NLS System.Globalization.Tests.CompareInfoTests.SortKeyTest with empty string and ZWJ fails on Windows Insiders Preview

[Gnbrkm41]

This is under the NLS configuration (System.Globalization.Nls.Tests), not sure if it's specific to the version of NLS that ships with the insider builds. Currently I'm running 10 Pro Insiders Build (Dev Channel) Build 20211, ko-KR.

runtime/src/libraries/System.Globalization/tests/CompareInfo/CompareInfoTests.cs

Lines 306 to 310 in 1b18efa

	// Zero-weight code points
	// In both NLS (Windows) and ICU the code point U+200C ZERO WIDTH NON-JOINER has a zero weight,
	// so it's compared as equal to the empty string. This means that we can't special-case GetHashCode("")
	// and return a fixed value; we actually need to call the underlying OS or ICU API to calculate the sort key.
	yield return new object[] { s_invariantCompare, "", "\u200c", CompareOptions.None, 0 };

Raw XML output:

<test name="System.Globalization.Tests.CompareInfoTests.SortKeyTest(compareInfo: CompareInfo - , string1: \&quot;\&quot;, string2: \&quot;‌\&quot;, options: None, expectedSign: 0)" type="System.Globalization.Tests.CompareInfoTests" method="SortKeyTest" time="0.0054359" result="Fail">
  <failure exception-type="Xunit.Sdk.EqualException">
    <message><![CDATA[Assert.Equal() Failure\r\nExpected: -1\r\nActual:   0]]></message>
    <stack-trace><![CDATA[   at System.Globalization.Tests.CompareInfoTests.SortKeyTest(CompareInfo compareInfo, String string1, String string2, CompareOptions options, Int32 expectedSign) in C:\Users\gotos\source\repos\runtime\src\libraries\System.Globalization\tests\CompareInfo\CompareInfoTests.cs:line 371]]></stack-trace>
  </failure>
</test>

cc @tarekgh

[ghost]

Tagging subscribers to this area: @tarekgh, @safern, @krwq
See info in area-owners.md if you want to be subscribed.

[safern]

It seems like what is being affected and returning a different value now is: Math.Sign(compareInfo.Compare(string1, string2, options))

@tarekgh could it be a bug introduced in Windows where "\u200c is not having a zero weight?

[tarekgh]

@Gnbrkm41 could you please try that with en-US instead of invariant? I am seeing many changes in Windows side for Invariant culture and this can be one of them.

[tarekgh]

I have confirmed this is a regression in Windows (repro on version 20201 too). I am talking to Windows team about it. but it looks we need to workaround it in our tests for now.