Test failure: System.Tests.DoubleTests.AtanPiTest on Android

[karelz]

Started failing on 6/22 in every Rolling run -- 7 failures in last 7 days

• 6/29: 14 failures in last 7 days

Affected test:

• System.Tests.DoubleTests.AtanPiTest

Platform:

• net7.0-Android-Release-x86-Mono_Release-Ubuntu.1804.Amd64.Android.29.Open

Failure:

Assert.Equal() Failure
Expected:                 -0.0
Actual:                   +0.0

   at System.AssertExtensions.Equal(Double expected, Double actual, Double variance) in /_/src/libraries/Common/tests/TestUtilities/System/AssertExtensions.cs:line 747
   at System.Tests.DoubleTests.AtanPiTest(Double value, Double expectedResult, Double allowedVariance) in /_/src/libraries/System.Runtime/tests/System/DoubleTests.cs:line 1475
   at System.Reflection.MethodInvoker.InterpretedInvoke(Object obj, Span`1 args, BindingFlags invokeAttr)

[ghost]

Tagging subscribers to this area: @dotnet/area-system-runtime
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Started failing on 6/22 in every Rolling run -- 7 failures in last 7 days

Affected test:

• System.Tests.DoubleTests.AtanPiTest

Failure:

Assert.Equal() Failure
Expected:                 -0.0
Actual:                   +0.0

   at System.AssertExtensions.Equal(Double expected, Double actual, Double variance) in /_/src/libraries/Common/tests/TestUtilities/System/AssertExtensions.cs:line 747
   at System.Tests.DoubleTests.AtanPiTest(Double value, Double expectedResult, Double allowedVariance) in /_/src/libraries/System.Runtime/tests/System/DoubleTests.cs:line 1475
   at System.Reflection.MethodInvoker.InterpretedInvoke(Object obj, Span`1 args, BindingFlags invokeAttr)

Author:	karelz
Assignees:	-
Labels:	area-System.Runtime, blocking-clean-ci, os-android
Milestone:	-

[karelz]

Due to large impact on CI, we should disable the tests ASAP

[dakersnar]

Investigating this now. I am not able to reproduce it locally on a non-android device. I immediately notice that AtanPi uses the Math.Atan() intrinsic, which could have incorrect behavior on android devices. What this boils down to is that we expect Atan(-0) to produce -0, but instead it is probably producing 0. I'm unsure if there is a flaw in our logic or a flaw in the intrinsic's behavior.

[dakersnar]

Here is the current implementation of AtanPi: return Math.Atan(x) / Pi;

We have a test verifying that Math.Atan(-0) returns -0. A simple example also shows that -0 divided by Pi should return -0. Given this, I believe that there is a bug in Mono for Math.AtanPi.

I'm unsure which mono area this should fall under so I'm putting it under "meta".