Index: clang/lib/CodeGen/CGCall.cpp
===================================================================
--- clang/lib/CodeGen/CGCall.cpp
+++ clang/lib/CodeGen/CGCall.cpp
@@ -3878,6 +3878,11 @@
   Address swiftErrorTemp = Address::invalid();
   Address swiftErrorArg = Address::invalid();
 
+  // When passing arguments using temporary allocas, we need to add the
+  // appropriate lifetime markers. This vector keeps track of all the lifetime
+  // markers that need to be ended right after the call.
+  SmallVector<CallLifetimeEnd, 2> CallLifetimeEndAfterCall;
+
   // Translate all of the arguments as necessary to match the IR lowering.
   assert(CallInfo.arg_size() == CallArgs.size() &&
          "Mismatch between function signature & arguments.");
@@ -3994,7 +3999,20 @@
           Address AI = CreateMemTempWithoutCast(
               I->Ty, ArgInfo.getIndirectAlign(), "byval-temp");
           IRCallArgs[FirstIRArg] = AI.getPointer();
+
+          // Emit lifetime markers for the temporary alloca.
+          uint64_t ByvalTempElementSize =
+              CGM.getDataLayout().getTypeAllocSize(AI.getElementType());
+          llvm::Value *LifetimeSize =
+              EmitLifetimeStart(ByvalTempElementSize, AI.getPointer());
+
+          // Generate the copy.
           I->copyInto(*this, AI);
+
+          // Add cleanup code to emit the end lifetime marker after the call.
+          if (LifetimeSize) // In case we disabled lifetime markers.
+            CallLifetimeEndAfterCall.emplace_back(AI, LifetimeSize);
+
         } else {
           // Skip the extra memcpy call.
           auto *T = V->getType()->getPointerElementType()->getPointerTo(
@@ -4562,6 +4580,11 @@
     }
   }
 
+  // Explicitly call CallLifetimeEnd::Emit just to re-use the code even though
+  // we can't use the full cleanup mechanism.
+  for (CallLifetimeEnd &LifetimeEnd : CallLifetimeEndAfterCall)
+    LifetimeEnd.Emit(*this, /*Flags=*/{});
+
   return Ret;
 }
 
Index: clang/test/CodeGen/aarch64-byval-temp.c
===================================================================
--- /dev/null
+++ clang/test/CodeGen/aarch64-byval-temp.c
@@ -0,0 +1,89 @@
+// RUN: %clang_cc1 -emit-llvm -triple arm64-- -o - %s -O0 | FileCheck %s --check-prefix=CHECK-O0
+// RUN: %clang_cc1 -emit-llvm -disable-llvm-optzns -triple arm64-- -o - %s -O3 | FileCheck %s --check-prefix=CHECK-O3
+
+struct large {
+    void* pointers[8];
+};
+
+void pass_large(struct large);
+
+// For arm64, we don't use byval to pass structs but instead we create
+// temporary allocas.
+//
+// Make sure we generate the appropriate lifetime markers for the temporary
+// allocas so that the optimizer can re-use stack slots if possible.
+void example() {
+    struct large l = {0};
+    pass_large(l);
+    pass_large(l);
+}
+// CHECK-O0-LABEL: define void @example(
+// The alloca for the struct on the stack.
+// CHECK-O0: %[[l:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+// The alloca for the temporary stack space that we use to pass the argument.
+// CHECK-O0-NEXT: %[[byvaltemp:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+// Another one to pass the argument to the second function call.
+// CHECK-O0-NEXT: %[[byvaltemp1:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+// First, memset `l` to 0.
+// CHECK-O0-NEXT: %[[bitcastl:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O0-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 %[[bitcastl]], i8 0, i64 64, i1 false)
+// Then, memcpy `l` to the temporary stack space.
+// CHECK-O0-NEXT: %[[src:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp]] to i8*
+// CHECK-O0-NEXT: %[[dst:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O0-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %[[src]], i8* align 8 %[[dst]], i64 64, i1 false)
+// Finally, call using a pointer to the temporary stack space.
+// CHECK-O0-NEXT: call void @pass_large(%struct.large* %[[byvaltemp]])
+// Now, do the same for the second call, using the second temporary alloca.
+// CHECK-O0-NEXT: %[[src:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp1]] to i8*
+// CHECK-O0-NEXT: %[[dst:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O0-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %[[src]], i8* align 8 %[[dst]], i64 64, i1 false)
+// CHECK-O0-NEXT: call void @pass_large(%struct.large* %[[byvaltemp1]])
+// CHECK-O0-NEXT: ret void
+//
+// At O3, we should have lifetime markers to help the optimizer re-use the temporary allocas.
+//
+// CHECK-O3-LABEL: define void @example(
+// The alloca for the struct on the stack.
+// CHECK-O3: %[[l:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+// The alloca for the temporary stack space that we use to pass the argument.
+// CHECK-O3-NEXT: %[[byvaltemp:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+// Another one to pass the argument to the second function call.
+// CHECK-O3-NEXT: %[[byvaltemp1:[0-9A-Za-z-]+]] = alloca %struct.large, align 8
+//
+// Mark the start of the lifetime for `l`
+// CHECK-O3-NEXT: %[[bitcastl:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.start.p0i8(i64 64, i8* %[[bitcastl]])
+//
+// First, memset `l` to 0.
+// CHECK-O3-NEXT: %[[bitcastl:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O3-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 %[[bitcastl]], i8 0, i64 64, i1 false)
+//
+// Lifetime of the first temporary starts here and ends right after the call.
+// CHECK-O3-NEXT: %[[bitcastbyvaltemp:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp]] to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.start.p0i8(i64 64, i8* %[[bitcastbyvaltemp]])
+//
+// Then, memcpy `l` to the temporary stack space.
+// CHECK-O3-NEXT: %[[src:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp]] to i8*
+// CHECK-O3-NEXT: %[[dst:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %[[src]], i8* align 8 %[[dst]], i64 64, i1 false)
+// Finally, call using a pointer to the temporary stack space.
+// CHECK-O3-NEXT: call void @pass_large(%struct.large* %[[byvaltemp]])
+//
+// The lifetime of the temporary used to pass a pointer to the struct ends here.
+// CHECK-O3-NEXT: %[[bitcastbyvaltemp:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp]] to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.end.p0i8(i64 64, i8* %[[bitcastbyvaltemp]])
+//
+// Now, do the same for the second call, using the second temporary alloca.
+// CHECK-O3-NEXT: %[[bitcastbyvaltemp:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp1]] to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.start.p0i8(i64 64, i8* %[[bitcastbyvaltemp]])
+// CHECK-O3-NEXT: %[[src:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp1]] to i8*
+// CHECK-O3-NEXT: %[[dst:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[l]] to i8*
+// CHECK-O3-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %[[src]], i8* align 8 %[[dst]], i64 64, i1 false)
+// CHECK-O3-NEXT: call void @pass_large(%struct.large* %[[byvaltemp1]])
+// CHECK-O3-NEXT: %[[bitcastbyvaltemp:[0-9A-Za-z-]+]] = bitcast %struct.large* %[[byvaltemp1]] to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.end.p0i8(i64 64, i8* %[[bitcastbyvaltemp]])
+//
+// Mark the end of the lifetime of `l`.
+// CHECK-O3-NEXT: %[[bitcastl:[0-9A-Za-z-]+]] = bitcast %struct.large* %l to i8*
+// CHECK-O3-NEXT: call void @llvm.lifetime.end.p0i8(i64 64, i8* %[[bitcastl]])
+// CHECK-O3-NEXT: ret void