[DirectX] assertion during memcpy expansion with struct with a struct field

[spall]

struct B {
  double L;
};

struct A{
  B C;
};

// Derived Struct truncate to scalar
export double call9(A A) {
  B B = A.C;
  return B.L;
}

emitMemcpyExpansion(llvm::IRBuilder<>&, llvm::Value*, llvm::Value*, llvm::ConstantInt*): Assertion `DstArrTy && "Expected Dst of memcpy to be a Pointer to an Array Type"' failed.

https://godbolt.org/z/7jv9scn81

[Icohedron]

This bug is due to a memcpy intrinsic being used on a struct. Currently, the memcpy legalization only applies to arrays.
For this specific repro case however, the struct is from a struct alloca:

...
%struct.B = type { double }

; Function Attrs: alwaysinline convergent mustprogress norecurse nounwind
define noundef nofpclass(nan inf) double @call9(A)(ptr noundef byval(%struct.A) align 1 %A) #0 !dbg !8 {
entry:
  %B = alloca %struct.B, align 1
  ...
  call void @llvm.memcpy.p0.p0.i32(ptr align 1 %B, ptr align 1 %C, i32 8, i1 false), !dbg !25
  ...
}
...

And struct allocas are not allowed (#147109) in DXIL. I am currently working on eliminating struct allocas from the IR, which will resolve this specific repro case without needing to implement memcpy legalization for structs.

However, we still need to implement a memcpy legalization for structs if memcpy can still appear on groupshared structs (which likely is the case once this shader https://godbolt.org/z/1eKPojcd3 from @bogner is able to compile in Clang).

[Icohedron]

Also for this repro case, the assertion error only occurs with -O0. With at least -O1, the current SROA implementation in LLVM is able to make this repro case compile successfully. My fix for #147109 will also enable SROA in -O0.

However, a slightly modified repro case will still cause the same assertion error even with optimizations enabled: https://godbolt.org/z/8eszqf79v

struct B {
  double L[10];
};

struct A{
  B C;
};

// Derived Struct truncate to scalar
export double call9(A A, int i) {
  B B = A.C;
  return B.L[i];
}

[Icohedron]

PR #161601 fixes the original repro, but not my modified repro (https://godbolt.org/z/8eszqf79v).
The issue is that the lowering of memcpy in the dxil legalize pass does not support structs, and one of the arguments to the memcpy is a function argument that is typed as a struct (%A).
This will be filed as a separate issue to be fixed, which I think @bogner was going to file?

[Icohedron]

Blocked, as #147109 can not be fixed with the current LLVM IR semantics.
See https://discourse.llvm.org/t/rfc-adding-instructions-to-to-carry-gep-type-traversal-information/88141