Miscompile when reading matrix from a structure in a ByteAddressBuffer

Description
A compute shader like this mis-compiles the read. It requires a fairly
particular setup as described in comments below

struct Box { float4x4 m; };   // a struct whose SOLE member is a float4x4

ByteAddressBuffer   src : register(t0);
RWByteAddressBuffer dst : register(u0);

[numthreads(1, 1, 1)]
void csmain()
{
    Box b = src.Load<Box>(0);   // 64-byte struct at offset 0

    // read the first matrix column through the struct member. 
    // DXC loads byte 0 for all four.
    dst.Store(0,  asuint(b.m._m00));
    dst.Store(4,  asuint(b.m._m10));
    dst.Store(8,  asuint(b.m._m20));
    dst.Store(12, asuint(b.m._m30));

    // CORRECT: copy the matrix to a local float4x4 first, then read. Offsets are right.
    float4x4 m = b.m;
    dst.Store(16, asuint(m._m00));
    dst.Store(20, asuint(m._m10));
    dst.Store(24, asuint(m._m20));
    dst.Store(28, asuint(m._m30));
}

Steps to Reproduce
https://godbolt.org/z/bG5reh3rh

Actual Behavior
The DXIL produced by dxc performs the first 4 reads at offset 0

Environment
I reproduced on Mac and Windows. I verified that it reproduces in v1.10.2605.24 and current HEAD.

Please see #8568 for reproducing test and a potential way to address it.

Start with the Godbolt reproduction to confirm the incorrect DXIL offsets, then read issue #8568 for the reproducing test and potential approach. Run that test and trace the compiler path for loading a matrix member from a struct returned by ByteAddressBuffer.Load. Done means the four direct reads use the correct matrix offsets rather than all reading byte 0.