Index: llvm/trunk/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ llvm/trunk/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -568,8 +568,6 @@
   const SITargetLowering& Lowering =
     *static_cast<const SITargetLowering*>(getTargetLowering());
 
-  // Write max value to m0 before each load operation
-
   assert(N->getOperand(0).getValueType() == MVT::Other && "Expected chain");
 
   SDValue M0 = Lowering.copyToM0(*CurDAG, N->getOperand(0), SDLoc(N),
@@ -587,10 +585,17 @@
 }
 
 SDNode *AMDGPUDAGToDAGISel::glueCopyToM0LDSInit(SDNode *N) const {
-  if (cast<MemSDNode>(N)->getAddressSpace() != AMDGPUAS::LOCAL_ADDRESS ||
-      !Subtarget->ldsRequiresM0Init())
-    return N;
-  return glueCopyToM0(N, CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
+  unsigned AS = cast<MemSDNode>(N)->getAddressSpace();
+  if (AS == AMDGPUAS::LOCAL_ADDRESS) {
+    if (Subtarget->ldsRequiresM0Init())
+      return glueCopyToM0(N, CurDAG->getTargetConstant(-1, SDLoc(N), MVT::i32));
+  } else if (AS == AMDGPUAS::REGION_ADDRESS) {
+    MachineFunction &MF = CurDAG->getMachineFunction();
+    unsigned Value = MF.getInfo<SIMachineFunctionInfo>()->getGDSSize();
+    return
+        glueCopyToM0(N, CurDAG->getTargetConstant(Value, SDLoc(N), MVT::i32));
+  }
+  return N;
 }
 
 MachineSDNode *AMDGPUDAGToDAGISel::buildSMovImm64(SDLoc &DL, uint64_t Imm,
Index: llvm/trunk/lib/Target/AMDGPU/AMDGPUInstructions.td
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ llvm/trunk/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -353,6 +353,10 @@
   return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
 }]>;
 
+class RegionAddress : CodePatPred<[{
+  return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUAS::REGION_ADDRESS;
+}]>;
+
 class GlobalAddress : CodePatPred<[{
   return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS;
 }]>;
@@ -402,6 +406,9 @@
 class LocalLoad <SDPatternOperator op> : LoadFrag <op>, LocalAddress;
 class LocalStore <SDPatternOperator op> : StoreFrag <op>, LocalAddress;
 
+class RegionLoad <SDPatternOperator op> : LoadFrag <op>, RegionAddress;
+class RegionStore <SDPatternOperator op> : StoreFrag <op>, RegionAddress;
+
 class GlobalLoad <SDPatternOperator op> : LoadFrag<op>, GlobalLoadAddress;
 class GlobalStore <SDPatternOperator op> : StoreFrag<op>, GlobalAddress;
 
@@ -497,6 +504,13 @@
   return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
 }]>;
 
+class region_binary_atomic_op<SDNode atomic_op> :
+  PatFrag<(ops node:$ptr, node:$value),
+    (atomic_op node:$ptr, node:$value), [{
+  return cast<MemSDNode>(N)->getAddressSpace() == AMDGPUAS::REGION_ADDRESS;
+}]>;
+
+
 def atomic_swap_local : local_binary_atomic_op<atomic_swap>;
 def atomic_load_add_local : local_binary_atomic_op<atomic_load_add>;
 def atomic_load_sub_local : local_binary_atomic_op<atomic_load_sub>;
@@ -521,6 +535,13 @@
       return AN->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
 }]>;
 
+class AtomicCmpSwapRegion <SDNode cmp_swap_node> : PatFrag<
+    (ops node:$ptr, node:$cmp, node:$swap),
+    (cmp_swap_node node:$ptr, node:$cmp, node:$swap), [{
+      AtomicSDNode *AN = cast<AtomicSDNode>(N);
+      return AN->getAddressSpace() == AMDGPUAS::REGION_ADDRESS;
+}]>;
+
 def atomic_cmp_swap_local : AtomicCmpSwapLocal <atomic_cmp_swap>;
 
