[clang] [clang][CodeGen][SPIR-V][AMDGPU] Tweak AMDGCNSPIRV ABI to allow for the correct handling of aggregates passed to kernels / functions. (PR #102776)

[Alex Voicu via cfe-commits]

================
@@ -64,6 +66,27 @@ void CommonSPIRABIInfo::setCCs() {
   RuntimeCC = llvm::CallingConv::SPIR_FUNC;
 }
 
+ABIArgInfo SPIRVABIInfo::classifyReturnType(QualType RetTy) const {
+  if (getTarget().getTriple().getVendor() != llvm::Triple::AMD)
+    return DefaultABIInfo::classifyReturnType(RetTy);
+  if (!isAggregateTypeForABI(RetTy) || getRecordArgABI(RetTy, getCXXABI()))
+    return DefaultABIInfo::classifyReturnType(RetTy);
+
+  if (const RecordType *RT = RetTy->getAs<RecordType>()) {
+    const RecordDecl *RD = RT->getDecl();
+    if (RD->hasFlexibleArrayMember())
+      return DefaultABIInfo::classifyReturnType(RetTy);
+  }
+
+  // TODO: The AMDGPU ABI is non-trivial to represent in SPIR-V; in order to
+  // avoid encoding various architecture specific bits here we return everything
+  // as direct to retain type info for things like aggregates, for later perusal
+  // when translating back to LLVM/lowering in the BE. This is also why we
+  // disable flattening as the outcomes can mismatch between SPIR-V and AMDGPU.
+  // This will be revisited / optimised in the future.
----------------
AlexVlx wrote:

To wit, it's necessary pain, in the end, but we've not yet figured out the right way to do it, since there are a few moving pieces involved. Perhaps the solution will end up to add `byref` to the SPIR-V spec; in the meanwhile, this is a sufficient, if temporary and slightly icky, solution.

https://github.com/llvm/llvm-project/pull/102776