[clang] [llvm] [HLSL][DXIL][SPIRV] Create llvm dot intrinsic and use for HLSL (PR #102872)

[Justin Bogner via cfe-commits]

bogner wrote:

> There is some discussion in the RFC, but I don't see a consensus on the "dot" intrinsic in particular. I personally haven't found the arguments in favor of it particularly compelling.
> 
> This really needs an RFC specific to that intrinsic (class), which includes a clear definition of the semantics of the intrinsics, why it needs to be a intrinsic and how it maps to different hardware (in particular, whether the chosen definition is actually sufficiently portable).

@pow2clk @farzonl and I discussed this a bit offline and we've come to agree with you here - while the RFC certainly does point out that there's demand for a generic dot intrinsic I think we were a bit overzealous to try to take it on in this context. What we need in HLSL is a fairly specific subset of the dot operation and it's probably more appropriate for us to just handle that with DirectX and SPIR-V specific intrinsics that do that rather than try to take on defining a fully generic dot operation here.

I'll update the RFC to drop the dot and rsqrt intrinsics, and if we later want to revisit those in a generic way they can be handled on their own merits and pitfalls.

https://github.com/llvm/llvm-project/pull/102872