| blob | afc46fbc23a67a355908f1d6051a06c9ecdd3334 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx600 < %s | FileCheck --check-prefixes=GFX6 %s
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx700 < %s | FileCheck --check-prefixes=GFX7 %s
4 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1010 < %s | FileCheck --check-prefixes=GFX10-WGP %s
5 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1010 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX10-CU %s
6 ; RUN: llc -mtriple=amdgcn-amd-amdpal -O0 -mcpu=gfx700 -amdgcn-skip-cache-invalidations < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
7 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX90A-NOTTGSPLIT %s
8 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx90a -mattr=+tgsplit < %s | FileCheck -check-prefixes=GFX90A-TGSPLIT %s
9 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx940 < %s | FileCheck -check-prefixes=GFX940-NOTTGSPLIT %s
10 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx940 -mattr=+tgsplit < %s | FileCheck -check-prefixes=GFX940-TGSPLIT %s
11 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1100 < %s | FileCheck --check-prefixes=GFX11-WGP %s
12 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1100 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX11-CU %s
13 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1200 < %s | FileCheck --check-prefixes=GFX12-WGP %s
14 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -O0 -mcpu=gfx1200 -mattr=+cumode < %s | FileCheck --check-prefixes=GFX12-CU %s
16 define amdgpu_kernel void @global_singlethread_unordered_load(
17 ; GFX6-LABEL: global_singlethread_unordered_load:
18 ; GFX6: ; %bb.0: ; %entry
19 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
20 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
21 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
22 ; GFX6-NEXT: s_mov_b32 s6, s9
23 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
24 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
25 ; GFX6-NEXT: s_mov_b32 s13, -1
26 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
27 ; GFX6-NEXT: s_mov_b32 s9, s6
28 ; GFX6-NEXT: s_mov_b32 s10, s13
29 ; GFX6-NEXT: s_mov_b32 s11, s12
30 ; GFX6-NEXT: s_mov_b32 s14, s5
31 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
32 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
33 ; GFX6-NEXT: s_mov_b32 s5, s14
34 ; GFX6-NEXT: s_mov_b32 s6, s13
35 ; GFX6-NEXT: s_mov_b32 s7, s12
36 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
37 ; GFX6-NEXT: s_waitcnt vmcnt(0)
38 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
39 ; GFX6-NEXT: s_endpgm
40 ;
41 ; GFX7-LABEL: global_singlethread_unordered_load:
42 ; GFX7: ; %bb.0: ; %entry
43 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
44 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
45 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
46 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
47 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
48 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
49 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
50 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
51 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
52 ; GFX7-NEXT: s_waitcnt vmcnt(0)
53 ; GFX7-NEXT: flat_store_dword v[0:1], v2
54 ; GFX7-NEXT: s_endpgm
55 ;
56 ; GFX10-WGP-LABEL: global_singlethread_unordered_load:
57 ; GFX10-WGP: ; %bb.0: ; %entry
58 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
59 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
60 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
61 ; GFX10-WGP-NEXT: s_nop 0
62 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
63 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
64 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
65 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
66 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
67 ; GFX10-WGP-NEXT: s_endpgm
68 ;
69 ; GFX10-CU-LABEL: global_singlethread_unordered_load:
70 ; GFX10-CU: ; %bb.0: ; %entry
71 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
72 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
73 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
74 ; GFX10-CU-NEXT: s_nop 0
75 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
76 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
77 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
78 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
79 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
80 ; GFX10-CU-NEXT: s_endpgm
81 ;
82 ; SKIP-CACHE-INV-LABEL: global_singlethread_unordered_load:
83 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
84 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
85 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
86 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
87 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
88 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
89 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
90 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
91 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
92 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
93 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
94 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
95 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
96 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
97 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
98 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
99 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
100 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
101 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
102 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
103 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
104 ; SKIP-CACHE-INV-NEXT: s_endpgm
105 ;
106 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_unordered_load:
107 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
108 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
109 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
110 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
111 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
112 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
113 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
114 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
115 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
116 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
117 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
118 ;
119 ; GFX90A-TGSPLIT-LABEL: global_singlethread_unordered_load:
120 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
121 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
122 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
123 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
124 ; GFX90A-TGSPLIT-NEXT: s_nop 0
125 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
126 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
127 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
128 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
129 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
130 ; GFX90A-TGSPLIT-NEXT: s_endpgm
131 ;
132 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_unordered_load:
133 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
134 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
135 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
136 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
137 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
138 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
139 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
140 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
141 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
142 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
143 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
144 ;
145 ; GFX940-TGSPLIT-LABEL: global_singlethread_unordered_load:
146 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
147 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
148 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
149 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
150 ; GFX940-TGSPLIT-NEXT: s_nop 0
151 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
152 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
153 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
154 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
155 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
156 ; GFX940-TGSPLIT-NEXT: s_endpgm
157 ;
158 ; GFX11-WGP-LABEL: global_singlethread_unordered_load:
159 ; GFX11-WGP: ; %bb.0: ; %entry
160 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
161 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
162 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
163 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
164 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
165 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
166 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
167 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
168 ; GFX11-WGP-NEXT: s_endpgm
169 ;
170 ; GFX11-CU-LABEL: global_singlethread_unordered_load:
171 ; GFX11-CU: ; %bb.0: ; %entry
172 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
173 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
174 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
175 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
176 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
177 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
178 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
179 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
180 ; GFX11-CU-NEXT: s_endpgm
181 ;
182 ; GFX12-WGP-LABEL: global_singlethread_unordered_load:
183 ; GFX12-WGP: ; %bb.0: ; %entry
184 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
185 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
186 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
187 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
188 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
189 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
190 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
191 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
192 ; GFX12-WGP-NEXT: s_endpgm
193 ;
194 ; GFX12-CU-LABEL: global_singlethread_unordered_load:
195 ; GFX12-CU: ; %bb.0: ; %entry
196 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
197 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
198 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
199 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
200 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
201 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
202 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
203 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
204 ; GFX12-CU-NEXT: s_endpgm
205 ptr addrspace(1) %in, ptr addrspace(1) %out) {
206 entry:
207 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread") unordered, align 4
208 store i32 %val, ptr addrspace(1) %out
209 ret void
210 }
212 define amdgpu_kernel void @global_singlethread_monotonic_load(
213 ; GFX6-LABEL: global_singlethread_monotonic_load:
214 ; GFX6: ; %bb.0: ; %entry
215 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
216 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
217 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
218 ; GFX6-NEXT: s_mov_b32 s6, s9
219 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
220 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
221 ; GFX6-NEXT: s_mov_b32 s13, -1
222 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
223 ; GFX6-NEXT: s_mov_b32 s9, s6
224 ; GFX6-NEXT: s_mov_b32 s10, s13
225 ; GFX6-NEXT: s_mov_b32 s11, s12
226 ; GFX6-NEXT: s_mov_b32 s14, s5
227 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
228 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
229 ; GFX6-NEXT: s_mov_b32 s5, s14
230 ; GFX6-NEXT: s_mov_b32 s6, s13
231 ; GFX6-NEXT: s_mov_b32 s7, s12
232 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
233 ; GFX6-NEXT: s_waitcnt vmcnt(0)
234 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
235 ; GFX6-NEXT: s_endpgm
236 ;
237 ; GFX7-LABEL: global_singlethread_monotonic_load:
238 ; GFX7: ; %bb.0: ; %entry
239 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
240 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
241 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
242 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
243 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
244 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
245 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
246 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
247 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
248 ; GFX7-NEXT: s_waitcnt vmcnt(0)
249 ; GFX7-NEXT: flat_store_dword v[0:1], v2
250 ; GFX7-NEXT: s_endpgm
251 ;
252 ; GFX10-WGP-LABEL: global_singlethread_monotonic_load:
253 ; GFX10-WGP: ; %bb.0: ; %entry
254 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
255 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
256 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
257 ; GFX10-WGP-NEXT: s_nop 0
258 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
259 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
260 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
261 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
262 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
263 ; GFX10-WGP-NEXT: s_endpgm
264 ;
265 ; GFX10-CU-LABEL: global_singlethread_monotonic_load:
266 ; GFX10-CU: ; %bb.0: ; %entry
267 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
268 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
269 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
270 ; GFX10-CU-NEXT: s_nop 0
271 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
272 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
273 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
274 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
275 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
276 ; GFX10-CU-NEXT: s_endpgm
277 ;
278 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_load:
279 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
280 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
281 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
282 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
283 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
284 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
285 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
286 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
287 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
288 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
289 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
290 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
291 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
292 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
293 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
294 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
295 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
296 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
297 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
298 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
299 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
300 ; SKIP-CACHE-INV-NEXT: s_endpgm
301 ;
302 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_load:
303 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
304 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
305 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
306 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
307 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
308 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
309 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
310 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
311 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
312 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
313 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
314 ;
315 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_load:
316 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
317 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
318 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
319 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
320 ; GFX90A-TGSPLIT-NEXT: s_nop 0
321 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
322 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
323 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
324 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
325 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
326 ; GFX90A-TGSPLIT-NEXT: s_endpgm
327 ;
328 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_load:
329 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
330 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
331 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
332 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
333 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
334 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
335 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
336 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
337 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
338 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
339 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
340 ;
341 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_load:
342 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
343 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
344 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
345 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
346 ; GFX940-TGSPLIT-NEXT: s_nop 0
347 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
348 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
349 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
350 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
351 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
352 ; GFX940-TGSPLIT-NEXT: s_endpgm
353 ;
354 ; GFX11-WGP-LABEL: global_singlethread_monotonic_load:
355 ; GFX11-WGP: ; %bb.0: ; %entry
356 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
357 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
358 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
359 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
360 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
361 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
362 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
363 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
364 ; GFX11-WGP-NEXT: s_endpgm
365 ;
366 ; GFX11-CU-LABEL: global_singlethread_monotonic_load:
367 ; GFX11-CU: ; %bb.0: ; %entry
368 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
369 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
370 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
371 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
372 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
373 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
374 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
375 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
376 ; GFX11-CU-NEXT: s_endpgm
377 ;
378 ; GFX12-WGP-LABEL: global_singlethread_monotonic_load:
379 ; GFX12-WGP: ; %bb.0: ; %entry
380 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
381 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
382 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
383 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
384 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
385 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
386 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
387 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
388 ; GFX12-WGP-NEXT: s_endpgm
389 ;
390 ; GFX12-CU-LABEL: global_singlethread_monotonic_load:
391 ; GFX12-CU: ; %bb.0: ; %entry
392 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
393 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
394 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
395 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
396 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
397 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
398 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
399 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
400 ; GFX12-CU-NEXT: s_endpgm
401 ptr addrspace(1) %in, ptr addrspace(1) %out) {
402 entry:
403 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread") monotonic, align 4
404 store i32 %val, ptr addrspace(1) %out
405 ret void
406 }
408 define amdgpu_kernel void @global_singlethread_acquire_load(
409 ; GFX6-LABEL: global_singlethread_acquire_load:
410 ; GFX6: ; %bb.0: ; %entry
411 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
412 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
413 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
414 ; GFX6-NEXT: s_mov_b32 s6, s9
415 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
416 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
417 ; GFX6-NEXT: s_mov_b32 s13, -1
418 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
419 ; GFX6-NEXT: s_mov_b32 s9, s6
420 ; GFX6-NEXT: s_mov_b32 s10, s13
421 ; GFX6-NEXT: s_mov_b32 s11, s12
422 ; GFX6-NEXT: s_mov_b32 s14, s5
423 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
424 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
425 ; GFX6-NEXT: s_mov_b32 s5, s14
426 ; GFX6-NEXT: s_mov_b32 s6, s13
427 ; GFX6-NEXT: s_mov_b32 s7, s12
428 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
429 ; GFX6-NEXT: s_waitcnt vmcnt(0)
430 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
431 ; GFX6-NEXT: s_endpgm
432 ;
433 ; GFX7-LABEL: global_singlethread_acquire_load:
434 ; GFX7: ; %bb.0: ; %entry
435 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
436 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
437 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
438 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
439 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
440 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
441 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
442 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
443 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
444 ; GFX7-NEXT: s_waitcnt vmcnt(0)
445 ; GFX7-NEXT: flat_store_dword v[0:1], v2
446 ; GFX7-NEXT: s_endpgm
447 ;
448 ; GFX10-WGP-LABEL: global_singlethread_acquire_load:
449 ; GFX10-WGP: ; %bb.0: ; %entry
450 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
451 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
452 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
453 ; GFX10-WGP-NEXT: s_nop 0
454 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
455 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
456 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
457 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
458 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
459 ; GFX10-WGP-NEXT: s_endpgm
460 ;
461 ; GFX10-CU-LABEL: global_singlethread_acquire_load:
462 ; GFX10-CU: ; %bb.0: ; %entry
463 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
464 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
465 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
466 ; GFX10-CU-NEXT: s_nop 0
467 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
468 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
469 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
470 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
471 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
472 ; GFX10-CU-NEXT: s_endpgm
473 ;
474 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_load:
475 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
476 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
477 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
478 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
479 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
480 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
481 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
482 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
483 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
484 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
485 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
486 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
487 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
488 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
489 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
490 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
491 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
492 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
493 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
494 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
495 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
496 ; SKIP-CACHE-INV-NEXT: s_endpgm
497 ;
498 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_load:
499 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
500 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
501 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
502 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
503 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
504 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
505 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
506 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
507 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
508 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
509 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
510 ;
511 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_load:
512 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
513 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
514 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
515 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
516 ; GFX90A-TGSPLIT-NEXT: s_nop 0
517 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
518 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
519 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
520 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
521 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
522 ; GFX90A-TGSPLIT-NEXT: s_endpgm
523 ;
524 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_load:
525 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
526 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
527 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
528 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
529 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
530 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
531 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
532 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
533 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
534 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
535 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
536 ;
537 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_load:
538 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
539 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
540 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
541 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
542 ; GFX940-TGSPLIT-NEXT: s_nop 0
543 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
544 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
545 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
546 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
547 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
548 ; GFX940-TGSPLIT-NEXT: s_endpgm
549 ;
550 ; GFX11-WGP-LABEL: global_singlethread_acquire_load:
551 ; GFX11-WGP: ; %bb.0: ; %entry
552 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
553 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
554 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
555 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
556 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
557 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
558 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
559 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
560 ; GFX11-WGP-NEXT: s_endpgm
561 ;
562 ; GFX11-CU-LABEL: global_singlethread_acquire_load:
563 ; GFX11-CU: ; %bb.0: ; %entry
564 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
565 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
566 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
567 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
568 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
569 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
570 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
571 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
572 ; GFX11-CU-NEXT: s_endpgm
573 ;
574 ; GFX12-WGP-LABEL: global_singlethread_acquire_load:
575 ; GFX12-WGP: ; %bb.0: ; %entry
576 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
577 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
578 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
579 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
580 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
581 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
582 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
583 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
584 ; GFX12-WGP-NEXT: s_endpgm
585 ;
586 ; GFX12-CU-LABEL: global_singlethread_acquire_load:
587 ; GFX12-CU: ; %bb.0: ; %entry
588 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
589 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
590 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
591 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
592 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
593 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
594 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
595 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
596 ; GFX12-CU-NEXT: s_endpgm
597 ptr addrspace(1) %in, ptr addrspace(1) %out) {
598 entry:
599 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread") acquire, align 4
600 store i32 %val, ptr addrspace(1) %out
601 ret void
602 }
604 define amdgpu_kernel void @global_singlethread_seq_cst_load(
605 ; GFX6-LABEL: global_singlethread_seq_cst_load:
606 ; GFX6: ; %bb.0: ; %entry
607 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
608 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
609 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
610 ; GFX6-NEXT: s_mov_b32 s6, s9
611 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
612 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
613 ; GFX6-NEXT: s_mov_b32 s13, -1
614 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
615 ; GFX6-NEXT: s_mov_b32 s9, s6
616 ; GFX6-NEXT: s_mov_b32 s10, s13
617 ; GFX6-NEXT: s_mov_b32 s11, s12
618 ; GFX6-NEXT: s_mov_b32 s14, s5
619 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
620 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
621 ; GFX6-NEXT: s_mov_b32 s5, s14
622 ; GFX6-NEXT: s_mov_b32 s6, s13
623 ; GFX6-NEXT: s_mov_b32 s7, s12
624 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
625 ; GFX6-NEXT: s_waitcnt vmcnt(0)
626 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
627 ; GFX6-NEXT: s_endpgm
628 ;
629 ; GFX7-LABEL: global_singlethread_seq_cst_load:
630 ; GFX7: ; %bb.0: ; %entry
631 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
632 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
633 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
634 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
635 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
636 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
637 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
638 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
639 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
640 ; GFX7-NEXT: s_waitcnt vmcnt(0)
641 ; GFX7-NEXT: flat_store_dword v[0:1], v2
642 ; GFX7-NEXT: s_endpgm
643 ;
644 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_load:
645 ; GFX10-WGP: ; %bb.0: ; %entry
646 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
647 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
648 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
649 ; GFX10-WGP-NEXT: s_nop 0
650 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
651 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
652 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
653 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
654 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
655 ; GFX10-WGP-NEXT: s_endpgm
656 ;
657 ; GFX10-CU-LABEL: global_singlethread_seq_cst_load:
658 ; GFX10-CU: ; %bb.0: ; %entry
659 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
660 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
661 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
662 ; GFX10-CU-NEXT: s_nop 0
663 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
664 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
665 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
666 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
667 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
668 ; GFX10-CU-NEXT: s_endpgm
669 ;
670 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_load:
671 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
672 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
673 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
674 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
675 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
676 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
677 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
678 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
679 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
680 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
681 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
682 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
683 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
684 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
685 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
686 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
687 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
688 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
689 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
690 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
691 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
692 ; SKIP-CACHE-INV-NEXT: s_endpgm
693 ;
694 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_load:
695 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
696 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
697 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
698 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
699 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
700 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
701 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
702 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
703 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
704 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
705 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
706 ;
707 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_load:
708 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
709 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
710 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
711 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
712 ; GFX90A-TGSPLIT-NEXT: s_nop 0
713 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
714 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
715 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
716 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
717 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
718 ; GFX90A-TGSPLIT-NEXT: s_endpgm
719 ;
720 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_load:
721 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
722 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
723 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
724 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
725 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
726 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
727 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
728 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
729 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
730 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
731 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
732 ;
733 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_load:
734 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
735 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
736 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
737 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
738 ; GFX940-TGSPLIT-NEXT: s_nop 0
739 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
740 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
741 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
742 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
743 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
744 ; GFX940-TGSPLIT-NEXT: s_endpgm
745 ;
746 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_load:
747 ; GFX11-WGP: ; %bb.0: ; %entry
748 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
749 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
750 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
751 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
752 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
753 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
754 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
755 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
756 ; GFX11-WGP-NEXT: s_endpgm
757 ;
758 ; GFX11-CU-LABEL: global_singlethread_seq_cst_load:
759 ; GFX11-CU: ; %bb.0: ; %entry
760 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
761 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
762 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
763 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
764 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
765 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
766 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
767 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
768 ; GFX11-CU-NEXT: s_endpgm
769 ;
770 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_load:
771 ; GFX12-WGP: ; %bb.0: ; %entry
772 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
773 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
774 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
775 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
776 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
777 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
778 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
779 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
780 ; GFX12-WGP-NEXT: s_endpgm
781 ;
782 ; GFX12-CU-LABEL: global_singlethread_seq_cst_load:
783 ; GFX12-CU: ; %bb.0: ; %entry
784 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
785 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
786 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
787 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
788 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
789 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
790 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
791 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
792 ; GFX12-CU-NEXT: s_endpgm
793 ptr addrspace(1) %in, ptr addrspace(1) %out) {
794 entry:
795 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread") seq_cst, align 4
796 store i32 %val, ptr addrspace(1) %out
797 ret void
798 }
800 define amdgpu_kernel void @global_singlethread_unordered_store(
801 ; GFX6-LABEL: global_singlethread_unordered_store:
802 ; GFX6: ; %bb.0: ; %entry
803 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
804 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
805 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
806 ; GFX6-NEXT: s_mov_b32 s11, s5
807 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
808 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
809 ; GFX6-NEXT: s_mov_b32 s10, -1
810 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
811 ; GFX6-NEXT: s_mov_b32 s5, s11
812 ; GFX6-NEXT: s_mov_b32 s6, s10
813 ; GFX6-NEXT: s_mov_b32 s7, s9
814 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
815 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
816 ; GFX6-NEXT: s_endpgm
817 ;
818 ; GFX7-LABEL: global_singlethread_unordered_store:
819 ; GFX7: ; %bb.0: ; %entry
820 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
821 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
822 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
823 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
824 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
825 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
826 ; GFX7-NEXT: flat_store_dword v[0:1], v2
827 ; GFX7-NEXT: s_endpgm
828 ;
829 ; GFX10-WGP-LABEL: global_singlethread_unordered_store:
830 ; GFX10-WGP: ; %bb.0: ; %entry
831 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
832 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
833 ; GFX10-WGP-NEXT: s_nop 0
834 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
835 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
836 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
837 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
838 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
839 ; GFX10-WGP-NEXT: s_endpgm
840 ;
841 ; GFX10-CU-LABEL: global_singlethread_unordered_store:
842 ; GFX10-CU: ; %bb.0: ; %entry
843 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
844 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
845 ; GFX10-CU-NEXT: s_nop 0
846 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
847 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
848 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
849 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
850 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
851 ; GFX10-CU-NEXT: s_endpgm
852 ;
853 ; SKIP-CACHE-INV-LABEL: global_singlethread_unordered_store:
854 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
855 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
856 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
857 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
858 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
859 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
860 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
861 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
862 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
863 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
864 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
865 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
866 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
867 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
868 ; SKIP-CACHE-INV-NEXT: s_endpgm
869 ;
870 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_unordered_store:
871 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
872 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
873 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
874 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
875 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
876 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
877 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
878 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
879 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
880 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
881 ;
882 ; GFX90A-TGSPLIT-LABEL: global_singlethread_unordered_store:
883 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
884 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
885 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
886 ; GFX90A-TGSPLIT-NEXT: s_nop 0
887 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
888 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
889 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
890 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
891 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
892 ; GFX90A-TGSPLIT-NEXT: s_endpgm
893 ;
894 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_unordered_store:
895 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
896 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
897 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
898 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
899 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
900 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
901 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
902 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
903 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
904 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
905 ;
906 ; GFX940-TGSPLIT-LABEL: global_singlethread_unordered_store:
907 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
908 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
909 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
910 ; GFX940-TGSPLIT-NEXT: s_nop 0
911 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
912 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
913 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
914 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
915 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
916 ; GFX940-TGSPLIT-NEXT: s_endpgm
917 ;
918 ; GFX11-WGP-LABEL: global_singlethread_unordered_store:
919 ; GFX11-WGP: ; %bb.0: ; %entry
920 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
921 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
922 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
923 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
924 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
925 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
926 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
927 ; GFX11-WGP-NEXT: s_endpgm
928 ;
929 ; GFX11-CU-LABEL: global_singlethread_unordered_store:
930 ; GFX11-CU: ; %bb.0: ; %entry
931 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
932 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
933 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
934 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
935 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
936 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
937 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
938 ; GFX11-CU-NEXT: s_endpgm
939 ;
940 ; GFX12-WGP-LABEL: global_singlethread_unordered_store:
941 ; GFX12-WGP: ; %bb.0: ; %entry
942 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
943 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
944 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
945 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
946 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
947 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
948 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
949 ; GFX12-WGP-NEXT: s_endpgm
950 ;
951 ; GFX12-CU-LABEL: global_singlethread_unordered_store:
952 ; GFX12-CU: ; %bb.0: ; %entry
953 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
954 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
955 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
956 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
957 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
958 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
959 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
960 ; GFX12-CU-NEXT: s_endpgm
961 i32 %in, ptr addrspace(1) %out) {
962 entry:
963 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread") unordered, align 4
964 ret void
965 }
967 define amdgpu_kernel void @global_singlethread_monotonic_store(
968 ; GFX6-LABEL: global_singlethread_monotonic_store:
969 ; GFX6: ; %bb.0: ; %entry
970 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
971 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
972 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
973 ; GFX6-NEXT: s_mov_b32 s11, s5
974 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
975 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
976 ; GFX6-NEXT: s_mov_b32 s10, -1
977 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
978 ; GFX6-NEXT: s_mov_b32 s5, s11
979 ; GFX6-NEXT: s_mov_b32 s6, s10
980 ; GFX6-NEXT: s_mov_b32 s7, s9
981 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
982 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
983 ; GFX6-NEXT: s_endpgm
984 ;
985 ; GFX7-LABEL: global_singlethread_monotonic_store:
986 ; GFX7: ; %bb.0: ; %entry
987 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
988 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
989 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
990 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
991 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
992 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
993 ; GFX7-NEXT: flat_store_dword v[0:1], v2
994 ; GFX7-NEXT: s_endpgm
995 ;
996 ; GFX10-WGP-LABEL: global_singlethread_monotonic_store:
997 ; GFX10-WGP: ; %bb.0: ; %entry
998 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
999 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
1000 ; GFX10-WGP-NEXT: s_nop 0
1001 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1002 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1003 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1004 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1005 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
1006 ; GFX10-WGP-NEXT: s_endpgm
1007 ;
1008 ; GFX10-CU-LABEL: global_singlethread_monotonic_store:
1009 ; GFX10-CU: ; %bb.0: ; %entry
1010 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
1011 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
1012 ; GFX10-CU-NEXT: s_nop 0
1013 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1014 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1015 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1016 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1017 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
1018 ; GFX10-CU-NEXT: s_endpgm
1019 ;
1020 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_store:
1021 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1022 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
1023 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
1024 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1025 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1026 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1027 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1028 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1029 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1030 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1031 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1032 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1033 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1034 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
1035 ; SKIP-CACHE-INV-NEXT: s_endpgm
1036 ;
1037 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_store:
1038 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1039 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1040 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1041 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1042 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1043 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1044 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1045 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1046 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1047 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1048 ;
1049 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_store:
1050 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1051 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1052 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1053 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1054 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1055 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1056 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1057 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1058 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1059 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1060 ;
1061 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_store:
1062 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1063 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1064 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1065 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1066 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1067 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1068 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1069 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1070 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1071 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1072 ;
1073 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_store:
1074 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1075 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1076 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1077 ; GFX940-TGSPLIT-NEXT: s_nop 0
1078 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1079 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1080 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1081 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1082 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1083 ; GFX940-TGSPLIT-NEXT: s_endpgm
1084 ;
1085 ; GFX11-WGP-LABEL: global_singlethread_monotonic_store:
1086 ; GFX11-WGP: ; %bb.0: ; %entry
1087 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1088 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1089 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1090 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1091 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1092 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1093 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1094 ; GFX11-WGP-NEXT: s_endpgm
1095 ;
1096 ; GFX11-CU-LABEL: global_singlethread_monotonic_store:
1097 ; GFX11-CU: ; %bb.0: ; %entry
1098 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1099 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1100 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1101 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1102 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1103 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1104 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1105 ; GFX11-CU-NEXT: s_endpgm
1106 ;
1107 ; GFX12-WGP-LABEL: global_singlethread_monotonic_store:
1108 ; GFX12-WGP: ; %bb.0: ; %entry
1109 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1110 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1111 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1112 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1113 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1114 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1115 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1116 ; GFX12-WGP-NEXT: s_endpgm
1117 ;
1118 ; GFX12-CU-LABEL: global_singlethread_monotonic_store:
1119 ; GFX12-CU: ; %bb.0: ; %entry
1120 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1121 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1122 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1123 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1124 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1125 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1126 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1127 ; GFX12-CU-NEXT: s_endpgm
1128 i32 %in, ptr addrspace(1) %out) {
1129 entry:
1130 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread") monotonic, align 4
1131 ret void
1132 }
1134 define amdgpu_kernel void @global_singlethread_release_store(
1135 ; GFX6-LABEL: global_singlethread_release_store:
1136 ; GFX6: ; %bb.0: ; %entry
1137 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
1138 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
1139 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1140 ; GFX6-NEXT: s_mov_b32 s11, s5
1141 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1142 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1143 ; GFX6-NEXT: s_mov_b32 s10, -1
1144 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1145 ; GFX6-NEXT: s_mov_b32 s5, s11
1146 ; GFX6-NEXT: s_mov_b32 s6, s10
1147 ; GFX6-NEXT: s_mov_b32 s7, s9
1148 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1149 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
1150 ; GFX6-NEXT: s_endpgm
1151 ;
1152 ; GFX7-LABEL: global_singlethread_release_store:
1153 ; GFX7: ; %bb.0: ; %entry
1154 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
1155 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
1156 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1157 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1158 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1159 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1160 ; GFX7-NEXT: flat_store_dword v[0:1], v2
1161 ; GFX7-NEXT: s_endpgm
1162 ;
1163 ; GFX10-WGP-LABEL: global_singlethread_release_store:
1164 ; GFX10-WGP: ; %bb.0: ; %entry
1165 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
1166 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
1167 ; GFX10-WGP-NEXT: s_nop 0
1168 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1169 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1170 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1171 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1172 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
1173 ; GFX10-WGP-NEXT: s_endpgm
1174 ;
1175 ; GFX10-CU-LABEL: global_singlethread_release_store:
1176 ; GFX10-CU: ; %bb.0: ; %entry
1177 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
1178 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
1179 ; GFX10-CU-NEXT: s_nop 0
1180 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1181 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1182 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1183 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1184 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
1185 ; GFX10-CU-NEXT: s_endpgm
1186 ;
1187 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_store:
1188 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1189 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
1190 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
1191 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1192 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1193 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1194 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1195 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1196 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1197 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1198 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1199 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1200 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1201 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
1202 ; SKIP-CACHE-INV-NEXT: s_endpgm
1203 ;
1204 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_store:
1205 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1206 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1207 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1208 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1209 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1210 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1211 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1212 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1213 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1214 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1215 ;
1216 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_store:
1217 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1218 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1219 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1220 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1221 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1222 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1223 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1224 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1225 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1226 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1227 ;
1228 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_store:
1229 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1230 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1231 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1232 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1233 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1234 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1235 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1236 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1237 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1238 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1239 ;
1240 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_store:
1241 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1242 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1243 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1244 ; GFX940-TGSPLIT-NEXT: s_nop 0
1245 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1246 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1247 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1248 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1249 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1250 ; GFX940-TGSPLIT-NEXT: s_endpgm
1251 ;
1252 ; GFX11-WGP-LABEL: global_singlethread_release_store:
1253 ; GFX11-WGP: ; %bb.0: ; %entry
1254 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1255 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1256 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1257 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1258 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1259 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1260 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1261 ; GFX11-WGP-NEXT: s_endpgm
1262 ;
1263 ; GFX11-CU-LABEL: global_singlethread_release_store:
1264 ; GFX11-CU: ; %bb.0: ; %entry
1265 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1266 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1267 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1268 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1269 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1270 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1271 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1272 ; GFX11-CU-NEXT: s_endpgm
1273 ;
1274 ; GFX12-WGP-LABEL: global_singlethread_release_store:
1275 ; GFX12-WGP: ; %bb.0: ; %entry
1276 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1277 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1278 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1279 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1280 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1281 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1282 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1283 ; GFX12-WGP-NEXT: s_endpgm
1284 ;
1285 ; GFX12-CU-LABEL: global_singlethread_release_store:
1286 ; GFX12-CU: ; %bb.0: ; %entry
