[mono] Try using tablegen-derived data to auto-generate LLVM IR fragments

[imhameed]

See https://github.com/dotnet/llvm-project/blob/release/9.x/clang/include/clang/Basic/arm_neon.td and https://github.com/dotnet/llvm-project/blob/release/9.x/clang/utils/TableGen/NeonEmitter.cpp.

[imhameed]

Some problems with this:

• While arm_neon.td does provide a complete description of everything necessary to generate something like arm_neon.h, each individual intrinsic is implemented in terms of clang builtins, and these don't necessarily directly correspond to individual LLVM IR operations. There's still lots of special-case hand-written C++ that turns clang builtins into IR sequences: https://github.com/dotnet/llvm-project/blob/release/9.x/clang/lib/CodeGen/CGBuiltin.cpp#L6906-L9277
• amd64 intrinsic headers and their related clang builtins are entirely hand-maintained: https://github.com/dotnet/llvm-project/blob/release/9.x/clang/lib/Headers/avx512bwintrin.h, https://github.com/dotnet/llvm-project/blob/release/9.x/clang/lib/CodeGen/CGBuiltin.cpp#L9972-L12192

Complete auto-generation would require transliterating these exceptions to some sort of "meta" IR construction DSL. It seems to me that it might be better to scrape each header for prototypes and generate a large file to be fed to clang -O2 -g0 -emit-llvm and then postprocess the resulting bitcode.

For now, I think I'll auto-generate the lower-effort stuff and hand-write the rest.