[RISC-V] Test TotalOrderIeee754ComparerTests: Fix NFloat - NaN testcases for RISC-V #97340
Conversation
ghost
added
the
community-contribution
|
Tagging subscribers to this area: @dotnet/area-system-numerics Issue Detailsfloat->double cast instruction produced by NFloat constructor does not preserve NaN sign and payload on RISC-V. Part of #84834
|
|
@dotnet-policy-service agree company="Samsung" |
| if (PlatformDetection.IsRiscV64Process) // float->double cast does not preserve NaN payload and sign on RISC-V | ||
| { | ||
| yield return new object[] { BitConverter.UInt32BitsToSingle(0xFFC00000), float.NegativeInfinity, 1 }; | ||
| yield return new object[] { BitConverter.UInt32BitsToSingle(0xFFC00000), -1.0f, 1 }; | ||
| yield return new object[] { BitConverter.UInt32BitsToSingle(0xFFC00000), BitConverter.UInt32BitsToSingle(0x7FC00000), 0 }; | ||
| yield return new object[] { BitConverter.UInt32BitsToSingle(0x7FC00000), BitConverter.UInt32BitsToSingle(0x7FC00001), 0 }; // implementation defined, not part of IEEE 754 totalOrder | ||
| } |
There was a problem hiding this comment.
I'm not sure this is the desirable/best fix.
The overall preservation of a NaN sign/payload is implementation defined; however, in the face of a NaN with sign and/or payload, IEEE 754 totalOrder describes exact handling/behavior that is required to be followed.
The test was using SingleTestData for simplicity purposes; however in the face of RiscV it is likely better to fix the test to take NFloat and to enumerate bitwise correct values depending on the underlying bitness of the platform. That way encountered NaN with sign/payload (such as from the developer doing a bitcast where RiscV doesn't normalize) can be validated.
There was a problem hiding this comment.
Do you mean to enumerate bitwise correct values only for NaN values or for all test values?
.../System.Runtime/tests/System.Runtime.Tests/System/Numerics/TotalOrderIeee754ComparerTests.cs
Outdated
Show resolved
Hide resolved
|
Looks like there are some CI failures related to this PR that need to be resolved still, specifically related to Failures look to be similar to |
Thank you for noticing, I have fixed this failure. |
|
@tannergooding, thank you for merge it! |
float->double cast instruction produced by NFloat constructor does not preserve NaN sign and payload on RISC-V.
That's why all NaN values in test data are same positive NaNs without payload.
Part of #84834
cc @tannergooding @gbalykov @t-mustafin @clamp03 @tomeksowi