The [InterceptsLocation] attribute generates a different "data" string for big-endian platforms

[steveharter]

Version Used: 4.8.0

Steps to Reproduce: See dotnet/runtime#105311

Expected Behavior: big-endian platforms have the same string value for the "data" parameter as little-endian platforms.

Actual Behavior: big-endian is different than little-endian due to byte swapping on char.

See also
dotnet/runtime#101079
dotnet/runtime#104693
#72133

[RikkiGibson]

The cause appears to be that SourceText.GetContentHash() does not produce the same hash on big-endian, even if the text is the same character-for-character.

[CyrusNajmabadi]

I did doc this as:

        /// This hash is safe to use across platforms and across processes, as long as the same version of Roslyn is
        /// used in all those locations. 

So this does seem to be a bug in the impl.

I'm curious if there's a fast way to do xxhash in an endian-independent fashion. @tannergooding @stephentoub ?

[stephentoub]

xxhash is not endian-dependent. Given the same sequence of bytes in, you get the same output sequence of bytes, regardless of whether on a big or little endian machine.

If you're getting a different answer, my guess is it's because a different sequence of bytes is being fed in, e.g. if you just reinterpret cast chars to bytes, that input sequence of bytes will be different based on endianness.

[CyrusNajmabadi]

e.g. if you just reinterpret cast chars to bytes, that input sequence of bytes will be different based on endianness.

Yes. That's what is going on. Is there a recommended way to work with chars and xxhash in such a way as to be endian independent and fast? Thanks!

[stephentoub]

e.g. if you just reinterpret cast chars to bytes, that input sequence of bytes will be different based on endianness.

Yes. That's what is going on. Is there a recommended way to work with chars and xxhash in such a way as to be endian independent and fast? Thanks!

You could reverse the endianness only if you're on a big-endian machine. Then there's no cost for the vast majority case where you're on little endian.

[CyrusNajmabadi]

Is there a recommended helper for that if I've got a span of bytes? Just wanting to get the best performance even in the uncommon case. Tnx!

[stephentoub]

Is there a recommended helper for that if I've got a span of bytes?

There's no such thing as endianness if all you have are bytes, since endiannes is about the order of bytes within a larger unit. You would need to know what that unit was so that you could appropriately reverse the bytes within each unit, e.g. you'd need to know that it was two-byte units (chars) coming in so that you could swap each pair.

Once you know that, there are a ton of ReverseEndianness APIs on BinaryPrimitives that can help, including ones that handle individual values as well as spans of values with vectorized/fast implementations, but if you're targeting netstandard2.0, that won't be as useful.