Merge branch 'main' into users/mingmingl-llvm/samplefdo-profile-formatusers/mingmingl-llvm/samplefdo-profile-format

1 files changed, 18467 insertions, 0 deletions

diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-cluster.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-cluster.ll
new file mode 100644
index 000000000000..8926893c68db
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-cluster.ll
@@ -0,0 +1,18467 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx600 < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx700 < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1010 < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1010 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -O0 -mcpu=gfx700 -amdgcn-skip-cache-invalidations < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX90A-NOTTGSPLIT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx90a -mattr=+tgsplit < %s | FileCheck -check-prefixes=GFX90A-TGSPLIT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX942-NOTTGSPLIT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx942 -mattr=+tgsplit < %s | FileCheck -check-prefixes=GFX942-TGSPLIT %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1100 < %s | FileCheck --check-prefixes=GFX11-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1100 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX11-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1200 < %s | FileCheck --check-prefixes=GFX12-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1200 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX12-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1250 < %s | FileCheck --check-prefixes=GFX1250 %s
+
+define amdgpu_kernel void @local_cluster_unordered_load(
+; GFX6-LABEL: local_cluster_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_unordered_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_unordered_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_unordered_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_unordered_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_unordered_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_unordered_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_unordered_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_unordered_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_unordered_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster") unordered, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_load(
+; GFX6-LABEL: local_cluster_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster") monotonic, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_load(
+; GFX6-LABEL: local_cluster_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster") acquire, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_load(
+; GFX6-LABEL: local_cluster_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster") seq_cst, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_unordered_store(
+; GFX6-LABEL: local_cluster_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_unordered_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_unordered_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_unordered_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_unordered_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_unordered_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_unordered_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_unordered_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_unordered_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_unordered_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_store(
+; GFX6-LABEL: local_cluster_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_store(
+; GFX6-LABEL: local_cluster_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_store(
+; GFX6-LABEL: local_cluster_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_atomicrmw(
+; GFX6-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_atomicrmw(
+; GFX6-LABEL: local_cluster_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_atomicrmw(
+; GFX6-LABEL: local_cluster_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") release
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_atomicrmw(
+; GFX6-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_atomicrmw(
+; GFX6-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") acquire
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") acq_rel
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster") seq_cst
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_monotonic_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_monotonic_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") monotonic seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acquire_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acquire_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acquire seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_release_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_release_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") release seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_acq_rel_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_acq_rel_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") acq_rel seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    buffer_wbinvl1_vol
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    buffer_inv sc0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    buffer_gl0_inv
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-WGP-NEXT:    s_wait_samplecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_storecnt 0x0
+; GFX12-WGP-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    global_inv scope:SCOPE_SE
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_seq_cst_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    s_wait_storecnt 0x0
+; GFX1250-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_unordered_load(
+; GFX6-LABEL: local_cluster_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_unordered_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_unordered_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_unordered_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_unordered_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_unordered_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_unordered_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_unordered_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_unordered_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_unordered_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster-one-as") unordered, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_load(
+; GFX6-LABEL: local_cluster_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster-one-as") monotonic, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_load(
+; GFX6-LABEL: local_cluster_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster-one-as") acquire, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_load(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr4
+; GFX6-NEXT:    ; kill: def $sgpr6 killed $sgpr5
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    ds_read_b32 v1, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    ds_read_b32 v1, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    ds_read_b32 v1, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    ds_read_b32 v1, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    ds_read_b32 v1, v0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    ds_load_b32 v1, v0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    ds_load_b32 v1, v0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    ds_load_b32 v1, v0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_load:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    ds_load_b32 v1, v0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %in, ptr addrspace(3) %out) {
+entry:
+  %val = load atomic i32, ptr addrspace(3) %in syncscope("cluster-one-as") seq_cst, align 4
+  store i32 %val, ptr addrspace(3) %out
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_unordered_store(
