[wasi][R2R] BFloat16 conversions from Int128/UInt128 produce wrong values

[lewing]

Description

In the CoreCLR WASI composite ReadyToRun library lane (LibraryTestsCoreCLR_WASI_R2R, trimmed), BFloat16 explicit conversions from Int128/UInt128 produce wrong values. The same tests pass in the WASI interpreter lane, so the R2R (RyuJIT wasm) code is wrong.

[FAIL] System.Numerics.Tests.BFloat16Tests.ExplicitConversion_FromInt128(i: 16563852592415126800007719339717296127, expected: 2.99E-26)
Expected: 2.9888E-26
Actual:   1.6532E+37
[FAIL] System.Numerics.Tests.BFloat16Tests.ExplicitConversion_FromInt128(i: 336347874021343232, expected: 0)
Expected:  +0.0
Actual:   3.3552E+17
[FAIL] System.Numerics.Tests.BFloat16Tests.ExplicitConversion_FromUInt128(...)

Reproduces locally (macOS arm64 host, wasmtime 49.0.1): System.Runtime.Tests built with /p:TestWasmReadyToRun=true /p:EnableAggressiveTrimming=true, class filter System.Numerics.Tests.BFloat16Tests: 1608 run, 3 failed.

Seen in Helix job eccd7d69-de48-4cc1-87fd-6036eb264d50 (#134813) and earlier runs. Quarantined on WASI R2R in #134813.

Note

This issue was drafted with GitHub Copilot.

[dotnet-policy-service]

Tagging subscribers to 'arch-wasm': @lewing, @pavelsavara
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.

[lewing]

cc @tannergooding @adamperlin