summaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll84
1 files changed, 84 insertions, 0 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
index 6a5a6e01c741..e7f7b1d196be 100644
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
@@ -952,6 +952,7 @@ define amdgpu_kernel void @global_singlethread_unordered_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -1118,6 +1119,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -1284,6 +1286,7 @@ define amdgpu_kernel void @global_singlethread_release_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -1450,6 +1453,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -1614,6 +1618,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -1778,6 +1783,7 @@ define amdgpu_kernel void @global_singlethread_acquire_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -1942,6 +1948,7 @@ define amdgpu_kernel void @global_singlethread_release_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -2106,6 +2113,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -2270,6 +2278,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -2462,6 +2471,7 @@ define amdgpu_kernel void @global_singlethread_acquire_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -2657,6 +2667,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -2852,6 +2863,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -3087,6 +3099,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -3320,6 +3333,7 @@ define amdgpu_kernel void @global_singlethread_acquire_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -3553,6 +3567,7 @@ define amdgpu_kernel void @global_singlethread_release_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -3786,6 +3801,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -4019,6 +4035,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -4252,6 +4269,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -4485,6 +4503,7 @@ define amdgpu_kernel void @global_singlethread_acquire_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -4718,6 +4737,7 @@ define amdgpu_kernel void @global_singlethread_release_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -4951,6 +4971,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -5184,6 +5205,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -5417,6 +5439,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -5650,6 +5673,7 @@ define amdgpu_kernel void @global_singlethread_acquire_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -5883,6 +5907,7 @@ define amdgpu_kernel void @global_singlethread_release_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -6116,6 +6141,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -6349,6 +6375,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -6612,6 +6639,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_monotonic_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -6879,6 +6907,7 @@ define amdgpu_kernel void @global_singlethread_acquire_monotonic_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -7146,6 +7175,7 @@ define amdgpu_kernel void @global_singlethread_release_monotonic_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -7413,6 +7443,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -7680,6 +7711,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -7947,6 +7979,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_acquire_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -8214,6 +8247,7 @@ define amdgpu_kernel void @global_singlethread_acquire_acquire_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -8481,6 +8515,7 @@ define amdgpu_kernel void @global_singlethread_release_acquire_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -8748,6 +8783,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_acquire_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -9015,6 +9051,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_acquire_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -9282,6 +9319,7 @@ define amdgpu_kernel void @global_singlethread_monotonic_seq_cst_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -9549,6 +9587,7 @@ define amdgpu_kernel void @global_singlethread_acquire_seq_cst_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -9816,6 +9855,7 @@ define amdgpu_kernel void @global_singlethread_release_seq_cst_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -10083,6 +10123,7 @@ define amdgpu_kernel void @global_singlethread_acq_rel_seq_cst_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -10350,6 +10391,7 @@ define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_ret_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -11301,6 +11343,7 @@ define amdgpu_kernel void @global_singlethread_one_as_unordered_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -11467,6 +11510,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -11633,6 +11677,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -11799,6 +11844,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_store(
; GFX1250-NEXT: v_mov_b32_e32 v0, 0
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
i32 %in, ptr addrspace(1) %out) {
@@ -11963,6 +12009,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -12127,6 +12174,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -12291,6 +12339,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -12455,6 +12504,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -12619,6 +12669,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in) {
@@ -12811,6 +12862,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -13006,6 +13058,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -13201,6 +13254,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_ret_atomicrmw(
; GFX1250-NEXT: s_load_b32 s2, s[4:5], 0x8
; GFX1250-NEXT: s_wait_kmcnt 0x0
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -13436,6 +13490,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_monotonic_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -13669,6 +13724,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -13902,6 +13958,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -14135,6 +14192,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -14368,6 +14426,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -14601,6 +14660,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -14834,6 +14894,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -15067,6 +15128,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -15300,6 +15362,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -15533,6 +15596,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -15766,6 +15830,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -15999,6 +16064,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -16232,6 +16298,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -16465,6 +16532,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -16698,6 +16766,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_cmpxchg(
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v0, v[2:3], s[0:1] offset:16
; GFX1250-NEXT: s_endpgm
ptr addrspace(1) %out, i32 %in, i32 %old) {
@@ -16961,6 +17030,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_monotonic_ret_cm
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -17228,6 +17298,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -17495,6 +17566,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_monotonic_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -17762,6 +17834,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -18029,6 +18102,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -18296,6 +18370,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_acquire_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -18563,6 +18638,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -18830,6 +18906,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_acquire_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -19097,6 +19174,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -19364,6 +19442,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -19631,6 +19710,7 @@ define amdgpu_kernel void @global_singlethread_one_as_monotonic_seq_cst_ret_cmpx
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -19898,6 +19978,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acquire_seq_cst_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -20165,6 +20246,7 @@ define amdgpu_kernel void @global_singlethread_one_as_release_seq_cst_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -20432,6 +20514,7 @@ define amdgpu_kernel void @global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
@@ -20699,6 +20782,7 @@ define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxch
; GFX1250-NEXT: v_mov_b32_e32 v1, s2
; GFX1250-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
; GFX1250-NEXT: v_mov_b32_e32 v3, v1
+; GFX1250-NEXT: s_wait_xcnt 0x0
; GFX1250-NEXT: global_atomic_cmpswap_b32 v1, v0, v[2:3], s[0:1] offset:16 th:TH_ATOMIC_RETURN
; GFX1250-NEXT: s_wait_loadcnt 0x0
; GFX1250-NEXT: global_store_b32 v0, v1, s[0:1]
generated by cgit v1.2.3 (git 2.25.1) at 2026-05-31 14:53:35 +0000