 multiclass global_binary_atomic_op<SDNode atomic_op> {
Index: llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -142,7 +142,7 @@
       unsigned Threshold = 0;
       if (AS == AMDGPUAS::PRIVATE_ADDRESS)
         Threshold = ThresholdPrivate;
-      else if (AS == AMDGPUAS::LOCAL_ADDRESS)
+      else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS)
         Threshold = ThresholdLocal;
       else
         continue;
@@ -160,7 +160,8 @@
         unsigned AllocaSize = Ty->isSized() ? DL.getTypeAllocSize(Ty) : 0;
         if (AllocaSize > MaxAlloca)
           continue;
-      } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
+      } else if (AS == AMDGPUAS::LOCAL_ADDRESS ||
+                 AS == AMDGPUAS::REGION_ADDRESS) {
         LocalGEPsSeen++;
         // Inhibit unroll for local memory if we have seen addressing not to
         // a variable, most likely we will be unable to combine it.
Index: llvm/trunk/lib/Target/AMDGPU/DSInstructions.td
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/DSInstructions.td
+++ llvm/trunk/lib/Target/AMDGPU/DSInstructions.td
@@ -601,9 +601,9 @@
   (DS_SWIZZLE_B32 $src, (as_i16imm $offset16), (i1 0))
 >;
 
-class DSReadPat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
+class DSReadPat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat <
   (vt (frag (DS1Addr1Offset i32:$ptr, i32:$offset))),
-  (inst $ptr, (as_i16imm $offset), (i1 0))
+  (inst $ptr, (as_i16imm $offset), (i1 gds))
 >;
 
 multiclass DSReadPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
@@ -657,9 +657,9 @@
 def : DSReadPat_D16<DS_READ_I8_D16, sextloadi8_d16_lo_local, v2f16>;
 }
 
-class DSWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
+class DSWritePat <DS_Pseudo inst, ValueType vt, PatFrag frag, int gds=0> : GCNPat <
   (frag vt:$value, (DS1Addr1Offset i32:$ptr, i32:$offset)),
-  (inst $ptr, $value, (as_i16imm $offset), (i1 0))
+  (inst $ptr, $value, (as_i16imm $offset), (i1 gds))
 >;
 
 multiclass DSWritePat_mc <DS_Pseudo inst, ValueType vt, string frag> {
@@ -734,75 +734,79 @@
 defm : DSWritePat_mc <DS_WRITE_B128, v4i32, "store_align16_local">;
 
 } // End AddedComplexity = 100
-class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
+class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat <
   (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
-  (inst $ptr, $value, (as_i16imm $offset), (i1 0))
+  (inst $ptr, $value, (as_i16imm $offset), (i1 gds))
 >;
 
 multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
   let OtherPredicates = [LDSRequiresM0Init] in {
-    def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_m0")>;
+    def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_local_m0")>;
   }
 
   let OtherPredicates = [NotLDSRequiresM0Init] in {
     def : DSAtomicRetPat<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt,
-                         !cast<PatFrag>(frag)>;
+                         !cast<PatFrag>(frag#"_local")>;
   }
+
+  def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_region_m0"), 1>;
 }
 
 
 
-class DSAtomicCmpXChg<DS_Pseudo inst, ValueType vt, PatFrag frag> : GCNPat <
+class DSAtomicCmpXChg<DS_Pseudo inst, ValueType vt, PatFrag frag, bit gds=0> : GCNPat <
   (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$cmp, vt:$swap),
-  (inst $ptr, $cmp, $swap, (as_i16imm $offset), (i1 0))
+  (inst $ptr, $cmp, $swap, (as_i16imm $offset), (i1 gds))
 >;
 
 multiclass DSAtomicCmpXChg_mc<DS_Pseudo inst, ValueType vt, string frag> {
   let OtherPredicates = [LDSRequiresM0Init] in {
-    def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_m0")>;
+    def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_local_m0")>;
   }
 
   let OtherPredicates = [NotLDSRequiresM0Init] in {
     def : DSAtomicCmpXChg<!cast<DS_Pseudo>(!cast<string>(inst)#"_gfx9"), vt,
-                          !cast<PatFrag>(frag)>;
+                          !cast<PatFrag>(frag#"_local")>;
   }
+
+  def : DSAtomicCmpXChg<inst, vt, !cast<PatFrag>(frag#"_region_m0"), 1>;
 }
 
