diff --git a/clang/include/clang/CIR/Dialect/IR/CIRTypes.td b/clang/include/clang/CIR/Dialect/IR/CIRTypes.td
index abbee1419613..88087f8915ad 100644
--- a/clang/include/clang/CIR/Dialect/IR/CIRTypes.td
+++ b/clang/include/clang/CIR/Dialect/IR/CIRTypes.td
@@ -128,6 +128,7 @@ def CIR_StructType : CIR_Type<"Struct", "struct",
     // for the struct size and alignment.
     mutable std::optional<unsigned> size{}, align{};
     mutable std::optional<bool> padded{};
+    mutable mlir::Type largestMember{};
     void computeSizeAndAlignment(const ::mlir::DataLayout &dataLayout) const;
   public:
     void dropAst();
@@ -141,12 +142,15 @@ def CIR_StructType : CIR_Type<"Struct", "struct",
       case RecordKind::Class:
         return "class." + name;
       case RecordKind::Union:
-        return "union "+ name;
+        return "union." + name;
       case RecordKind::Struct:
         return "struct." + name;
       }
     }
 
+    /// Return the member with the largest bit-length.
+    mlir::Type getLargestMember(const ::mlir::DataLayout &dataLayout) const;
+
     /// Return whether this is a class declaration.
     bool isClass() const { return getKind() == RecordKind::Class; }
 
diff --git a/clang/lib/CIR/Dialect/IR/CIRTypes.cpp b/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
index 47691a7fc463..241f570b8719 100644
--- a/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
+++ b/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
@@ -18,6 +18,7 @@
 #include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/DialectImplementation.h"
+#include "mlir/Interfaces/DataLayoutInterfaces.h"
 #include "mlir/Support/LogicalResult.h"
 
 #include "llvm/ADT/STLExtras.h"
@@ -94,6 +95,15 @@ void BoolType::print(mlir::AsmPrinter &printer) const {}
 // StructType Definitions
 //===----------------------------------------------------------------------===//
 
+/// Return the largest member of in the type.
+///
+/// Recurses into union members never returning a union as the largest member.
+Type StructType::getLargestMember(const ::mlir::DataLayout &dataLayout) const {
+  if (!largestMember)
+    computeSizeAndAlignment(dataLayout);
+  return largestMember;
+}
+
 Type StructType::parse(mlir::AsmParser &parser) {
   const auto loc = parser.getCurrentLocation();
   llvm::SmallVector<mlir::Type> members;
@@ -278,7 +288,7 @@ void StructType::computeSizeAndAlignment(
     const ::mlir::DataLayout &dataLayout) const {
   assert(!isOpaque() && "Cannot get layout of opaque structs");
   // Do not recompute.
-  if (size || align || padded)
+  if (size || align || padded || largestMember)
     return;
 
   // This is a similar algorithm to LLVM's StructLayout.
@@ -287,11 +297,25 @@ void StructType::computeSizeAndAlignment(
   [[maybe_unused]] bool isPadded = false;
   unsigned numElements = getNumElements();
   auto members = getMembers();
+  unsigned largestMemberSize = 0;
 
   // Loop over each of the elements, placing them in memory.
   for (unsigned i = 0, e = numElements; i != e; ++i) {
     auto ty = members[i];
 
+    // Found a nested union: recurse into it to fetch its largest member.
+    auto structMember = ty.dyn_cast<StructType>();
+    if (structMember && structMember.isUnion()) {
+      auto candidate = structMember.getLargestMember(dataLayout);
+      if (dataLayout.getTypeSize(candidate) > largestMemberSize) {
+        largestMember = candidate;
+        largestMemberSize = dataLayout.getTypeSize(largestMember);
+      }
+    } else if (dataLayout.getTypeSize(ty) > largestMemberSize) {
+      largestMember = ty;
+      largestMemberSize = dataLayout.getTypeSize(largestMember);
+    }
+
     // This matches LLVM since it uses the ABI instead of preferred alignment.
     const llvm::Align tyAlign =
         llvm::Align(getPacked() ? 1 : dataLayout.getTypeABIAlignment(ty));
@@ -312,6 +336,14 @@ void StructType::computeSizeAndAlignment(
     structSize += dataLayout.getTypeSize(ty);
   }
 