1287 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1288 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1289 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1290 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1291 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1292 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1293 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1294 ; GFX12-CU-NEXT: s_endpgm
1295 i32 %in, ptr addrspace(1) %out) {
1296 entry:
1297 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread") release, align 4
1298 ret void
1299 }
1301 define amdgpu_kernel void @global_singlethread_seq_cst_store(
1302 ; GFX6-LABEL: global_singlethread_seq_cst_store:
1303 ; GFX6: ; %bb.0: ; %entry
1304 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
1305 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
1306 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1307 ; GFX6-NEXT: s_mov_b32 s11, s5
1308 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1309 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1310 ; GFX6-NEXT: s_mov_b32 s10, -1
1311 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1312 ; GFX6-NEXT: s_mov_b32 s5, s11
1313 ; GFX6-NEXT: s_mov_b32 s6, s10
1314 ; GFX6-NEXT: s_mov_b32 s7, s9
1315 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1316 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
1317 ; GFX6-NEXT: s_endpgm
1318 ;
1319 ; GFX7-LABEL: global_singlethread_seq_cst_store:
1320 ; GFX7: ; %bb.0: ; %entry
1321 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
1322 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
1323 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1324 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1325 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1326 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1327 ; GFX7-NEXT: flat_store_dword v[0:1], v2
1328 ; GFX7-NEXT: s_endpgm
1329 ;
1330 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_store:
1331 ; GFX10-WGP: ; %bb.0: ; %entry
1332 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
1333 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
1334 ; GFX10-WGP-NEXT: s_nop 0
1335 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1336 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1337 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1338 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1339 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
1340 ; GFX10-WGP-NEXT: s_endpgm
1341 ;
1342 ; GFX10-CU-LABEL: global_singlethread_seq_cst_store:
1343 ; GFX10-CU: ; %bb.0: ; %entry
1344 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
1345 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
1346 ; GFX10-CU-NEXT: s_nop 0
1347 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1348 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1349 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1350 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1351 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
1352 ; GFX10-CU-NEXT: s_endpgm
1353 ;
1354 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_store:
1355 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1356 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
1357 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
1358 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1359 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1360 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1361 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1362 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1363 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1364 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1365 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1366 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1367 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1368 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
1369 ; SKIP-CACHE-INV-NEXT: s_endpgm
1370 ;
1371 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_store:
1372 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1373 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1374 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1375 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1376 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1377 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1378 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1379 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1380 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1381 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1382 ;
1383 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_store:
1384 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1385 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
1386 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
1387 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1388 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
1389 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1390 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1391 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1392 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
1393 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1394 ;
1395 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_store:
1396 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1397 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1398 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1399 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1400 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1401 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1402 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1403 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1404 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1405 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1406 ;
1407 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_store:
1408 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1409 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
1410 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
1411 ; GFX940-TGSPLIT-NEXT: s_nop 0
1412 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
1413 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1414 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1415 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1416 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
1417 ; GFX940-TGSPLIT-NEXT: s_endpgm
1418 ;
1419 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_store:
1420 ; GFX11-WGP: ; %bb.0: ; %entry
1421 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1422 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1423 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1424 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1425 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1426 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1427 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1428 ; GFX11-WGP-NEXT: s_endpgm
1429 ;
1430 ; GFX11-CU-LABEL: global_singlethread_seq_cst_store:
1431 ; GFX11-CU: ; %bb.0: ; %entry
1432 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1433 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1434 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1435 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1436 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1437 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1438 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1439 ; GFX11-CU-NEXT: s_endpgm
1440 ;
1441 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_store:
1442 ; GFX12-WGP: ; %bb.0: ; %entry
1443 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
1444 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
1445 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1446 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1447 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1448 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1449 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
1450 ; GFX12-WGP-NEXT: s_endpgm
1451 ;
1452 ; GFX12-CU-LABEL: global_singlethread_seq_cst_store:
1453 ; GFX12-CU: ; %bb.0: ; %entry
1454 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
1455 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
1456 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
1457 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1458 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1459 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1460 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
1461 ; GFX12-CU-NEXT: s_endpgm
1462 i32 %in, ptr addrspace(1) %out) {
1463 entry:
1464 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread") seq_cst, align 4
1465 ret void
1466 }
1468 define amdgpu_kernel void @global_singlethread_monotonic_atomicrmw(
1469 ; GFX6-LABEL: global_singlethread_monotonic_atomicrmw:
1470 ; GFX6: ; %bb.0: ; %entry
1471 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1472 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
1473 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1474 ; GFX6-NEXT: s_mov_b32 s11, s5
1475 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1476 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1477 ; GFX6-NEXT: s_mov_b32 s10, -1
1478 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1479 ; GFX6-NEXT: s_mov_b32 s5, s11
1480 ; GFX6-NEXT: s_mov_b32 s6, s10
1481 ; GFX6-NEXT: s_mov_b32 s7, s9
1482 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1483 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
1484 ; GFX6-NEXT: s_endpgm
1485 ;
1486 ; GFX7-LABEL: global_singlethread_monotonic_atomicrmw:
1487 ; GFX7: ; %bb.0: ; %entry
1488 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
1489 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
1490 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
1491 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1492 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1493 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1494 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1495 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
1496 ; GFX7-NEXT: s_endpgm
1497 ;
1498 ; GFX10-WGP-LABEL: global_singlethread_monotonic_atomicrmw:
1499 ; GFX10-WGP: ; %bb.0: ; %entry
1500 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1501 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1502 ; GFX10-WGP-NEXT: s_nop 0
1503 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
1504 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1505 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1506 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
1507 ; GFX10-WGP-NEXT: s_endpgm
1508 ;
1509 ; GFX10-CU-LABEL: global_singlethread_monotonic_atomicrmw:
1510 ; GFX10-CU: ; %bb.0: ; %entry
1511 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1512 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1513 ; GFX10-CU-NEXT: s_nop 0
1514 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
1515 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1516 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1517 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
1518 ; GFX10-CU-NEXT: s_endpgm
1519 ;
1520 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_atomicrmw:
1521 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1522 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1523 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
1524 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1525 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1526 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1527 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1528 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1529 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1530 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1531 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1532 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1533 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1534 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
1535 ; SKIP-CACHE-INV-NEXT: s_endpgm
1536 ;
1537 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_atomicrmw:
1538 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1539 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1540 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1541 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1542 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1543 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1544 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1545 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1546 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1547 ;
1548 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_atomicrmw:
1549 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1550 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1551 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1552 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1553 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1554 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1555 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1556 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1557 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1558 ;
1559 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_atomicrmw:
1560 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1561 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1562 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1563 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1564 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1565 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1566 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1567 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1568 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1569 ;
1570 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_atomicrmw:
1571 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1572 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1573 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1574 ; GFX940-TGSPLIT-NEXT: s_nop 0
1575 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1576 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1577 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1578 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1579 ; GFX940-TGSPLIT-NEXT: s_endpgm
1580 ;
1581 ; GFX11-WGP-LABEL: global_singlethread_monotonic_atomicrmw:
1582 ; GFX11-WGP: ; %bb.0: ; %entry
1583 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1584 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1585 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1586 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1587 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1588 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1589 ; GFX11-WGP-NEXT: s_endpgm
1590 ;
1591 ; GFX11-CU-LABEL: global_singlethread_monotonic_atomicrmw:
1592 ; GFX11-CU: ; %bb.0: ; %entry
1593 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1594 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1595 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1596 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1597 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1598 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1599 ; GFX11-CU-NEXT: s_endpgm
1600 ;
1601 ; GFX12-WGP-LABEL: global_singlethread_monotonic_atomicrmw:
1602 ; GFX12-WGP: ; %bb.0: ; %entry
1603 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1604 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1605 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1606 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1607 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1608 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1609 ; GFX12-WGP-NEXT: s_endpgm
1610 ;
1611 ; GFX12-CU-LABEL: global_singlethread_monotonic_atomicrmw:
1612 ; GFX12-CU: ; %bb.0: ; %entry
1613 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1614 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1615 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1616 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1617 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1618 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1619 ; GFX12-CU-NEXT: s_endpgm
1620 ptr addrspace(1) %out, i32 %in) {
1621 entry:
1622 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") monotonic
1623 ret void
1624 }
1626 define amdgpu_kernel void @global_singlethread_acquire_atomicrmw(
1627 ; GFX6-LABEL: global_singlethread_acquire_atomicrmw:
1628 ; GFX6: ; %bb.0: ; %entry
1629 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1630 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
1631 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1632 ; GFX6-NEXT: s_mov_b32 s11, s5
1633 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1634 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1635 ; GFX6-NEXT: s_mov_b32 s10, -1
1636 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1637 ; GFX6-NEXT: s_mov_b32 s5, s11
1638 ; GFX6-NEXT: s_mov_b32 s6, s10
1639 ; GFX6-NEXT: s_mov_b32 s7, s9
1640 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1641 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
1642 ; GFX6-NEXT: s_endpgm
1643 ;
1644 ; GFX7-LABEL: global_singlethread_acquire_atomicrmw:
1645 ; GFX7: ; %bb.0: ; %entry
1646 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
1647 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
1648 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
1649 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1650 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1651 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1652 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1653 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
1654 ; GFX7-NEXT: s_endpgm
1655 ;
1656 ; GFX10-WGP-LABEL: global_singlethread_acquire_atomicrmw:
1657 ; GFX10-WGP: ; %bb.0: ; %entry
1658 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1659 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1660 ; GFX10-WGP-NEXT: s_nop 0
1661 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
1662 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1663 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1664 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
1665 ; GFX10-WGP-NEXT: s_endpgm
1666 ;
1667 ; GFX10-CU-LABEL: global_singlethread_acquire_atomicrmw:
1668 ; GFX10-CU: ; %bb.0: ; %entry
1669 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1670 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1671 ; GFX10-CU-NEXT: s_nop 0
1672 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
1673 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1674 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1675 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
1676 ; GFX10-CU-NEXT: s_endpgm
1677 ;
1678 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_atomicrmw:
1679 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1680 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1681 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
1682 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1683 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1684 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1685 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1686 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1687 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1688 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1689 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1690 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1691 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1692 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
1693 ; SKIP-CACHE-INV-NEXT: s_endpgm
1694 ;
1695 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_atomicrmw:
1696 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1697 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1698 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1699 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1700 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1701 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1702 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1703 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1704 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1705 ;
1706 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_atomicrmw:
1707 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1708 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1709 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1710 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1711 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1712 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1713 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1714 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1715 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1716 ;
1717 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_atomicrmw:
1718 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1719 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1720 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1721 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1722 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1723 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1724 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1725 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1726 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1727 ;
1728 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_atomicrmw:
1729 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1730 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1731 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1732 ; GFX940-TGSPLIT-NEXT: s_nop 0
1733 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1734 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1735 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1736 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1737 ; GFX940-TGSPLIT-NEXT: s_endpgm
1738 ;
1739 ; GFX11-WGP-LABEL: global_singlethread_acquire_atomicrmw:
1740 ; GFX11-WGP: ; %bb.0: ; %entry
1741 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1742 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1743 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1744 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1745 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1746 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1747 ; GFX11-WGP-NEXT: s_endpgm
1748 ;
1749 ; GFX11-CU-LABEL: global_singlethread_acquire_atomicrmw:
1750 ; GFX11-CU: ; %bb.0: ; %entry
1751 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1752 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1753 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1754 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1755 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1756 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1757 ; GFX11-CU-NEXT: s_endpgm
1758 ;
1759 ; GFX12-WGP-LABEL: global_singlethread_acquire_atomicrmw:
1760 ; GFX12-WGP: ; %bb.0: ; %entry
1761 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1762 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1763 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1764 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1765 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1766 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1767 ; GFX12-WGP-NEXT: s_endpgm
1768 ;
1769 ; GFX12-CU-LABEL: global_singlethread_acquire_atomicrmw:
1770 ; GFX12-CU: ; %bb.0: ; %entry
1771 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1772 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1773 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1774 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1775 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1776 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1777 ; GFX12-CU-NEXT: s_endpgm
1778 ptr addrspace(1) %out, i32 %in) {
1779 entry:
1780 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") acquire
1781 ret void
1782 }
1784 define amdgpu_kernel void @global_singlethread_release_atomicrmw(
1785 ; GFX6-LABEL: global_singlethread_release_atomicrmw:
1786 ; GFX6: ; %bb.0: ; %entry
1787 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1788 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
1789 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1790 ; GFX6-NEXT: s_mov_b32 s11, s5
1791 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1792 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1793 ; GFX6-NEXT: s_mov_b32 s10, -1
1794 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1795 ; GFX6-NEXT: s_mov_b32 s5, s11
1796 ; GFX6-NEXT: s_mov_b32 s6, s10
1797 ; GFX6-NEXT: s_mov_b32 s7, s9
1798 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1799 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
1800 ; GFX6-NEXT: s_endpgm
1801 ;
1802 ; GFX7-LABEL: global_singlethread_release_atomicrmw:
1803 ; GFX7: ; %bb.0: ; %entry
1804 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
1805 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
1806 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
1807 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1808 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1809 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1810 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1811 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
1812 ; GFX7-NEXT: s_endpgm
1813 ;
1814 ; GFX10-WGP-LABEL: global_singlethread_release_atomicrmw:
1815 ; GFX10-WGP: ; %bb.0: ; %entry
1816 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1817 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1818 ; GFX10-WGP-NEXT: s_nop 0
1819 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
1820 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1821 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1822 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
1823 ; GFX10-WGP-NEXT: s_endpgm
1824 ;
1825 ; GFX10-CU-LABEL: global_singlethread_release_atomicrmw:
1826 ; GFX10-CU: ; %bb.0: ; %entry
1827 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1828 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1829 ; GFX10-CU-NEXT: s_nop 0
1830 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
1831 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1832 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1833 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
1834 ; GFX10-CU-NEXT: s_endpgm
1835 ;
1836 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_atomicrmw:
1837 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1838 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1839 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
1840 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1841 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
1842 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1843 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
1844 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
1845 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1846 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
1847 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
1848 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
1849 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
1850 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
1851 ; SKIP-CACHE-INV-NEXT: s_endpgm
1852 ;
1853 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_atomicrmw:
1854 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
1855 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1856 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1857 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
1858 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1859 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1860 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1861 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1862 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
1863 ;
1864 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_atomicrmw:
1865 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
1866 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1867 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1868 ; GFX90A-TGSPLIT-NEXT: s_nop 0
1869 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
1870 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1871 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
1872 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
1873 ; GFX90A-TGSPLIT-NEXT: s_endpgm
1874 ;
1875 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_atomicrmw:
1876 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
1877 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1878 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1879 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
1880 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1881 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1882 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1883 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1884 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
1885 ;
1886 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_atomicrmw:
1887 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
1888 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
1889 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1890 ; GFX940-TGSPLIT-NEXT: s_nop 0
1891 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
1892 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
1893 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
1894 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
1895 ; GFX940-TGSPLIT-NEXT: s_endpgm
1896 ;
1897 ; GFX11-WGP-LABEL: global_singlethread_release_atomicrmw:
1898 ; GFX11-WGP: ; %bb.0: ; %entry
1899 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
1900 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1901 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1902 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
1903 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
1904 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1905 ; GFX11-WGP-NEXT: s_endpgm
1906 ;
1907 ; GFX11-CU-LABEL: global_singlethread_release_atomicrmw:
1908 ; GFX11-CU: ; %bb.0: ; %entry
1909 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
1910 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1911 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1912 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
1913 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
1914 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1915 ; GFX11-CU-NEXT: s_endpgm
1916 ;
1917 ; GFX12-WGP-LABEL: global_singlethread_release_atomicrmw:
1918 ; GFX12-WGP: ; %bb.0: ; %entry
1919 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
1920 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1921 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
1922 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
1923 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
1924 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1925 ; GFX12-WGP-NEXT: s_endpgm
1926 ;
1927 ; GFX12-CU-LABEL: global_singlethread_release_atomicrmw:
1928 ; GFX12-CU: ; %bb.0: ; %entry
1929 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
1930 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
1931 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
1932 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
1933 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
1934 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
1935 ; GFX12-CU-NEXT: s_endpgm
1936 ptr addrspace(1) %out, i32 %in) {
1937 entry:
1938 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") release
1939 ret void
1940 }
1942 define amdgpu_kernel void @global_singlethread_acq_rel_atomicrmw(
1943 ; GFX6-LABEL: global_singlethread_acq_rel_atomicrmw:
1944 ; GFX6: ; %bb.0: ; %entry
1945 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1946 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
1947 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
1948 ; GFX6-NEXT: s_mov_b32 s11, s5
1949 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
1950 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
1951 ; GFX6-NEXT: s_mov_b32 s10, -1
1952 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1953 ; GFX6-NEXT: s_mov_b32 s5, s11
1954 ; GFX6-NEXT: s_mov_b32 s6, s10
1955 ; GFX6-NEXT: s_mov_b32 s7, s9
1956 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
1957 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
1958 ; GFX6-NEXT: s_endpgm
1959 ;
1960 ; GFX7-LABEL: global_singlethread_acq_rel_atomicrmw:
1961 ; GFX7: ; %bb.0: ; %entry
1962 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
1963 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
1964 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
1965 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
1966 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
1967 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
1968 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
1969 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
1970 ; GFX7-NEXT: s_endpgm
1971 ;
1972 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_atomicrmw:
1973 ; GFX10-WGP: ; %bb.0: ; %entry
1974 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
1975 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1976 ; GFX10-WGP-NEXT: s_nop 0
1977 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
1978 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
1979 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
1980 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
1981 ; GFX10-WGP-NEXT: s_endpgm
1982 ;
1983 ; GFX10-CU-LABEL: global_singlethread_acq_rel_atomicrmw:
1984 ; GFX10-CU: ; %bb.0: ; %entry
1985 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
1986 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
1987 ; GFX10-CU-NEXT: s_nop 0
1988 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
1989 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
1990 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
1991 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
1992 ; GFX10-CU-NEXT: s_endpgm
1993 ;
1994 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_atomicrmw:
1995 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
1996 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
1997 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
1998 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
1999 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
2000 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2001 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
2002 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2003 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2004 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
2005 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
2006 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
2007 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2008 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
2009 ; SKIP-CACHE-INV-NEXT: s_endpgm
2010 ;
2011 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_atomicrmw:
2012 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2013 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2014 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2015 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
2016 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2017 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2018 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2019 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
2020 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2021 ;
2022 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_atomicrmw:
2023 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2024 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2025 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2026 ; GFX90A-TGSPLIT-NEXT: s_nop 0
2027 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2028 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2029 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2030 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
2031 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2032 ;
2033 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_atomicrmw:
2034 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2035 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2036 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2037 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
2038 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2039 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2040 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2041 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
2042 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2043 ;
2044 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_atomicrmw:
2045 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2046 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2047 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2048 ; GFX940-TGSPLIT-NEXT: s_nop 0
2049 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2050 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2051 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2052 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
2053 ; GFX940-TGSPLIT-NEXT: s_endpgm
2054 ;
2055 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_atomicrmw:
2056 ; GFX11-WGP: ; %bb.0: ; %entry
2057 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2058 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2059 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2060 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2061 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
2062 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2063 ; GFX11-WGP-NEXT: s_endpgm
2064 ;
2065 ; GFX11-CU-LABEL: global_singlethread_acq_rel_atomicrmw:
2066 ; GFX11-CU: ; %bb.0: ; %entry
2067 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2068 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2069 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2070 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2071 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
2072 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2073 ; GFX11-CU-NEXT: s_endpgm
2074 ;
2075 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_atomicrmw:
2076 ; GFX12-WGP: ; %bb.0: ; %entry
2077 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
2078 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2079 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2080 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
2081 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
2082 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2083 ; GFX12-WGP-NEXT: s_endpgm
2084 ;
2085 ; GFX12-CU-LABEL: global_singlethread_acq_rel_atomicrmw:
2086 ; GFX12-CU: ; %bb.0: ; %entry
2087 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
2088 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2089 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2090 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
2091 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
2092 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2093 ; GFX12-CU-NEXT: s_endpgm
2094 ptr addrspace(1) %out, i32 %in) {
2095 entry:
2096 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") acq_rel
2097 ret void
2098 }
2100 define amdgpu_kernel void @global_singlethread_seq_cst_atomicrmw(
2101 ; GFX6-LABEL: global_singlethread_seq_cst_atomicrmw:
2102 ; GFX6: ; %bb.0: ; %entry
2103 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2104 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
2105 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
2106 ; GFX6-NEXT: s_mov_b32 s11, s5
2107 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
2108 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
2109 ; GFX6-NEXT: s_mov_b32 s10, -1
2110 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
2111 ; GFX6-NEXT: s_mov_b32 s5, s11
2112 ; GFX6-NEXT: s_mov_b32 s6, s10
2113 ; GFX6-NEXT: s_mov_b32 s7, s9
2114 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
2115 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
2116 ; GFX6-NEXT: s_endpgm
2117 ;
2118 ; GFX7-LABEL: global_singlethread_seq_cst_atomicrmw:
2119 ; GFX7: ; %bb.0: ; %entry
2120 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
2121 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
2122 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
2123 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
2124 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
2125 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
2126 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
2127 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
2128 ; GFX7-NEXT: s_endpgm
2129 ;
2130 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_atomicrmw:
2131 ; GFX10-WGP: ; %bb.0: ; %entry
2132 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
2133 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2134 ; GFX10-WGP-NEXT: s_nop 0
2135 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
2136 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
2137 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
2138 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
2139 ; GFX10-WGP-NEXT: s_endpgm
2140 ;
2141 ; GFX10-CU-LABEL: global_singlethread_seq_cst_atomicrmw:
2142 ; GFX10-CU: ; %bb.0: ; %entry
2143 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
2144 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2145 ; GFX10-CU-NEXT: s_nop 0
2146 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
2147 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
2148 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
2149 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
2150 ; GFX10-CU-NEXT: s_endpgm
2151 ;
2152 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_atomicrmw:
2153 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2154 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2155 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
2156 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2157 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
2158 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2159 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
2160 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2161 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2162 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
2163 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
2164 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
2165 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2166 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
2167 ; SKIP-CACHE-INV-NEXT: s_endpgm
2168 ;
2169 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_atomicrmw:
2170 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2171 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2172 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2173 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
2174 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2175 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2176 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2177 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
2178 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2179 ;
2180 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_atomicrmw:
2181 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2182 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2183 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2184 ; GFX90A-TGSPLIT-NEXT: s_nop 0
2185 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2186 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2187 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2188 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
2189 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2190 ;
2191 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_atomicrmw:
2192 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2193 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2194 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2195 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
2196 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2197 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2198 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2199 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
2200 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2201 ;
2202 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_atomicrmw:
2203 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2204 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2205 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2206 ; GFX940-TGSPLIT-NEXT: s_nop 0
2207 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2208 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2209 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2210 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
2211 ; GFX940-TGSPLIT-NEXT: s_endpgm
2212 ;
2213 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_atomicrmw:
2214 ; GFX11-WGP: ; %bb.0: ; %entry
2215 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2216 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2217 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2218 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2219 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
2220 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2221 ; GFX11-WGP-NEXT: s_endpgm
2222 ;
2223 ; GFX11-CU-LABEL: global_singlethread_seq_cst_atomicrmw:
2224 ; GFX11-CU: ; %bb.0: ; %entry
2225 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2226 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2227 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2228 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2229 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
2230 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2231 ; GFX11-CU-NEXT: s_endpgm
2232 ;
2233 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_atomicrmw:
2234 ; GFX12-WGP: ; %bb.0: ; %entry
2235 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
2236 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2237 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2238 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
2239 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
2240 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2241 ; GFX12-WGP-NEXT: s_endpgm
2242 ;
2243 ; GFX12-CU-LABEL: global_singlethread_seq_cst_atomicrmw:
2244 ; GFX12-CU: ; %bb.0: ; %entry
2245 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
2246 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2247 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2248 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
2249 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
2250 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
2251 ; GFX12-CU-NEXT: s_endpgm
2252 ptr addrspace(1) %out, i32 %in) {
2253 entry:
2254 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") seq_cst
2255 ret void
2256 }
2258 define amdgpu_kernel void @global_singlethread_acquire_ret_atomicrmw(
2259 ; GFX6-LABEL: global_singlethread_acquire_ret_atomicrmw:
2260 ; GFX6: ; %bb.0: ; %entry
2261 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2262 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
2263 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
2264 ; GFX6-NEXT: s_mov_b32 s11, s5
2265 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
2266 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
2267 ; GFX6-NEXT: s_mov_b32 s10, -1
2268 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
2269 ; GFX6-NEXT: s_mov_b32 s5, s11
2270 ; GFX6-NEXT: s_mov_b32 s6, s10
2271 ; GFX6-NEXT: s_mov_b32 s7, s9
2272 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
2273 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
2274 ; GFX6-NEXT: s_waitcnt vmcnt(0)
2275 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
2276 ; GFX6-NEXT: s_endpgm
2277 ;
2278 ; GFX7-LABEL: global_singlethread_acquire_ret_atomicrmw:
2279 ; GFX7: ; %bb.0: ; %entry
2280 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2281 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
2282 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
2283 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2284 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2285 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
2286 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
2287 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2288 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2289 ; GFX7-NEXT: s_waitcnt vmcnt(0)
2290 ; GFX7-NEXT: flat_store_dword v[0:1], v2
2291 ; GFX7-NEXT: s_endpgm
2292 ;
2293 ; GFX10-WGP-LABEL: global_singlethread_acquire_ret_atomicrmw:
2294 ; GFX10-WGP: ; %bb.0: ; %entry
2295 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
2296 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2297 ; GFX10-WGP-NEXT: s_nop 0
2298 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
2299 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
2300 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
2301 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2302 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
2303 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
2304 ; GFX10-WGP-NEXT: s_endpgm
2305 ;
2306 ; GFX10-CU-LABEL: global_singlethread_acquire_ret_atomicrmw:
2307 ; GFX10-CU: ; %bb.0: ; %entry
2308 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
2309 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2310 ; GFX10-CU-NEXT: s_nop 0
2311 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
2312 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
2313 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
2314 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2315 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
2316 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
2317 ; GFX10-CU-NEXT: s_endpgm
2318 ;
2319 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_ret_atomicrmw:
2320 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2321 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2322 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
2323 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2324 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
2325 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2326 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
2327 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2328 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2329 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
2330 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
2331 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
2332 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2333 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
2334 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2335 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2336 ; SKIP-CACHE-INV-NEXT: s_endpgm
2337 ;
2338 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_ret_atomicrmw:
2339 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2340 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2341 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2342 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
2343 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2344 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2345 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2346 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2347 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2348 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2349 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2350 ;
2351 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_ret_atomicrmw:
2352 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2353 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2354 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2355 ; GFX90A-TGSPLIT-NEXT: s_nop 0
2356 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2357 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2358 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2359 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2360 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2361 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2362 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2363 ;
2364 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_ret_atomicrmw:
2365 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2366 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2367 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2368 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
2369 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2370 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2371 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2372 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2373 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2374 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2375 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2376 ;
2377 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_ret_atomicrmw:
2378 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2379 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2380 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2381 ; GFX940-TGSPLIT-NEXT: s_nop 0
2382 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2383 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2384 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2385 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2386 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2387 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2388 ; GFX940-TGSPLIT-NEXT: s_endpgm
2389 ;
2390 ; GFX11-WGP-LABEL: global_singlethread_acquire_ret_atomicrmw:
2391 ; GFX11-WGP: ; %bb.0: ; %entry
2392 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2393 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2394 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2395 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2396 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
2397 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2398 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
2399 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2400 ; GFX11-WGP-NEXT: s_endpgm
2401 ;
2402 ; GFX11-CU-LABEL: global_singlethread_acquire_ret_atomicrmw:
2403 ; GFX11-CU: ; %bb.0: ; %entry
2404 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2405 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2406 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2407 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2408 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
2409 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2410 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
2411 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2412 ; GFX11-CU-NEXT: s_endpgm
2413 ;
2414 ; GFX12-WGP-LABEL: global_singlethread_acquire_ret_atomicrmw:
2415 ; GFX12-WGP: ; %bb.0: ; %entry
2416 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
2417 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2418 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2419 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
2420 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
2421 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2422 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
2423 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2424 ; GFX12-WGP-NEXT: s_endpgm
2425 ;
2426 ; GFX12-CU-LABEL: global_singlethread_acquire_ret_atomicrmw:
2427 ; GFX12-CU: ; %bb.0: ; %entry
2428 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
2429 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2430 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2431 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
2432 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
2433 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2434 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
2435 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2436 ; GFX12-CU-NEXT: s_endpgm
2437 ptr addrspace(1) %out, i32 %in) {
2438 entry:
2439 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") acquire
2440 store i32 %val, ptr addrspace(1) %out, align 4
2441 ret void
2442 }
2444 define amdgpu_kernel void @global_singlethread_acq_rel_ret_atomicrmw(
2445 ; GFX6-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2446 ; GFX6: ; %bb.0: ; %entry
2447 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2448 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
2449 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
2450 ; GFX6-NEXT: s_mov_b32 s11, s5
2451 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
2452 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
2453 ; GFX6-NEXT: s_mov_b32 s10, -1
2454 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
2455 ; GFX6-NEXT: s_mov_b32 s5, s11
2456 ; GFX6-NEXT: s_mov_b32 s6, s10
2457 ; GFX6-NEXT: s_mov_b32 s7, s9
2458 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
2459 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
2460 ; GFX6-NEXT: s_waitcnt vmcnt(0)
2461 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
2462 ; GFX6-NEXT: s_endpgm
2463 ;
2464 ; GFX7-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2465 ; GFX7: ; %bb.0: ; %entry
2466 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2467 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
2468 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
2469 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2470 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2471 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
2472 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
2473 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2474 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2475 ; GFX7-NEXT: s_waitcnt vmcnt(0)
2476 ; GFX7-NEXT: flat_store_dword v[0:1], v2
2477 ; GFX7-NEXT: s_endpgm
2478 ;
2479 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2480 ; GFX10-WGP: ; %bb.0: ; %entry
2481 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
2482 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2483 ; GFX10-WGP-NEXT: s_nop 0
2484 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
2485 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
2486 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
2487 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2488 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
2489 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
2490 ; GFX10-WGP-NEXT: s_endpgm
2491 ;
2492 ; GFX10-CU-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2493 ; GFX10-CU: ; %bb.0: ; %entry
2494 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
2495 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2496 ; GFX10-CU-NEXT: s_nop 0
2497 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
2498 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
2499 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
2500 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2501 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
2502 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
2503 ; GFX10-CU-NEXT: s_endpgm
2504 ;
2505 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2506 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2507 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2508 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
2509 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2510 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
2511 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2512 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
2513 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2514 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2515 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
2516 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
2517 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
2518 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2519 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
2520 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2521 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2522 ; SKIP-CACHE-INV-NEXT: s_endpgm
2523 ;
2524 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2525 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2526 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2527 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2528 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
2529 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2530 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2531 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2532 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2533 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2534 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2535 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2536 ;
2537 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2538 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2539 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2540 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2541 ; GFX90A-TGSPLIT-NEXT: s_nop 0
2542 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2543 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2544 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2545 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2546 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2547 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2548 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2549 ;
2550 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2551 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2552 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2553 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2554 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
2555 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2556 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2557 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2558 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2559 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2560 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2561 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2562 ;
2563 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2564 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2565 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2566 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2567 ; GFX940-TGSPLIT-NEXT: s_nop 0
2568 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2569 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2570 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2571 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2572 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2573 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2574 ; GFX940-TGSPLIT-NEXT: s_endpgm
2575 ;
2576 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2577 ; GFX11-WGP: ; %bb.0: ; %entry
2578 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2579 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2580 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2581 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2582 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
2583 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2584 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