 
 
 // 32-bit atomics.
-defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B32, i32, "atomic_swap_local">;
-defm : DSAtomicRetPat_mc<DS_ADD_RTN_U32, i32, "atomic_load_add_local">;
-defm : DSAtomicRetPat_mc<DS_SUB_RTN_U32, i32, "atomic_load_sub_local">;
-defm : DSAtomicRetPat_mc<DS_INC_RTN_U32, i32, "atomic_inc_local">;
-defm : DSAtomicRetPat_mc<DS_DEC_RTN_U32, i32, "atomic_dec_local">;
-defm : DSAtomicRetPat_mc<DS_AND_RTN_B32, i32, "atomic_load_and_local">;
-defm : DSAtomicRetPat_mc<DS_OR_RTN_B32, i32, "atomic_load_or_local">;
-defm : DSAtomicRetPat_mc<DS_XOR_RTN_B32, i32, "atomic_load_xor_local">;
-defm : DSAtomicRetPat_mc<DS_MIN_RTN_I32, i32, "atomic_load_min_local">;
-defm : DSAtomicRetPat_mc<DS_MAX_RTN_I32, i32, "atomic_load_max_local">;
-defm : DSAtomicRetPat_mc<DS_MIN_RTN_U32, i32, "atomic_load_umin_local">;
-defm : DSAtomicRetPat_mc<DS_MAX_RTN_U32, i32, "atomic_load_umax_local">;
-defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B32, i32, "atomic_cmp_swap_local">;
-defm : DSAtomicRetPat_mc<DS_MIN_RTN_F32, f32, "atomic_load_fmin_local">;
-defm : DSAtomicRetPat_mc<DS_MAX_RTN_F32, f32, "atomic_load_fmax_local">;
-defm : DSAtomicRetPat_mc<DS_ADD_RTN_F32, f32, "atomic_load_fadd_local">;
+defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B32, i32, "atomic_swap">;
+defm : DSAtomicRetPat_mc<DS_ADD_RTN_U32, i32, "atomic_load_add">;
+defm : DSAtomicRetPat_mc<DS_SUB_RTN_U32, i32, "atomic_load_sub">;
+defm : DSAtomicRetPat_mc<DS_INC_RTN_U32, i32, "atomic_inc">;
+defm : DSAtomicRetPat_mc<DS_DEC_RTN_U32, i32, "atomic_dec">;
+defm : DSAtomicRetPat_mc<DS_AND_RTN_B32, i32, "atomic_load_and">;
+defm : DSAtomicRetPat_mc<DS_OR_RTN_B32, i32, "atomic_load_or">;
+defm : DSAtomicRetPat_mc<DS_XOR_RTN_B32, i32, "atomic_load_xor">;
+defm : DSAtomicRetPat_mc<DS_MIN_RTN_I32, i32, "atomic_load_min">;
+defm : DSAtomicRetPat_mc<DS_MAX_RTN_I32, i32, "atomic_load_max">;
+defm : DSAtomicRetPat_mc<DS_MIN_RTN_U32, i32, "atomic_load_umin">;
+defm : DSAtomicRetPat_mc<DS_MAX_RTN_U32, i32, "atomic_load_umax">;
+defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B32, i32, "atomic_cmp_swap">;
+defm : DSAtomicRetPat_mc<DS_MIN_RTN_F32, f32, "atomic_load_fmin">;
+defm : DSAtomicRetPat_mc<DS_MAX_RTN_F32, f32, "atomic_load_fmax">;
+defm : DSAtomicRetPat_mc<DS_ADD_RTN_F32, f32, "atomic_load_fadd">;
 
