[Clang][LLVM][Coroutines] Prevent __coro_gro from outliving __promise

clang/lib/CodeGen/CGCoroutine.cpp

@@ -535,6 +535,11 @@ struct GetReturnObjectManager {
     Builder.CreateStore(Builder.getFalse(), GroActiveFlag);
 
     GroEmission = CGF.EmitAutoVarAlloca(*GroVarDecl);
+    auto *GroAlloca = dyn_cast_or_null<llvm::AllocaInst>(
+        GroEmission.getOriginalAllocatedAddress().getPointer());
+    assert(GroAlloca && "expected alloca to be emitted");
+    GroAlloca->setMetadata(llvm::LLVMContext::MD_coro_outside_frame,
+                           llvm::MDNode::get(CGF.CGM.getLLVMContext(), {}));
 
     // Remember the top of EHStack before emitting the cleanup.
     auto old_top = CGF.EHStack.stable_begin();

clang/test/CodeGenCoroutines/coro-gro.cpp

@@ -30,12 +30,13 @@ void doSomething() noexcept;
 int f() {
   // CHECK: %[[RetVal:.+]] = alloca i32
   // CHECK: %[[GroActive:.+]] = alloca i1
+  // CHECK: %[[CoroGro:.+]] = alloca %struct.GroType, {{.*}} !coro.outside.frame ![[OutFrameMetadata:.+]]
 
   // CHECK: %[[Size:.+]] = call i64 @llvm.coro.size.i64()
   // CHECK: call noalias noundef nonnull ptr @_Znwm(i64 noundef %[[Size]])
   // CHECK: store i1 false, ptr %[[GroActive]]
   // CHECK: call void @_ZNSt16coroutine_traitsIiJEE12promise_typeC1Ev(
-  // CHECK: call void @_ZNSt16coroutine_traitsIiJEE12promise_type17get_return_objectEv(
+  // CHECK: call void @_ZNSt16coroutine_traitsIiJEE12promise_type17get_return_objectEv({{.*}} %[[CoroGro]]
   // CHECK: store i1 true, ptr %[[GroActive]]
 
   Cleanup cleanup;
@@ -104,3 +105,4 @@ invoker g() {
   // CHECK: call void @_ZN7invoker15invoker_promise17get_return_objectEv({{.*}} %[[AggRes]]
   co_return;
 }
+// CHECK: ![[OutFrameMetadata]] = !{}

llvm/docs/Coroutines.rst

@@ -1775,6 +1775,25 @@ CoroCleanup
 This pass runs late to lower all coroutine related intrinsics not replaced by
 earlier passes.
 
+Metadata
+========
+
+'``coro.outside.frame``' Metadata
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``coro.outside.frame`` metadata may be attached to an alloca instruction to
+to signify that it shouldn't be promoted to the coroutine frame, useful for
+filtering allocas out by the frontend when emitting internal control mechanisms.
+Additionally, this metadata is only used as a flag, so the associated
+node must be empty.
+
+.. code-block:: text
+
+  %__coro_gro = alloca %struct.GroType, align 1, !coro.outside.frame !0
+
+  ...
+  !0 = !{}
+
 Areas Requiring Attention
 =========================
 #. When coro.suspend returns -1, the coroutine is suspended, and it's possible

llvm/include/llvm/IR/FixedMetadataKinds.def

@@ -50,3 +50,4 @@ LLVM_FIXED_MD_KIND(MD_callsite, "callsite", 35)
 LLVM_FIXED_MD_KIND(MD_kcfi_type, "kcfi_type", 36)
 LLVM_FIXED_MD_KIND(MD_pcsections, "pcsections", 37)
 LLVM_FIXED_MD_KIND(MD_DIAssignID, "DIAssignID", 38)
+LLVM_FIXED_MD_KIND(MD_coro_outside_frame, "coro.outside.frame", 39)

llvm/lib/Transforms/Coroutines/CoroFrame.cpp

@@ -2804,6 +2804,11 @@ static void collectFrameAlloca(AllocaInst *AI, coro::Shape &Shape,
   if (AI == Shape.SwitchLowering.PromiseAlloca)
     return;
 
+  // The __coro_gro alloca should outlive the promise, make sure we
+  // keep it outside the frame.
+  if (MDNode *MD = AI->getMetadata(LLVMContext::MD_coro_outside_frame))
+    return;
+
   // The code that uses lifetime.start intrinsic does not work for functions
   // with loops without exit. Disable it on ABIs we know to generate such
   // code.

llvm/test/Transforms/Coroutines/coro-alloca-outside-frame.ll

@@ -0,0 +1,61 @@
+; Tests that CoroSplit can succesfully skip allocas that shall not live on the frame
+; RUN: opt < %s -passes='cgscc(coro-split),simplifycfg,early-cse' -S -o %t.ll
+; RUN: FileCheck --input-file=%t.ll %s
+
+define ptr @f(i1 %n) presplitcoroutine {
+entry:
+  %x = alloca i64, !coro.outside.frame !{}
+  %y = alloca i64
+  %id = call token @llvm.coro.id(i32 0, ptr null, ptr null, ptr null)
+  %size = call i32 @llvm.coro.size.i32()
+  %alloc = call ptr @malloc(i32 %size)
+  %hdl = call ptr @llvm.coro.begin(token %id, ptr %alloc)
+  br i1 %n, label %flag_true, label %flag_false
+
+flag_true:
+  br label %merge
+
+flag_false:
+  br label %merge
+
+merge:
+  %alias_phi = phi ptr [ %x, %flag_true ], [ %y, %flag_false ]
+  %sp1 = call i8 @llvm.coro.suspend(token none, i1 false)
+  switch i8 %sp1, label %suspend [i8 0, label %resume
+                                  i8 1, label %cleanup]
+resume:
+  call void @print(ptr %alias_phi)
+  br label %cleanup
+
+cleanup:
+  %mem = call ptr @llvm.coro.free(token %id, ptr %hdl)
+  call void @free(ptr %mem)
+  br label %suspend
+
+suspend:
+  call i1 @llvm.coro.end(ptr %hdl, i1 0, token none)
+  ret ptr %hdl
+}
+
+; %y and %alias_phi would all go to the frame, but not %x
+; CHECK:       %f.Frame = type { ptr, ptr, i64, ptr, i1 }
+; CHECK-LABEL: @f(
+; CHECK:         %x = alloca i64, align 8, !coro.outside.frame !0
+; CHECK-NOT:     %x.reload.addr = getelementptr inbounds %f.Frame, ptr %hdl, i32 0, i32 2
+; CHECK:         %y.reload.addr = getelementptr inbounds %f.Frame, ptr %hdl, i32 0, i32 2
+; CHECK:         %alias_phi = phi ptr [ %y.reload.addr, %merge.from.flag_false ], [ %x, %entry ]
+
+declare ptr @llvm.coro.free(token, ptr)
+declare i32 @llvm.coro.size.i32()
+declare i8  @llvm.coro.suspend(token, i1)
+declare void @llvm.coro.resume(ptr)
+declare void @llvm.coro.destroy(ptr)
+
+declare token @llvm.coro.id(i32, ptr, ptr, ptr)
+declare i1 @llvm.coro.alloc(token)
+declare ptr @llvm.coro.begin(token, ptr)
+declare i1 @llvm.coro.end(ptr, i1, token)
+
+declare void @print(ptr)
+declare noalias ptr @malloc(i32)
+declare void @free(ptr)