2585 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2586 ; GFX11-WGP-NEXT: s_endpgm
2587 ;
2588 ; GFX11-CU-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2589 ; GFX11-CU: ; %bb.0: ; %entry
2590 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2591 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2592 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2593 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2594 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
2595 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2596 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
2597 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2598 ; GFX11-CU-NEXT: s_endpgm
2599 ;
2600 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2601 ; GFX12-WGP: ; %bb.0: ; %entry
2602 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
2603 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2604 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2605 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
2606 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
2607 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2608 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
2609 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2610 ; GFX12-WGP-NEXT: s_endpgm
2611 ;
2612 ; GFX12-CU-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
2613 ; GFX12-CU: ; %bb.0: ; %entry
2614 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
2615 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2616 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2617 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
2618 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
2619 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2620 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
2621 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2622 ; GFX12-CU-NEXT: s_endpgm
2623 ptr addrspace(1) %out, i32 %in) {
2624 entry:
2625 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") acq_rel
2626 store i32 %val, ptr addrspace(1) %out, align 4
2627 ret void
2628 }
2630 define amdgpu_kernel void @global_singlethread_seq_cst_ret_atomicrmw(
2631 ; GFX6-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2632 ; GFX6: ; %bb.0: ; %entry
2633 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2634 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
2635 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
2636 ; GFX6-NEXT: s_mov_b32 s11, s5
2637 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
2638 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
2639 ; GFX6-NEXT: s_mov_b32 s10, -1
2640 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
2641 ; GFX6-NEXT: s_mov_b32 s5, s11
2642 ; GFX6-NEXT: s_mov_b32 s6, s10
2643 ; GFX6-NEXT: s_mov_b32 s7, s9
2644 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
2645 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
2646 ; GFX6-NEXT: s_waitcnt vmcnt(0)
2647 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
2648 ; GFX6-NEXT: s_endpgm
2649 ;
2650 ; GFX7-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2651 ; GFX7: ; %bb.0: ; %entry
2652 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2653 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
2654 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
2655 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2656 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2657 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
2658 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
2659 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2660 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2661 ; GFX7-NEXT: s_waitcnt vmcnt(0)
2662 ; GFX7-NEXT: flat_store_dword v[0:1], v2
2663 ; GFX7-NEXT: s_endpgm
2664 ;
2665 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2666 ; GFX10-WGP: ; %bb.0: ; %entry
2667 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
2668 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2669 ; GFX10-WGP-NEXT: s_nop 0
2670 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
2671 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
2672 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
2673 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2674 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
2675 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
2676 ; GFX10-WGP-NEXT: s_endpgm
2677 ;
2678 ; GFX10-CU-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2679 ; GFX10-CU: ; %bb.0: ; %entry
2680 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
2681 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2682 ; GFX10-CU-NEXT: s_nop 0
2683 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
2684 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
2685 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
2686 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2687 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
2688 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
2689 ; GFX10-CU-NEXT: s_endpgm
2690 ;
2691 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2692 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2693 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2694 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
2695 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2696 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
2697 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2698 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
2699 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
2700 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2701 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
2702 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
2703 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
2704 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
2705 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
2706 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
2707 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
2708 ; SKIP-CACHE-INV-NEXT: s_endpgm
2709 ;
2710 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2711 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2712 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2713 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2714 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
2715 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2716 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2717 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2718 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2719 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2720 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2721 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2722 ;
2723 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2724 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2725 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2726 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2727 ; GFX90A-TGSPLIT-NEXT: s_nop 0
2728 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
2729 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2730 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2731 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
2732 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2733 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
2734 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2735 ;
2736 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2737 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2738 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2739 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2740 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
2741 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2742 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2743 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2744 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2745 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
2746 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2747 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2748 ;
2749 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2750 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2751 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2752 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2753 ; GFX940-TGSPLIT-NEXT: s_nop 0
2754 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
2755 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2756 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2757 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
2758 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
2759 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
2760 ; GFX940-TGSPLIT-NEXT: s_endpgm
2761 ;
2762 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2763 ; GFX11-WGP: ; %bb.0: ; %entry
2764 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2765 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2766 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2767 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2768 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
2769 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2770 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
2771 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2772 ; GFX11-WGP-NEXT: s_endpgm
2773 ;
2774 ; GFX11-CU-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2775 ; GFX11-CU: ; %bb.0: ; %entry
2776 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2777 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2778 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2779 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2780 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
2781 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
2782 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
2783 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2784 ; GFX11-CU-NEXT: s_endpgm
2785 ;
2786 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2787 ; GFX12-WGP: ; %bb.0: ; %entry
2788 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
2789 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2790 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
2791 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
2792 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
2793 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2794 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
2795 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
2796 ; GFX12-WGP-NEXT: s_endpgm
2797 ;
2798 ; GFX12-CU-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
2799 ; GFX12-CU: ; %bb.0: ; %entry
2800 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
2801 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
2802 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
2803 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
2804 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
2805 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
2806 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
2807 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
2808 ; GFX12-CU-NEXT: s_endpgm
2809 ptr addrspace(1) %out, i32 %in) {
2810 entry:
2811 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread") seq_cst
2812 store i32 %val, ptr addrspace(1) %out, align 4
2813 ret void
2814 }
2816 define amdgpu_kernel void @global_singlethread_monotonic_monotonic_cmpxchg(
2817 ; GFX6-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2818 ; GFX6: ; %bb.0: ; %entry
2819 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
2820 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
2821 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
2822 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
2823 ; GFX6-NEXT: s_mov_b32 s12, s5
2824 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
2825 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
2826 ; GFX6-NEXT: s_mov_b32 s11, -1
2827 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
2828 ; GFX6-NEXT: s_mov_b32 s5, s12
2829 ; GFX6-NEXT: s_mov_b32 s6, s11
2830 ; GFX6-NEXT: s_mov_b32 s7, s10
2831 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
2832 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
2833 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
2834 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
2835 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
2836 ; GFX6-NEXT: s_endpgm
2837 ;
2838 ; GFX7-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2839 ; GFX7: ; %bb.0: ; %entry
2840 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
2841 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
2842 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
2843 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
2844 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
2845 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
2846 ; GFX7-NEXT: s_mov_b32 s4, s8
2847 ; GFX7-NEXT: s_mov_b32 s5, s9
2848 ; GFX7-NEXT: s_mov_b32 s9, s10
2849 ; GFX7-NEXT: s_mov_b32 s8, s11
2850 ; GFX7-NEXT: s_add_u32 s4, s4, s9
2851 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
2852 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
2853 ; GFX7-NEXT: s_mov_b32 s5, s8
2854 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
2855 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
2856 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
2857 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
2858 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
2859 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
2860 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
2861 ; GFX7-NEXT: s_endpgm
2862 ;
2863 ; GFX10-WGP-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2864 ; GFX10-WGP: ; %bb.0: ; %entry
2865 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
2866 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
2867 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
2868 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
2869 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
2870 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
2871 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
2872 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
2873 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
2874 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
2875 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
2876 ; GFX10-WGP-NEXT: s_endpgm
2877 ;
2878 ; GFX10-CU-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2879 ; GFX10-CU: ; %bb.0: ; %entry
2880 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
2881 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
2882 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
2883 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
2884 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
2885 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
2886 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
2887 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
2888 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
2889 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
2890 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
2891 ; GFX10-CU-NEXT: s_endpgm
2892 ;
2893 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2894 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
2895 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
2896 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
2897 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
2898 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
2899 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
2900 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
2901 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
2902 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
2903 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
2904 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
2905 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
2906 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
2907 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
2908 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
2909 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
2910 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
2911 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
2912 ; SKIP-CACHE-INV-NEXT: s_endpgm
2913 ;
2914 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2915 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
2916 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
2917 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2918 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
2919 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
2920 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
2921 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2922 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
2923 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2924 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
2925 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
2926 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
2927 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
2928 ;
2929 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2930 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
2931 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
2932 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2933 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
2934 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
2935 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
2936 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2937 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
2938 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
2939 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
2940 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
2941 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
2942 ; GFX90A-TGSPLIT-NEXT: s_endpgm
2943 ;
2944 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2945 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
2946 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
2947 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2948 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
2949 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
2950 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
2951 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2952 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
2953 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2954 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
2955 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
2956 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
2957 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
2958 ;
2959 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2960 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
2961 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
2962 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
2963 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
2964 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
2965 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
2966 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
2967 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
2968 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
2969 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
2970 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
2971 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
2972 ; GFX940-TGSPLIT-NEXT: s_endpgm
2973 ;
2974 ; GFX11-WGP-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2975 ; GFX11-WGP: ; %bb.0: ; %entry
2976 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
2977 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
2978 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
2979 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
2980 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
2981 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
2982 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
2983 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
2984 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
2985 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
2986 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
2987 ; GFX11-WGP-NEXT: s_endpgm
2988 ;
2989 ; GFX11-CU-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
2990 ; GFX11-CU: ; %bb.0: ; %entry
2991 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
2992 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
2993 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
2994 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
2995 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
2996 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
2997 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
2998 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
2999 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3000 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
3001 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3002 ; GFX11-CU-NEXT: s_endpgm
3003 ;
3004 ; GFX12-WGP-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
3005 ; GFX12-WGP: ; %bb.0: ; %entry
3006 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3007 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
3008 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3009 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3010 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3011 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
3012 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
3013 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
3014 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3015 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
3016 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3017 ; GFX12-WGP-NEXT: s_endpgm
3018 ;
3019 ; GFX12-CU-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
3020 ; GFX12-CU: ; %bb.0: ; %entry
3021 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3022 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
3023 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3024 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3025 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3026 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
3027 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
3028 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
3029 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3030 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
3031 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3032 ; GFX12-CU-NEXT: s_endpgm
3033 ptr addrspace(1) %out, i32 %in, i32 %old) {
3034 entry:
3035 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
3036 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
3037 ret void
3038 }
3040 define amdgpu_kernel void @global_singlethread_acquire_monotonic_cmpxchg(
3041 ; GFX6-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3042 ; GFX6: ; %bb.0: ; %entry
3043 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
3044 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
3045 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
3046 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
3047 ; GFX6-NEXT: s_mov_b32 s12, s5
3048 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
3049 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
3050 ; GFX6-NEXT: s_mov_b32 s11, -1
3051 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
3052 ; GFX6-NEXT: s_mov_b32 s5, s12
3053 ; GFX6-NEXT: s_mov_b32 s6, s11
3054 ; GFX6-NEXT: s_mov_b32 s7, s10
3055 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
3056 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
3057 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3058 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
3059 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3060 ; GFX6-NEXT: s_endpgm
3061 ;
3062 ; GFX7-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3063 ; GFX7: ; %bb.0: ; %entry
3064 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
3065 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
3066 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
3067 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
3068 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
3069 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
3070 ; GFX7-NEXT: s_mov_b32 s4, s8
3071 ; GFX7-NEXT: s_mov_b32 s5, s9
3072 ; GFX7-NEXT: s_mov_b32 s9, s10
3073 ; GFX7-NEXT: s_mov_b32 s8, s11
3074 ; GFX7-NEXT: s_add_u32 s4, s4, s9
3075 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
3076 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
3077 ; GFX7-NEXT: s_mov_b32 s5, s8
3078 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
3079 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
3080 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3081 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
3082 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
3083 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
3084 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3085 ; GFX7-NEXT: s_endpgm
3086 ;
3087 ; GFX10-WGP-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3088 ; GFX10-WGP: ; %bb.0: ; %entry
3089 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
3090 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
3091 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3092 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
3093 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
3094 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
3095 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
3096 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
3097 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3098 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
3099 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3100 ; GFX10-WGP-NEXT: s_endpgm
3101 ;
3102 ; GFX10-CU-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3103 ; GFX10-CU: ; %bb.0: ; %entry
3104 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
3105 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
3106 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3107 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
3108 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
3109 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
3110 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
3111 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
3112 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3113 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
3114 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3115 ; GFX10-CU-NEXT: s_endpgm
3116 ;
3117 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3118 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3119 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
3120 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
3121 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
3122 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3123 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
3124 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
3125 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
3126 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
3127 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
3128 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
3129 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
3130 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
3131 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
3132 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
3133 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3134 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
3135 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
3136 ; SKIP-CACHE-INV-NEXT: s_endpgm
3137 ;
3138 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3139 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
3140 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3141 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3142 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3143 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3144 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3145 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3146 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3147 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3148 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3149 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3150 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3151 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
3152 ;
3153 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3154 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
3155 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3156 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3157 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3158 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3159 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3160 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3161 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3162 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3163 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3164 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3165 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3166 ; GFX90A-TGSPLIT-NEXT: s_endpgm
3167 ;
3168 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3169 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
3170 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3171 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3172 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3173 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3174 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3175 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3176 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3177 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3178 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3179 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3180 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3181 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
3182 ;
3183 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3184 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
3185 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3186 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3187 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3188 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3189 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3190 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3191 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3192 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3193 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3194 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3195 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3196 ; GFX940-TGSPLIT-NEXT: s_endpgm
3197 ;
3198 ; GFX11-WGP-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3199 ; GFX11-WGP: ; %bb.0: ; %entry
3200 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3201 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
3202 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3203 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3204 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3205 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
3206 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
3207 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
3208 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3209 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
3210 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3211 ; GFX11-WGP-NEXT: s_endpgm
3212 ;
3213 ; GFX11-CU-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3214 ; GFX11-CU: ; %bb.0: ; %entry
3215 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3216 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
3217 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3218 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3219 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3220 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
3221 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
3222 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
3223 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3224 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
3225 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3226 ; GFX11-CU-NEXT: s_endpgm
3227 ;
3228 ; GFX12-WGP-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3229 ; GFX12-WGP: ; %bb.0: ; %entry
3230 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3231 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
3232 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3233 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3234 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3235 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
3236 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
3237 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
3238 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3239 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
3240 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3241 ; GFX12-WGP-NEXT: s_endpgm
3242 ;
3243 ; GFX12-CU-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
3244 ; GFX12-CU: ; %bb.0: ; %entry
3245 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3246 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
3247 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3248 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3249 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3250 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
3251 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
3252 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
3253 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3254 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
3255 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3256 ; GFX12-CU-NEXT: s_endpgm
3257 ptr addrspace(1) %out, i32 %in, i32 %old) {
3258 entry:
3259 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
3260 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
3261 ret void
3262 }
3264 define amdgpu_kernel void @global_singlethread_release_monotonic_cmpxchg(
3265 ; GFX6-LABEL: global_singlethread_release_monotonic_cmpxchg:
3266 ; GFX6: ; %bb.0: ; %entry
3267 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
3268 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
3269 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
3270 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
3271 ; GFX6-NEXT: s_mov_b32 s12, s5
3272 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
3273 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
3274 ; GFX6-NEXT: s_mov_b32 s11, -1
3275 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
3276 ; GFX6-NEXT: s_mov_b32 s5, s12
3277 ; GFX6-NEXT: s_mov_b32 s6, s11
3278 ; GFX6-NEXT: s_mov_b32 s7, s10
3279 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
3280 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
3281 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3282 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
3283 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3284 ; GFX6-NEXT: s_endpgm
3285 ;
3286 ; GFX7-LABEL: global_singlethread_release_monotonic_cmpxchg:
3287 ; GFX7: ; %bb.0: ; %entry
3288 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
3289 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
3290 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
3291 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
3292 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
3293 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
3294 ; GFX7-NEXT: s_mov_b32 s4, s8
3295 ; GFX7-NEXT: s_mov_b32 s5, s9
3296 ; GFX7-NEXT: s_mov_b32 s9, s10
3297 ; GFX7-NEXT: s_mov_b32 s8, s11
3298 ; GFX7-NEXT: s_add_u32 s4, s4, s9
3299 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
3300 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
3301 ; GFX7-NEXT: s_mov_b32 s5, s8
3302 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
3303 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
3304 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3305 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
3306 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
3307 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
3308 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3309 ; GFX7-NEXT: s_endpgm
3310 ;
3311 ; GFX10-WGP-LABEL: global_singlethread_release_monotonic_cmpxchg:
3312 ; GFX10-WGP: ; %bb.0: ; %entry
3313 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
3314 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
3315 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3316 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
3317 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
3318 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
3319 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
3320 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
3321 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3322 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
3323 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3324 ; GFX10-WGP-NEXT: s_endpgm
3325 ;
3326 ; GFX10-CU-LABEL: global_singlethread_release_monotonic_cmpxchg:
3327 ; GFX10-CU: ; %bb.0: ; %entry
3328 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
3329 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
3330 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3331 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
3332 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
3333 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
3334 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
3335 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
3336 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3337 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
3338 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3339 ; GFX10-CU-NEXT: s_endpgm
3340 ;
3341 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_monotonic_cmpxchg:
3342 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3343 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
3344 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
3345 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
3346 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3347 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
3348 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
3349 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
3350 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
3351 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
3352 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
3353 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
3354 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
3355 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
3356 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
3357 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3358 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
3359 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
3360 ; SKIP-CACHE-INV-NEXT: s_endpgm
3361 ;
3362 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_monotonic_cmpxchg:
3363 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
3364 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3365 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3366 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3367 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3368 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3369 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3370 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3371 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3372 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3373 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3374 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3375 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
3376 ;
3377 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_monotonic_cmpxchg:
3378 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
3379 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3380 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3381 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3382 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3383 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3384 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3385 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3386 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3387 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3388 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3389 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3390 ; GFX90A-TGSPLIT-NEXT: s_endpgm
3391 ;
3392 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_monotonic_cmpxchg:
3393 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
3394 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3395 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3396 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3397 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3398 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3399 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3400 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3401 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3402 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3403 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3404 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3405 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
3406 ;
3407 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_monotonic_cmpxchg:
3408 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
3409 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3410 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3411 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3412 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3413 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3414 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3415 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3416 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3417 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3418 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3419 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3420 ; GFX940-TGSPLIT-NEXT: s_endpgm
3421 ;
3422 ; GFX11-WGP-LABEL: global_singlethread_release_monotonic_cmpxchg:
3423 ; GFX11-WGP: ; %bb.0: ; %entry
3424 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3425 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
3426 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3427 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3428 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3429 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
3430 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
3431 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
3432 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3433 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
3434 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3435 ; GFX11-WGP-NEXT: s_endpgm
3436 ;
3437 ; GFX11-CU-LABEL: global_singlethread_release_monotonic_cmpxchg:
3438 ; GFX11-CU: ; %bb.0: ; %entry
3439 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3440 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
3441 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3442 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3443 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3444 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
3445 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
3446 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
3447 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3448 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
3449 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3450 ; GFX11-CU-NEXT: s_endpgm
3451 ;
3452 ; GFX12-WGP-LABEL: global_singlethread_release_monotonic_cmpxchg:
3453 ; GFX12-WGP: ; %bb.0: ; %entry
3454 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3455 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
3456 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3457 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3458 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3459 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
3460 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
3461 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
3462 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3463 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
3464 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3465 ; GFX12-WGP-NEXT: s_endpgm
3466 ;
3467 ; GFX12-CU-LABEL: global_singlethread_release_monotonic_cmpxchg:
3468 ; GFX12-CU: ; %bb.0: ; %entry
3469 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3470 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
3471 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3472 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3473 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3474 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
3475 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
3476 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
3477 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3478 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
3479 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3480 ; GFX12-CU-NEXT: s_endpgm
3481 ptr addrspace(1) %out, i32 %in, i32 %old) {
3482 entry:
3483 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
3484 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
3485 ret void
3486 }
3488 define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_cmpxchg(
3489 ; GFX6-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3490 ; GFX6: ; %bb.0: ; %entry
3491 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
3492 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
3493 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
3494 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
3495 ; GFX6-NEXT: s_mov_b32 s12, s5
3496 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
3497 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
3498 ; GFX6-NEXT: s_mov_b32 s11, -1
3499 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
3500 ; GFX6-NEXT: s_mov_b32 s5, s12
3501 ; GFX6-NEXT: s_mov_b32 s6, s11
3502 ; GFX6-NEXT: s_mov_b32 s7, s10
3503 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
3504 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
3505 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3506 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
3507 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3508 ; GFX6-NEXT: s_endpgm
3509 ;
3510 ; GFX7-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3511 ; GFX7: ; %bb.0: ; %entry
3512 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
3513 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
3514 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
3515 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
3516 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
3517 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
3518 ; GFX7-NEXT: s_mov_b32 s4, s8
3519 ; GFX7-NEXT: s_mov_b32 s5, s9
3520 ; GFX7-NEXT: s_mov_b32 s9, s10
3521 ; GFX7-NEXT: s_mov_b32 s8, s11
3522 ; GFX7-NEXT: s_add_u32 s4, s4, s9
3523 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
3524 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
3525 ; GFX7-NEXT: s_mov_b32 s5, s8
3526 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
3527 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
3528 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3529 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
3530 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
3531 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
3532 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3533 ; GFX7-NEXT: s_endpgm
3534 ;
3535 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3536 ; GFX10-WGP: ; %bb.0: ; %entry
3537 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
3538 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
3539 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3540 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
3541 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
3542 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
3543 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
3544 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
3545 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3546 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
3547 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3548 ; GFX10-WGP-NEXT: s_endpgm
3549 ;
3550 ; GFX10-CU-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3551 ; GFX10-CU: ; %bb.0: ; %entry
3552 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
3553 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
3554 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3555 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
3556 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
3557 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
3558 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
3559 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
3560 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3561 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
3562 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3563 ; GFX10-CU-NEXT: s_endpgm
3564 ;
3565 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3566 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3567 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
3568 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
3569 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
3570 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3571 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
3572 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
3573 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
3574 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
3575 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
3576 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
3577 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
3578 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
3579 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
3580 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
3581 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3582 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
3583 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
3584 ; SKIP-CACHE-INV-NEXT: s_endpgm
3585 ;
3586 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3587 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
3588 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3589 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3590 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3591 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3592 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3593 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3594 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3595 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3596 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3597 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3598 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3599 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
3600 ;
3601 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3602 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
3603 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3604 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3605 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3606 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3607 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3608 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3609 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3610 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3611 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3612 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3613 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3614 ; GFX90A-TGSPLIT-NEXT: s_endpgm
3615 ;
3616 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3617 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
3618 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3619 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3620 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3621 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3622 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3623 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3624 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3625 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3626 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3627 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3628 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3629 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
3630 ;
3631 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3632 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
3633 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3634 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3635 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3636 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3637 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3638 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3639 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3640 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3641 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3642 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3643 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3644 ; GFX940-TGSPLIT-NEXT: s_endpgm
3645 ;
3646 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3647 ; GFX11-WGP: ; %bb.0: ; %entry
3648 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3649 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
3650 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3651 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3652 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3653 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
3654 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
3655 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
3656 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3657 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
3658 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3659 ; GFX11-WGP-NEXT: s_endpgm
3660 ;
3661 ; GFX11-CU-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3662 ; GFX11-CU: ; %bb.0: ; %entry
3663 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3664 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
3665 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3666 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3667 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3668 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
3669 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
3670 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
3671 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3672 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
3673 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3674 ; GFX11-CU-NEXT: s_endpgm
3675 ;
3676 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3677 ; GFX12-WGP: ; %bb.0: ; %entry
3678 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3679 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
3680 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3681 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3682 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3683 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
3684 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
3685 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
3686 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3687 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
3688 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3689 ; GFX12-WGP-NEXT: s_endpgm
3690 ;
3691 ; GFX12-CU-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
3692 ; GFX12-CU: ; %bb.0: ; %entry
3693 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3694 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
3695 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3696 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3697 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3698 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
3699 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
3700 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
3701 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3702 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
3703 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3704 ; GFX12-CU-NEXT: s_endpgm
3705 ptr addrspace(1) %out, i32 %in, i32 %old) {
3706 entry:
3707 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
3708 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
3709 ret void
3710 }
3712 define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_cmpxchg(
3713 ; GFX6-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3714 ; GFX6: ; %bb.0: ; %entry
3715 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
3716 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
3717 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
3718 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
3719 ; GFX6-NEXT: s_mov_b32 s12, s5
3720 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
3721 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
3722 ; GFX6-NEXT: s_mov_b32 s11, -1
3723 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
3724 ; GFX6-NEXT: s_mov_b32 s5, s12
3725 ; GFX6-NEXT: s_mov_b32 s6, s11
3726 ; GFX6-NEXT: s_mov_b32 s7, s10
3727 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
3728 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
3729 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3730 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
3731 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3732 ; GFX6-NEXT: s_endpgm
3733 ;
3734 ; GFX7-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3735 ; GFX7: ; %bb.0: ; %entry
3736 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
3737 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
3738 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
3739 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
3740 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
3741 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
3742 ; GFX7-NEXT: s_mov_b32 s4, s8
3743 ; GFX7-NEXT: s_mov_b32 s5, s9
3744 ; GFX7-NEXT: s_mov_b32 s9, s10
3745 ; GFX7-NEXT: s_mov_b32 s8, s11
3746 ; GFX7-NEXT: s_add_u32 s4, s4, s9
3747 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
3748 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
3749 ; GFX7-NEXT: s_mov_b32 s5, s8
3750 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
3751 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
3752 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3753 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
3754 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
3755 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
3756 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3757 ; GFX7-NEXT: s_endpgm
3758 ;
3759 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3760 ; GFX10-WGP: ; %bb.0: ; %entry
3761 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
3762 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
3763 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3764 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
3765 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
3766 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
3767 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
3768 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
3769 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3770 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
3771 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3772 ; GFX10-WGP-NEXT: s_endpgm
3773 ;
3774 ; GFX10-CU-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3775 ; GFX10-CU: ; %bb.0: ; %entry
3776 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
3777 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
3778 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3779 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
3780 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
3781 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
3782 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
3783 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
3784 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3785 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
3786 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3787 ; GFX10-CU-NEXT: s_endpgm
3788 ;
3789 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3790 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
3791 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
3792 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
3793 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
3794 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
3795 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
3796 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
3797 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
3798 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
3799 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
3800 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
3801 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
3802 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
3803 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
3804 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
3805 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3806 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
3807 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
3808 ; SKIP-CACHE-INV-NEXT: s_endpgm
3809 ;
3810 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3811 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
3812 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3813 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3814 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3815 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3816 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3817 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3818 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3819 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3820 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3821 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3822 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3823 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
3824 ;
3825 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3826 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
3827 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
3828 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3829 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3830 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
3831 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
3832 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3833 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
3834 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