 // 64-bit atomics.
-defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B64, i64, "atomic_swap_local">;
-defm : DSAtomicRetPat_mc<DS_ADD_RTN_U64, i64, "atomic_load_add_local">;
-defm : DSAtomicRetPat_mc<DS_SUB_RTN_U64, i64, "atomic_load_sub_local">;
-defm : DSAtomicRetPat_mc<DS_INC_RTN_U64, i64, "atomic_inc_local">;
-defm : DSAtomicRetPat_mc<DS_DEC_RTN_U64, i64, "atomic_dec_local">;
-defm : DSAtomicRetPat_mc<DS_AND_RTN_B64, i64, "atomic_load_and_local">;
-defm : DSAtomicRetPat_mc<DS_OR_RTN_B64, i64, "atomic_load_or_local">;
-defm : DSAtomicRetPat_mc<DS_XOR_RTN_B64, i64, "atomic_load_xor_local">;
-defm : DSAtomicRetPat_mc<DS_MIN_RTN_I64, i64, "atomic_load_min_local">;
-defm : DSAtomicRetPat_mc<DS_MAX_RTN_I64, i64, "atomic_load_max_local">;
-defm : DSAtomicRetPat_mc<DS_MIN_RTN_U64, i64, "atomic_load_umin_local">;
-defm : DSAtomicRetPat_mc<DS_MAX_RTN_U64, i64, "atomic_load_umax_local">;
+defm : DSAtomicRetPat_mc<DS_WRXCHG_RTN_B64, i64, "atomic_swap">;
+defm : DSAtomicRetPat_mc<DS_ADD_RTN_U64, i64, "atomic_load_add">;
+defm : DSAtomicRetPat_mc<DS_SUB_RTN_U64, i64, "atomic_load_sub">;
+defm : DSAtomicRetPat_mc<DS_INC_RTN_U64, i64, "atomic_inc">;
+defm : DSAtomicRetPat_mc<DS_DEC_RTN_U64, i64, "atomic_dec">;
+defm : DSAtomicRetPat_mc<DS_AND_RTN_B64, i64, "atomic_load_and">;
+defm : DSAtomicRetPat_mc<DS_OR_RTN_B64, i64, "atomic_load_or">;
+defm : DSAtomicRetPat_mc<DS_XOR_RTN_B64, i64, "atomic_load_xor">;
+defm : DSAtomicRetPat_mc<DS_MIN_RTN_I64, i64, "atomic_load_min">;
+defm : DSAtomicRetPat_mc<DS_MAX_RTN_I64, i64, "atomic_load_max">;
+defm : DSAtomicRetPat_mc<DS_MIN_RTN_U64, i64, "atomic_load_umin">;
+defm : DSAtomicRetPat_mc<DS_MAX_RTN_U64, i64, "atomic_load_umax">;
 
-defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B64, i64, "atomic_cmp_swap_local">;
+defm : DSAtomicCmpXChg_mc<DS_CMPST_RTN_B64, i64, "atomic_cmp_swap">;
 
 def : Pat <
   (SIds_ordered_count i32:$value, i16:$offset),
Index: llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1173,7 +1173,7 @@
   } else if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
     unsigned MaxPrivateBits = 8 * getSubtarget()->getMaxPrivateElementSize();
     return (MemVT.getSizeInBits() <= MaxPrivateBits);
-  } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
+  } else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
     return (MemVT.getSizeInBits() <= 2 * 32);
   }
   return true;
@@ -7135,7 +7135,7 @@
     default:
       llvm_unreachable("unsupported private_element_size");
     }
-  } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
+  } else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
     // Use ds_read_b128 if possible.
     if (Subtarget->useDS128() && Load->getAlignment() >= 16 &&
         MemVT.getStoreSize() == 16)
@@ -7557,7 +7557,7 @@
     default:
       llvm_unreachable("unsupported private_element_size");
     }
-  } else if (AS == AMDGPUAS::LOCAL_ADDRESS) {
+  } else if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
     // Use ds_write_b128 if possible.
     if (Subtarget->useDS128() && Store->getAlignment() >= 16 &&
         VT.getStoreSize() == 16 && NumElements != 3)
Index: llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.td
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.td
+++ llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.td
@@ -505,6 +505,7 @@
   >;
 
   def _local_m0 : local_binary_atomic_op <!cast<SDNode>(NAME#"_glue")>;
+  def _region_m0 : region_binary_atomic_op <!cast<SDNode>(NAME#"_glue")>;
 }
 
