Index: lib/Target/AMDGPU/AMDGPUConvertAtomicLibCalls.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUConvertAtomicLibCalls.cpp
+++ lib/Target/AMDGPU/AMDGPUConvertAtomicLibCalls.cpp
@@ -542,7 +542,7 @@
     dyn_cast<AtomicRMWInst>(NI)->setSynchScope(
         (SynchronizationScope)OCL1XAtomicScope.getValue());
     if (NeedCast) {
-      NI = Builder.CreateBitCast(NI, F->getType());
+      NI = Builder.CreateBitCast(NI, F->getReturnType());
     }
   } else if (Type == CMPXCHG) {
     NI = Builder.CreateAtomicCmpXchg(
Index: test/CodeGen/AMDGPU/opencl-1.2-atomics.ll
===================================================================
--- test/CodeGen/AMDGPU/opencl-1.2-atomics.ll
+++ test/CodeGen/AMDGPU/opencl-1.2-atomics.ll
@@ -3,15 +3,20 @@
 ; CHECK: atomicrmw add i32 addrspace(1)* null, i32 0 synchscope(2) monotonic
 ; CHECK: cmpxchg i32 addrspace(1)* null, i32 0, i32 0 synchscope(2) monotonic monotonic
 ; CHECK: atomicrmw xchg i32 addrspace(1)* null, i32 0 synchscope(2) monotonic
+; CHECK: [[A:%[0-9]*]] = atomicrmw xchg i32 addrspace(3)* null, i32 0 synchscope(2) monotonic
+; CHECK: bitcast i32 [[A]] to float
+
 define amdgpu_kernel void @test() {
 entry:
   %call0 = call i32 @_Z8atom_incPU3AS1Vi(i32 addrspace(1)* null)
   %call1 = call i32 @_Z10atomic_addPU3AS1Vii(i32 addrspace(1)* null, i32 0)
   %call2 = call i32 @_Z14atomic_cmpxchgPU3AS1Viii(i32 addrspace(1)* null, i32 0, i32 0)
   %call3 = call i32 @_Z11atomic_xchgPU3AS1Vii(i32 addrspace(1)* null, i32 0)
+  %call4 = call float @_Z11atomic_xchgPU3AS3Vff(float addrspace(3)* null, float 0.0)
   ret void
 }
 declare i32 @_Z8atom_incPU3AS1Vi(i32 addrspace(1)*)
 declare i32 @_Z10atomic_addPU3AS1Vii(i32 addrspace(1)*, i32)
 declare i32 @_Z14atomic_cmpxchgPU3AS1Viii(i32 addrspace(1)*, i32, i32)
 declare i32 @_Z11atomic_xchgPU3AS1Vii(i32 addrspace(1)*, i32)
+declare float @_Z11atomic_xchgPU3AS3Vff(float addrspace(3)*, float)