Index: llvm/trunk/include/llvm/IR/Intrinsics.td
===================================================================
--- llvm/trunk/include/llvm/IR/Intrinsics.td
+++ llvm/trunk/include/llvm/IR/Intrinsics.td
@@ -615,9 +615,16 @@
                                      llvm_anyptr_ty],
                                     [IntrArgMemOnly, NoCapture<2>]>;
 
+// invariant.group.barrier can't be marked with 'readnone' (IntrNoMem),
+// because it would cause CSE of two barriers with the same argument.
+// Readonly and argmemonly says that barrier only reads its argument and
+// it can be CSE only if memory didn't change between 2 barriers call,
+// which is valid.
+// The argument also can't be marked with 'returned' attribute, because
+// it would remove barrier. 
 def int_invariant_group_barrier : Intrinsic<[llvm_ptr_ty],
                                             [llvm_ptr_ty],
-                                            [IntrNoMem]>;
+                                            [IntrReadMem, IntrArgMemOnly]>;
 
 //===------------------------ Stackmap Intrinsics -------------------------===//
 //
Index: llvm/trunk/test/Analysis/MemorySSA/invariant-groups.ll
===================================================================
--- llvm/trunk/test/Analysis/MemorySSA/invariant-groups.ll
+++ llvm/trunk/test/Analysis/MemorySSA/invariant-groups.ll
@@ -16,6 +16,8 @@
   store i32 1, i32* @g, align 4
 
   %1 = bitcast i32* %a to i8*
+; CHECK: MemoryUse(2)
+; CHECK-NEXT: %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a32 = bitcast i8* %a8 to i32*
 
@@ -33,6 +35,8 @@
   store i32 0, i32* %a, align 4, !invariant.group !0
 
   %1 = bitcast i32* %a to i8*
+; CHECK: MemoryUse(1)
+; CHECK-NEXT: %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a32 = bitcast i8* %a8 to i32*
 
@@ -50,6 +54,8 @@
   %v = load i32, i32* %a, align 4, !invariant.group !0
 
   %1 = bitcast i32* %a to i8*
+; CHECK: MemoryUse(liveOnEntry)
+; CHECK-NEXT: %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a32 = bitcast i8* %a8 to i32*
 
@@ -79,6 +85,8 @@
 ; CHECK-NEXT: store i32 1
   store i32 1, i32* @g, align 4
   %1 = bitcast i32* %a to i8*
+; CHECK: MemoryUse(2)
+; CHECK-NEXT: %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a8 = call i8* @llvm.invariant.group.barrier(i8* %1)
   %a32 = bitcast i8* %a8 to i32*
 
@@ -135,6 +143,9 @@
 ; CHECK: 2 = MemoryDef(1)
 ; CHECK-NEXT: call void @clobber
   call void @clobber8(i8* %p)
+
+; CHECK: MemoryUse(2)
+; CHECK-NEXT: %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   br i1 undef, label %Loop.Body, label %Loop.End
 
@@ -179,6 +190,9 @@
 ; CHECK: 2 = MemoryDef(1)
 ; CHECK-NEXT: call void @clobber
   call void @clobber8(i8* %p)
+
+; CHECK: MemoryUse(2)
+; CHECK-NEXT: %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   br i1 undef, label %Loop.Body, label %Loop.End
 
@@ -238,6 +252,8 @@
 ; CHECK: 2 = MemoryDef(1)
 ; CHECK-NEXT: call void @clobber
   call void @clobber8(i8* %p)
+; CHECK: MemoryUse(2)
+; CHECK-NEXT: %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   %after = call i8* @llvm.invariant.group.barrier(i8* %p)
   br i1 undef, label %Loop.Pre, label %Loop.End
 
Index: llvm/trunk/test/Other/invariant.group.barrier.ll
===================================================================
--- llvm/trunk/test/Other/invariant.group.barrier.ll
+++ llvm/trunk/test/Other/invariant.group.barrier.ll
@@ -0,0 +1,62 @@
+; RUN: opt -S -early-cse < %s | FileCheck %s
+; RUN: opt -S -gvn < %s | FileCheck %s
+; RUN: opt -S -newgvn < %s | FileCheck %s
+; RUN: opt -S -O3 < %s | FileCheck %s
+
+; These tests checks if passes with CSE functionality can do CSE on
+; invariant.group.barrier, that is prohibited if there is a memory clobber
+; between barriers call.
+
+; CHECK-LABEL: define i8 @optimizable()
+define i8 @optimizable() {
+entry:
+    %ptr = alloca i8
+    store i8 42, i8* %ptr, !invariant.group !0
+; CHECK: call i8* @llvm.invariant.group.barrier
+    %ptr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
+; CHECK-NOT: call i8* @llvm.invariant.group.barrier
+    %ptr3 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
+; CHECK: call void @clobber(i8* {{.*}}%ptr)
+    call void @clobber(i8* %ptr)
+
+; CHECK: call void @use(i8* {{.*}}%ptr2)
+    call void @use(i8* %ptr2)
+; CHECK: call void @use(i8* {{.*}}%ptr2)
+    call void @use(i8* %ptr3)
+; CHECK: load i8, i8* %ptr2, {{.*}}!invariant.group
+    %v = load i8, i8* %ptr3, !invariant.group !0
+
+    ret i8 %v
+}
+
+; CHECK-LABEL: define i8 @unoptimizable()
+define i8 @unoptimizable() {
+entry:
+    %ptr = alloca i8
+    store i8 42, i8* %ptr, !invariant.group !0
+; CHECK: call i8* @llvm.invariant.group.barrier
+    %ptr2 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
+    call void @clobber(i8* %ptr)
+; CHECK: call i8* @llvm.invariant.group.barrier
+    %ptr3 = call i8* @llvm.invariant.group.barrier(i8* %ptr)
+; CHECK: call void @clobber(i8* {{.*}}%ptr)
+    call void @clobber(i8* %ptr)
+; CHECK: call void @use(i8* {{.*}}%ptr2)
+    call void @use(i8* %ptr2)
+; CHECK: call void @use(i8* {{.*}}%ptr3)
+    call void @use(i8* %ptr3)
+; CHECK: load i8, i8* %ptr3, {{.*}}!invariant.group
+    %v = load i8, i8* %ptr3, !invariant.group !0
+
+    ret i8 %v
+}
+
+declare void @use(i8* readonly)
+
+declare void @clobber(i8*)
+; CHECK: Function Attrs: argmemonly nounwind readonly
+; CHECK-NEXT: declare i8* @llvm.invariant.group.barrier(i8*)
+declare i8* @llvm.invariant.group.barrier(i8*)
+
+!0 = !{}
+