[C++][R] strptime fails to parse with %b or %B on Windows

[asfimport]

The following works on macOS

library(arrow, warn.conflicts = FALSE)

a <- Array$create("2021 September 13")
call_function("strptime", a, options = list(format = "%Y %B %d", unit = 0L))
#> Array
#> <timestamp[s]>
#> [
#>   2021-09-13 00:00:00
#> ]

but fails on Windows, with:

#> Error: Invalid: Failed to parse string: '2021 September 13' as a scalar of type timestamp[s]

I came across this while implementing the parse_date_time() bindings, when the Windows CI jobs were failing to parse date with the months expressed either as name or short name. https://github.com/apache/arrow/runs/6270529412?check_suite_focus=true#step:20:20856

Reporter: Dragoș Moldovan-Grünfeld / @dragosmg
Watchers: Rok Mihevc / @rok

Related issues:

• [C++] Strptime issues umbrella (is a child of)
• [C++][Python] strptime fails to parse with %p on Windows (is related to)

_{Note: This issue was originally created as ARROW-16443. Please see the migration documentation for further details.}

[asfimport]

Rok Mihevc / @rok:
A relevant c++ test would in scalar_string_test.cc would be:

StrptimeOptions options("%Y %B %d", TimeUnit::SECOND, /*error_is_null=*/false);
this->CheckUnary("strptime", R"(["2021 September 13"])", timestamp(TimeUnit::SECOND), R"(["2021-09-13"])", &options); 

[asfimport]

Rok Mihevc / @rok:
This could be a locale issue. We also noticed this behaviour on a windows machine with:

LC_TIME=English_United States.1252

[asfimport]

Dragoș Moldovan-Grünfeld / @dragosmg:
The test-dplyr-funcs-datetime.R file where the unit test failing on the CI mentioned in the ticket description forces LC_TIME to be "C".
Snippet:

# base::strptime() defaults to local timezone
# but arrow's strptime defaults to UTC.
# So that tests are consistent, set the local timezone to UTC
# TODO: consider reevaluating this workaround after ARROW-12980
withr::local_timezone("UTC")

if (tolower(Sys.info()[["sysname"]]) == "windows") {
  withr::local_locale(LC_TIME = "C")
}

[pearu]

PR #50146 (primarily about CSV date/time parsing) includes a fix for this: the vendored musl strptime used on Windows compiles out all name-based specifiers because nl_langinfo() is unavailable there. The PR adds static C-locale name tables, making %b/%B (and %a/%A/%h) parse on Windows, case-insensitively, matching musl's C-locale behavior on Linux. Validated on Windows CI (MinGW CLANG64/MINGW64, MSVC Python wheel, RTools, MATLAB, and ARM64 MSVC). The issue's reproducer ("2021 September 13" with "%Y %B %d") parses with that change.