+  // For unions, the size and aligment is that of the largest element.
+  if (isUnion()) {
+    size = largestMemberSize;
+    align = structAlignment.value();
+    padded = false;
+    return;
+  }
+
   // Add padding to the end of the struct so that it could be put in an array
   // and all array elements would be aligned correctly.
   if (!llvm::isAligned(structAlignment, structSize)) {
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
index 011984399c83..ffa3f099d703 100644
--- a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
+++ b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
@@ -54,10 +54,12 @@
 #include "clang/CIR/Dialect/IR/CIRTypes.h"
 #include "clang/CIR/Passes.h"
 #include "llvm/ADT/APInt.h"
+#include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/Sequence.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
+#include "llvm/IR/DataLayout.h"
 #include "llvm/IR/DebugInfoMetadata.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/Support/Casting.h"
@@ -1306,8 +1308,8 @@ class CIRGlobalOpLowering
           sourceSymbol.getSymName());
       llvm::SmallVector<mlir::LLVM::GEPArg> offset{0};
       auto gepOp = rewriter.create<mlir::LLVM::GEPOp>(
-          loc, llvmType, sourceSymbol.getType(),
-          addressOfOp.getResult(), offset);
+          loc, llvmType, sourceSymbol.getType(), addressOfOp.getResult(),
+          offset);
       rewriter.create<mlir::LLVM::ReturnOp>(loc, gepOp.getResult());
       return mlir::success();
     } else if (isa<mlir::cir::ZeroAttr, mlir::cir::NullAttr>(init.value())) {
@@ -1721,14 +1723,30 @@ class CIRGetMemberOpLowering
   matchAndRewrite(mlir::cir::GetMemberOp op, OpAdaptor adaptor,
                   mlir::ConversionPatternRewriter &rewriter) const override {
     auto llResTy = getTypeConverter()->convertType(op.getType());
-    // Since the base address is a pointer to structs, the first offset is
-    // always zero. The second offset tell us which member it will access.
-    llvm::SmallVector<mlir::LLVM::GEPArg> offset{0, op.getIndex()};
-    const auto elementTy = getTypeConverter()->convertType(
-        op.getAddr().getType().getPointee());
-    rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(
-        op, llResTy, elementTy, adaptor.getAddr(), offset);
-    return mlir::success();
+    const auto structTy =
+        op.getAddrTy().getPointee().cast<mlir::cir::StructType>();
+    assert(structTy && "expected struct type");
+
+    switch (structTy.getKind()) {
+    case mlir::cir::StructType::Struct: {
+      // Since the base address is a pointer to an aggregate, the first offset
+      // is always zero. The second offset tell us which member it will access.
+      llvm::SmallVector<mlir::LLVM::GEPArg, 2> offset{0, op.getIndex()};
+      const auto elementTy = getTypeConverter()->convertType(structTy);
+      rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(op, llResTy, elementTy,
+                                                     adaptor.getAddr(), offset);
+      return mlir::success();
+    }
+    case mlir::cir::StructType::Union:
+      // Union members share the address space, so we just need a bitcast to
+      // conform to type-checking.
+      rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(op, llResTy,
+                                                         adaptor.getAddr());
+      return mlir::success();
+    default:
+      return op.emitError()
+             << "struct kind '" << structTy.getKind() << "' is NYI";
+    }
   }
 };
 
@@ -1789,7 +1807,8 @@ void populateCIRToLLVMConversionPatterns(mlir::RewritePatternSet &patterns,
 }
 
 namespace {
-void prepareTypeConverter(mlir::LLVMTypeConverter &converter) {
+void prepareTypeConverter(mlir::LLVMTypeConverter &converter,
+                          mlir::DataLayout &dataLayout) {
   converter.addConversion([&](mlir::cir::PointerType type) -> mlir::Type {
     return mlir::LLVM::LLVMPointerType::get(&converter.getContext());
   });
@@ -1814,9 +1833,24 @@ void prepareTypeConverter(mlir::LLVMTypeConverter &converter) {
     return mlir::LLVM::LLVMFunctionType::get(result, arguments, varArg);
   });
   converter.addConversion([&](mlir::cir::StructType type) -> mlir::Type {
+    // FIXME(cir): create separate unions, struct, and classes types.
+    // Convert struct members.
     llvm::SmallVector<mlir::Type> llvmMembers;
-    for (auto ty : type.getMembers())
-      llvmMembers.push_back(converter.convertType(ty));
+    switch (type.getKind()) {
+    case mlir::cir::StructType::Class:
+      // TODO(cir): This should be properly validated.
+    case mlir::cir::StructType::Struct:
+      for (auto ty : type.getMembers())
+        llvmMembers.push_back(converter.convertType(ty));
+      break;
+    // Unions are lowered as only the largest member.
+    case mlir::cir::StructType::Union: {
+      auto largestMember = type.getLargestMember(dataLayout);
+      if (largestMember)
+        llvmMembers.push_back(converter.convertType(largestMember));
+      break;
+    }
+    }
 
     // Struct has a name: lower as an identified struct.
     mlir::LLVM::LLVMStructType llvmStruct;
@@ -1847,7 +1881,7 @@ static void buildCtorList(mlir::ModuleOp module) {
         assert(attr.isa<mlir::cir::GlobalCtorAttr>() &&
                "must be a GlobalCtorAttr");
         if (auto ctorAttr = attr.cast<mlir::cir::GlobalCtorAttr>()) {
-           // default priority is 65536
+          // default priority is 65536
           int priority = 65536;
           if (ctorAttr.getPriority())
             priority = *ctorAttr.getPriority();
@@ -1885,15 +1919,15 @@ static void buildCtorList(mlir::ModuleOp module) {
   newGlobalOp.getRegion().push_back(new mlir::Block());
   builder.setInsertionPointToEnd(newGlobalOp.getInitializerBlock());
 
