ExplicitConversion_FromSingle failing due to NaN != NaN

[stephentoub]

Build Information

Build: https://dev.azure.com/dnceng-public/public/_build/results?buildId=768837&view=results
Build error leg or test failing: System.Tests.HalfTests.ExplicitConversion_FromSingle
Pull request:

Error Message

Fill the error message using step by step known issues guidance.

{
  "ErrorMessage": "",
  "ErrorPattern": "FAIL.*System.Tests.HalfTests.ExplicitConversion_FromSingle",
  "BuildRetry": false,
  "ExcludeConsoleLog": false
}

Known issue validation

Build: 🔎 https://dev.azure.com/dnceng-public/public/_build/results?buildId=768837
Error message validated: [FAIL.*System.Tests.HalfTests.ExplicitConversion_FromSingle]
Result validation: ✅ Known issue matched with the provided build.
Validation performed at: 8/8/2024 7:39:18 PM UTC

[14:40:21] info: [FAIL] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[14:40:21] info: Assert.Equal() Failure: Values differ
[14:40:21] info: Expected:   NaN
[14:40:21] info: Actual:     NaN
[14:40:21] info:    at System.AssertExtensions.Equal(Half expected, Half actual)
[14:40:21] info:    at System.Tests.HalfTests.ExplicitConversion_FromSingle(Single f, Half expected)
[14:40:21] info:    at System.Object.InvokeStub_HalfTests.ExplicitConversion_FromSingle(Object , Span`1 )
[14:40:21] info:    at System.Reflection.MethodBaseInvoker.InvokeWithFewArgs(Object obj, BindingFlags invokeAttr, Binder binder, Object[] parameters, CultureInfo culture)
[14:40:21] info: [FAIL] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[14:40:21] info: Assert.Equal() Failure: Values differ
[14:40:21] info: Expected:   NaN
[14:40:21] info: Actual:     NaN
[14:40:21] info:    at System.AssertExtensions.Equal(Half expected, Half actual)
[14:40:21] info:    at System.Tests.HalfTests.ExplicitConversion_FromSingle(Single f, Half expected)
[14:40:21] info:    at System.Object.InvokeStub_HalfTests.ExplicitConversion_FromSingle(Object , Span`1 )
[14:40:21] info:    at System.Reflection.MethodBaseInvoker.InvokeWithFewArgs(Object obj, BindingFlags invokeAttr, Binder binder, Object[] parameters, CultureInfo culture)

Report

Build	Definition	Test	Pull Request
868508	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#109602
865074	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#107230
864857	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#109645
862539	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#109501
858442	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#109261
855817	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#108992
854066	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution
853397	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#109229
843247	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#108788
842800	dotnet/runtime	WasmTestOnChrome-ST-System.Runtime.Tests.WorkItemExecution	#105206

Summary

24-Hour Hit Count	7-Day Hit Count	1-Month Count
1	3	10

[dotnet-policy-service]

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

[dotnet-policy-service]

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

[ericstj]

This issue is matching to tests jobs which log test names. Need to further constrain the match.

[mdh1418]

Just hit this in #106040 for WasmTestOnChrome-ST-System.Runtime.Tests
config: net9.0-browser-Release-wasm-Mono_Release-WasmTestOnChrome

[18:27:21] info: [STRT] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[18:27:21] info: [FAIL] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[18:27:21] info: Assert.Equal() Failure: Values differ
[18:27:21] info: Expected:   NaN
[18:27:21] info: Actual:     NaN
[18:27:21] info:    at System.AssertExtensions.Equal(Half expected, Half actual)
[18:27:21] info:    at System.Tests.HalfTests.ExplicitConversion_FromSingle(Single f, Half expected)
[18:27:21] info:    at System.Object.InvokeStub_HalfTests.ExplicitConversion_FromSingle(Object , Span`1 )
[18:27:21] info:    at System.Reflection.MethodBaseInvoker.InvokeWithFewArgs(Object obj, BindingFlags invokeAttr, Binder binder, Object[] parameters, CultureInfo culture)
[18:27:21] info: [STRT] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[18:27:21] info: [FAIL] System.Tests.HalfTests.ExplicitConversion_FromSingle(f: NaN, expected: NaN)
[18:27:21] info: Assert.Equal() Failure: Values differ
[18:27:21] info: Expected:   NaN
[18:27:21] info: Actual:     NaN
[18:27:21] info:    at System.AssertExtensions.Equal(Half expected, Half actual)
[18:27:21] info:    at System.Tests.HalfTests.ExplicitConversion_FromSingle(Single f, Half expected)
[18:27:21] info:    at System.Object.InvokeStub_HalfTests.ExplicitConversion_FromSingle(Object , Span`1 )
[18:27:21] info:    at System.Reflection.MethodBaseInvoker.InvokeWithFewArgs(Object obj, BindingFlags invokeAttr, Binder binder, Object[] parameters, CultureInfo culture)

Not sure why it didn't match in build analysis

[tannergooding]

The test here should likely be filtered out on WASM for the time being, with a tracking issue against it.

