diff --git a/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp b/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
index fbf901c47a7c..7afa993f65de 100644
--- a/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenItaniumCXXABI.cpp
@@ -2261,7 +2261,17 @@ mlir::Value CIRGenItaniumCXXABI::getCXXDestructorImplicitParam(
 
 void CIRGenItaniumCXXABI::emitRethrow(CIRGenFunction &CGF, bool isNoReturn) {
   // void __cxa_rethrow();
-  llvm_unreachable("NYI");
+
+  if (isNoReturn) {
+    auto &builder = CGF.getBuilder();
+    assert(CGF.currSrcLoc && "expected source location");
+    auto loc = *CGF.currSrcLoc;
+    builder.create<cir::ThrowOp>(loc, mlir::Value{}, mlir::FlatSymbolRefAttr{},
+                                 mlir::FlatSymbolRefAttr{});
+    builder.create<cir::UnreachableOp>(loc);
+  } else {
+    llvm_unreachable("NYI");
+  }
 }
 
 void CIRGenItaniumCXXABI::emitThrow(CIRGenFunction &CGF,
diff --git a/clang/lib/CIR/Dialect/Transforms/FlattenCFG.cpp b/clang/lib/CIR/Dialect/Transforms/FlattenCFG.cpp
index 2eb4b6134686..e912a9aa6c8f 100644
--- a/clang/lib/CIR/Dialect/Transforms/FlattenCFG.cpp
+++ b/clang/lib/CIR/Dialect/Transforms/FlattenCFG.cpp
@@ -426,11 +426,14 @@ class CIRTryOpFlattening : public mlir::OpRewritePattern<cir::TryOp> {
                      SmallVectorImpl<mlir::Block *> &landingPads) const {
     // Replace the tryOp return with a branch that jumps out of the body.
     rewriter.setInsertionPointToEnd(afterBody);
-    auto tryBodyYield = cast<cir::YieldOp>(afterBody->getTerminator());
 
     mlir::Block *beforeCatch = rewriter.getInsertionBlock();
     rewriter.setInsertionPointToEnd(beforeCatch);
-    rewriter.replaceOpWithNewOp<cir::BrOp>(tryBodyYield, afterTry);
+
+    // Check if the terminator is a YieldOp because there could be another
+    // terminator, e.g. unreachable
+    if (auto tryBodyYield = dyn_cast<cir::YieldOp>(afterBody->getTerminator()))
+      rewriter.replaceOpWithNewOp<cir::BrOp>(tryBodyYield, afterTry);
 
     // Start the landing pad by getting the inflight exception information.
     mlir::Block *nextDispatcher =
diff --git a/clang/lib/CIR/Dialect/Transforms/LoweringPrepare.cpp b/clang/lib/CIR/Dialect/Transforms/LoweringPrepare.cpp
index 20239d843d2d..f6f3e58e0322 100644
--- a/clang/lib/CIR/Dialect/Transforms/LoweringPrepare.cpp
+++ b/clang/lib/CIR/Dialect/Transforms/LoweringPrepare.cpp
@@ -86,6 +86,7 @@ struct LoweringPreparePass : public LoweringPrepareBase<LoweringPreparePass> {
   void lowerToMemCpy(StoreOp op);
   void lowerArrayDtor(ArrayDtor op);
   void lowerArrayCtor(ArrayCtor op);
+  void lowerThrowOp(ThrowOp op);
 
   /// Collect annotations of global values in the module
   void addGlobalAnnotations(mlir::Operation *op, mlir::ArrayAttr annotations);
@@ -1133,6 +1134,25 @@ void LoweringPreparePass::lowerIterEndOp(IterEndOp op) {
   op.erase();
 }
 