-  mlir::Value result = builder.create<mlir::LLVM::UndefOp>(
-      loc, CtorStructArrayTy);
+  mlir::Value result =
+      builder.create<mlir::LLVM::UndefOp>(loc, CtorStructArrayTy);
 
   for (uint64_t I = 0; I < globalCtors.size(); I++) {
     auto fn = globalCtors[I];
     mlir::Value structInit =
         builder.create<mlir::LLVM::UndefOp>(loc, CtorStructTy);
-    mlir::Value initPriority =
-        builder.create<mlir::LLVM::ConstantOp>(loc, CtorStructFields[0], fn.second);
+    mlir::Value initPriority = builder.create<mlir::LLVM::ConstantOp>(
+        loc, CtorStructFields[0], fn.second);
     mlir::Value initFuncAddr = builder.create<mlir::LLVM::AddressOfOp>(
         loc, CtorStructFields[1], fn.first);
     mlir::Value initAssociate =
@@ -1914,9 +1948,9 @@ static void buildCtorList(mlir::ModuleOp module) {
 
 void ConvertCIRToLLVMPass::runOnOperation() {
   auto module = getOperation();
-
+  mlir::DataLayout dataLayout(module);
   mlir::LLVMTypeConverter converter(&getContext());
-  prepareTypeConverter(converter);
+  prepareTypeConverter(converter, dataLayout);
 
   mlir::RewritePatternSet patterns(&getContext());
 
diff --git a/clang/test/CIR/Lowering/unions.cir b/clang/test/CIR/Lowering/unions.cir
new file mode 100644
index 000000000000..c5ee736c4a7d
--- /dev/null
+++ b/clang/test/CIR/Lowering/unions.cir
@@ -0,0 +1,42 @@
+// RUN: cir-opt %s -cir-to-llvm -o %t.mlir
+// RUN: FileCheck --input-file=%t.mlir %s
+
+!s16i = !cir.int<s, 16>
+!s32i = !cir.int<s, 32>
+#true = #cir.bool<true> : !cir.bool
+!ty_22U122 = !cir.struct<union "U1" {!cir.bool, !s16i, !s32i} #cir.recdecl.ast>
+!ty_22U222 = !cir.struct<union "U2" {f64, !ty_22U122} #cir.recdecl.ast>
+!ty_22U322 = !cir.struct<union "U3" {!s16i, !ty_22U122} #cir.recdecl.ast>
+module {
+  // Should lower union to struct with only the largest member.
+  cir.global external @u1 = #cir.zero : !ty_22U122
+  // CHECK: llvm.mlir.global external @u1() {addr_space = 0 : i32} : !llvm.struct<"union.U1", (i32)>
+
+  // Should recursively find the largest member if there are nested unions.
+  cir.global external @u2 = #cir.zero : !ty_22U222
+  cir.global external @u3 = #cir.zero : !ty_22U322
+  // CHECK: llvm.mlir.global external @u2() {addr_space = 0 : i32} : !llvm.struct<"union.U2", (f64)>
+  // CHECK: llvm.mlir.global external @u3() {addr_space = 0 : i32} : !llvm.struct<"union.U3", (i32)>
+
+  // CHECK: llvm.func @test
+  cir.func @test(%arg0: !cir.ptr<!ty_22U122>) {
+
+    // Should store directly to the union's base address.
+    %5 = cir.const(#true) : !cir.bool
+    %6 = cir.get_member %arg0[0] {name = "b"} : !cir.ptr<!ty_22U122> -> !cir.ptr<!cir.bool>
+    cir.store %5, %6 : !cir.bool, cir.ptr <!cir.bool>
+    // CHECK: %[[#VAL:]] = llvm.mlir.constant(1 : i8) : i8
+    // The bitcast it just to bypass the type checker. It will be replaced by an opaque pointer.
+    // CHECK: %[[#ADDR:]] = llvm.bitcast %{{.+}} : !llvm.ptr to !llvm.ptr
+    // CHECK: llvm.store %[[#VAL]], %[[#ADDR]] : i8, !llvm.ptr
+
+    // Should load direclty from the union's base address.
+    %7 = cir.get_member %arg0[0] {name = "b"} : !cir.ptr<!ty_22U122> -> !cir.ptr<!cir.bool>
+    %8 = cir.load %7 : cir.ptr <!cir.bool>, !cir.bool
+    // The bitcast it just to bypass the type checker. It will be replaced by an opaque pointer.
+    // CHECK: %[[#BASE:]] = llvm.bitcast %{{.+}} : !llvm.ptr to !llvm.ptr
+    // CHECK: %{{.+}} = llvm.load %[[#BASE]] : !llvm.ptr
+
+    cir.return
+  }
+}