 defm atomic_load_add : SIAtomicM0Glue2 <"LOAD_ADD">;
@@ -528,6 +529,7 @@
 >;
 
 def atomic_cmp_swap_local_m0 : AtomicCmpSwapLocal<atomic_cmp_swap_glue>;
+def atomic_cmp_swap_region_m0 : AtomicCmpSwapRegion<atomic_cmp_swap_glue>;
 
 
 def as_i1imm : SDNodeXForm<imm, [{
Index: llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h
+++ llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.h
@@ -262,6 +262,7 @@
   unsigned GITPtrHigh;
 
   unsigned HighBitsOf32BitAddress;
+  unsigned GDSSize;
 
   // Current recorded maximum possible occupancy.
   unsigned Occupancy;
@@ -489,6 +490,10 @@
     return HighBitsOf32BitAddress;
   }
 
+  unsigned getGDSSize() const {
+    return GDSSize;
+  }
+
   unsigned getNumUserSGPRs() const {
     return NumUserSGPRs;
   }
Index: llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
===================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ llvm/trunk/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -46,7 +46,8 @@
     ImplicitBufferPtr(false),
     ImplicitArgPtr(false),
     GITPtrHigh(0xffffffff),
-    HighBitsOf32BitAddress(0) {
+    HighBitsOf32BitAddress(0),
+    GDSSize(0) {
   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
   const Function &F = MF.getFunction();
   FlatWorkGroupSizes = ST.getFlatWorkGroupSizes(F);
@@ -159,6 +160,10 @@
   S = A.getValueAsString();
   if (!S.empty())
     S.consumeInteger(0, HighBitsOf32BitAddress);
+
+  S = F.getFnAttribute("amdgpu-gds-size").getValueAsString();
+  if (!S.empty())
+    S.consumeInteger(0, GDSSize);
 }
 
 void SIMachineFunctionInfo::limitOccupancy(const MachineFunction &MF) {
Index: llvm/trunk/test/CodeGen/AMDGPU/gds-atomic.ll
===================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/gds-atomic.ll
+++ llvm/trunk/test/CodeGen/AMDGPU/gds-atomic.ll
@@ -0,0 +1,128 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+
+; FUNC-LABEL: {{^}}atomic_add_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_add_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_add_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw volatile add i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_add_ret_gds_const_offset:
+; GCN: s_movk_i32 m0, 0x80
+; GCN: ds_add_rtn_u32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:20 gds
+define amdgpu_kernel void @atomic_add_ret_gds_const_offset(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #0 {
+  %gep = getelementptr i32, i32 addrspace(2)* %gds, i32 5
+  %val = atomicrmw volatile add i32 addrspace(2)* %gep, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_sub_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_sub_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_sub_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw sub i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_and_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_and_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_and_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw and i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_or_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_or_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_or_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw or i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_xor_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_xor_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_xor_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw xor i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umin_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_min_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_umin_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw umin i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_umax_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_max_rtn_u32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_umax_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw umax i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_imin_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_min_rtn_i32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_imin_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw min i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_imax_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_max_rtn_i32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_imax_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw max i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_xchg_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_wrxchg_rtn_b32 v{{[0-9]+}}, v[[OFF]], v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_xchg_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = atomicrmw xchg i32 addrspace(2)* %gds, i32 5 acq_rel
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}atomic_cmpxchg_ret_gds:
+; GCN-DAG: v_mov_b32_e32 v[[OFF:[0-9]+]], s
+; GCN-DAG: s_movk_i32 m0, 0x1000
+; GCN: ds_cmpst_rtn_b32 v{{[0-9]+}}, v[[OFF:[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}} gds
+define amdgpu_kernel void @atomic_cmpxchg_ret_gds(i32 addrspace(1)* %out, i32 addrspace(2)* %gds) #1 {
+  %val = cmpxchg i32 addrspace(2)* %gds, i32 0, i32 1 acquire acquire
+  %x = extractvalue { i32, i1 } %val, 0
+  store i32 %x, i32 addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind "amdgpu-gds-size"="128" }
+attributes #1 = { nounwind "amdgpu-gds-size"="4096" }