WASM isn't technically doing anything wrong here, normalizing NaN is fully allowed by the IEEE 754 spec. However, it is undesirable and not recommended for most cases and should typically be avoided if possible.

WASM also has instructions that should guarantee the underlying bits are preserved, so it should be possible for it to be preserved and match the behavior of other targets (both for RyuJIT and Mono).

[akoeplinger]

Hm the weird thing here is that this seems to be an intermittent failure. I've checked and we definitely have passing runs of the test on the exact same config.

[tannergooding]

Is there perhaps a difference in the WASM version or configuration options for the failing platform as compared to others?

Maybe some machine specific issue that's only triggering in some scenarios?

[akoeplinger]

the tests are running in containers on the same helix queue so even the VM/machine should be the same...

Passing run: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-106013-merge-6f4bade3cf974edcb8/WasmTestOnChrome-ST-System.Runtime.Tests/1/console.3fb2056a.log?helixlogtype=result

Failing run: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-106040-merge-0b77eabb49f848cab5/WasmTestOnChrome-ST-System.Runtime.Tests/1/console.5edeb661.log?helixlogtype=result

The only difference is the order of the tests, but I'm having a hard time imagining how that could be relevant

[akoeplinger]

hm one interesting thing is that the failing run reports two tests that failed i.e. five total, but the passing run only ran four NaN tests... that seems just a log quirk, the testResult.xml only contains four NaN tests and two of them failing.

[akoeplinger]

According to Kusto test data this started happening on or before 2024-06-06.
I also found that it's happening on Windows-based containers too so it's not related to the underlying OS.

Given it's only happening intermittently and doesn't appear to be a blocker for 9.0 I'm moving it to 10.0

[lambdageek]

Couple of things could be responsible which might be explained by the test order:

1. jiterp might be kicking in
2. browser JIT might be kicking in

But I'd like to push back against hte idea that bitwise comparison of two NaNs is a good idea. there are many NaN bitpatterns; all are semantically indistinguishable. Reasonable code should not depend on the bit pattern staying identical after computation

[tannergooding]

But I'd like to push back against hte idea that bitwise comparison of two NaNs is a good idea. there are many NaN bitpatterns; all are semantically indistinguishable. Reasonable code should not depend on the bit pattern staying identical after computation

This is a core thing frequently depended upon for SIMD, NaN boxing, etc. It also follows the IEEE 754 "recommended" guidelines, even if it's not strictly required by the IEEE 754 spec.

The WASM 1.0 and 2.0 specs similarly:

• distinguish canonical NaN from arithmetic NaN (a nan with payload n)
• provide explicit APIs for getting the bits of any NaN
  • fbits(+/-nan(n)) is defined as fsign(+/-) * 1^expon(n) * ibits(sinif(n))
  • which is to say, its doing BitConverter.SingleToInt32Bits or BitConverter.DoubleToInt64Bits or vice-versa
• recommend that operators propagate NaN payloads (matching the IEEE 754 recommendation)
• provides explicit support for specifying the payload of a NaN

However, the WASM specs do notably explicitly allow non-determinism for fneg, fabs, and fcopysign; explicitly specifying that the sign and payload of the result are non-deterministic

So while a given implementation is technically allowed to do "other things" (such as always canonicalizing NaN), its highly irregular to do so and that's why we have tests that validate such conversions are preserving bits and propagating payloads as expected. -- It's also notably cheaper to propagate and to preserve NaN as is then it is to canonicalize (particularly for things like fbits), so there's really no reason for a browser to deviate here

[pavelsavara]

@tannergooding @lambdageek what are the next steps here ?

[carlossanlop]

Should the milestone be 10? This is hitting 9.0 PRs too. Example:

• 9.0 PR: [release/9.0] Update dependencies from dotnet/roslyn #108326
• Job: https://dev.azure.com/dnceng-public/public/_build/results?buildId=839877&view=logs&j=d4e38924-13a0-58bd-9074-6a4810543e7c&t=7a581d54-fbdb-5b9d-4377-cc346f489112&l=104
• Log: https://helixre107v0xd1eu3ibi6ka.blob.core.windows.net/dotnet-runtime-refs-pull-108326-merge-06a2d595e69e44cd8b/WasmTestOnV8-ST-System.Runtime.Tests/1/console.46ef387c.log?helixlogtype=result