3835 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3836 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3837 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
3838 ; GFX90A-TGSPLIT-NEXT: s_endpgm
3839 ;
3840 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3841 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
3842 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3843 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3844 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3845 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3846 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3847 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3848 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3849 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3850 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3851 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3852 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3853 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
3854 ;
3855 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3856 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
3857 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
3858 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
3859 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
3860 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
3861 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
3862 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
3863 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
3864 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
3865 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3866 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
3867 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
3868 ; GFX940-TGSPLIT-NEXT: s_endpgm
3869 ;
3870 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3871 ; GFX11-WGP: ; %bb.0: ; %entry
3872 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3873 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
3874 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3875 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3876 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3877 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
3878 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
3879 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
3880 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3881 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
3882 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3883 ; GFX11-WGP-NEXT: s_endpgm
3884 ;
3885 ; GFX11-CU-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3886 ; GFX11-CU: ; %bb.0: ; %entry
3887 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3888 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
3889 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3890 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3891 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3892 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
3893 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
3894 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
3895 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3896 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
3897 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3898 ; GFX11-CU-NEXT: s_endpgm
3899 ;
3900 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3901 ; GFX12-WGP: ; %bb.0: ; %entry
3902 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
3903 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
3904 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3905 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
3906 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
3907 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
3908 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
3909 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
3910 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3911 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
3912 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3913 ; GFX12-WGP-NEXT: s_endpgm
3914 ;
3915 ; GFX12-CU-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
3916 ; GFX12-CU: ; %bb.0: ; %entry
3917 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
3918 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
3919 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
3920 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
3921 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
3922 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
3923 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
3924 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
3925 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3926 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
3927 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
3928 ; GFX12-CU-NEXT: s_endpgm
3929 ptr addrspace(1) %out, i32 %in, i32 %old) {
3930 entry:
3931 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
3932 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
3933 ret void
3934 }
3936 define amdgpu_kernel void @global_singlethread_monotonic_acquire_cmpxchg(
3937 ; GFX6-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
3938 ; GFX6: ; %bb.0: ; %entry
3939 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
3940 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
3941 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
3942 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
3943 ; GFX6-NEXT: s_mov_b32 s12, s5
3944 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
3945 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
3946 ; GFX6-NEXT: s_mov_b32 s11, -1
3947 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
3948 ; GFX6-NEXT: s_mov_b32 s5, s12
3949 ; GFX6-NEXT: s_mov_b32 s6, s11
3950 ; GFX6-NEXT: s_mov_b32 s7, s10
3951 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
3952 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
3953 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
3954 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
3955 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
3956 ; GFX6-NEXT: s_endpgm
3957 ;
3958 ; GFX7-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
3959 ; GFX7: ; %bb.0: ; %entry
3960 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
3961 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
3962 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
3963 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
3964 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
3965 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
3966 ; GFX7-NEXT: s_mov_b32 s4, s8
3967 ; GFX7-NEXT: s_mov_b32 s5, s9
3968 ; GFX7-NEXT: s_mov_b32 s9, s10
3969 ; GFX7-NEXT: s_mov_b32 s8, s11
3970 ; GFX7-NEXT: s_add_u32 s4, s4, s9
3971 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
3972 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
3973 ; GFX7-NEXT: s_mov_b32 s5, s8
3974 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
3975 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
3976 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
3977 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
3978 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
3979 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
3980 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
3981 ; GFX7-NEXT: s_endpgm
3982 ;
3983 ; GFX10-WGP-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
3984 ; GFX10-WGP: ; %bb.0: ; %entry
3985 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
3986 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
3987 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
3988 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
3989 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
3990 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
3991 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
3992 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
3993 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
3994 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
3995 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
3996 ; GFX10-WGP-NEXT: s_endpgm
3997 ;
3998 ; GFX10-CU-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
3999 ; GFX10-CU: ; %bb.0: ; %entry
4000 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
4001 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
4002 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4003 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
4004 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
4005 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
4006 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
4007 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
4008 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4009 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
4010 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4011 ; GFX10-CU-NEXT: s_endpgm
4012 ;
4013 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4014 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4015 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
4016 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
4017 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
4018 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4019 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
4020 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
4021 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
4022 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
4023 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
4024 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
4025 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
4026 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
4027 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
4028 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
4029 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4030 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
4031 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
4032 ; SKIP-CACHE-INV-NEXT: s_endpgm
4033 ;
4034 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4035 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
4036 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4037 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4038 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4039 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4040 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4041 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4042 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4043 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4044 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4045 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4046 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4047 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
4048 ;
4049 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4050 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
4051 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4052 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4053 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4054 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4055 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4056 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4057 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4058 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4059 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4060 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4061 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4062 ; GFX90A-TGSPLIT-NEXT: s_endpgm
4063 ;
4064 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4065 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
4066 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4067 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4068 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4069 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4070 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4071 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4072 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4073 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4074 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4075 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4076 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4077 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
4078 ;
4079 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4080 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
4081 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4082 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4083 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4084 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4085 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4086 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4087 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4088 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4089 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4090 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4091 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4092 ; GFX940-TGSPLIT-NEXT: s_endpgm
4093 ;
4094 ; GFX11-WGP-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4095 ; GFX11-WGP: ; %bb.0: ; %entry
4096 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4097 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
4098 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4099 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4100 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4101 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
4102 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
4103 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
4104 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4105 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
4106 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4107 ; GFX11-WGP-NEXT: s_endpgm
4108 ;
4109 ; GFX11-CU-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4110 ; GFX11-CU: ; %bb.0: ; %entry
4111 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4112 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
4113 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4114 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4115 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4116 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
4117 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
4118 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
4119 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4120 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
4121 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4122 ; GFX11-CU-NEXT: s_endpgm
4123 ;
4124 ; GFX12-WGP-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4125 ; GFX12-WGP: ; %bb.0: ; %entry
4126 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4127 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
4128 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4129 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4130 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4131 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
4132 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
4133 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
4134 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4135 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
4136 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4137 ; GFX12-WGP-NEXT: s_endpgm
4138 ;
4139 ; GFX12-CU-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
4140 ; GFX12-CU: ; %bb.0: ; %entry
4141 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4142 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
4143 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4144 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4145 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4146 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
4147 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
4148 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
4149 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4150 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
4151 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4152 ; GFX12-CU-NEXT: s_endpgm
4153 ptr addrspace(1) %out, i32 %in, i32 %old) {
4154 entry:
4155 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
4156 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic acquire
4157 ret void
4158 }
4160 define amdgpu_kernel void @global_singlethread_acquire_acquire_cmpxchg(
4161 ; GFX6-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4162 ; GFX6: ; %bb.0: ; %entry
4163 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
4164 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
4165 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
4166 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
4167 ; GFX6-NEXT: s_mov_b32 s12, s5
4168 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
4169 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
4170 ; GFX6-NEXT: s_mov_b32 s11, -1
4171 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
4172 ; GFX6-NEXT: s_mov_b32 s5, s12
4173 ; GFX6-NEXT: s_mov_b32 s6, s11
4174 ; GFX6-NEXT: s_mov_b32 s7, s10
4175 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
4176 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
4177 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4178 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
4179 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4180 ; GFX6-NEXT: s_endpgm
4181 ;
4182 ; GFX7-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4183 ; GFX7: ; %bb.0: ; %entry
4184 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
4185 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
4186 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
4187 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
4188 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
4189 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
4190 ; GFX7-NEXT: s_mov_b32 s4, s8
4191 ; GFX7-NEXT: s_mov_b32 s5, s9
4192 ; GFX7-NEXT: s_mov_b32 s9, s10
4193 ; GFX7-NEXT: s_mov_b32 s8, s11
4194 ; GFX7-NEXT: s_add_u32 s4, s4, s9
4195 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
4196 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
4197 ; GFX7-NEXT: s_mov_b32 s5, s8
4198 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
4199 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
4200 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4201 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
4202 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
4203 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
4204 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
4205 ; GFX7-NEXT: s_endpgm
4206 ;
4207 ; GFX10-WGP-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4208 ; GFX10-WGP: ; %bb.0: ; %entry
4209 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
4210 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
4211 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4212 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
4213 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
4214 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
4215 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
4216 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
4217 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4218 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
4219 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4220 ; GFX10-WGP-NEXT: s_endpgm
4221 ;
4222 ; GFX10-CU-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4223 ; GFX10-CU: ; %bb.0: ; %entry
4224 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
4225 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
4226 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4227 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
4228 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
4229 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
4230 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
4231 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
4232 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4233 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
4234 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4235 ; GFX10-CU-NEXT: s_endpgm
4236 ;
4237 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4238 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4239 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
4240 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
4241 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
4242 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4243 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
4244 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
4245 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
4246 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
4247 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
4248 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
4249 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
4250 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
4251 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
4252 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
4253 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4254 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
4255 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
4256 ; SKIP-CACHE-INV-NEXT: s_endpgm
4257 ;
4258 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4259 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
4260 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4261 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4262 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4263 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4264 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4265 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4266 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4267 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4268 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4269 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4270 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4271 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
4272 ;
4273 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4274 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
4275 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4276 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4277 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4278 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4279 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4280 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4281 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4282 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4283 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4284 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4285 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4286 ; GFX90A-TGSPLIT-NEXT: s_endpgm
4287 ;
4288 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4289 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
4290 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4291 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4292 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4293 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4294 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4295 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4296 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4297 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4298 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4299 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4300 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4301 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
4302 ;
4303 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4304 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
4305 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4306 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4307 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4308 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4309 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4310 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4311 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4312 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4313 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4314 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4315 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4316 ; GFX940-TGSPLIT-NEXT: s_endpgm
4317 ;
4318 ; GFX11-WGP-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4319 ; GFX11-WGP: ; %bb.0: ; %entry
4320 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4321 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
4322 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4323 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4324 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4325 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
4326 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
4327 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
4328 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4329 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
4330 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4331 ; GFX11-WGP-NEXT: s_endpgm
4332 ;
4333 ; GFX11-CU-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4334 ; GFX11-CU: ; %bb.0: ; %entry
4335 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4336 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
4337 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4338 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4339 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4340 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
4341 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
4342 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
4343 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4344 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
4345 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4346 ; GFX11-CU-NEXT: s_endpgm
4347 ;
4348 ; GFX12-WGP-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4349 ; GFX12-WGP: ; %bb.0: ; %entry
4350 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4351 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
4352 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4353 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4354 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4355 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
4356 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
4357 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
4358 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4359 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
4360 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4361 ; GFX12-WGP-NEXT: s_endpgm
4362 ;
4363 ; GFX12-CU-LABEL: global_singlethread_acquire_acquire_cmpxchg:
4364 ; GFX12-CU: ; %bb.0: ; %entry
4365 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4366 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
4367 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4368 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4369 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4370 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
4371 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
4372 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
4373 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4374 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
4375 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4376 ; GFX12-CU-NEXT: s_endpgm
4377 ptr addrspace(1) %out, i32 %in, i32 %old) {
4378 entry:
4379 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
4380 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
4381 ret void
4382 }
4384 define amdgpu_kernel void @global_singlethread_release_acquire_cmpxchg(
4385 ; GFX6-LABEL: global_singlethread_release_acquire_cmpxchg:
4386 ; GFX6: ; %bb.0: ; %entry
4387 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
4388 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
4389 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
4390 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
4391 ; GFX6-NEXT: s_mov_b32 s12, s5
4392 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
4393 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
4394 ; GFX6-NEXT: s_mov_b32 s11, -1
4395 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
4396 ; GFX6-NEXT: s_mov_b32 s5, s12
4397 ; GFX6-NEXT: s_mov_b32 s6, s11
4398 ; GFX6-NEXT: s_mov_b32 s7, s10
4399 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
4400 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
4401 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4402 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
4403 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4404 ; GFX6-NEXT: s_endpgm
4405 ;
4406 ; GFX7-LABEL: global_singlethread_release_acquire_cmpxchg:
4407 ; GFX7: ; %bb.0: ; %entry
4408 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
4409 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
4410 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
4411 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
4412 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
4413 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
4414 ; GFX7-NEXT: s_mov_b32 s4, s8
4415 ; GFX7-NEXT: s_mov_b32 s5, s9
4416 ; GFX7-NEXT: s_mov_b32 s9, s10
4417 ; GFX7-NEXT: s_mov_b32 s8, s11
4418 ; GFX7-NEXT: s_add_u32 s4, s4, s9
4419 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
4420 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
4421 ; GFX7-NEXT: s_mov_b32 s5, s8
4422 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
4423 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
4424 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4425 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
4426 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
4427 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
4428 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
4429 ; GFX7-NEXT: s_endpgm
4430 ;
4431 ; GFX10-WGP-LABEL: global_singlethread_release_acquire_cmpxchg:
4432 ; GFX10-WGP: ; %bb.0: ; %entry
4433 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
4434 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
4435 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4436 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
4437 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
4438 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
4439 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
4440 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
4441 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4442 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
4443 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4444 ; GFX10-WGP-NEXT: s_endpgm
4445 ;
4446 ; GFX10-CU-LABEL: global_singlethread_release_acquire_cmpxchg:
4447 ; GFX10-CU: ; %bb.0: ; %entry
4448 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
4449 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
4450 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4451 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
4452 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
4453 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
4454 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
4455 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
4456 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4457 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
4458 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4459 ; GFX10-CU-NEXT: s_endpgm
4460 ;
4461 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_acquire_cmpxchg:
4462 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4463 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
4464 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
4465 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
4466 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4467 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
4468 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
4469 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
4470 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
4471 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
4472 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
4473 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
4474 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
4475 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
4476 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
4477 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4478 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
4479 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
4480 ; SKIP-CACHE-INV-NEXT: s_endpgm
4481 ;
4482 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_acquire_cmpxchg:
4483 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
4484 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4485 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4486 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4487 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4488 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4489 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4490 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4491 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4492 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4493 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4494 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4495 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
4496 ;
4497 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_acquire_cmpxchg:
4498 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
4499 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4500 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4501 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4502 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4503 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4504 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4505 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4506 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4507 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4508 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4509 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4510 ; GFX90A-TGSPLIT-NEXT: s_endpgm
4511 ;
4512 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_acquire_cmpxchg:
4513 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
4514 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4515 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4516 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4517 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4518 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4519 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4520 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4521 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4522 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4523 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4524 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4525 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
4526 ;
4527 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_acquire_cmpxchg:
4528 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
4529 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4530 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4531 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4532 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4533 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4534 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4535 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4536 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4537 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4538 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4539 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4540 ; GFX940-TGSPLIT-NEXT: s_endpgm
4541 ;
4542 ; GFX11-WGP-LABEL: global_singlethread_release_acquire_cmpxchg:
4543 ; GFX11-WGP: ; %bb.0: ; %entry
4544 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4545 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
4546 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4547 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4548 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4549 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
4550 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
4551 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
4552 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4553 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
4554 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4555 ; GFX11-WGP-NEXT: s_endpgm
4556 ;
4557 ; GFX11-CU-LABEL: global_singlethread_release_acquire_cmpxchg:
4558 ; GFX11-CU: ; %bb.0: ; %entry
4559 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4560 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
4561 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4562 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4563 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4564 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
4565 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
4566 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
4567 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4568 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
4569 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4570 ; GFX11-CU-NEXT: s_endpgm
4571 ;
4572 ; GFX12-WGP-LABEL: global_singlethread_release_acquire_cmpxchg:
4573 ; GFX12-WGP: ; %bb.0: ; %entry
4574 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4575 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
4576 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4577 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4578 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4579 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
4580 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
4581 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
4582 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4583 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
4584 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4585 ; GFX12-WGP-NEXT: s_endpgm
4586 ;
4587 ; GFX12-CU-LABEL: global_singlethread_release_acquire_cmpxchg:
4588 ; GFX12-CU: ; %bb.0: ; %entry
4589 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4590 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
4591 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4592 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4593 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4594 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
4595 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
4596 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
4597 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4598 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
4599 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4600 ; GFX12-CU-NEXT: s_endpgm
4601 ptr addrspace(1) %out, i32 %in, i32 %old) {
4602 entry:
4603 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
4604 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
4605 ret void
4606 }
4608 define amdgpu_kernel void @global_singlethread_acq_rel_acquire_cmpxchg(
4609 ; GFX6-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4610 ; GFX6: ; %bb.0: ; %entry
4611 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
4612 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
4613 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
4614 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
4615 ; GFX6-NEXT: s_mov_b32 s12, s5
4616 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
4617 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
4618 ; GFX6-NEXT: s_mov_b32 s11, -1
4619 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
4620 ; GFX6-NEXT: s_mov_b32 s5, s12
4621 ; GFX6-NEXT: s_mov_b32 s6, s11
4622 ; GFX6-NEXT: s_mov_b32 s7, s10
4623 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
4624 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
4625 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4626 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
4627 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4628 ; GFX6-NEXT: s_endpgm
4629 ;
4630 ; GFX7-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4631 ; GFX7: ; %bb.0: ; %entry
4632 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
4633 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
4634 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
4635 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
4636 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
4637 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
4638 ; GFX7-NEXT: s_mov_b32 s4, s8
4639 ; GFX7-NEXT: s_mov_b32 s5, s9
4640 ; GFX7-NEXT: s_mov_b32 s9, s10
4641 ; GFX7-NEXT: s_mov_b32 s8, s11
4642 ; GFX7-NEXT: s_add_u32 s4, s4, s9
4643 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
4644 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
4645 ; GFX7-NEXT: s_mov_b32 s5, s8
4646 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
4647 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
4648 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4649 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
4650 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
4651 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
4652 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
4653 ; GFX7-NEXT: s_endpgm
4654 ;
4655 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4656 ; GFX10-WGP: ; %bb.0: ; %entry
4657 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
4658 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
4659 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4660 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
4661 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
4662 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
4663 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
4664 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
4665 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4666 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
4667 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4668 ; GFX10-WGP-NEXT: s_endpgm
4669 ;
4670 ; GFX10-CU-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4671 ; GFX10-CU: ; %bb.0: ; %entry
4672 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
4673 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
4674 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4675 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
4676 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
4677 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
4678 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
4679 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
4680 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4681 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
4682 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4683 ; GFX10-CU-NEXT: s_endpgm
4684 ;
4685 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4686 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4687 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
4688 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
4689 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
4690 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4691 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
4692 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
4693 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
4694 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
4695 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
4696 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
4697 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
4698 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
4699 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
4700 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
4701 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4702 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
4703 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
4704 ; SKIP-CACHE-INV-NEXT: s_endpgm
4705 ;
4706 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4707 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
4708 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4709 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4710 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4711 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4712 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4713 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4714 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4715 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4716 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4717 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4718 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4719 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
4720 ;
4721 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4722 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
4723 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4724 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4725 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4726 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4727 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4728 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4729 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4730 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4731 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4732 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4733 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4734 ; GFX90A-TGSPLIT-NEXT: s_endpgm
4735 ;
4736 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4737 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
4738 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4739 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4740 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4741 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4742 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4743 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4744 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4745 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4746 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4747 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4748 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4749 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
4750 ;
4751 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4752 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
4753 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4754 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4755 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4756 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4757 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4758 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4759 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4760 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4761 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4762 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4763 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4764 ; GFX940-TGSPLIT-NEXT: s_endpgm
4765 ;
4766 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4767 ; GFX11-WGP: ; %bb.0: ; %entry
4768 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4769 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
4770 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4771 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4772 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4773 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
4774 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
4775 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
4776 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4777 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
4778 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4779 ; GFX11-WGP-NEXT: s_endpgm
4780 ;
4781 ; GFX11-CU-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4782 ; GFX11-CU: ; %bb.0: ; %entry
4783 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4784 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
4785 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4786 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4787 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4788 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
4789 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
4790 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
4791 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4792 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
4793 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4794 ; GFX11-CU-NEXT: s_endpgm
4795 ;
4796 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4797 ; GFX12-WGP: ; %bb.0: ; %entry
4798 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4799 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
4800 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4801 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4802 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4803 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
4804 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
4805 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
4806 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4807 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
4808 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4809 ; GFX12-WGP-NEXT: s_endpgm
4810 ;
4811 ; GFX12-CU-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
4812 ; GFX12-CU: ; %bb.0: ; %entry
4813 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
4814 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
4815 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4816 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
4817 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
4818 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
4819 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
4820 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
4821 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4822 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
4823 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
4824 ; GFX12-CU-NEXT: s_endpgm
4825 ptr addrspace(1) %out, i32 %in, i32 %old) {
4826 entry:
4827 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
4828 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
4829 ret void
4830 }
4832 define amdgpu_kernel void @global_singlethread_seq_cst_acquire_cmpxchg(
4833 ; GFX6-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4834 ; GFX6: ; %bb.0: ; %entry
4835 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
4836 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
4837 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
4838 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
4839 ; GFX6-NEXT: s_mov_b32 s12, s5
4840 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
4841 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
4842 ; GFX6-NEXT: s_mov_b32 s11, -1
4843 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
4844 ; GFX6-NEXT: s_mov_b32 s5, s12
4845 ; GFX6-NEXT: s_mov_b32 s6, s11
4846 ; GFX6-NEXT: s_mov_b32 s7, s10
4847 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
4848 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
4849 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4850 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
4851 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
4852 ; GFX6-NEXT: s_endpgm
4853 ;
4854 ; GFX7-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4855 ; GFX7: ; %bb.0: ; %entry
4856 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
4857 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
4858 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
4859 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
4860 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
4861 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
4862 ; GFX7-NEXT: s_mov_b32 s4, s8
4863 ; GFX7-NEXT: s_mov_b32 s5, s9
4864 ; GFX7-NEXT: s_mov_b32 s9, s10
4865 ; GFX7-NEXT: s_mov_b32 s8, s11
4866 ; GFX7-NEXT: s_add_u32 s4, s4, s9
4867 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
4868 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
4869 ; GFX7-NEXT: s_mov_b32 s5, s8
4870 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
4871 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
4872 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4873 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
4874 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
4875 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
4876 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
4877 ; GFX7-NEXT: s_endpgm
4878 ;
4879 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4880 ; GFX10-WGP: ; %bb.0: ; %entry
4881 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
4882 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
4883 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4884 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
4885 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
4886 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
4887 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
4888 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
4889 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4890 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
4891 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4892 ; GFX10-WGP-NEXT: s_endpgm
4893 ;
4894 ; GFX10-CU-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4895 ; GFX10-CU: ; %bb.0: ; %entry
4896 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
4897 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
4898 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4899 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
4900 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
4901 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
4902 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
4903 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
4904 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
4905 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
4906 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
4907 ; GFX10-CU-NEXT: s_endpgm
4908 ;
4909 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4910 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
4911 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
4912 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
4913 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
4914 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
4915 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
4916 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
4917 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
4918 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
4919 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
4920 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
4921 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
4922 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
4923 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
4924 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
4925 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
4926 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
4927 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
4928 ; SKIP-CACHE-INV-NEXT: s_endpgm
4929 ;
4930 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4931 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
4932 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4933 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4934 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4935 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4936 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4937 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4938 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4939 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4940 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4941 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4942 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4943 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
4944 ;
4945 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4946 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
4947 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
4948 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4949 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
4950 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
4951 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
4952 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4953 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
4954 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
4955 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4956 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4957 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
4958 ; GFX90A-TGSPLIT-NEXT: s_endpgm
4959 ;
4960 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4961 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
4962 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4963 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4964 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4965 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4966 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4967 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4968 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4969 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4970 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4971 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4972 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4973 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
4974 ;
4975 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4976 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
4977 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
4978 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
4979 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
4980 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
4981 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
4982 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
4983 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
4984 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
4985 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
4986 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
4987 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
4988 ; GFX940-TGSPLIT-NEXT: s_endpgm
4989 ;
4990 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
4991 ; GFX11-WGP: ; %bb.0: ; %entry
4992 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
4993 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
4994 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
4995 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
4996 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
4997 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
4998 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
4999 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
5000 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5001 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
5002 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5003 ; GFX11-WGP-NEXT: s_endpgm
5004 ;
5005 ; GFX11-CU-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
5006 ; GFX11-CU: ; %bb.0: ; %entry
5007 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5008 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
5009 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5010 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5011 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5012 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
5013 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
5014 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
5015 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5016 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
5017 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5018 ; GFX11-CU-NEXT: s_endpgm
5019 ;
5020 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
5021 ; GFX12-WGP: ; %bb.0: ; %entry
5022 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5023 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
5024 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5025 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5026 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5027 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
5028 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
5029 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
5030 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5031 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
5032 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5033 ; GFX12-WGP-NEXT: s_endpgm
5034 ;
5035 ; GFX12-CU-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
5036 ; GFX12-CU: ; %bb.0: ; %entry
5037 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5038 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
5039 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5040 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5041 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5042 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
5043 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
5044 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
5045 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5046 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
5047 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5048 ; GFX12-CU-NEXT: s_endpgm
5049 ptr addrspace(1) %out, i32 %in, i32 %old) {
5050 entry:
5051 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
5052 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
5053 ret void
5054 }
5056 define amdgpu_kernel void @global_singlethread_monotonic_seq_cst_cmpxchg(
5057 ; GFX6-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5058 ; GFX6: ; %bb.0: ; %entry
5059 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
5060 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
5061 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
5062 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
5063 ; GFX6-NEXT: s_mov_b32 s12, s5
5064 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
5065 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
5066 ; GFX6-NEXT: s_mov_b32 s11, -1
5067 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
5068 ; GFX6-NEXT: s_mov_b32 s5, s12
5069 ; GFX6-NEXT: s_mov_b32 s6, s11
5070 ; GFX6-NEXT: s_mov_b32 s7, s10
5071 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
5072 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
5073 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5074 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
5075 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
5076 ; GFX6-NEXT: s_endpgm
5077 ;
5078 ; GFX7-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5079 ; GFX7: ; %bb.0: ; %entry
5080 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
5081 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
