LLVM error when accessing an array of matrices using a vector subscript on the final column.

Description
Error in the LLVM when trying to compile code that uses an array of matrices and trying to index the final element by the subscript .a. Switching this to [3] fixes the issue.

Steps to Reproduce
Gobolt.org link
Command line: -T cs_6_6 -E main -Od

void swap(inout uint lhs, inout uint rhs)
{
    uint tmp = lhs;
    lhs = rhs;
    rhs = tmp;
}

[numthreads(64,1,1)]
void main(uint GI : SV_GroupIndex, uint3 groupID : SV_GroupID)
{
    uint2x4 a[2];
    uint2x4 b = a[0];
    
    swap(b[0].a, b[1].a); // Work
    swap(a[0][0].a, a[0][1].a); // Doesn't work
}

Actual Behavior

error: validation errors
error: Module bitcode is invalid.
error: Instruction does not dominate all uses!
  %41 = load i32, i32* %40
  store i32 %41, i32* %38

Validation failed.


Compiler returned: 5

Environment
DXC version 1.6.2112 to DXC version 1.8.2502
Windows 11, version 22631.5039

Reproduce the failure with the linked Godbolt HLSL example using DXC and the command line -T cs_6_6 -E main -Od. Compare the failing a[0][0].a access with the working equivalent accesses, then verify that the compiler no longer reports invalid LLVM bitcode or dominance errors.