[clang] [CIR] Upstream proper function alias lowering (PR #150520)

[Henrich Lauko via cfe-commits]

================
@@ -919,13 +919,46 @@ rewriteCallOrInvoke(mlir::Operation *op, mlir::ValueRange callOperands,
                            memoryEffects, noUnwind, willReturn);
 
   mlir::LLVM::LLVMFunctionType llvmFnTy;
+
+  // Temporary to handle the case where we need to prepend an operand if the
+  // callee is an alias.
+  SmallVector<mlir::Value> adjustedCallOperands;
+
   if (calleeAttr) { // direct call
-    mlir::FunctionOpInterface fn =
-        mlir::SymbolTable::lookupNearestSymbolFrom<mlir::FunctionOpInterface>(
-            op, calleeAttr);
-    assert(fn && "Did not find function for call");
-    llvmFnTy = cast<mlir::LLVM::LLVMFunctionType>(
-        converter->convertType(fn.getFunctionType()));
+    mlir::Operation *callee =
+        mlir::SymbolTable::lookupNearestSymbolFrom(op, calleeAttr);
+    if (auto fn = dyn_cast<mlir::FunctionOpInterface>(callee)) {
+      llvmFnTy = cast<mlir::LLVM::LLVMFunctionType>(
+          converter->convertType(fn.getFunctionType()));
+    } else if (auto alias = cast<mlir::LLVM::AliasOp>(callee)) {
+      // If the callee wasan alias. In that case,
+      // we need to prepend the address of the alias to the operands. The
+      // way aliases work in the LLVM dialect is a little counter-intuitive.
+      // The AliasOp itself is a pseudo-function that returns the address of
+      // the global value being aliased, but when we generate the call we
+      // need to insert an operation that gets the address of the AliasOp.
+      // This all gets sorted out when the LLVM dialect is lowered to LLVM IR.
+      auto symAttr = cast<mlir::FlatSymbolRefAttr>(calleeAttr);
----------------
xlauko wrote:

```suggestion
      auto symAttr = mlir::cast<mlir::FlatSymbolRefAttr>(calleeAttr);
```

https://github.com/llvm/llvm-project/pull/150520