[clang] [llvm] [HLSL][SPIR-V] Add SV_DispatchThreadID semantic support (PR #82536)

[Justin Bogner via llvm-commits]

================
@@ -342,8 +343,19 @@ llvm::Value *CGHLSLRuntime::emitInputSemantic(IRBuilder<> &B,
     return B.CreateCall(FunctionCallee(DxGroupIndex));
   }
   if (D.hasAttr<HLSLSV_DispatchThreadIDAttr>()) {
-    llvm::Function *DxThreadID = CGM.getIntrinsic(Intrinsic::dx_thread_id);
-    return buildVectorInput(B, DxThreadID, Ty);
+    llvm::Function *ThreadIDIntrinsic;
+    switch (CGM.getTarget().getTriple().getArch()) {
+    case llvm::Triple::dxil:
+      ThreadIDIntrinsic = CGM.getIntrinsic(Intrinsic::dx_thread_id);
+      break;
+    case llvm::Triple::spirv:
+      ThreadIDIntrinsic = CGM.getIntrinsic(Intrinsic::spv_thread_id);
+      break;
+    default:
+      llvm_unreachable("Input semantic not supported by target");
+      break;
+    }
+    return buildVectorInput(B, ThreadIDIntrinsic, Ty);
----------------
bogner wrote:

So this pattern is going to get really annoying really fast. I think there are two approaches to how we can handle this in a better way:
1. Use TargetLibraryInfo or something similar, where backends can register their own support for getting the right intrinsics for this kind of thing.
2. Add a tool in llvm/lib/Frontend/HLSL that knows the answers for SPIR-V and DXIL

I suspect (1) is a more robust solution, but it has a couple of problems. Notably it's a big thing to design and it isn't entirely clear that it's all that useful outside of DXIL and SPIR-V, especially in the short term. I don't think we should do this now.

For (2) though, we could probably just have a .def file that maps the DXIL and SPIR-V intrinsics to a generic enum, or alternatively just a set of getters for the intrinsics we care about. Then this just looks like HLSLIntrinsic::get(threadid) or HLSLIntrinsic::getThreadID() or something to that effect. WDYT?

https://github.com/llvm/llvm-project/pull/82536