diff --git a/clang/lib/CodeGen/CGCoroutine.cpp b/clang/lib/CodeGen/CGCoroutine.cpp
--- a/clang/lib/CodeGen/CGCoroutine.cpp
+++ b/clang/lib/CodeGen/CGCoroutine.cpp
@@ -245,10 +245,39 @@
   }
 
   LValueOrRValue Res;
-  if (forLValue)
+  if (forLValue) {
     Res.LV = CGF.EmitLValue(S.getResumeExpr());
-  else
+  } else {
+    // If the result of co_await is stored into a memory address, we want to
+    // make sure the lifetime of that memory address does not start too early,
+    // causing more space to be allocated in the frame rather than on stack.
+    // The lifetime of the return value of co_await should start here right
+    // before we attempt to assign it.
+    auto adjustLifetimeStart = [&]() {
+      // aggSlot points to the instruction that allocated the object. Latter
+      // instructions use it in this pattern:
+      //   %tmpX = alloca %.., align 4
+      //   %0 = bitcast %...* %tmpX to i8*
+      //   call void @llvm.lifetime.start.p0i8(i64 ..., i8* nonnull %0) #2
+      // Hence we trace back through uses to eventually locate the lifetime
+      // start intrinsics marker, and move it down to the current insertion
+      // point.
+      auto *AllocaInst =
+          dyn_cast_or_null<llvm::AllocaInst>(aggSlot.getPointer());
+      if (!AllocaInst || AllocaInst->getNumUses() != 1)
+        return;
+      auto *CastInst = dyn_cast<llvm::CastInst>(*AllocaInst->users().begin());
+      if (!CastInst || CastInst->getNumUses() != 1)
+        return;
+      if (auto *LifetimeInst =
+              dyn_cast<llvm::IntrinsicInst>(*CastInst->users().begin())) {
+        LifetimeInst->removeFromParent();
+        CGF.Builder.Insert(LifetimeInst);
+      }
+    };
+    adjustLifetimeStart();
     Res.RV = CGF.EmitAnyExpr(S.getResumeExpr(), aggSlot, ignoreResult);
+  }
 
   if (TryStmt) {
     Builder.CreateFlagStore(false, Coro.ResumeEHVar);