[HLSL][DirectX] AddUint64 vec4 case resulting in invalid DXIL.

[spall]

https://godbolt.org/z/To9e81PWs

StructuredBuffer<uint4> In : register(t0);
RWStructuredBuffer<uint4> Out : register(u1);

[numthreads(1,1,1)]
void main() {
  Out[3] = AddUint64(In[0], In[0]);
}

This code, when compiled, produces invalid DXIL

# .---command stdout------------
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%.upto077112 = insertelement <2 x i1> undef, i1 %.elem1, i32 0' in block 'entry' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%58 = insertelement <2 x i1> %.upto077112, i1 %.elem17, i32 1' in block 'entry' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%.i09113 = extractelement <2 x i1> %58, i32 0' in block 'entry' of function 'main'.
# | Function: main: error: Instructions must be of an allowed type.
# | note: at '%.i110114 = extractelement <2 x i1> %58, i32 1' in block 'entry' of function 'main'.
# | Validation failed.
# |
# `-----------------------------
# .---command stderr------------
# | clang-dxc: error: dxv command failed with exit code 1 (use -v to see invocation)
# `-----------------------------
# error: command failed with exit status: 1

These 'insertelement' and 'extractelement' instructions are created by the scalarizer pass, but are unnecessary and could be eliminated.

[Icohedron]

I think the Scalarizer may have a related bug w.r.t working with vectors of i1s.
It may be related to an interesting behavior in the vectorSplit function I found when eliminating a crash in the Scalarizer when working with a non-zero min-bits parameter (#128538 (comment)) and that is that the vectorSplit function will always return an empty VectorSplit when it receives a vector of bools (i1).

[Icohedron]

The issue is that the scalarization of ExtractValue instruction is using the incorrect type for a non-zero index.

llvm-project/llvm/lib/Transforms/Scalar/Scalarizer.cpp

Line 1081 in 7b5d8a0

Type *VecType = cast<FixedVectorType>(OpTy->getContainedType(0));

OpTy->getContainedType(0) should be OpTy->getContainedType(EVI.getIndices()[0])

However making this change seems to cause a crash bug with the Scalarizer which I am still determining how to fix.