5082 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
5083 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
5084 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
5085 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
5086 ; GFX7-NEXT: s_mov_b32 s4, s8
5087 ; GFX7-NEXT: s_mov_b32 s5, s9
5088 ; GFX7-NEXT: s_mov_b32 s9, s10
5089 ; GFX7-NEXT: s_mov_b32 s8, s11
5090 ; GFX7-NEXT: s_add_u32 s4, s4, s9
5091 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
5092 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
5093 ; GFX7-NEXT: s_mov_b32 s5, s8
5094 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
5095 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
5096 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5097 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
5098 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
5099 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
5100 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
5101 ; GFX7-NEXT: s_endpgm
5102 ;
5103 ; GFX10-WGP-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5104 ; GFX10-WGP: ; %bb.0: ; %entry
5105 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
5106 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
5107 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5108 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
5109 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
5110 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
5111 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
5112 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
5113 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5114 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
5115 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5116 ; GFX10-WGP-NEXT: s_endpgm
5117 ;
5118 ; GFX10-CU-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5119 ; GFX10-CU: ; %bb.0: ; %entry
5120 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
5121 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
5122 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5123 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
5124 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
5125 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
5126 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
5127 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
5128 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5129 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
5130 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5131 ; GFX10-CU-NEXT: s_endpgm
5132 ;
5133 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5134 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
5135 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
5136 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
5137 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
5138 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
5139 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
5140 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
5141 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
5142 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
5143 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
5144 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
5145 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
5146 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
5147 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
5148 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
5149 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5150 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
5151 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
5152 ; SKIP-CACHE-INV-NEXT: s_endpgm
5153 ;
5154 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5155 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
5156 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5157 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5158 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5159 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5160 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5161 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5162 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5163 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5164 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5165 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5166 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5167 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
5168 ;
5169 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5170 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
5171 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5172 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5173 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5174 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5175 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5176 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5177 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5178 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5179 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5180 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5181 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5182 ; GFX90A-TGSPLIT-NEXT: s_endpgm
5183 ;
5184 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5185 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
5186 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5187 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5188 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5189 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5190 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5191 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5192 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5193 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5194 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5195 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5196 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5197 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
5198 ;
5199 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5200 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
5201 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5202 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5203 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5204 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5205 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5206 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5207 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5208 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5209 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5210 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5211 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5212 ; GFX940-TGSPLIT-NEXT: s_endpgm
5213 ;
5214 ; GFX11-WGP-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5215 ; GFX11-WGP: ; %bb.0: ; %entry
5216 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5217 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
5218 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5219 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5220 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5221 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
5222 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
5223 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
5224 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5225 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
5226 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5227 ; GFX11-WGP-NEXT: s_endpgm
5228 ;
5229 ; GFX11-CU-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5230 ; GFX11-CU: ; %bb.0: ; %entry
5231 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5232 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
5233 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5234 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5235 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5236 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
5237 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
5238 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
5239 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5240 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
5241 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5242 ; GFX11-CU-NEXT: s_endpgm
5243 ;
5244 ; GFX12-WGP-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5245 ; GFX12-WGP: ; %bb.0: ; %entry
5246 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5247 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
5248 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5249 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5250 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5251 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
5252 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
5253 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
5254 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5255 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
5256 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5257 ; GFX12-WGP-NEXT: s_endpgm
5258 ;
5259 ; GFX12-CU-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
5260 ; GFX12-CU: ; %bb.0: ; %entry
5261 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5262 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
5263 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5264 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5265 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5266 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
5267 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
5268 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
5269 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5270 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
5271 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5272 ; GFX12-CU-NEXT: s_endpgm
5273 ptr addrspace(1) %out, i32 %in, i32 %old) {
5274 entry:
5275 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
5276 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic seq_cst
5277 ret void
5278 }
5280 define amdgpu_kernel void @global_singlethread_acquire_seq_cst_cmpxchg(
5281 ; GFX6-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5282 ; GFX6: ; %bb.0: ; %entry
5283 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
5284 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
5285 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
5286 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
5287 ; GFX6-NEXT: s_mov_b32 s12, s5
5288 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
5289 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
5290 ; GFX6-NEXT: s_mov_b32 s11, -1
5291 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
5292 ; GFX6-NEXT: s_mov_b32 s5, s12
5293 ; GFX6-NEXT: s_mov_b32 s6, s11
5294 ; GFX6-NEXT: s_mov_b32 s7, s10
5295 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
5296 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
5297 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5298 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
5299 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
5300 ; GFX6-NEXT: s_endpgm
5301 ;
5302 ; GFX7-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5303 ; GFX7: ; %bb.0: ; %entry
5304 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
5305 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
5306 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
5307 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
5308 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
5309 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
5310 ; GFX7-NEXT: s_mov_b32 s4, s8
5311 ; GFX7-NEXT: s_mov_b32 s5, s9
5312 ; GFX7-NEXT: s_mov_b32 s9, s10
5313 ; GFX7-NEXT: s_mov_b32 s8, s11
5314 ; GFX7-NEXT: s_add_u32 s4, s4, s9
5315 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
5316 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
5317 ; GFX7-NEXT: s_mov_b32 s5, s8
5318 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
5319 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
5320 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5321 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
5322 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
5323 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
5324 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
5325 ; GFX7-NEXT: s_endpgm
5326 ;
5327 ; GFX10-WGP-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5328 ; GFX10-WGP: ; %bb.0: ; %entry
5329 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
5330 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
5331 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5332 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
5333 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
5334 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
5335 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
5336 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
5337 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5338 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
5339 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5340 ; GFX10-WGP-NEXT: s_endpgm
5341 ;
5342 ; GFX10-CU-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5343 ; GFX10-CU: ; %bb.0: ; %entry
5344 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
5345 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
5346 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5347 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
5348 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
5349 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
5350 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
5351 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
5352 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5353 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
5354 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5355 ; GFX10-CU-NEXT: s_endpgm
5356 ;
5357 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5358 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
5359 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
5360 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
5361 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
5362 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
5363 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
5364 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
5365 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
5366 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
5367 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
5368 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
5369 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
5370 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
5371 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
5372 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
5373 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5374 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
5375 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
5376 ; SKIP-CACHE-INV-NEXT: s_endpgm
5377 ;
5378 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5379 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
5380 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5381 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5382 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5383 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5384 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5385 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5386 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5387 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5388 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5389 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5390 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5391 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
5392 ;
5393 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5394 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
5395 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5396 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5397 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5398 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5399 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5400 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5401 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5402 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5403 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5404 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5405 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5406 ; GFX90A-TGSPLIT-NEXT: s_endpgm
5407 ;
5408 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5409 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
5410 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5411 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5412 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5413 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5414 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5415 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5416 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5417 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5418 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5419 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5420 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5421 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
5422 ;
5423 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5424 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
5425 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5426 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5427 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5428 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5429 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5430 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5431 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5432 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5433 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5434 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5435 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5436 ; GFX940-TGSPLIT-NEXT: s_endpgm
5437 ;
5438 ; GFX11-WGP-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5439 ; GFX11-WGP: ; %bb.0: ; %entry
5440 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5441 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
5442 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5443 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5444 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5445 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
5446 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
5447 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
5448 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5449 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
5450 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5451 ; GFX11-WGP-NEXT: s_endpgm
5452 ;
5453 ; GFX11-CU-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5454 ; GFX11-CU: ; %bb.0: ; %entry
5455 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5456 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
5457 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5458 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5459 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5460 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
5461 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
5462 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
5463 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5464 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
5465 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5466 ; GFX11-CU-NEXT: s_endpgm
5467 ;
5468 ; GFX12-WGP-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5469 ; GFX12-WGP: ; %bb.0: ; %entry
5470 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5471 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
5472 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5473 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5474 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5475 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
5476 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
5477 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
5478 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5479 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
5480 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5481 ; GFX12-WGP-NEXT: s_endpgm
5482 ;
5483 ; GFX12-CU-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
5484 ; GFX12-CU: ; %bb.0: ; %entry
5485 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5486 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
5487 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5488 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5489 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5490 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
5491 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
5492 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
5493 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5494 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
5495 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5496 ; GFX12-CU-NEXT: s_endpgm
5497 ptr addrspace(1) %out, i32 %in, i32 %old) {
5498 entry:
5499 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
5500 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire seq_cst
5501 ret void
5502 }
5504 define amdgpu_kernel void @global_singlethread_release_seq_cst_cmpxchg(
5505 ; GFX6-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5506 ; GFX6: ; %bb.0: ; %entry
5507 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
5508 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
5509 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
5510 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
5511 ; GFX6-NEXT: s_mov_b32 s12, s5
5512 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
5513 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
5514 ; GFX6-NEXT: s_mov_b32 s11, -1
5515 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
5516 ; GFX6-NEXT: s_mov_b32 s5, s12
5517 ; GFX6-NEXT: s_mov_b32 s6, s11
5518 ; GFX6-NEXT: s_mov_b32 s7, s10
5519 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
5520 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
5521 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5522 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
5523 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
5524 ; GFX6-NEXT: s_endpgm
5525 ;
5526 ; GFX7-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5527 ; GFX7: ; %bb.0: ; %entry
5528 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
5529 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
5530 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
5531 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
5532 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
5533 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
5534 ; GFX7-NEXT: s_mov_b32 s4, s8
5535 ; GFX7-NEXT: s_mov_b32 s5, s9
5536 ; GFX7-NEXT: s_mov_b32 s9, s10
5537 ; GFX7-NEXT: s_mov_b32 s8, s11
5538 ; GFX7-NEXT: s_add_u32 s4, s4, s9
5539 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
5540 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
5541 ; GFX7-NEXT: s_mov_b32 s5, s8
5542 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
5543 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
5544 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5545 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
5546 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
5547 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
5548 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
5549 ; GFX7-NEXT: s_endpgm
5550 ;
5551 ; GFX10-WGP-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5552 ; GFX10-WGP: ; %bb.0: ; %entry
5553 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
5554 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
5555 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5556 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
5557 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
5558 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
5559 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
5560 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
5561 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5562 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
5563 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5564 ; GFX10-WGP-NEXT: s_endpgm
5565 ;
5566 ; GFX10-CU-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5567 ; GFX10-CU: ; %bb.0: ; %entry
5568 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
5569 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
5570 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5571 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
5572 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
5573 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
5574 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
5575 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
5576 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5577 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
5578 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5579 ; GFX10-CU-NEXT: s_endpgm
5580 ;
5581 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5582 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
5583 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
5584 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
5585 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
5586 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
5587 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
5588 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
5589 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
5590 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
5591 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
5592 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
5593 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
5594 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
5595 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
5596 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
5597 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5598 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
5599 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
5600 ; SKIP-CACHE-INV-NEXT: s_endpgm
5601 ;
5602 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5603 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
5604 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5605 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5606 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5607 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5608 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5609 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5610 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5611 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5612 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5613 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5614 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5615 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
5616 ;
5617 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5618 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
5619 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5620 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5621 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5622 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5623 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5624 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5625 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5626 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5627 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5628 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5629 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5630 ; GFX90A-TGSPLIT-NEXT: s_endpgm
5631 ;
5632 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5633 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
5634 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5635 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5636 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5637 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5638 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5639 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5640 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5641 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5642 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5643 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5644 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5645 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
5646 ;
5647 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5648 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
5649 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5650 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5651 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5652 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5653 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5654 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5655 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5656 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5657 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5658 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5659 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5660 ; GFX940-TGSPLIT-NEXT: s_endpgm
5661 ;
5662 ; GFX11-WGP-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5663 ; GFX11-WGP: ; %bb.0: ; %entry
5664 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5665 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
5666 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5667 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5668 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5669 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
5670 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
5671 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
5672 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5673 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
5674 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5675 ; GFX11-WGP-NEXT: s_endpgm
5676 ;
5677 ; GFX11-CU-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5678 ; GFX11-CU: ; %bb.0: ; %entry
5679 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5680 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
5681 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5682 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5683 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5684 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
5685 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
5686 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
5687 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5688 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
5689 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5690 ; GFX11-CU-NEXT: s_endpgm
5691 ;
5692 ; GFX12-WGP-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5693 ; GFX12-WGP: ; %bb.0: ; %entry
5694 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5695 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
5696 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5697 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5698 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5699 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
5700 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
5701 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
5702 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5703 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
5704 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5705 ; GFX12-WGP-NEXT: s_endpgm
5706 ;
5707 ; GFX12-CU-LABEL: global_singlethread_release_seq_cst_cmpxchg:
5708 ; GFX12-CU: ; %bb.0: ; %entry
5709 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5710 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
5711 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5712 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5713 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5714 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
5715 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
5716 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
5717 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5718 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
5719 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5720 ; GFX12-CU-NEXT: s_endpgm
5721 ptr addrspace(1) %out, i32 %in, i32 %old) {
5722 entry:
5723 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
5724 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release seq_cst
5725 ret void
5726 }
5728 define amdgpu_kernel void @global_singlethread_acq_rel_seq_cst_cmpxchg(
5729 ; GFX6-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5730 ; GFX6: ; %bb.0: ; %entry
5731 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
5732 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
5733 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
5734 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
5735 ; GFX6-NEXT: s_mov_b32 s12, s5
5736 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
5737 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
5738 ; GFX6-NEXT: s_mov_b32 s11, -1
5739 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
5740 ; GFX6-NEXT: s_mov_b32 s5, s12
5741 ; GFX6-NEXT: s_mov_b32 s6, s11
5742 ; GFX6-NEXT: s_mov_b32 s7, s10
5743 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
5744 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
5745 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5746 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
5747 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
5748 ; GFX6-NEXT: s_endpgm
5749 ;
5750 ; GFX7-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5751 ; GFX7: ; %bb.0: ; %entry
5752 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
5753 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
5754 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
5755 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
5756 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
5757 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
5758 ; GFX7-NEXT: s_mov_b32 s4, s8
5759 ; GFX7-NEXT: s_mov_b32 s5, s9
5760 ; GFX7-NEXT: s_mov_b32 s9, s10
5761 ; GFX7-NEXT: s_mov_b32 s8, s11
5762 ; GFX7-NEXT: s_add_u32 s4, s4, s9
5763 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
5764 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
5765 ; GFX7-NEXT: s_mov_b32 s5, s8
5766 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
5767 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
5768 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5769 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
5770 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
5771 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
5772 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
5773 ; GFX7-NEXT: s_endpgm
5774 ;
5775 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5776 ; GFX10-WGP: ; %bb.0: ; %entry
5777 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
5778 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
5779 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5780 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
5781 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
5782 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
5783 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
5784 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
5785 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5786 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
5787 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5788 ; GFX10-WGP-NEXT: s_endpgm
5789 ;
5790 ; GFX10-CU-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5791 ; GFX10-CU: ; %bb.0: ; %entry
5792 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
5793 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
5794 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5795 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
5796 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
5797 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
5798 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
5799 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
5800 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5801 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
5802 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
5803 ; GFX10-CU-NEXT: s_endpgm
5804 ;
5805 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5806 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
5807 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
5808 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
5809 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
5810 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
5811 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
5812 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
5813 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
5814 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
5815 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
5816 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
5817 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
5818 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
5819 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
5820 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
5821 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5822 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
5823 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
5824 ; SKIP-CACHE-INV-NEXT: s_endpgm
5825 ;
5826 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5827 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
5828 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5829 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5830 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5831 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5832 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5833 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5834 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5835 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5836 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5837 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5838 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5839 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
5840 ;
5841 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5842 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
5843 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
5844 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5845 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
5846 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
5847 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
5848 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5849 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
5850 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
5851 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5852 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5853 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
5854 ; GFX90A-TGSPLIT-NEXT: s_endpgm
5855 ;
5856 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5857 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
5858 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5859 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5860 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5861 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5862 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5863 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5864 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5865 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5866 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5867 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5868 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5869 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
5870 ;
5871 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5872 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
5873 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
5874 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
5875 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
5876 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
5877 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
5878 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
5879 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
5880 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
5881 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5882 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
5883 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
5884 ; GFX940-TGSPLIT-NEXT: s_endpgm
5885 ;
5886 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5887 ; GFX11-WGP: ; %bb.0: ; %entry
5888 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5889 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
5890 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5891 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5892 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5893 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
5894 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
5895 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
5896 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5897 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
5898 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5899 ; GFX11-WGP-NEXT: s_endpgm
5900 ;
5901 ; GFX11-CU-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5902 ; GFX11-CU: ; %bb.0: ; %entry
5903 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5904 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
5905 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5906 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5907 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5908 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
5909 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
5910 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
5911 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5912 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
5913 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5914 ; GFX11-CU-NEXT: s_endpgm
5915 ;
5916 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5917 ; GFX12-WGP: ; %bb.0: ; %entry
5918 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
5919 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
5920 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5921 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
5922 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
5923 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
5924 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
5925 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
5926 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5927 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
5928 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5929 ; GFX12-WGP-NEXT: s_endpgm
5930 ;
5931 ; GFX12-CU-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
5932 ; GFX12-CU: ; %bb.0: ; %entry
5933 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
5934 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
5935 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
5936 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
5937 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
5938 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
5939 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
5940 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
5941 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
5942 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
5943 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
5944 ; GFX12-CU-NEXT: s_endpgm
5945 ptr addrspace(1) %out, i32 %in, i32 %old) {
5946 entry:
5947 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
5948 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel seq_cst
5949 ret void
5950 }
5952 define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_cmpxchg(
5953 ; GFX6-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
5954 ; GFX6: ; %bb.0: ; %entry
5955 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
5956 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
5957 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
5958 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
5959 ; GFX6-NEXT: s_mov_b32 s12, s5
5960 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
5961 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
5962 ; GFX6-NEXT: s_mov_b32 s11, -1
5963 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
5964 ; GFX6-NEXT: s_mov_b32 s5, s12
5965 ; GFX6-NEXT: s_mov_b32 s6, s11
5966 ; GFX6-NEXT: s_mov_b32 s7, s10
5967 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
5968 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
5969 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
5970 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
5971 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
5972 ; GFX6-NEXT: s_endpgm
5973 ;
5974 ; GFX7-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
5975 ; GFX7: ; %bb.0: ; %entry
5976 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
5977 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
5978 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
5979 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
5980 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
5981 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
5982 ; GFX7-NEXT: s_mov_b32 s4, s8
5983 ; GFX7-NEXT: s_mov_b32 s5, s9
5984 ; GFX7-NEXT: s_mov_b32 s9, s10
5985 ; GFX7-NEXT: s_mov_b32 s8, s11
5986 ; GFX7-NEXT: s_add_u32 s4, s4, s9
5987 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
5988 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
5989 ; GFX7-NEXT: s_mov_b32 s5, s8
5990 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
5991 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
5992 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
5993 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
5994 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
5995 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
5996 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
5997 ; GFX7-NEXT: s_endpgm
5998 ;
5999 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6000 ; GFX10-WGP: ; %bb.0: ; %entry
6001 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
6002 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
6003 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6004 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
6005 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
6006 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
6007 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
6008 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
6009 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6010 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
6011 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
6012 ; GFX10-WGP-NEXT: s_endpgm
6013 ;
6014 ; GFX10-CU-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6015 ; GFX10-CU: ; %bb.0: ; %entry
6016 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
6017 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
6018 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6019 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
6020 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
6021 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
6022 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
6023 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
6024 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6025 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
6026 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
6027 ; GFX10-CU-NEXT: s_endpgm
6028 ;
6029 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6030 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
6031 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
6032 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
6033 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
6034 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
6035 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
6036 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
6037 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
6038 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
6039 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
6040 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
6041 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
6042 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
6043 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
6044 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
6045 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6046 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
6047 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
6048 ; SKIP-CACHE-INV-NEXT: s_endpgm
6049 ;
6050 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6051 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
6052 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6053 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6054 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6055 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6056 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6057 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6058 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6059 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6060 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6061 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6062 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
6063 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
6064 ;
6065 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6066 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
6067 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6068 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6069 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6070 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6071 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6072 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6073 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6074 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6075 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6076 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6077 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
6078 ; GFX90A-TGSPLIT-NEXT: s_endpgm
6079 ;
6080 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6081 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
6082 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6083 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6084 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6085 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6086 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6087 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6088 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6089 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6090 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6091 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6092 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
6093 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
6094 ;
6095 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6096 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
6097 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6098 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6099 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6100 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6101 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6102 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6103 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6104 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6105 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6106 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6107 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
6108 ; GFX940-TGSPLIT-NEXT: s_endpgm
6109 ;
6110 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6111 ; GFX11-WGP: ; %bb.0: ; %entry
6112 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6113 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
6114 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6115 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6116 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6117 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
6118 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
6119 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
6120 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6121 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
6122 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
6123 ; GFX11-WGP-NEXT: s_endpgm
6124 ;
6125 ; GFX11-CU-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6126 ; GFX11-CU: ; %bb.0: ; %entry
6127 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6128 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
6129 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6130 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6131 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6132 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
6133 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
6134 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
6135 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6136 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
6137 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
6138 ; GFX11-CU-NEXT: s_endpgm
6139 ;
6140 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6141 ; GFX12-WGP: ; %bb.0: ; %entry
6142 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6143 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
6144 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6145 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6146 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6147 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
6148 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
6149 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
6150 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6151 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
6152 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
6153 ; GFX12-WGP-NEXT: s_endpgm
6154 ;
6155 ; GFX12-CU-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
6156 ; GFX12-CU: ; %bb.0: ; %entry
6157 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6158 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
6159 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6160 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6161 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6162 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
6163 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
6164 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
6165 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6166 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
6167 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
6168 ; GFX12-CU-NEXT: s_endpgm
6169 ptr addrspace(1) %out, i32 %in, i32 %old) {
6170 entry:
6171 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
6172 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
6173 ret void
6174 }
6176 define amdgpu_kernel void @global_singlethread_monotonic_monotonic_ret_cmpxchg(
6177 ; GFX6-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6178 ; GFX6: ; %bb.0: ; %entry
6179 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6180 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
6181 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
6182 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
6183 ; GFX6-NEXT: s_mov_b32 s12, s5
6184 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
6185 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
6186 ; GFX6-NEXT: s_mov_b32 s11, -1
6187 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
6188 ; GFX6-NEXT: s_mov_b32 s5, s12
6189 ; GFX6-NEXT: s_mov_b32 s6, s11
6190 ; GFX6-NEXT: s_mov_b32 s7, s10
6191 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
6192 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
6193 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6194 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
6195 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
6196 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6197 ; GFX6-NEXT: s_waitcnt vmcnt(0)
6198 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
6199 ; GFX6-NEXT: s_endpgm
6200 ;
6201 ; GFX7-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6202 ; GFX7: ; %bb.0: ; %entry
6203 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6204 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
6205 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
6206 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
6207 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
6208 ; GFX7-NEXT: s_mov_b32 s6, s4
6209 ; GFX7-NEXT: s_mov_b32 s7, s5
6210 ; GFX7-NEXT: s_mov_b32 s11, s12
6211 ; GFX7-NEXT: s_mov_b32 s10, s13
6212 ; GFX7-NEXT: s_add_u32 s6, s6, s11
6213 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
6214 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
6215 ; GFX7-NEXT: s_mov_b32 s7, s10
6216 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
6217 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
6218 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6219 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
6220 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
6221 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
6222 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
6223 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
6224 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
6225 ; GFX7-NEXT: s_waitcnt vmcnt(0)
6226 ; GFX7-NEXT: flat_store_dword v[0:1], v2
6227 ; GFX7-NEXT: s_endpgm
6228 ;
6229 ; GFX10-WGP-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6230 ; GFX10-WGP: ; %bb.0: ; %entry
6231 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
6232 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
6233 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6234 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
6235 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
6236 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
6237 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
6238 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
6239 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6240 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
6241 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6242 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
6243 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
6244 ; GFX10-WGP-NEXT: s_endpgm
6245 ;
6246 ; GFX10-CU-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6247 ; GFX10-CU: ; %bb.0: ; %entry
6248 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
6249 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
6250 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6251 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
6252 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
6253 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
6254 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
6255 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
6256 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6257 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
6258 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6259 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
6260 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
6261 ; GFX10-CU-NEXT: s_endpgm
6262 ;
6263 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6264 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
6265 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
6266 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
6267 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
6268 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
6269 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
6270 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
6271 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
6272 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
6273 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
6274 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
6275 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
6276 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
6277 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
6278 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
6279 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6280 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
6281 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
6282 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6283 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
6284 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
6285 ; SKIP-CACHE-INV-NEXT: s_endpgm
6286 ;
6287 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6288 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
6289 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6290 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6291 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6292 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6293 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6294 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6295 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6296 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6297 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6298 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6299 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6300 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6301 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6302 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
6303 ;
6304 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6305 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
6306 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6307 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6308 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6309 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6310 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6311 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6312 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6313 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6314 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6315 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6316 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6317 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6318 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6319 ; GFX90A-TGSPLIT-NEXT: s_endpgm
6320 ;
6321 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6322 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