+; GFX6-LABEL: local_cluster_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_unordered_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_unordered_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_unordered_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_unordered_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_unordered_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_unordered_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_unordered_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_unordered_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_unordered_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_store(
+; GFX6-LABEL: local_cluster_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_store(
+; GFX6-LABEL: local_cluster_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_store(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_store:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    i32 %in, ptr addrspace(3) %out) {
+entry:
+  store atomic i32 %in, ptr addrspace(3) %out syncscope("cluster-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr5
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s5
+; GFX6-NEXT:    v_mov_b32_e32 v1, s4
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s5
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s4
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s4
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s0
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s0
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") acquire
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") acq_rel
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_ret_atomicrmw:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg ptr addrspace(3) %out, i32 %in syncscope("cluster-one-as") seq_cst
+  store i32 %val, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_monotonic_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_acquire_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr4 killed $sgpr6
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s6
+; GFX6-NEXT:    v_mov_b32_e32 v1, s5
+; GFX6-NEXT:    v_mov_b32_e32 v2, s4
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s6
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v2, s4
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s5
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s4
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s1
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s0
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX12-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s1
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s0
+; GFX1250-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_monotonic_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") monotonic seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acquire_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acquire_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acquire seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_release_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_release_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") release seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_acq_rel_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") acq_rel seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX6-NEXT:    ; kill: def $sgpr5 killed $sgpr4
+; GFX6-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX6-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    v_mov_b32_e32 v1, s6
+; GFX6-NEXT:    v_mov_b32_e32 v2, s5
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX7-NEXT:    s_load_dword s5, s[8:9], 0x1
+; GFX7-NEXT:    s_load_dword s6, s[8:9], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v1, s6
+; GFX7-NEXT:    v_mov_b32_e32 v2, s5
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-WGP-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX10-CU-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX10-CU-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s6
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s5
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[4:5], 0x1
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[4:5], 0x2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+;
+; GFX90A-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX90A-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-NOTTGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX90A-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX90A-TGSPLIT:       ; %bb.0: ; %entry
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s4, s[8:9], 0x0
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s5, s[8:9], 0x4
+; GFX90A-TGSPLIT-NEXT:    s_load_dword s6, s[8:9], 0x8
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s6
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s5
+; GFX90A-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX90A-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s4
+; GFX90A-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX90A-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX90A-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-NOTTGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX942-NOTTGSPLIT:       ; %bb.0: ; %entry
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-NOTTGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-NOTTGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-NOTTGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-NOTTGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-NOTTGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-NOTTGSPLIT-NEXT:    s_endpgm
+;
+; GFX942-TGSPLIT-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX942-TGSPLIT:       ; %bb.0: ; %entry
+; GFX942-TGSPLIT-NEXT:    s_load_dword s0, s[4:5], 0x0
+; GFX942-TGSPLIT-NEXT:    s_load_dword s1, s[4:5], 0x4
+; GFX942-TGSPLIT-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v1, s2
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v2, s1
+; GFX942-TGSPLIT-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX942-TGSPLIT-NEXT:    v_mov_b32_e32 v0, s0
+; GFX942-TGSPLIT-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942-TGSPLIT-NEXT:    ds_write_b32 v0, v1
+; GFX942-TGSPLIT-NEXT:    s_endpgm
+;
+; GFX11-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX11-WGP:       ; %bb.0: ; %entry
+; GFX11-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX11-WGP-NEXT:    s_endpgm
+;
+; GFX11-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX11-CU:       ; %bb.0: ; %entry
+; GFX11-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX11-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX11-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    ds_store_b32 v0, v1
+; GFX11-CU-NEXT:    s_endpgm
+;
+; GFX12-WGP-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX12-WGP:       ; %bb.0: ; %entry
+; GFX12-WGP-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-WGP-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-WGP-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-WGP-NEXT:    s_wait_kmcnt 0x0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-WGP-NEXT:    s_wait_dscnt 0x0
+; GFX12-WGP-NEXT:    ds_store_b32 v0, v1
+; GFX12-WGP-NEXT:    s_endpgm
+;
+; GFX12-CU-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX12-CU:       ; %bb.0: ; %entry
+; GFX12-CU-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX12-CU-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX12-CU-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX12-CU-NEXT:    s_wait_kmcnt 0x0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX12-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX12-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-CU-NEXT:    s_wait_dscnt 0x0
+; GFX12-CU-NEXT:    ds_store_b32 v0, v1
+; GFX12-CU-NEXT:    s_endpgm
+;
+; GFX1250-LABEL: local_cluster_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX1250:       ; %bb.0: ; %entry
+; GFX1250-NEXT:    s_load_b32 s0, s[4:5], 0x0
+; GFX1250-NEXT:    s_load_b32 s2, s[4:5], 0x4
+; GFX1250-NEXT:    s_load_b32 s1, s[4:5], 0x8
+; GFX1250-NEXT:    s_wait_kmcnt 0x0
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    v_mov_b32_e32 v1, s2
+; GFX1250-NEXT:    v_mov_b32_e32 v2, s1
+; GFX1250-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
+; GFX1250-NEXT:    s_wait_dscnt 0x0
+; GFX1250-NEXT:    ds_store_b32 v0, v1
+; GFX1250-NEXT:    s_endpgm
+    ptr addrspace(3) %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, ptr addrspace(3) %out, i32 4
+  %val = cmpxchg volatile ptr addrspace(3) %gep, i32 %old, i32 %in syncscope("cluster-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, ptr addrspace(3) %out, align 4
+  ret void
+}