[HLSL] Investigate whether we can avoid memset and memcpy in object construction and type casting

[bogner]

Clang's codegen for C++ constructors doesn't seem to check the target for whether we should generate memcpy for members, and there are also memcpys generated for certain types of cast. We should investigate whether this can and should be made target specific (Especially after #138787), and if so we can eliminate DXIL's custom legalization of memset and avoid implementing memcpy legalization.

[Icohedron]

I'll leave a note here that memcpy appears to emitted when passing in arrays to function arguments: https://godbolt.org/z/nbW5M176d

[Icohedron]

I have instrumented all instances of "CreateMemCpy" in Clang, compiled all the DML shaders with this instrumented Clang, and found that all memcpy intrinsics emitted by Clang came from this line:

llvm-project/clang/lib/CodeGen/CGExprAgg.cpp

Line 2397 in 5f69d68

auto *Inst = Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, isVolatile);

This line is part of the function EmitAggregateCopy which handles assignment between two aggregate-typed LValues.

DXC notably handled this case with its own call to a EmitHLSLAggregateCopy function:
https://github.com/microsoft/DirectXShaderCompiler/blob/b78ac50593248bbbf04ba51cb48c3e1f94b416c6/tools/clang/lib/CodeGen/CGExprAgg.cpp#L1542-L1548

The EmitHLSLAggregateCopy function in DXC is located here: https://github.com/microsoft/DirectXShaderCompiler/blob/b4baabb7da9e483b624e12a86ae29df7d162d4f2/tools/clang/lib/CodeGen/CGHLSLMS.cpp#L5715
But it notably still emits memcpy intrinsics for structs and arrays.