6323 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6324 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6325 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6326 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6327 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6328 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6329 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6330 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6331 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6332 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6333 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6334 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6335 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6336 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
6337 ;
6338 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6339 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
6340 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6341 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6342 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6343 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6344 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6345 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6346 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6347 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6348 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6349 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6350 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6351 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6352 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6353 ; GFX940-TGSPLIT-NEXT: s_endpgm
6354 ;
6355 ; GFX11-WGP-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6356 ; GFX11-WGP: ; %bb.0: ; %entry
6357 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6358 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
6359 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6360 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6361 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6362 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
6363 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
6364 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
6365 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6366 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
6367 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6368 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
6369 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6370 ; GFX11-WGP-NEXT: s_endpgm
6371 ;
6372 ; GFX11-CU-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6373 ; GFX11-CU: ; %bb.0: ; %entry
6374 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6375 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
6376 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6377 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6378 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6379 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
6380 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
6381 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
6382 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6383 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
6384 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6385 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
6386 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6387 ; GFX11-CU-NEXT: s_endpgm
6388 ;
6389 ; GFX12-WGP-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6390 ; GFX12-WGP: ; %bb.0: ; %entry
6391 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6392 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
6393 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6394 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6395 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6396 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
6397 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
6398 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
6399 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6400 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
6401 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6402 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
6403 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6404 ; GFX12-WGP-NEXT: s_endpgm
6405 ;
6406 ; GFX12-CU-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
6407 ; GFX12-CU: ; %bb.0: ; %entry
6408 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6409 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
6410 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6411 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6412 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6413 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
6414 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
6415 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
6416 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6417 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
6418 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6419 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
6420 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6421 ; GFX12-CU-NEXT: s_endpgm
6422 ptr addrspace(1) %out, i32 %in, i32 %old) {
6423 entry:
6424 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
6425 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
6426 %val0 = extractvalue { i32, i1 } %val, 0
6427 store i32 %val0, ptr addrspace(1) %out, align 4
6428 ret void
6429 }
6431 define amdgpu_kernel void @global_singlethread_acquire_monotonic_ret_cmpxchg(
6432 ; GFX6-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6433 ; GFX6: ; %bb.0: ; %entry
6434 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6435 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
6436 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
6437 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
6438 ; GFX6-NEXT: s_mov_b32 s12, s5
6439 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
6440 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
6441 ; GFX6-NEXT: s_mov_b32 s11, -1
6442 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
6443 ; GFX6-NEXT: s_mov_b32 s5, s12
6444 ; GFX6-NEXT: s_mov_b32 s6, s11
6445 ; GFX6-NEXT: s_mov_b32 s7, s10
6446 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
6447 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
6448 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6449 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
6450 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
6451 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6452 ; GFX6-NEXT: s_waitcnt vmcnt(0)
6453 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
6454 ; GFX6-NEXT: s_endpgm
6455 ;
6456 ; GFX7-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6457 ; GFX7: ; %bb.0: ; %entry
6458 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6459 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
6460 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
6461 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
6462 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
6463 ; GFX7-NEXT: s_mov_b32 s6, s4
6464 ; GFX7-NEXT: s_mov_b32 s7, s5
6465 ; GFX7-NEXT: s_mov_b32 s11, s12
6466 ; GFX7-NEXT: s_mov_b32 s10, s13
6467 ; GFX7-NEXT: s_add_u32 s6, s6, s11
6468 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
6469 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
6470 ; GFX7-NEXT: s_mov_b32 s7, s10
6471 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
6472 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
6473 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6474 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
6475 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
6476 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
6477 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
6478 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
6479 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
6480 ; GFX7-NEXT: s_waitcnt vmcnt(0)
6481 ; GFX7-NEXT: flat_store_dword v[0:1], v2
6482 ; GFX7-NEXT: s_endpgm
6483 ;
6484 ; GFX10-WGP-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6485 ; GFX10-WGP: ; %bb.0: ; %entry
6486 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
6487 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
6488 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6489 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
6490 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
6491 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
6492 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
6493 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
6494 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6495 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
6496 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6497 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
6498 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
6499 ; GFX10-WGP-NEXT: s_endpgm
6500 ;
6501 ; GFX10-CU-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6502 ; GFX10-CU: ; %bb.0: ; %entry
6503 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
6504 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
6505 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6506 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
6507 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
6508 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
6509 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
6510 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
6511 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6512 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
6513 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6514 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
6515 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
6516 ; GFX10-CU-NEXT: s_endpgm
6517 ;
6518 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6519 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
6520 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
6521 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
6522 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
6523 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
6524 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
6525 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
6526 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
6527 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
6528 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
6529 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
6530 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
6531 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
6532 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
6533 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
6534 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6535 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
6536 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
6537 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6538 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
6539 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
6540 ; SKIP-CACHE-INV-NEXT: s_endpgm
6541 ;
6542 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6543 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
6544 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6545 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6546 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6547 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6548 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6549 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6550 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6551 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6552 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6553 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6554 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6555 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6556 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6557 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
6558 ;
6559 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6560 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
6561 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6562 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6563 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6564 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6565 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6566 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6567 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6568 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6569 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6570 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6571 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6572 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6573 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6574 ; GFX90A-TGSPLIT-NEXT: s_endpgm
6575 ;
6576 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6577 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
6578 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6579 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6580 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6581 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6582 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6583 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6584 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6585 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6586 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6587 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6588 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6589 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6590 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6591 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
6592 ;
6593 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6594 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
6595 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6596 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6597 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6598 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6599 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6600 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6601 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6602 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6603 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6604 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6605 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6606 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6607 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6608 ; GFX940-TGSPLIT-NEXT: s_endpgm
6609 ;
6610 ; GFX11-WGP-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6611 ; GFX11-WGP: ; %bb.0: ; %entry
6612 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6613 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
6614 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6615 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6616 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6617 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
6618 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
6619 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
6620 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6621 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
6622 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6623 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
6624 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6625 ; GFX11-WGP-NEXT: s_endpgm
6626 ;
6627 ; GFX11-CU-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6628 ; GFX11-CU: ; %bb.0: ; %entry
6629 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6630 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
6631 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6632 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6633 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6634 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
6635 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
6636 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
6637 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6638 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
6639 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6640 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
6641 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6642 ; GFX11-CU-NEXT: s_endpgm
6643 ;
6644 ; GFX12-WGP-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6645 ; GFX12-WGP: ; %bb.0: ; %entry
6646 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6647 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
6648 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6649 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6650 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6651 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
6652 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
6653 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
6654 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6655 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
6656 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6657 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
6658 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6659 ; GFX12-WGP-NEXT: s_endpgm
6660 ;
6661 ; GFX12-CU-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
6662 ; GFX12-CU: ; %bb.0: ; %entry
6663 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6664 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
6665 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6666 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6667 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6668 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
6669 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
6670 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
6671 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6672 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
6673 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6674 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
6675 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6676 ; GFX12-CU-NEXT: s_endpgm
6677 ptr addrspace(1) %out, i32 %in, i32 %old) {
6678 entry:
6679 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
6680 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
6681 %val0 = extractvalue { i32, i1 } %val, 0
6682 store i32 %val0, ptr addrspace(1) %out, align 4
6683 ret void
6684 }
6686 define amdgpu_kernel void @global_singlethread_release_monotonic_ret_cmpxchg(
6687 ; GFX6-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6688 ; GFX6: ; %bb.0: ; %entry
6689 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6690 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
6691 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
6692 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
6693 ; GFX6-NEXT: s_mov_b32 s12, s5
6694 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
6695 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
6696 ; GFX6-NEXT: s_mov_b32 s11, -1
6697 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
6698 ; GFX6-NEXT: s_mov_b32 s5, s12
6699 ; GFX6-NEXT: s_mov_b32 s6, s11
6700 ; GFX6-NEXT: s_mov_b32 s7, s10
6701 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
6702 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
6703 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6704 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
6705 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
6706 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6707 ; GFX6-NEXT: s_waitcnt vmcnt(0)
6708 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
6709 ; GFX6-NEXT: s_endpgm
6710 ;
6711 ; GFX7-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6712 ; GFX7: ; %bb.0: ; %entry
6713 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6714 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
6715 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
6716 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
6717 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
6718 ; GFX7-NEXT: s_mov_b32 s6, s4
6719 ; GFX7-NEXT: s_mov_b32 s7, s5
6720 ; GFX7-NEXT: s_mov_b32 s11, s12
6721 ; GFX7-NEXT: s_mov_b32 s10, s13
6722 ; GFX7-NEXT: s_add_u32 s6, s6, s11
6723 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
6724 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
6725 ; GFX7-NEXT: s_mov_b32 s7, s10
6726 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
6727 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
6728 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6729 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
6730 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
6731 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
6732 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
6733 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
6734 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
6735 ; GFX7-NEXT: s_waitcnt vmcnt(0)
6736 ; GFX7-NEXT: flat_store_dword v[0:1], v2
6737 ; GFX7-NEXT: s_endpgm
6738 ;
6739 ; GFX10-WGP-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6740 ; GFX10-WGP: ; %bb.0: ; %entry
6741 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
6742 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
6743 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6744 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
6745 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
6746 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
6747 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
6748 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
6749 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6750 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
6751 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6752 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
6753 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
6754 ; GFX10-WGP-NEXT: s_endpgm
6755 ;
6756 ; GFX10-CU-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6757 ; GFX10-CU: ; %bb.0: ; %entry
6758 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
6759 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
6760 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6761 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
6762 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
6763 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
6764 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
6765 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
6766 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6767 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
6768 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
6769 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
6770 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
6771 ; GFX10-CU-NEXT: s_endpgm
6772 ;
6773 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6774 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
6775 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
6776 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
6777 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
6778 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
6779 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
6780 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
6781 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
6782 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
6783 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
6784 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
6785 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
6786 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
6787 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
6788 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
6789 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6790 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
6791 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
6792 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6793 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
6794 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
6795 ; SKIP-CACHE-INV-NEXT: s_endpgm
6796 ;
6797 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6798 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
6799 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6800 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6801 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6802 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6803 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6804 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6805 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6806 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6807 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6808 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6809 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6810 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6811 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6812 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
6813 ;
6814 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6815 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
6816 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
6817 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6818 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6819 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
6820 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
6821 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6822 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
6823 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
6824 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6825 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6826 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
6827 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6828 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
6829 ; GFX90A-TGSPLIT-NEXT: s_endpgm
6830 ;
6831 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6832 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
6833 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6834 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6835 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6836 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6837 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6838 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6839 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6840 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6841 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6842 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6843 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6844 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
6845 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6846 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
6847 ;
6848 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6849 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
6850 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
6851 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
6852 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
6853 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
6854 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
6855 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
6856 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
6857 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
6858 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6859 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
6860 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
6861 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
6862 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
6863 ; GFX940-TGSPLIT-NEXT: s_endpgm
6864 ;
6865 ; GFX11-WGP-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6866 ; GFX11-WGP: ; %bb.0: ; %entry
6867 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6868 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
6869 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6870 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6871 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6872 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
6873 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
6874 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
6875 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6876 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
6877 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6878 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
6879 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6880 ; GFX11-WGP-NEXT: s_endpgm
6881 ;
6882 ; GFX11-CU-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6883 ; GFX11-CU: ; %bb.0: ; %entry
6884 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6885 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
6886 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6887 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6888 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6889 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
6890 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
6891 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
6892 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6893 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
6894 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
6895 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
6896 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6897 ; GFX11-CU-NEXT: s_endpgm
6898 ;
6899 ; GFX12-WGP-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6900 ; GFX12-WGP: ; %bb.0: ; %entry
6901 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
6902 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
6903 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6904 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
6905 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
6906 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
6907 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
6908 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
6909 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6910 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
6911 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6912 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
6913 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
6914 ; GFX12-WGP-NEXT: s_endpgm
6915 ;
6916 ; GFX12-CU-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
6917 ; GFX12-CU: ; %bb.0: ; %entry
6918 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
6919 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
6920 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
6921 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
6922 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
6923 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
6924 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
6925 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
6926 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
6927 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
6928 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
6929 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
6930 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
6931 ; GFX12-CU-NEXT: s_endpgm
6932 ptr addrspace(1) %out, i32 %in, i32 %old) {
6933 entry:
6934 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
6935 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
6936 %val0 = extractvalue { i32, i1 } %val, 0
6937 store i32 %val0, ptr addrspace(1) %out, align 4
6938 ret void
6939 }
6941 define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_ret_cmpxchg(
6942 ; GFX6-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
6943 ; GFX6: ; %bb.0: ; %entry
6944 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6945 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
6946 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
6947 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
6948 ; GFX6-NEXT: s_mov_b32 s12, s5
6949 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
6950 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
6951 ; GFX6-NEXT: s_mov_b32 s11, -1
6952 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
6953 ; GFX6-NEXT: s_mov_b32 s5, s12
6954 ; GFX6-NEXT: s_mov_b32 s6, s11
6955 ; GFX6-NEXT: s_mov_b32 s7, s10
6956 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
6957 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
6958 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
6959 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
6960 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
6961 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
6962 ; GFX6-NEXT: s_waitcnt vmcnt(0)
6963 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
6964 ; GFX6-NEXT: s_endpgm
6965 ;
6966 ; GFX7-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
6967 ; GFX7: ; %bb.0: ; %entry
6968 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
6969 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
6970 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
6971 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
6972 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
6973 ; GFX7-NEXT: s_mov_b32 s6, s4
6974 ; GFX7-NEXT: s_mov_b32 s7, s5
6975 ; GFX7-NEXT: s_mov_b32 s11, s12
6976 ; GFX7-NEXT: s_mov_b32 s10, s13
6977 ; GFX7-NEXT: s_add_u32 s6, s6, s11
6978 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
6979 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
6980 ; GFX7-NEXT: s_mov_b32 s7, s10
6981 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
6982 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
6983 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
6984 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
6985 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
6986 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
6987 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
6988 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
6989 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
6990 ; GFX7-NEXT: s_waitcnt vmcnt(0)
6991 ; GFX7-NEXT: flat_store_dword v[0:1], v2
6992 ; GFX7-NEXT: s_endpgm
6993 ;
6994 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
6995 ; GFX10-WGP: ; %bb.0: ; %entry
6996 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
6997 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
6998 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
6999 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
7000 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
7001 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
7002 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
7003 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
7004 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7005 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
7006 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7007 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
7008 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
7009 ; GFX10-WGP-NEXT: s_endpgm
7010 ;
7011 ; GFX10-CU-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7012 ; GFX10-CU: ; %bb.0: ; %entry
7013 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
7014 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
7015 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7016 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
7017 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
7018 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
7019 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
7020 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
7021 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7022 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
7023 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7024 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
7025 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
7026 ; GFX10-CU-NEXT: s_endpgm
7027 ;
7028 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7029 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
7030 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
7031 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
7032 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
7033 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
7034 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
7035 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
7036 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
7037 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
7038 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
7039 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
7040 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
7041 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
7042 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
7043 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
7044 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7045 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
7046 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
7047 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7048 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
7049 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
7050 ; SKIP-CACHE-INV-NEXT: s_endpgm
7051 ;
7052 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7053 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
7054 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7055 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7056 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7057 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7058 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7059 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7060 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7061 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7062 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7063 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7064 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7065 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7066 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7067 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
7068 ;
7069 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7070 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
7071 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7072 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7073 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7074 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7075 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7076 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7077 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7078 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7079 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7080 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7081 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7082 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7083 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7084 ; GFX90A-TGSPLIT-NEXT: s_endpgm
7085 ;
7086 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7087 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
7088 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7089 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7090 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7091 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7092 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7093 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7094 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7095 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7096 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7097 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7098 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7099 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7100 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7101 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
7102 ;
7103 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7104 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
7105 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7106 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7107 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7108 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7109 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7110 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7111 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7112 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7113 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7114 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7115 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7116 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7117 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7118 ; GFX940-TGSPLIT-NEXT: s_endpgm
7119 ;
7120 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7121 ; GFX11-WGP: ; %bb.0: ; %entry
7122 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7123 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
7124 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7125 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7126 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7127 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
7128 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
7129 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
7130 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7131 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
7132 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7133 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
7134 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7135 ; GFX11-WGP-NEXT: s_endpgm
7136 ;
7137 ; GFX11-CU-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7138 ; GFX11-CU: ; %bb.0: ; %entry
7139 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7140 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
7141 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7142 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7143 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7144 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
7145 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
7146 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
7147 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7148 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
7149 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7150 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
7151 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7152 ; GFX11-CU-NEXT: s_endpgm
7153 ;
7154 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7155 ; GFX12-WGP: ; %bb.0: ; %entry
7156 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7157 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
7158 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7159 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7160 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7161 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
7162 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
7163 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
7164 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7165 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
7166 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7167 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
7168 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7169 ; GFX12-WGP-NEXT: s_endpgm
7170 ;
7171 ; GFX12-CU-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
7172 ; GFX12-CU: ; %bb.0: ; %entry
7173 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7174 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
7175 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7176 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7177 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7178 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
7179 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
7180 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
7181 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7182 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
7183 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7184 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
7185 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7186 ; GFX12-CU-NEXT: s_endpgm
7187 ptr addrspace(1) %out, i32 %in, i32 %old) {
7188 entry:
7189 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
7190 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
7191 %val0 = extractvalue { i32, i1 } %val, 0
7192 store i32 %val0, ptr addrspace(1) %out, align 4
7193 ret void
7194 }
7196 define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_ret_cmpxchg(
7197 ; GFX6-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7198 ; GFX6: ; %bb.0: ; %entry
7199 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7200 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
7201 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
7202 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
7203 ; GFX6-NEXT: s_mov_b32 s12, s5
7204 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
7205 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
7206 ; GFX6-NEXT: s_mov_b32 s11, -1
7207 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
7208 ; GFX6-NEXT: s_mov_b32 s5, s12
7209 ; GFX6-NEXT: s_mov_b32 s6, s11
7210 ; GFX6-NEXT: s_mov_b32 s7, s10
7211 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
7212 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
7213 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7214 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
7215 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
7216 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7217 ; GFX6-NEXT: s_waitcnt vmcnt(0)
7218 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
7219 ; GFX6-NEXT: s_endpgm
7220 ;
7221 ; GFX7-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7222 ; GFX7: ; %bb.0: ; %entry
7223 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7224 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
7225 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
7226 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
7227 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
7228 ; GFX7-NEXT: s_mov_b32 s6, s4
7229 ; GFX7-NEXT: s_mov_b32 s7, s5
7230 ; GFX7-NEXT: s_mov_b32 s11, s12
7231 ; GFX7-NEXT: s_mov_b32 s10, s13
7232 ; GFX7-NEXT: s_add_u32 s6, s6, s11
7233 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
7234 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
7235 ; GFX7-NEXT: s_mov_b32 s7, s10
7236 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
7237 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
7238 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7239 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
7240 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
7241 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
7242 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
7243 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
7244 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
7245 ; GFX7-NEXT: s_waitcnt vmcnt(0)
7246 ; GFX7-NEXT: flat_store_dword v[0:1], v2
7247 ; GFX7-NEXT: s_endpgm
7248 ;
7249 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7250 ; GFX10-WGP: ; %bb.0: ; %entry
7251 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
7252 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
7253 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7254 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
7255 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
7256 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
7257 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
7258 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
7259 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7260 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
7261 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7262 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
7263 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
7264 ; GFX10-WGP-NEXT: s_endpgm
7265 ;
7266 ; GFX10-CU-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7267 ; GFX10-CU: ; %bb.0: ; %entry
7268 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
7269 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
7270 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7271 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
7272 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
7273 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
7274 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
7275 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
7276 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7277 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
7278 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7279 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
7280 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
7281 ; GFX10-CU-NEXT: s_endpgm
7282 ;
7283 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7284 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
7285 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
7286 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
7287 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
7288 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
7289 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
7290 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
7291 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
7292 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
7293 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
7294 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
7295 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
7296 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
7297 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
7298 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
7299 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7300 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
7301 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
7302 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7303 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
7304 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
7305 ; SKIP-CACHE-INV-NEXT: s_endpgm
7306 ;
7307 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7308 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
7309 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7310 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7311 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7312 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7313 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7314 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7315 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7316 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7317 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7318 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7319 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7320 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7321 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7322 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
7323 ;
7324 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7325 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
7326 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7327 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7328 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7329 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7330 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7331 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7332 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7333 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7334 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7335 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7336 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7337 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7338 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7339 ; GFX90A-TGSPLIT-NEXT: s_endpgm
7340 ;
7341 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7342 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
7343 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7344 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7345 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7346 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7347 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7348 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7349 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7350 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7351 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7352 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7353 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7354 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7355 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7356 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
7357 ;
7358 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7359 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
7360 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7361 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7362 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7363 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7364 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7365 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7366 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7367 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7368 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7369 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7370 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7371 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7372 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7373 ; GFX940-TGSPLIT-NEXT: s_endpgm
7374 ;
7375 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7376 ; GFX11-WGP: ; %bb.0: ; %entry
7377 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7378 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
7379 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7380 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7381 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7382 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
7383 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
7384 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
7385 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7386 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
7387 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7388 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
7389 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7390 ; GFX11-WGP-NEXT: s_endpgm
7391 ;
7392 ; GFX11-CU-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7393 ; GFX11-CU: ; %bb.0: ; %entry
7394 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7395 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
7396 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7397 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7398 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7399 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
7400 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
7401 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
7402 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7403 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
7404 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7405 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
7406 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7407 ; GFX11-CU-NEXT: s_endpgm
7408 ;
7409 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7410 ; GFX12-WGP: ; %bb.0: ; %entry
7411 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7412 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
7413 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7414 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7415 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7416 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
7417 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
7418 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
7419 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7420 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
7421 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7422 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
7423 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7424 ; GFX12-WGP-NEXT: s_endpgm
7425 ;
7426 ; GFX12-CU-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
7427 ; GFX12-CU: ; %bb.0: ; %entry
7428 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7429 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
7430 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7431 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7432 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7433 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
7434 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
7435 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
7436 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7437 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
7438 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7439 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
7440 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7441 ; GFX12-CU-NEXT: s_endpgm
7442 ptr addrspace(1) %out, i32 %in, i32 %old) {
7443 entry:
7444 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
7445 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
7446 %val0 = extractvalue { i32, i1 } %val, 0
7447 store i32 %val0, ptr addrspace(1) %out, align 4
7448 ret void
7449 }
7451 define amdgpu_kernel void @global_singlethread_monotonic_acquire_ret_cmpxchg(
7452 ; GFX6-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7453 ; GFX6: ; %bb.0: ; %entry
7454 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7455 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
7456 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
7457 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
7458 ; GFX6-NEXT: s_mov_b32 s12, s5
7459 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
7460 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
7461 ; GFX6-NEXT: s_mov_b32 s11, -1
7462 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
7463 ; GFX6-NEXT: s_mov_b32 s5, s12
7464 ; GFX6-NEXT: s_mov_b32 s6, s11
7465 ; GFX6-NEXT: s_mov_b32 s7, s10
7466 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
7467 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
7468 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7469 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
7470 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
7471 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7472 ; GFX6-NEXT: s_waitcnt vmcnt(0)
7473 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
7474 ; GFX6-NEXT: s_endpgm
7475 ;
7476 ; GFX7-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7477 ; GFX7: ; %bb.0: ; %entry
7478 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7479 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
7480 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
7481 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
7482 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
7483 ; GFX7-NEXT: s_mov_b32 s6, s4
7484 ; GFX7-NEXT: s_mov_b32 s7, s5
7485 ; GFX7-NEXT: s_mov_b32 s11, s12
7486 ; GFX7-NEXT: s_mov_b32 s10, s13
7487 ; GFX7-NEXT: s_add_u32 s6, s6, s11
7488 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
7489 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
7490 ; GFX7-NEXT: s_mov_b32 s7, s10
7491 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
7492 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
7493 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7494 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
7495 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
7496 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
7497 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
7498 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
7499 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
7500 ; GFX7-NEXT: s_waitcnt vmcnt(0)
7501 ; GFX7-NEXT: flat_store_dword v[0:1], v2
7502 ; GFX7-NEXT: s_endpgm
7503 ;
7504 ; GFX10-WGP-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7505 ; GFX10-WGP: ; %bb.0: ; %entry
7506 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
7507 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
7508 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7509 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
7510 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
7511 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
7512 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
7513 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
7514 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7515 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
7516 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7517 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
7518 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
7519 ; GFX10-WGP-NEXT: s_endpgm
7520 ;
7521 ; GFX10-CU-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7522 ; GFX10-CU: ; %bb.0: ; %entry
7523 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
7524 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
7525 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7526 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
7527 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
7528 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
7529 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
7530 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
7531 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7532 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
7533 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7534 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
7535 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
7536 ; GFX10-CU-NEXT: s_endpgm
7537 ;
7538 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7539 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
7540 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
7541 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
7542 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
7543 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
7544 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
7545 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
7546 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
7547 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
7548 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
7549 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
7550 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
7551 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
7552 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
7553 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
7554 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7555 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
