s390x: XxHash3Tests.Hash_OneShot_Expected and Hash_Streaming_Expected tests fail

[akoeplinger]

e.g. https://dev.azure.com/dnceng-public/public/_build/results?buildId=76403&view=ms.vss-test-web.build-test-results-tab&runId=1580894&resultId=130159&paneView=debug

Assert.Equal() Failure
Expected: 3402425285238347205
Actual:   18319083679080722056


Stack trace
   at System.IO.Hashing.Tests.XxHash3Tests.Hash_OneShot_Expected() in /_/src/libraries/System.IO.Hashing/tests/XxHash3Tests.cs:line 39
   at System.Reflection.MethodInvoker.InterpretedInvoke(Object obj, Span`1 args, BindingFlags invokeAttr) in /_/src/mono/System.Private.CoreLib/src/System/Reflection/MethodInvoker.Mono.cs:line 33

/cc @nealef @uweigand

[uweigand]

It seems the failures start occurring once the string length gets larger than 240 bytes. This is when the DeriveSecretFromSeed routine gets first involved, and on initial glance this doesn't appear to be endian safe:

for (int i = 0; i < SecretLengthBytes; i += sizeof(ulong) * 2)
{
    Unsafe.WriteUnaligned(destinationSecret + i, Unsafe.ReadUnaligned<ulong>(defaultSecret + i) + seed);
    Unsafe.WriteUnaligned(destinationSecret + i + 8, Unsafe.ReadUnaligned<ulong>(defaultSecret + i + 8) - seed);
}

The original C implementation has here:

        for (i=0; i < nbRounds; i++) {
            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
       }

Looks like this needs byte-swaps at the loads and stores on a big-endian system.

[uweigand]

CC @stephentoub