+void LoweringPreparePass::lowerThrowOp(ThrowOp op) {
+  CIRBaseBuilderTy builder(getContext());
+
+  if (op.rethrows()) {
+    auto voidTy = cir::VoidType::get(builder.getContext());
+    auto fnType = cir::FuncType::get({}, voidTy);
+    auto fnName = "__cxa_rethrow";
+
+    builder.setInsertionPointToStart(&theModule.getBodyRegion().front());
+    FuncOp f = buildRuntimeFunction(builder, fnName, op.getLoc(), fnType);
+
+    builder.setInsertionPointAfter(op.getOperation());
+    auto call = builder.createTryCallOp(op.getLoc(), f, {});
+
+    op->replaceAllUsesWith(call);
+    op->erase();
+  }
+}
+
 void LoweringPreparePass::addGlobalAnnotations(mlir::Operation *op,
                                                mlir::ArrayAttr annotations) {
   auto globalValue = cast<mlir::SymbolOpInterface>(op);
@@ -1195,6 +1215,8 @@ void LoweringPreparePass::runOnOp(Operation *op) {
     }
     if (std::optional<mlir::ArrayAttr> annotations = fnOp.getAnnotations())
       addGlobalAnnotations(fnOp, annotations.value());
+  } else if (auto throwOp = dyn_cast<cir::ThrowOp>(op)) {
+    lowerThrowOp(throwOp);
   }
 }
 
@@ -1211,7 +1233,7 @@ void LoweringPreparePass::runOnOperation() {
   op->walk([&](Operation *op) {
     if (isa<UnaryOp, BinOp, CastOp, ComplexBinOp, CmpThreeWayOp, VAArgOp,
             GlobalOp, DynamicCastOp, StdFindOp, IterEndOp, IterBeginOp,
-            ArrayCtor, ArrayDtor, cir::FuncOp, StoreOp>(op))
+            ArrayCtor, ArrayDtor, cir::FuncOp, StoreOp, ThrowOp>(op))
       opsToTransform.push_back(op);
   });
 