7556 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
7557 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7558 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
7559 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
7560 ; SKIP-CACHE-INV-NEXT: s_endpgm
7561 ;
7562 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7563 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
7564 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7565 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7566 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7567 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7568 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7569 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7570 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7571 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7572 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7573 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7574 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7575 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7576 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7577 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
7578 ;
7579 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7580 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
7581 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7582 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7583 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7584 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7585 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7586 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7587 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7588 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7589 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7590 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7591 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7592 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7593 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7594 ; GFX90A-TGSPLIT-NEXT: s_endpgm
7595 ;
7596 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7597 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
7598 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7599 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7600 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7601 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7602 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7603 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7604 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7605 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7606 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7607 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7608 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7609 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7610 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7611 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
7612 ;
7613 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7614 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
7615 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7616 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7617 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7618 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7619 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7620 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7621 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7622 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7623 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7624 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7625 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7626 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7627 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7628 ; GFX940-TGSPLIT-NEXT: s_endpgm
7629 ;
7630 ; GFX11-WGP-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7631 ; GFX11-WGP: ; %bb.0: ; %entry
7632 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7633 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
7634 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7635 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7636 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7637 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
7638 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
7639 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
7640 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7641 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
7642 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7643 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
7644 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7645 ; GFX11-WGP-NEXT: s_endpgm
7646 ;
7647 ; GFX11-CU-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7648 ; GFX11-CU: ; %bb.0: ; %entry
7649 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7650 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
7651 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7652 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7653 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7654 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
7655 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
7656 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
7657 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7658 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
7659 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7660 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
7661 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7662 ; GFX11-CU-NEXT: s_endpgm
7663 ;
7664 ; GFX12-WGP-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7665 ; GFX12-WGP: ; %bb.0: ; %entry
7666 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7667 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
7668 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7669 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7670 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7671 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
7672 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
7673 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
7674 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7675 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
7676 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7677 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
7678 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7679 ; GFX12-WGP-NEXT: s_endpgm
7680 ;
7681 ; GFX12-CU-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
7682 ; GFX12-CU: ; %bb.0: ; %entry
7683 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7684 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
7685 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7686 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7687 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7688 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
7689 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
7690 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
7691 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7692 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
7693 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7694 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
7695 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7696 ; GFX12-CU-NEXT: s_endpgm
7697 ptr addrspace(1) %out, i32 %in, i32 %old) {
7698 entry:
7699 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
7700 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic acquire
7701 %val0 = extractvalue { i32, i1 } %val, 0
7702 store i32 %val0, ptr addrspace(1) %out, align 4
7703 ret void
7704 }
7706 define amdgpu_kernel void @global_singlethread_acquire_acquire_ret_cmpxchg(
7707 ; GFX6-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7708 ; GFX6: ; %bb.0: ; %entry
7709 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7710 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
7711 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
7712 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
7713 ; GFX6-NEXT: s_mov_b32 s12, s5
7714 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
7715 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
7716 ; GFX6-NEXT: s_mov_b32 s11, -1
7717 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
7718 ; GFX6-NEXT: s_mov_b32 s5, s12
7719 ; GFX6-NEXT: s_mov_b32 s6, s11
7720 ; GFX6-NEXT: s_mov_b32 s7, s10
7721 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
7722 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
7723 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7724 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
7725 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
7726 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7727 ; GFX6-NEXT: s_waitcnt vmcnt(0)
7728 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
7729 ; GFX6-NEXT: s_endpgm
7730 ;
7731 ; GFX7-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7732 ; GFX7: ; %bb.0: ; %entry
7733 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7734 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
7735 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
7736 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
7737 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
7738 ; GFX7-NEXT: s_mov_b32 s6, s4
7739 ; GFX7-NEXT: s_mov_b32 s7, s5
7740 ; GFX7-NEXT: s_mov_b32 s11, s12
7741 ; GFX7-NEXT: s_mov_b32 s10, s13
7742 ; GFX7-NEXT: s_add_u32 s6, s6, s11
7743 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
7744 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
7745 ; GFX7-NEXT: s_mov_b32 s7, s10
7746 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
7747 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
7748 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7749 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
7750 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
7751 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
7752 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
7753 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
7754 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
7755 ; GFX7-NEXT: s_waitcnt vmcnt(0)
7756 ; GFX7-NEXT: flat_store_dword v[0:1], v2
7757 ; GFX7-NEXT: s_endpgm
7758 ;
7759 ; GFX10-WGP-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7760 ; GFX10-WGP: ; %bb.0: ; %entry
7761 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
7762 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
7763 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7764 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
7765 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
7766 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
7767 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
7768 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
7769 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7770 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
7771 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7772 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
7773 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
7774 ; GFX10-WGP-NEXT: s_endpgm
7775 ;
7776 ; GFX10-CU-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7777 ; GFX10-CU: ; %bb.0: ; %entry
7778 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
7779 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
7780 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7781 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
7782 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
7783 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
7784 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
7785 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
7786 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7787 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
7788 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
7789 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
7790 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
7791 ; GFX10-CU-NEXT: s_endpgm
7792 ;
7793 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7794 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
7795 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
7796 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
7797 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
7798 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
7799 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
7800 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
7801 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
7802 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
7803 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
7804 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
7805 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
7806 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
7807 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
7808 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
7809 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7810 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
7811 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
7812 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7813 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
7814 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
7815 ; SKIP-CACHE-INV-NEXT: s_endpgm
7816 ;
7817 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7818 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
7819 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7820 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7821 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7822 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7823 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7824 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7825 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7826 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7827 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7828 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7829 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7830 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7831 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7832 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
7833 ;
7834 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7835 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
7836 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
7837 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7838 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
7839 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
7840 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
7841 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7842 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
7843 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
7844 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7845 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7846 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
7847 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7848 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
7849 ; GFX90A-TGSPLIT-NEXT: s_endpgm
7850 ;
7851 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7852 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
7853 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7854 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7855 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7856 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7857 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7858 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7859 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7860 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7861 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7862 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7863 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7864 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
7865 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7866 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
7867 ;
7868 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7869 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
7870 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
7871 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
7872 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
7873 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
7874 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
7875 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
7876 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
7877 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
7878 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
7879 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
7880 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
7881 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
7882 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
7883 ; GFX940-TGSPLIT-NEXT: s_endpgm
7884 ;
7885 ; GFX11-WGP-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7886 ; GFX11-WGP: ; %bb.0: ; %entry
7887 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7888 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
7889 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7890 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7891 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7892 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
7893 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
7894 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
7895 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7896 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
7897 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7898 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
7899 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7900 ; GFX11-WGP-NEXT: s_endpgm
7901 ;
7902 ; GFX11-CU-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7903 ; GFX11-CU: ; %bb.0: ; %entry
7904 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7905 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
7906 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7907 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7908 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7909 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
7910 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
7911 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
7912 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7913 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
7914 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
7915 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
7916 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7917 ; GFX11-CU-NEXT: s_endpgm
7918 ;
7919 ; GFX12-WGP-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7920 ; GFX12-WGP: ; %bb.0: ; %entry
7921 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
7922 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
7923 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7924 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
7925 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
7926 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
7927 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
7928 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
7929 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7930 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
7931 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7932 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
7933 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
7934 ; GFX12-WGP-NEXT: s_endpgm
7935 ;
7936 ; GFX12-CU-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
7937 ; GFX12-CU: ; %bb.0: ; %entry
7938 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
7939 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
7940 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
7941 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
7942 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
7943 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
7944 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
7945 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
7946 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
7947 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
7948 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
7949 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
7950 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
7951 ; GFX12-CU-NEXT: s_endpgm
7952 ptr addrspace(1) %out, i32 %in, i32 %old) {
7953 entry:
7954 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
7955 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
7956 %val0 = extractvalue { i32, i1 } %val, 0
7957 store i32 %val0, ptr addrspace(1) %out, align 4
7958 ret void
7959 }
7961 define amdgpu_kernel void @global_singlethread_release_acquire_ret_cmpxchg(
7962 ; GFX6-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
7963 ; GFX6: ; %bb.0: ; %entry
7964 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7965 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
7966 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
7967 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
7968 ; GFX6-NEXT: s_mov_b32 s12, s5
7969 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
7970 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
7971 ; GFX6-NEXT: s_mov_b32 s11, -1
7972 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
7973 ; GFX6-NEXT: s_mov_b32 s5, s12
7974 ; GFX6-NEXT: s_mov_b32 s6, s11
7975 ; GFX6-NEXT: s_mov_b32 s7, s10
7976 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
7977 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
7978 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
7979 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
7980 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
7981 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
7982 ; GFX6-NEXT: s_waitcnt vmcnt(0)
7983 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
7984 ; GFX6-NEXT: s_endpgm
7985 ;
7986 ; GFX7-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
7987 ; GFX7: ; %bb.0: ; %entry
7988 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
7989 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
7990 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
7991 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
7992 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
7993 ; GFX7-NEXT: s_mov_b32 s6, s4
7994 ; GFX7-NEXT: s_mov_b32 s7, s5
7995 ; GFX7-NEXT: s_mov_b32 s11, s12
7996 ; GFX7-NEXT: s_mov_b32 s10, s13
7997 ; GFX7-NEXT: s_add_u32 s6, s6, s11
7998 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
7999 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
8000 ; GFX7-NEXT: s_mov_b32 s7, s10
8001 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
8002 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
8003 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8004 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
8005 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
8006 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
8007 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
8008 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
8009 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
8010 ; GFX7-NEXT: s_waitcnt vmcnt(0)
8011 ; GFX7-NEXT: flat_store_dword v[0:1], v2
8012 ; GFX7-NEXT: s_endpgm
8013 ;
8014 ; GFX10-WGP-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8015 ; GFX10-WGP: ; %bb.0: ; %entry
8016 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
8017 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
8018 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8019 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
8020 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
8021 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
8022 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
8023 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
8024 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8025 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
8026 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8027 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
8028 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
8029 ; GFX10-WGP-NEXT: s_endpgm
8030 ;
8031 ; GFX10-CU-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8032 ; GFX10-CU: ; %bb.0: ; %entry
8033 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
8034 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
8035 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8036 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
8037 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
8038 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
8039 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
8040 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
8041 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8042 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
8043 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8044 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
8045 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
8046 ; GFX10-CU-NEXT: s_endpgm
8047 ;
8048 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8049 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
8050 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
8051 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
8052 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
8053 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
8054 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
8055 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
8056 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
8057 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
8058 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
8059 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
8060 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
8061 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
8062 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
8063 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
8064 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8065 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
8066 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
8067 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8068 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
8069 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
8070 ; SKIP-CACHE-INV-NEXT: s_endpgm
8071 ;
8072 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8073 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
8074 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8075 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8076 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8077 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8078 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8079 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8080 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8081 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8082 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8083 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8084 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8085 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8086 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8087 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
8088 ;
8089 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8090 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
8091 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8092 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8093 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8094 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8095 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8096 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8097 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8098 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8099 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8100 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8101 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8102 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8103 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8104 ; GFX90A-TGSPLIT-NEXT: s_endpgm
8105 ;
8106 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8107 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
8108 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8109 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8110 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8111 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8112 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8113 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8114 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8115 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8116 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8117 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8118 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8119 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8120 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8121 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
8122 ;
8123 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8124 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
8125 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8126 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8127 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8128 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8129 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8130 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8131 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8132 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8133 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8134 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8135 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8136 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8137 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8138 ; GFX940-TGSPLIT-NEXT: s_endpgm
8139 ;
8140 ; GFX11-WGP-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8141 ; GFX11-WGP: ; %bb.0: ; %entry
8142 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8143 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
8144 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8145 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8146 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8147 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
8148 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
8149 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
8150 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8151 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
8152 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8153 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
8154 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8155 ; GFX11-WGP-NEXT: s_endpgm
8156 ;
8157 ; GFX11-CU-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8158 ; GFX11-CU: ; %bb.0: ; %entry
8159 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8160 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
8161 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8162 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8163 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8164 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
8165 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
8166 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
8167 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8168 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
8169 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8170 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
8171 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8172 ; GFX11-CU-NEXT: s_endpgm
8173 ;
8174 ; GFX12-WGP-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8175 ; GFX12-WGP: ; %bb.0: ; %entry
8176 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8177 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
8178 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8179 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8180 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8181 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
8182 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
8183 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
8184 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8185 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
8186 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8187 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
8188 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8189 ; GFX12-WGP-NEXT: s_endpgm
8190 ;
8191 ; GFX12-CU-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
8192 ; GFX12-CU: ; %bb.0: ; %entry
8193 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8194 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
8195 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8196 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8197 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8198 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
8199 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
8200 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
8201 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8202 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
8203 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8204 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
8205 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8206 ; GFX12-CU-NEXT: s_endpgm
8207 ptr addrspace(1) %out, i32 %in, i32 %old) {
8208 entry:
8209 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
8210 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
8211 %val0 = extractvalue { i32, i1 } %val, 0
8212 store i32 %val0, ptr addrspace(1) %out, align 4
8213 ret void
8214 }
8216 define amdgpu_kernel void @global_singlethread_acq_rel_acquire_ret_cmpxchg(
8217 ; GFX6-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8218 ; GFX6: ; %bb.0: ; %entry
8219 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8220 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
8221 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
8222 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
8223 ; GFX6-NEXT: s_mov_b32 s12, s5
8224 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
8225 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
8226 ; GFX6-NEXT: s_mov_b32 s11, -1
8227 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
8228 ; GFX6-NEXT: s_mov_b32 s5, s12
8229 ; GFX6-NEXT: s_mov_b32 s6, s11
8230 ; GFX6-NEXT: s_mov_b32 s7, s10
8231 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
8232 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
8233 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8234 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
8235 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
8236 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8237 ; GFX6-NEXT: s_waitcnt vmcnt(0)
8238 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
8239 ; GFX6-NEXT: s_endpgm
8240 ;
8241 ; GFX7-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8242 ; GFX7: ; %bb.0: ; %entry
8243 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8244 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
8245 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
8246 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
8247 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
8248 ; GFX7-NEXT: s_mov_b32 s6, s4
8249 ; GFX7-NEXT: s_mov_b32 s7, s5
8250 ; GFX7-NEXT: s_mov_b32 s11, s12
8251 ; GFX7-NEXT: s_mov_b32 s10, s13
8252 ; GFX7-NEXT: s_add_u32 s6, s6, s11
8253 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
8254 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
8255 ; GFX7-NEXT: s_mov_b32 s7, s10
8256 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
8257 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
8258 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8259 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
8260 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
8261 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
8262 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
8263 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
8264 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
8265 ; GFX7-NEXT: s_waitcnt vmcnt(0)
8266 ; GFX7-NEXT: flat_store_dword v[0:1], v2
8267 ; GFX7-NEXT: s_endpgm
8268 ;
8269 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8270 ; GFX10-WGP: ; %bb.0: ; %entry
8271 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
8272 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
8273 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8274 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
8275 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
8276 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
8277 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
8278 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
8279 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8280 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
8281 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8282 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
8283 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
8284 ; GFX10-WGP-NEXT: s_endpgm
8285 ;
8286 ; GFX10-CU-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8287 ; GFX10-CU: ; %bb.0: ; %entry
8288 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
8289 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
8290 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8291 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
8292 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
8293 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
8294 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
8295 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
8296 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8297 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
8298 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8299 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
8300 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
8301 ; GFX10-CU-NEXT: s_endpgm
8302 ;
8303 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8304 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
8305 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
8306 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
8307 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
8308 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
8309 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
8310 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
8311 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
8312 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
8313 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
8314 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
8315 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
8316 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
8317 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
8318 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
8319 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8320 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
8321 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
8322 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8323 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
8324 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
8325 ; SKIP-CACHE-INV-NEXT: s_endpgm
8326 ;
8327 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8328 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
8329 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8330 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8331 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8332 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8333 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8334 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8335 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8336 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8337 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8338 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8339 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8340 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8341 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8342 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
8343 ;
8344 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8345 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
8346 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8347 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8348 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8349 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8350 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8351 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8352 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8353 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8354 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8355 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8356 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8357 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8358 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8359 ; GFX90A-TGSPLIT-NEXT: s_endpgm
8360 ;
8361 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8362 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
8363 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8364 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8365 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8366 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8367 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8368 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8369 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8370 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8371 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8372 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8373 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8374 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8375 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8376 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
8377 ;
8378 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8379 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
8380 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8381 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8382 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8383 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8384 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8385 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8386 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8387 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8388 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8389 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8390 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8391 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8392 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8393 ; GFX940-TGSPLIT-NEXT: s_endpgm
8394 ;
8395 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8396 ; GFX11-WGP: ; %bb.0: ; %entry
8397 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8398 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
8399 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8400 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8401 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8402 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
8403 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
8404 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
8405 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8406 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
8407 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8408 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
8409 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8410 ; GFX11-WGP-NEXT: s_endpgm
8411 ;
8412 ; GFX11-CU-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8413 ; GFX11-CU: ; %bb.0: ; %entry
8414 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8415 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
8416 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8417 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8418 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8419 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
8420 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
8421 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
8422 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8423 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
8424 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8425 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
8426 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8427 ; GFX11-CU-NEXT: s_endpgm
8428 ;
8429 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8430 ; GFX12-WGP: ; %bb.0: ; %entry
8431 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8432 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
8433 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8434 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8435 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8436 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
8437 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
8438 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
8439 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8440 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
8441 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8442 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
8443 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8444 ; GFX12-WGP-NEXT: s_endpgm
8445 ;
8446 ; GFX12-CU-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
8447 ; GFX12-CU: ; %bb.0: ; %entry
8448 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8449 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
8450 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8451 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8452 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8453 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
8454 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
8455 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
8456 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8457 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
8458 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8459 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
8460 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8461 ; GFX12-CU-NEXT: s_endpgm
8462 ptr addrspace(1) %out, i32 %in, i32 %old) {
8463 entry:
8464 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
8465 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
8466 %val0 = extractvalue { i32, i1 } %val, 0
8467 store i32 %val0, ptr addrspace(1) %out, align 4
8468 ret void
8469 }
8471 define amdgpu_kernel void @global_singlethread_seq_cst_acquire_ret_cmpxchg(
8472 ; GFX6-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8473 ; GFX6: ; %bb.0: ; %entry
8474 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8475 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
8476 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
8477 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
8478 ; GFX6-NEXT: s_mov_b32 s12, s5
8479 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
8480 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
8481 ; GFX6-NEXT: s_mov_b32 s11, -1
8482 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
8483 ; GFX6-NEXT: s_mov_b32 s5, s12
8484 ; GFX6-NEXT: s_mov_b32 s6, s11
8485 ; GFX6-NEXT: s_mov_b32 s7, s10
8486 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
8487 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
8488 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8489 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
8490 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
8491 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8492 ; GFX6-NEXT: s_waitcnt vmcnt(0)
8493 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
8494 ; GFX6-NEXT: s_endpgm
8495 ;
8496 ; GFX7-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8497 ; GFX7: ; %bb.0: ; %entry
8498 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8499 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
8500 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
8501 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
8502 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
8503 ; GFX7-NEXT: s_mov_b32 s6, s4
8504 ; GFX7-NEXT: s_mov_b32 s7, s5
8505 ; GFX7-NEXT: s_mov_b32 s11, s12
8506 ; GFX7-NEXT: s_mov_b32 s10, s13
8507 ; GFX7-NEXT: s_add_u32 s6, s6, s11
8508 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
8509 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
8510 ; GFX7-NEXT: s_mov_b32 s7, s10
8511 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
8512 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
8513 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8514 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
8515 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
8516 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
8517 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
8518 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
8519 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
8520 ; GFX7-NEXT: s_waitcnt vmcnt(0)
8521 ; GFX7-NEXT: flat_store_dword v[0:1], v2
8522 ; GFX7-NEXT: s_endpgm
8523 ;
8524 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8525 ; GFX10-WGP: ; %bb.0: ; %entry
8526 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
8527 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
8528 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8529 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
8530 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
8531 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
8532 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
8533 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
8534 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8535 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
8536 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8537 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
8538 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
8539 ; GFX10-WGP-NEXT: s_endpgm
8540 ;
8541 ; GFX10-CU-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8542 ; GFX10-CU: ; %bb.0: ; %entry
8543 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
8544 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
8545 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8546 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
8547 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
8548 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
8549 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
8550 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
8551 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8552 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
8553 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8554 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
8555 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
8556 ; GFX10-CU-NEXT: s_endpgm
8557 ;
8558 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8559 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
8560 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
8561 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
8562 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
8563 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
8564 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
8565 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
8566 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
8567 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
8568 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
8569 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
8570 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
8571 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
8572 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
8573 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
8574 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8575 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
8576 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
8577 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8578 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
8579 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
8580 ; SKIP-CACHE-INV-NEXT: s_endpgm
8581 ;
8582 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8583 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
8584 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8585 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8586 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8587 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8588 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8589 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8590 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8591 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8592 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8593 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8594 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8595 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8596 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8597 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
8598 ;
8599 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8600 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
8601 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8602 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8603 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8604 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8605 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8606 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8607 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8608 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8609 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8610 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8611 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8612 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8613 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8614 ; GFX90A-TGSPLIT-NEXT: s_endpgm
8615 ;
8616 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8617 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
8618 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8619 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8620 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8621 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8622 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8623 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8624 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8625 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8626 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8627 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8628 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8629 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8630 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8631 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
8632 ;
8633 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8634 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
8635 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8636 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8637 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8638 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8639 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8640 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8641 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8642 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8643 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8644 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8645 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8646 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8647 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8648 ; GFX940-TGSPLIT-NEXT: s_endpgm
8649 ;
8650 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8651 ; GFX11-WGP: ; %bb.0: ; %entry
8652 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8653 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
8654 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8655 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8656 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8657 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
8658 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
8659 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
8660 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8661 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
8662 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8663 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
8664 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8665 ; GFX11-WGP-NEXT: s_endpgm
8666 ;
8667 ; GFX11-CU-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8668 ; GFX11-CU: ; %bb.0: ; %entry
8669 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8670 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
8671 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8672 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8673 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8674 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
8675 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
8676 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
8677 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8678 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
8679 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8680 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
8681 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8682 ; GFX11-CU-NEXT: s_endpgm
8683 ;
8684 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8685 ; GFX12-WGP: ; %bb.0: ; %entry
8686 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8687 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
8688 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8689 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8690 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8691 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
8692 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
8693 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
8694 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8695 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
8696 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8697 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
8698 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8699 ; GFX12-WGP-NEXT: s_endpgm
8700 ;
8701 ; GFX12-CU-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
8702 ; GFX12-CU: ; %bb.0: ; %entry
8703 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8704 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
8705 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8706 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8707 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8708 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
8709 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
8710 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
8711 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8712 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
8713 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8714 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
8715 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8716 ; GFX12-CU-NEXT: s_endpgm
8717 ptr addrspace(1) %out, i32 %in, i32 %old) {
8718 entry:
8719 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
8720 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
8721 %val0 = extractvalue { i32, i1 } %val, 0
8722 store i32 %val0, ptr addrspace(1) %out, align 4
8723 ret void
8724 }
8726 define amdgpu_kernel void @global_singlethread_monotonic_seq_cst_ret_cmpxchg(
8727 ; GFX6-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8728 ; GFX6: ; %bb.0: ; %entry
8729 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8730 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
8731 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
8732 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
8733 ; GFX6-NEXT: s_mov_b32 s12, s5
8734 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
8735 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
8736 ; GFX6-NEXT: s_mov_b32 s11, -1
8737 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
8738 ; GFX6-NEXT: s_mov_b32 s5, s12
8739 ; GFX6-NEXT: s_mov_b32 s6, s11
8740 ; GFX6-NEXT: s_mov_b32 s7, s10
8741 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
8742 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
8743 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8744 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
8745 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
8746 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8747 ; GFX6-NEXT: s_waitcnt vmcnt(0)
8748 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
8749 ; GFX6-NEXT: s_endpgm
8750 ;
8751 ; GFX7-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8752 ; GFX7: ; %bb.0: ; %entry
8753 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8754 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
8755 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
8756 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
8757 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
8758 ; GFX7-NEXT: s_mov_b32 s6, s4
8759 ; GFX7-NEXT: s_mov_b32 s7, s5
8760 ; GFX7-NEXT: s_mov_b32 s11, s12
8761 ; GFX7-NEXT: s_mov_b32 s10, s13
8762 ; GFX7-NEXT: s_add_u32 s6, s6, s11
8763 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
8764 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
8765 ; GFX7-NEXT: s_mov_b32 s7, s10
8766 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
8767 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
8768 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8769 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
8770 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
8771 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
8772 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
8773 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
8774 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
8775 ; GFX7-NEXT: s_waitcnt vmcnt(0)
8776 ; GFX7-NEXT: flat_store_dword v[0:1], v2
8777 ; GFX7-NEXT: s_endpgm
8778 ;
8779 ; GFX10-WGP-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8780 ; GFX10-WGP: ; %bb.0: ; %entry
8781 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
8782 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
8783 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8784 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
8785 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
8786 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
8787 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
8788 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
