diff --git a/llvm/test/CodeGen/AMDGPU/invariant-image-load.ll b/llvm/test/CodeGen/AMDGPU/invariant-image-load.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/invariant-image-load.ll
@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx900 < %s | FileCheck %s -check-prefix=GFX9
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx900 < %s | FileCheck %s -check-prefix=GFX9
+
+; Marking the image loads as invariant should allow both loads to be hoisted
+; above both stores.
+define amdgpu_ps void @test(<8 x i32> inreg %load, <8 x i32> inreg %store) {
+; GFX9-LABEL: test:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    image_load v1, v0, s[0:7] dmask:0x1 unorm
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    image_store v1, v0, s[8:15] dmask:0x1 unorm
+; GFX9-NEXT:    v_mov_b32_e32 v0, 1
+; GFX9-NEXT:    image_load v1, v0, s[0:7] dmask:0x1 unorm
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    image_store v1, v0, s[8:15] dmask:0x1 unorm
+; GFX9-NEXT:    s_endpgm
+  %data0 = call float @llvm.amdgcn.image.load.1d.f32.i32(i32 1, i32 0, <8 x i32> %load, i32 0, i32 0), !invariant.load !0
+  call void @llvm.amdgcn.image.store.1d.f32.i32(float %data0, i32 1, i32 0, <8 x i32> %store, i32 0, i32 0)
+  %data1 = call float @llvm.amdgcn.image.load.1d.f32.i32(i32 1, i32 1, <8 x i32> %load, i32 0, i32 0), !invariant.load !0
+  call void @llvm.amdgcn.image.store.1d.f32.i32(float %data1, i32 1, i32 1, <8 x i32> %store, i32 0, i32 0)
+  ret void
+}
+
+declare float @llvm.amdgcn.image.load.1d.f32.i32(i32 immarg, i32, <8 x i32>, i32 immarg, i32 immarg)
+declare void @llvm.amdgcn.image.store.1d.f32.i32(float, i32 immarg, i32, <8 x i32>, i32 immarg, i32 immarg)
+
+!0 = !{}