diff --git a/clang/test/CIR/CodeGen/throw.cpp b/clang/test/CIR/CodeGen/throw.cpp
index c2395c3725c3..87b34d231767 100644
--- a/clang/test/CIR/CodeGen/throw.cpp
+++ b/clang/test/CIR/CodeGen/throw.cpp
@@ -21,4 +21,180 @@ double d(int a, int b) {
 // LLVM: %[[ADDR:.*]] = call ptr @__cxa_allocate_exception(i64 8)
 // LLVM: store ptr @.str, ptr %[[ADDR]], align 8
 // LLVM: call void @__cxa_throw(ptr %[[ADDR]], ptr @_ZTIPKc, ptr null)
-// LLVM: unreachable
\ No newline at end of file
+// LLVM: unreachable
+
+struct S {
+  S() {}
+};
+
+void refoo1() {
+  int r = 1;
+  try {
+    S s;
+    throw;
+  } catch (...) {
+    ++r;
+  }
+}
+
+// CIR-LABEL: @_Z6refoo1v()
+// CIR:   %[[V0:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["r", init] {alignment = 4 : i64}
+// CIR:   %[[V1:.*]] = cir.const #cir.int<1> : !s32i
+// CIR:   cir.store %[[V1]], %[[V0]] : !s32i, !cir.ptr<!s32i>
+// CIR:   cir.scope {
+// CIR:     %[[V2:.*]] = cir.alloca !ty_S, !cir.ptr<!ty_S>, ["s", init] {alignment = 1 : i64}
+// CIR:     cir.try {
+// CIR:       cir.call exception @_ZN1SC2Ev(%[[V2]]) : (!cir.ptr<!ty_S>) -> ()
+// CIR:       cir.call exception @__cxa_rethrow() : () -> ()
+// CIR:       cir.unreachable
+// CIR:     } catch [type #cir.all {
+// CIR:       %[[V3:.*]] = cir.catch_param -> !cir.ptr<!void>
+// CIR:       %[[V4:.*]] = cir.load %[[V0]] : !cir.ptr<!s32i>, !s32i
+// CIR:       %[[V5:.*]] = cir.unary(inc, %[[V4]]) : !s32i, !s32i
+// CIR:       cir.store %[[V5]], %[[V0]] : !s32i, !cir.ptr<!s32i>
+// CIR:       cir.yield
+// CIR:     }]
+// CIR:   }
+// CIR:   cir.return
+// CIR: }
+
+// LLVM: define dso_local void @_Z6refoo1v()
+// LLVM:   %[[V1:.*]] = alloca %struct.S, i64 1, align 1
+// LLVM:   %[[V2:.*]] = alloca i32, i64 1, align 4
+// LLVM:   store i32 1, ptr %[[V2]], align 4
+// LLVM:   br label %[[B3:.*]]
+// LLVM: [[B3]]:
+// LLVM:   br label %[[B4:.*]]
+// LLVM: [[B4]]:
+// LLVM:   invoke void @_ZN1SC2Ev(ptr %[[V1]])
+// LLVM:           to label %[[B5:.*]] unwind label %[[B7:.*]]
+// LLVM: [[B5]]:
+// LLVM:   invoke void @__cxa_rethrow()
+// LLVM:           to label %[[B6:.*]] unwind label %[[B11:.*]]
+// LLVM: [[B6]]:
+// LLVM:   unreachable
+// LLVM: [[B7]]:
+// LLVM:   %[[V8:.*]] = landingpad { ptr, i32 }
+// LLVM:           catch ptr null
+// LLVM:   %[[V9:.*]] = extractvalue { ptr, i32 } %[[V8]], 0
+// LLVM:   %[[V10:.*]] = extractvalue { ptr, i32 } %[[V8]], 1
+// LLVM:   br label %[[B15:.*]]
+// LLVM: [[B11]]:
+// LLVM:   %[[V12:.*]] = landingpad { ptr, i32 }
+// LLVM:           catch ptr null
+// LLVM:   %[[V13:.*]] = extractvalue { ptr, i32 } %[[V12]], 0
+// LLVM:   %[[V14:.*]] = extractvalue { ptr, i32 } %[[V12]], 1
+// LLVM:   br label %[[B15:.*]]
+// LLVM: [[B15]]:
+// LLVM:   %[[V16:.*]] = phi ptr [ %[[V9]], %[[B7]] ], [ %[[V13]], %[[B11]] ]
+// LLVM:   %[[V17:.*]] = call ptr @__cxa_begin_catch(ptr %[[V16]])
+// LLVM:   %[[V18:.*]] = load i32, ptr %[[V2]], align 4
+// LLVM:   %[[V19:.*]] = add i32 %[[V18]], 1
+// LLVM:   store i32 %[[V19]], ptr %[[V2]], align 4
+// LLVM:   call void @__cxa_end_catch()
+
+void refoo2() {
+  int r = 1;
+  try {
+    for (int i = 0; i < 5; i++) {
+      S s;
+      throw;
+    }
+    S s;
+  } catch (...) {
+    ++r;
+  }
+}
+
+// CIR-LABEL: @_Z6refoo2v()
+// CIR:   %[[V0:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["r", init] {alignment = 4 : i64}
+// CIR:   %[[V1:.*]] = cir.const #cir.int<1> : !s32i
+// CIR:   cir.store %[[V1]], %[[V0]] : !s32i, !cir.ptr<!s32i>
+// CIR:   cir.scope {
+// CIR:     %[[V2:.*]] = cir.alloca !ty_S, !cir.ptr<!ty_S>, ["s", init] {alignment = 1 : i64}
+// CIR:     cir.try {
+// CIR:       cir.scope {
