[clang] [CIR] Wire AMDGPU into the call-convention lowering pass (PR #220197)

[Steffen Larsen via cfe-commits]

================
@@ -664,6 +668,26 @@ convertABIArgInfo(const llvm::abi::ArgInfo &info, MLIRContext *ctx,
     // than emit a read of the wrong bytes.
     if (offset && coerceIsRegisterTuple)
       return std::nullopt;
+    // A pointer moved to another address space keeps its pointee, which
+    // abiTypeToCIR cannot recover.
+    if (auto origPtr = dyn_cast_if_present<cir::PointerType>(origTy)) {
+      if (const auto *coercePtr =
+              dyn_cast_if_present<llvm::abi::PointerType>(coerceAbi)) {
+        auto origAS = dyn_cast_if_present<cir::TargetAddressSpaceAttr>(
+            origPtr.getAddrSpace());
+        unsigned origASValue = origAS ? origAS.getValue() : 0;
+        if (!offset && coercePtr->getAddrSpace() != origASValue) {
----------------
steffenlarsen wrote:

I'm a little confused as to why we need to do the conversion into `origASValue` and not just pass `origPtr.getAddrSpace()` directly?

https://github.com/llvm/llvm-project/pull/220197