[RFC] Better support for typed pointers in an opaque pointer world

[nikic]

@beanz

One area that I think would be beneficial both for DXIL and SPIR-V is to retain some limited support for typed pointers indefinitely.

Both SPIR-V and DXIL will require pointer type analysis in some form or another to rematerialized pointer types. That is way easier if you can retain the ability to represent typed pointers in LLVM’s type system. To make this possible the DirectX backend has a DXILPointerType which is a subclass of llvm::Type so that it can be used to generate FunctionType objects and other objects that depend on encompassed types.

And this is fine by me, as long as DXILPointerType never makes it into ordinary IR. If SPIR-V would like to use it as well, I would also be fine with moving the implementation of that type into LLVM core (as long as it is not exposed via IR or bitcode).

@beanz

Before the typed pointer support gets removed it would be nice to either have a plan for representing a shared typed pointer infrastructure for SPIR-V & DXIL, or to just recognize that we’ll likely be duplicating that logic.

This seems like a pure code organization problem for the SPIR-V and DXIL backends, which does not materially affect the rest of the project, and does not need to block typed pointer removal. Both backends currently implement their own solutions to this problem, and if it can be shared, that would be nice. But I don’t think that changes anything about the big picture (which is that any “typed pointer” knowledge has to be encapsulated in a way that is transparent to anyone not working specifically on DXIL or SPIR-V).