[DirectX] GEPs to multi-dimensional arrays are incorrectly flattened

[Icohedron]

The DXIL Flatten Arrays pass is incorrectly transforming getelementptr instructions operating on mutli-dimensional arrays.

The DXIL Flatten Arrays pass needs to be modified to maintain correctness when transforming GEPs on multi-dimensional arrays.

https://godbolt.org/z/evPPPWEfT

IR dump before flatten arrays (-mllvm -stop-before=dxil-flatten-arrays):

@g = local_unnamed_addr addrspace(3) global [2 x [2 x float]] zeroinitializer, align 4, !dbg !0
...
  %2 = tail call i32 @llvm.dx.thread.id(i32 0)
  %3 = getelementptr inbounds nuw [2 x [2 x float]], ptr addrspace(3) @g, i32 0, i32 %2, i32 1, !dbg !91
  %4 = load float, ptr addrspace(3) %3, align 4, !dbg !91, !tbaa !92

IR dump after flatten arrays:

@g.1dim = local_unnamed_addr addrspace(3) global [4 x float] zeroinitializer, align 4
...
  %2 = tail call i32 @llvm.dx.thread.id(i32 0)
  %3 = getelementptr inbounds [4 x float], ptr addrspace(3) @g.1dim, i32 1
  %4 = load float, ptr addrspace(3) %3, align 4, !dbg !91

As can be seen, the GEP after flatten arrays is incorrect (such as %2 becoming unused) and out-of-bounds (because g.1dim is only one array, so i32 1 is OOB).

A similar case can be seen for local arrays: https://godbolt.org/z/M7xaxGhMa

[farzonl]

There is still an outstanding issue I haven't figured out around the fact we dropped %2.
%2 = tail call i32 @llvm.dx.thread.id(i32 0)

I think there is likely a bug in our index calculating code where we don't take the genInstructionFlattenIndices or if we do something is happening and those instructions are not preserved. Need to do some debugging to figure it out.

llvm-project/llvm/lib/Target/DirectX/DXILFlattenArrays.cpp

Lines 267 to 271 in 8ca220f

	if (GEPInfo.AllIndicesAreConstInt)
	FlatIndex = genConstFlattenIndices(GEPInfo.Indices, GEPInfo.Dims, Builder);
	else
	FlatIndex =
	genInstructionFlattenIndices(GEPInfo.Indices, GEPInfo.Dims, Builder);