[HLSL][Resources] `rawBufferStore` datarace in concurrent writing to `uint4`

[inbelic]

When trying to writing to a single component of uint4 it will create a local copy of the uint4, update the single component and then store the entire uint4 back into memory. If multiple threads are writing to the uint4 components concurrently, then this will cause a datarace.

I assume the issue is in the codegen of the accessor to Out[0][ThreadID.y] for assignment.

Godbolt reproduction: https://godbolt.org/z/na1n9vqvh

This is consistent with the runtime output where only one component of the uint4 is written here for the GroupMemoryBarrierWithGroupSync tests.

[damyanp]

@bogner to do initial investigation/triage

[bogner]

This looks tricky. What's happening here is that dxc does optimizations after scalarization, so it only sees one element of the vector being written and removes the other writes. In clang we preserve the vector throughout the optimizer, so what we see is a read of a 4-element vector, a modification to the read copy, and a writeback of the whole thing - the optimizer leaves this alone as you would expect.