but should leave it in the Python code, which can be reused in other implementation. From this point of view, #128444 and #136022 are equivalent.

[encukou]

I think #136022 removing def _supports_c99 is OK. The Python code still uses _time.strftime. If other implementations replace _time, they will need to update the Python code as well. There's a test so they don't forget.