+// CIR:         %[[V3:.*]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["i", init] {alignment = 4 : i64}
+// CIR:         %[[V4:.*]] = cir.const #cir.int<0> : !s32i
+// CIR:         cir.store %[[V4]], %[[V3]] : !s32i, !cir.ptr<!s32i>
+// CIR:         cir.for : cond {
+// CIR:           %[[V5:.*]] = cir.load %[[V3]] : !cir.ptr<!s32i>, !s32i
+// CIR:           %[[V6:.*]] = cir.const #cir.int<5> : !s32i
+// CIR:           %[[V7:.*]] = cir.cmp(lt, %[[V5]], %[[V6]]) : !s32i, !cir.bool
+// CIR:           cir.condition(%[[V7]])
+// CIR:         } body {
+// CIR:           cir.scope {
+// CIR:             %[[V5:.*]] = cir.alloca !ty_S, !cir.ptr<!ty_S>, ["s", init] {alignment = 1 : i64}
+// CIR:             cir.call exception @_ZN1SC2Ev(%[[V5]]) : (!cir.ptr<!ty_S>) -> ()
+// CIR:             cir.call exception @__cxa_rethrow() : () -> ()
+// CIR:             cir.unreachable
+// CIR:           }
+// CIR:           cir.yield
+// CIR:         } step {
+// CIR:           %[[V5:.*]] = cir.load %[[V3]] : !cir.ptr<!s32i>, !s32i
+// CIR:           %[[V6:.*]] = cir.unary(inc, %[[V5]]) : !s32i, !s32i
+// CIR:           cir.store %[[V6]], %[[V3]] : !s32i, !cir.ptr<!s32i>
+// CIR:           cir.yield
+// CIR:         }
+// CIR:       }
+// CIR:       cir.call exception @_ZN1SC2Ev(%[[V2]]) : (!cir.ptr<!ty_S>) -> ()
+// CIR:       cir.yield
+// CIR:     } catch [type #cir.all {
+// CIR:       %[[V3:.*]] = cir.catch_param -> !cir.ptr<!void>
+// CIR:       %[[V4:.*]] = cir.load %[[V0]] : !cir.ptr<!s32i>, !s32i
+// CIR:       %[[V5:.*]] = cir.unary(inc, %[[V4]]) : !s32i, !s32i
+// CIR:       cir.store %[[V5]], %[[V0]] : !s32i, !cir.ptr<!s32i>
+// CIR:       cir.yield
+// CIR:     }]
+// CIR:   }
+// CIR:   cir.return
+// CIR: }
+
+// LLVM: {{.*}}:
+// LLVM:   invoke void @_ZN1SC2Ev(ptr %[[V3:.*]])
+// LLVM:           to label %[[B13:.*]] unwind label %[[B22:.*]]
+// LLVM: [[B13]]:
+// LLVM:   invoke void @__cxa_rethrow()
+// LLVM:           to label %[[B14:.*]] unwind label %[[B26:.*]]
+// LLVM: [[B14]]:
+// LLVM:   unreachable
+// LLVM: [[B15]]:
+// LLVM:   br label %[[B16:.*]]
+// LLVM: [[B16]]:
+// LLVM:   %[[V17]] = load i32, ptr {{.*}}, align 4
+// LLVM:   %[[V18]] = add i32 %[[V17]], 1
+// LLVM:   store i32 %[[V18]], ptr {{.*}}, align 4
+// LLVM:   br label {{.*}}
+// LLVM: %[[B19:.*]]
+// LLVM:   br label %[[B20:.*]]
+// LLVM: [[B20]]:
+// LLVM:   invoke void @_ZN1SC2Ev(ptr {{.*}})
+// LLVM:           to label %[[B21:.*]] unwind label %[[B30:.*]]
+// LLVM: [[B21]]:
+// LLVM:   br label {{.*}}
+// LLVM: [[B22]]:
+// LLVM:   %[[V23:.*]] = landingpad { ptr, i32 }
+// LLVM:           catch ptr null
+// LLVM:   %[[V24:.*]] = extractvalue { ptr, i32 } %[[V23]], 0
+// LLVM:   %[[V25:.*]] = extractvalue { ptr, i32 } %[[V23]], 1
+// LLVM:   br label %[[B34:.*]]
+// LLVM: [[B26]]:
+// LLVM:   %[[V27:.*]] = landingpad { ptr, i32 }
+// LLVM:           catch ptr null
+// LLVM:   %[[V28:.*]] = extractvalue { ptr, i32 } %[[V27]], 0
+// LLVM:   %[[V29:.*]] = extractvalue { ptr, i32 } %[[V27]], 1
+// LLVM:   br label %[[B34:.*]]
+// LLVM: [[B30]]:
+// LLVM:   %[[V31:.*]] = landingpad { ptr, i32 }
+// LLVM:           catch ptr null
+// LLVM:   %[[V32:.*]] = extractvalue { ptr, i32 } %[[V31]], 0
+// LLVM:   %[[V33:.*]] = extractvalue { ptr, i32 } %[[V31]], 1
+// LLVM:   br label %[[B34:.*]]
+// LLVM: [[B34]]:
+// LLVM:   %[[V35:.*]] = phi ptr [ %[[V32]], %[[B30]] ], [ %[[V24]], %[[B22]] ], [ %[[V28]], %[[B26]] ]
+// LLVM:   %[[V36:.*]] = call ptr @__cxa_begin_catch(ptr %[[V35]])
+// LLVM:   %[[V37:.*]] = load i32, ptr {{.*}}, align 4
+// LLVM:   %[[V38:.*]] = add i32 %[[V37]], 1
+// LLVM:   store i32 %[[V38]], ptr {{.*}}, align 4
+// LLVM:   call void @__cxa_end_catch()
+// LLVM:   br label {{.*}}