8789 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8790 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
8791 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8792 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
8793 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
8794 ; GFX10-WGP-NEXT: s_endpgm
8795 ;
8796 ; GFX10-CU-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8797 ; GFX10-CU: ; %bb.0: ; %entry
8798 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
8799 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
8800 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8801 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
8802 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
8803 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
8804 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
8805 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
8806 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8807 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
8808 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
8809 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
8810 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
8811 ; GFX10-CU-NEXT: s_endpgm
8812 ;
8813 ; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8814 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
8815 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
8816 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
8817 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
8818 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
8819 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
8820 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
8821 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
8822 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
8823 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
8824 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
8825 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
8826 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
8827 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
8828 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
8829 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8830 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
8831 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
8832 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
8833 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
8834 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
8835 ; SKIP-CACHE-INV-NEXT: s_endpgm
8836 ;
8837 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8838 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
8839 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8840 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8841 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8842 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8843 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8844 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8845 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8846 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8847 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8848 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8849 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8850 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8851 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8852 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
8853 ;
8854 ; GFX90A-TGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8855 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
8856 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
8857 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8858 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
8859 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
8860 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
8861 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8862 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
8863 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
8864 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8865 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8866 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
8867 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8868 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
8869 ; GFX90A-TGSPLIT-NEXT: s_endpgm
8870 ;
8871 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8872 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
8873 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8874 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8875 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8876 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8877 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8878 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8879 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8880 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8881 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8882 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8883 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8884 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
8885 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8886 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
8887 ;
8888 ; GFX940-TGSPLIT-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8889 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
8890 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
8891 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
8892 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
8893 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
8894 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
8895 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
8896 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
8897 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
8898 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
8899 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
8900 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
8901 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
8902 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
8903 ; GFX940-TGSPLIT-NEXT: s_endpgm
8904 ;
8905 ; GFX11-WGP-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8906 ; GFX11-WGP: ; %bb.0: ; %entry
8907 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8908 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
8909 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8910 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8911 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8912 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
8913 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
8914 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
8915 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8916 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
8917 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8918 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
8919 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8920 ; GFX11-WGP-NEXT: s_endpgm
8921 ;
8922 ; GFX11-CU-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8923 ; GFX11-CU: ; %bb.0: ; %entry
8924 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8925 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
8926 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8927 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8928 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8929 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
8930 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
8931 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
8932 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8933 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
8934 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
8935 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
8936 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8937 ; GFX11-CU-NEXT: s_endpgm
8938 ;
8939 ; GFX12-WGP-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8940 ; GFX12-WGP: ; %bb.0: ; %entry
8941 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
8942 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
8943 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8944 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
8945 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
8946 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
8947 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
8948 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
8949 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8950 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
8951 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8952 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
8953 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
8954 ; GFX12-WGP-NEXT: s_endpgm
8955 ;
8956 ; GFX12-CU-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
8957 ; GFX12-CU: ; %bb.0: ; %entry
8958 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
8959 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
8960 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
8961 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
8962 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
8963 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
8964 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
8965 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
8966 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
8967 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
8968 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
8969 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
8970 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
8971 ; GFX12-CU-NEXT: s_endpgm
8972 ptr addrspace(1) %out, i32 %in, i32 %old) {
8973 entry:
8974 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
8975 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") monotonic seq_cst
8976 %val0 = extractvalue { i32, i1 } %val, 0
8977 store i32 %val0, ptr addrspace(1) %out, align 4
8978 ret void
8979 }
8981 define amdgpu_kernel void @global_singlethread_acquire_seq_cst_ret_cmpxchg(
8982 ; GFX6-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
8983 ; GFX6: ; %bb.0: ; %entry
8984 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
8985 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
8986 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
8987 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
8988 ; GFX6-NEXT: s_mov_b32 s12, s5
8989 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
8990 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
8991 ; GFX6-NEXT: s_mov_b32 s11, -1
8992 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
8993 ; GFX6-NEXT: s_mov_b32 s5, s12
8994 ; GFX6-NEXT: s_mov_b32 s6, s11
8995 ; GFX6-NEXT: s_mov_b32 s7, s10
8996 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
8997 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
8998 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
8999 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
9000 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
9001 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9002 ; GFX6-NEXT: s_waitcnt vmcnt(0)
9003 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
9004 ; GFX6-NEXT: s_endpgm
9005 ;
9006 ; GFX7-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9007 ; GFX7: ; %bb.0: ; %entry
9008 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9009 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
9010 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
9011 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
9012 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
9013 ; GFX7-NEXT: s_mov_b32 s6, s4
9014 ; GFX7-NEXT: s_mov_b32 s7, s5
9015 ; GFX7-NEXT: s_mov_b32 s11, s12
9016 ; GFX7-NEXT: s_mov_b32 s10, s13
9017 ; GFX7-NEXT: s_add_u32 s6, s6, s11
9018 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
9019 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
9020 ; GFX7-NEXT: s_mov_b32 s7, s10
9021 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
9022 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
9023 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9024 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
9025 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
9026 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
9027 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
9028 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
9029 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
9030 ; GFX7-NEXT: s_waitcnt vmcnt(0)
9031 ; GFX7-NEXT: flat_store_dword v[0:1], v2
9032 ; GFX7-NEXT: s_endpgm
9033 ;
9034 ; GFX10-WGP-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9035 ; GFX10-WGP: ; %bb.0: ; %entry
9036 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
9037 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
9038 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9039 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
9040 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
9041 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
9042 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
9043 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
9044 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9045 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
9046 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9047 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
9048 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
9049 ; GFX10-WGP-NEXT: s_endpgm
9050 ;
9051 ; GFX10-CU-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9052 ; GFX10-CU: ; %bb.0: ; %entry
9053 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
9054 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
9055 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9056 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
9057 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
9058 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
9059 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
9060 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
9061 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9062 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
9063 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9064 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
9065 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
9066 ; GFX10-CU-NEXT: s_endpgm
9067 ;
9068 ; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9069 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
9070 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
9071 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
9072 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
9073 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
9074 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
9075 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
9076 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
9077 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
9078 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
9079 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
9080 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
9081 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
9082 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
9083 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
9084 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9085 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
9086 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
9087 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9088 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
9089 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
9090 ; SKIP-CACHE-INV-NEXT: s_endpgm
9091 ;
9092 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9093 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
9094 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9095 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9096 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9097 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9098 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9099 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9100 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9101 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9102 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9103 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9104 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9105 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9106 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9107 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
9108 ;
9109 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9110 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
9111 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9112 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9113 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9114 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9115 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9116 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9117 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9118 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9119 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9120 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9121 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9122 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9123 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9124 ; GFX90A-TGSPLIT-NEXT: s_endpgm
9125 ;
9126 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9127 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
9128 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9129 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9130 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9131 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9132 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9133 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9134 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9135 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9136 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9137 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9138 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9139 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9140 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9141 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
9142 ;
9143 ; GFX940-TGSPLIT-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9144 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
9145 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9146 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9147 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9148 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9149 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9150 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9151 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9152 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9153 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9154 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9155 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9156 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9157 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9158 ; GFX940-TGSPLIT-NEXT: s_endpgm
9159 ;
9160 ; GFX11-WGP-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9161 ; GFX11-WGP: ; %bb.0: ; %entry
9162 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9163 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
9164 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9165 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9166 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9167 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
9168 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
9169 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
9170 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9171 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
9172 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9173 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
9174 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9175 ; GFX11-WGP-NEXT: s_endpgm
9176 ;
9177 ; GFX11-CU-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9178 ; GFX11-CU: ; %bb.0: ; %entry
9179 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9180 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
9181 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9182 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9183 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9184 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
9185 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
9186 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
9187 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9188 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
9189 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9190 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
9191 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9192 ; GFX11-CU-NEXT: s_endpgm
9193 ;
9194 ; GFX12-WGP-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9195 ; GFX12-WGP: ; %bb.0: ; %entry
9196 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9197 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
9198 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9199 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9200 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9201 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
9202 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
9203 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
9204 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9205 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
9206 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9207 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
9208 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9209 ; GFX12-WGP-NEXT: s_endpgm
9210 ;
9211 ; GFX12-CU-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
9212 ; GFX12-CU: ; %bb.0: ; %entry
9213 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9214 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
9215 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9216 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9217 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9218 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
9219 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
9220 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
9221 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9222 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
9223 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9224 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
9225 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9226 ; GFX12-CU-NEXT: s_endpgm
9227 ptr addrspace(1) %out, i32 %in, i32 %old) {
9228 entry:
9229 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
9230 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acquire seq_cst
9231 %val0 = extractvalue { i32, i1 } %val, 0
9232 store i32 %val0, ptr addrspace(1) %out, align 4
9233 ret void
9234 }
9236 define amdgpu_kernel void @global_singlethread_release_seq_cst_ret_cmpxchg(
9237 ; GFX6-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9238 ; GFX6: ; %bb.0: ; %entry
9239 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9240 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
9241 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
9242 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
9243 ; GFX6-NEXT: s_mov_b32 s12, s5
9244 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
9245 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
9246 ; GFX6-NEXT: s_mov_b32 s11, -1
9247 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
9248 ; GFX6-NEXT: s_mov_b32 s5, s12
9249 ; GFX6-NEXT: s_mov_b32 s6, s11
9250 ; GFX6-NEXT: s_mov_b32 s7, s10
9251 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
9252 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
9253 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9254 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
9255 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
9256 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9257 ; GFX6-NEXT: s_waitcnt vmcnt(0)
9258 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
9259 ; GFX6-NEXT: s_endpgm
9260 ;
9261 ; GFX7-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9262 ; GFX7: ; %bb.0: ; %entry
9263 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9264 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
9265 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
9266 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
9267 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
9268 ; GFX7-NEXT: s_mov_b32 s6, s4
9269 ; GFX7-NEXT: s_mov_b32 s7, s5
9270 ; GFX7-NEXT: s_mov_b32 s11, s12
9271 ; GFX7-NEXT: s_mov_b32 s10, s13
9272 ; GFX7-NEXT: s_add_u32 s6, s6, s11
9273 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
9274 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
9275 ; GFX7-NEXT: s_mov_b32 s7, s10
9276 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
9277 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
9278 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9279 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
9280 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
9281 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
9282 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
9283 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
9284 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
9285 ; GFX7-NEXT: s_waitcnt vmcnt(0)
9286 ; GFX7-NEXT: flat_store_dword v[0:1], v2
9287 ; GFX7-NEXT: s_endpgm
9288 ;
9289 ; GFX10-WGP-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9290 ; GFX10-WGP: ; %bb.0: ; %entry
9291 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
9292 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
9293 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9294 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
9295 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
9296 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
9297 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
9298 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
9299 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9300 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
9301 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9302 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
9303 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
9304 ; GFX10-WGP-NEXT: s_endpgm
9305 ;
9306 ; GFX10-CU-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9307 ; GFX10-CU: ; %bb.0: ; %entry
9308 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
9309 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
9310 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9311 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
9312 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
9313 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
9314 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
9315 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
9316 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9317 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
9318 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9319 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
9320 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
9321 ; GFX10-CU-NEXT: s_endpgm
9322 ;
9323 ; SKIP-CACHE-INV-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9324 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
9325 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
9326 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
9327 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
9328 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
9329 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
9330 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
9331 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
9332 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
9333 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
9334 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
9335 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
9336 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
9337 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
9338 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
9339 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9340 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
9341 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
9342 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9343 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
9344 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
9345 ; SKIP-CACHE-INV-NEXT: s_endpgm
9346 ;
9347 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9348 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
9349 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9350 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9351 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9352 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9353 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9354 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9355 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9356 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9357 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9358 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9359 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9360 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9361 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9362 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
9363 ;
9364 ; GFX90A-TGSPLIT-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9365 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
9366 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9367 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9368 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9369 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9370 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9371 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9372 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9373 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9374 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9375 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9376 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9377 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9378 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9379 ; GFX90A-TGSPLIT-NEXT: s_endpgm
9380 ;
9381 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9382 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
9383 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9384 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9385 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9386 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9387 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9388 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9389 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9390 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9391 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9392 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9393 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9394 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9395 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9396 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
9397 ;
9398 ; GFX940-TGSPLIT-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9399 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
9400 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9401 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9402 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9403 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9404 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9405 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9406 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9407 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9408 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9409 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9410 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9411 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9412 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9413 ; GFX940-TGSPLIT-NEXT: s_endpgm
9414 ;
9415 ; GFX11-WGP-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9416 ; GFX11-WGP: ; %bb.0: ; %entry
9417 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9418 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
9419 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9420 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9421 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9422 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
9423 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
9424 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
9425 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9426 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
9427 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9428 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
9429 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9430 ; GFX11-WGP-NEXT: s_endpgm
9431 ;
9432 ; GFX11-CU-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9433 ; GFX11-CU: ; %bb.0: ; %entry
9434 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9435 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
9436 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9437 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9438 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9439 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
9440 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
9441 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
9442 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9443 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
9444 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9445 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
9446 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9447 ; GFX11-CU-NEXT: s_endpgm
9448 ;
9449 ; GFX12-WGP-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9450 ; GFX12-WGP: ; %bb.0: ; %entry
9451 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9452 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
9453 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9454 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9455 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9456 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
9457 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
9458 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
9459 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9460 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
9461 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9462 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
9463 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9464 ; GFX12-WGP-NEXT: s_endpgm
9465 ;
9466 ; GFX12-CU-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
9467 ; GFX12-CU: ; %bb.0: ; %entry
9468 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9469 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
9470 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9471 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9472 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9473 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
9474 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
9475 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
9476 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9477 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
9478 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9479 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
9480 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9481 ; GFX12-CU-NEXT: s_endpgm
9482 ptr addrspace(1) %out, i32 %in, i32 %old) {
9483 entry:
9484 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
9485 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") release seq_cst
9486 %val0 = extractvalue { i32, i1 } %val, 0
9487 store i32 %val0, ptr addrspace(1) %out, align 4
9488 ret void
9489 }
9491 define amdgpu_kernel void @global_singlethread_acq_rel_seq_cst_ret_cmpxchg(
9492 ; GFX6-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9493 ; GFX6: ; %bb.0: ; %entry
9494 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9495 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
9496 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
9497 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
9498 ; GFX6-NEXT: s_mov_b32 s12, s5
9499 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
9500 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
9501 ; GFX6-NEXT: s_mov_b32 s11, -1
9502 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
9503 ; GFX6-NEXT: s_mov_b32 s5, s12
9504 ; GFX6-NEXT: s_mov_b32 s6, s11
9505 ; GFX6-NEXT: s_mov_b32 s7, s10
9506 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
9507 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
9508 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9509 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
9510 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
9511 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9512 ; GFX6-NEXT: s_waitcnt vmcnt(0)
9513 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
9514 ; GFX6-NEXT: s_endpgm
9515 ;
9516 ; GFX7-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9517 ; GFX7: ; %bb.0: ; %entry
9518 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9519 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
9520 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
9521 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
9522 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
9523 ; GFX7-NEXT: s_mov_b32 s6, s4
9524 ; GFX7-NEXT: s_mov_b32 s7, s5
9525 ; GFX7-NEXT: s_mov_b32 s11, s12
9526 ; GFX7-NEXT: s_mov_b32 s10, s13
9527 ; GFX7-NEXT: s_add_u32 s6, s6, s11
9528 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
9529 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
9530 ; GFX7-NEXT: s_mov_b32 s7, s10
9531 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
9532 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
9533 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9534 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
9535 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
9536 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
9537 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
9538 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
9539 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
9540 ; GFX7-NEXT: s_waitcnt vmcnt(0)
9541 ; GFX7-NEXT: flat_store_dword v[0:1], v2
9542 ; GFX7-NEXT: s_endpgm
9543 ;
9544 ; GFX10-WGP-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9545 ; GFX10-WGP: ; %bb.0: ; %entry
9546 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
9547 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
9548 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9549 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
9550 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
9551 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
9552 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
9553 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
9554 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9555 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
9556 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9557 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
9558 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
9559 ; GFX10-WGP-NEXT: s_endpgm
9560 ;
9561 ; GFX10-CU-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9562 ; GFX10-CU: ; %bb.0: ; %entry
9563 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
9564 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
9565 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9566 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
9567 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
9568 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
9569 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
9570 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
9571 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9572 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
9573 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9574 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
9575 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
9576 ; GFX10-CU-NEXT: s_endpgm
9577 ;
9578 ; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9579 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
9580 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
9581 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
9582 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
9583 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
9584 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
9585 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
9586 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
9587 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
9588 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
9589 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
9590 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
9591 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
9592 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
9593 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
9594 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9595 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
9596 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
9597 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9598 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
9599 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
9600 ; SKIP-CACHE-INV-NEXT: s_endpgm
9601 ;
9602 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9603 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
9604 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9605 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9606 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9607 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9608 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9609 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9610 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9611 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9612 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9613 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9614 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9615 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9616 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9617 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
9618 ;
9619 ; GFX90A-TGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9620 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
9621 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9622 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9623 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9624 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9625 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9626 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9627 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9628 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9629 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9630 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9631 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9632 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9633 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9634 ; GFX90A-TGSPLIT-NEXT: s_endpgm
9635 ;
9636 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9637 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
9638 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9639 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9640 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9641 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9642 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9643 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9644 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9645 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9646 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9647 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9648 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9649 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9650 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9651 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
9652 ;
9653 ; GFX940-TGSPLIT-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9654 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
9655 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9656 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9657 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9658 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9659 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9660 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9661 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9662 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9663 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9664 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9665 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9666 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9667 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9668 ; GFX940-TGSPLIT-NEXT: s_endpgm
9669 ;
9670 ; GFX11-WGP-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9671 ; GFX11-WGP: ; %bb.0: ; %entry
9672 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9673 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
9674 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9675 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9676 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9677 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
9678 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
9679 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
9680 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9681 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
9682 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9683 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
9684 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9685 ; GFX11-WGP-NEXT: s_endpgm
9686 ;
9687 ; GFX11-CU-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9688 ; GFX11-CU: ; %bb.0: ; %entry
9689 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9690 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
9691 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9692 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9693 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9694 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
9695 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
9696 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
9697 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9698 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
9699 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9700 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
9701 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9702 ; GFX11-CU-NEXT: s_endpgm
9703 ;
9704 ; GFX12-WGP-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9705 ; GFX12-WGP: ; %bb.0: ; %entry
9706 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9707 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
9708 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9709 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9710 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9711 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
9712 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
9713 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
9714 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9715 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
9716 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9717 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
9718 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9719 ; GFX12-WGP-NEXT: s_endpgm
9720 ;
9721 ; GFX12-CU-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
9722 ; GFX12-CU: ; %bb.0: ; %entry
9723 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9724 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
9725 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9726 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9727 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9728 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
9729 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
9730 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
9731 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9732 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
9733 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9734 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
9735 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9736 ; GFX12-CU-NEXT: s_endpgm
9737 ptr addrspace(1) %out, i32 %in, i32 %old) {
9738 entry:
9739 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
9740 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel seq_cst
9741 %val0 = extractvalue { i32, i1 } %val, 0
9742 store i32 %val0, ptr addrspace(1) %out, align 4
9743 ret void
9744 }
9746 define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_ret_cmpxchg(
9747 ; GFX6-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9748 ; GFX6: ; %bb.0: ; %entry
9749 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9750 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
9751 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
9752 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
9753 ; GFX6-NEXT: s_mov_b32 s12, s5
9754 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
9755 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
9756 ; GFX6-NEXT: s_mov_b32 s11, -1
9757 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
9758 ; GFX6-NEXT: s_mov_b32 s5, s12
9759 ; GFX6-NEXT: s_mov_b32 s6, s11
9760 ; GFX6-NEXT: s_mov_b32 s7, s10
9761 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
9762 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
9763 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9764 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
9765 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
9766 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9767 ; GFX6-NEXT: s_waitcnt vmcnt(0)
9768 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
9769 ; GFX6-NEXT: s_endpgm
9770 ;
9771 ; GFX7-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9772 ; GFX7: ; %bb.0: ; %entry
9773 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
9774 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
9775 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
9776 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
9777 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
9778 ; GFX7-NEXT: s_mov_b32 s6, s4
9779 ; GFX7-NEXT: s_mov_b32 s7, s5
9780 ; GFX7-NEXT: s_mov_b32 s11, s12
9781 ; GFX7-NEXT: s_mov_b32 s10, s13
9782 ; GFX7-NEXT: s_add_u32 s6, s6, s11
9783 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
9784 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
9785 ; GFX7-NEXT: s_mov_b32 s7, s10
9786 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
9787 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
9788 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9789 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
9790 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
9791 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
9792 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
9793 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
9794 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
9795 ; GFX7-NEXT: s_waitcnt vmcnt(0)
9796 ; GFX7-NEXT: flat_store_dword v[0:1], v2
9797 ; GFX7-NEXT: s_endpgm
9798 ;
9799 ; GFX10-WGP-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9800 ; GFX10-WGP: ; %bb.0: ; %entry
9801 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
9802 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
9803 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9804 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
9805 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
9806 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
9807 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
9808 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
9809 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9810 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
9811 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9812 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
9813 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
9814 ; GFX10-WGP-NEXT: s_endpgm
9815 ;
9816 ; GFX10-CU-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9817 ; GFX10-CU: ; %bb.0: ; %entry
9818 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
9819 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
9820 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9821 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
9822 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
9823 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
9824 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
9825 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
9826 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9827 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
9828 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
9829 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
9830 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
9831 ; GFX10-CU-NEXT: s_endpgm
9832 ;
9833 ; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9834 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
9835 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
9836 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
9837 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
9838 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
9839 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
9840 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
9841 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
9842 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
9843 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
9844 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
9845 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
9846 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
9847 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
9848 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
9849 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
9850 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
9851 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
9852 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
9853 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
9854 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
9855 ; SKIP-CACHE-INV-NEXT: s_endpgm
9856 ;
9857 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9858 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
9859 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9860 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9861 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9862 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9863 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9864 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9865 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9866 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9867 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9868 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9869 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9870 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9871 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9872 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
9873 ;
9874 ; GFX90A-TGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9875 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
9876 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
9877 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9878 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
9879 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
9880 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
9881 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9882 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
9883 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
9884 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9885 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9886 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
9887 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9888 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
9889 ; GFX90A-TGSPLIT-NEXT: s_endpgm
9890 ;
9891 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9892 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
9893 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9894 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9895 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9896 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9897 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9898 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9899 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9900 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9901 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9902 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9903 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9904 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
9905 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9906 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
9907 ;
9908 ; GFX940-TGSPLIT-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9909 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
9910 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
9911 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
9912 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
9913 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
9914 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
9915 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
9916 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
9917 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
9918 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
9919 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
9920 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
9921 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
9922 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
9923 ; GFX940-TGSPLIT-NEXT: s_endpgm
9924 ;
9925 ; GFX11-WGP-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9926 ; GFX11-WGP: ; %bb.0: ; %entry
9927 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9928 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
9929 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9930 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9931 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9932 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
9933 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
9934 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
9935 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9936 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
9937 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9938 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
9939 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9940 ; GFX11-WGP-NEXT: s_endpgm
9941 ;
9942 ; GFX11-CU-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9943 ; GFX11-CU: ; %bb.0: ; %entry
9944 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9945 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
9946 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9947 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9948 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9949 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
9950 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
9951 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
9952 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9953 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
9954 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
9955 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
9956 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9957 ; GFX11-CU-NEXT: s_endpgm
9958 ;
9959 ; GFX12-WGP-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9960 ; GFX12-WGP: ; %bb.0: ; %entry
9961 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
9962 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
9963 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9964 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
9965 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
9966 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
9967 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
9968 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
9969 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9970 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
9971 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9972 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
9973 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
9974 ; GFX12-WGP-NEXT: s_endpgm
9975 ;
9976 ; GFX12-CU-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
9977 ; GFX12-CU: ; %bb.0: ; %entry
9978 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
9979 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
9980 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
9981 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
9982 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
9983 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
9984 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
9985 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
9986 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
9987 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
9988 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
9989 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
9990 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
9991 ; GFX12-CU-NEXT: s_endpgm
9992 ptr addrspace(1) %out, i32 %in, i32 %old) {
9993 entry:
9994 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
9995 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
9996 %val0 = extractvalue { i32, i1 } %val, 0
9997 store i32 %val0, ptr addrspace(1) %out, align 4
9998 ret void
9999 }
10001 define amdgpu_kernel void @global_singlethread_one_as_unordered_load(
10002 ; GFX6-LABEL: global_singlethread_one_as_unordered_load:
10003 ; GFX6: ; %bb.0: ; %entry
10004 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
10005 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10006 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10007 ; GFX6-NEXT: s_mov_b32 s6, s9
10008 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
10009 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
10010 ; GFX6-NEXT: s_mov_b32 s13, -1
10011 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
10012 ; GFX6-NEXT: s_mov_b32 s9, s6
10013 ; GFX6-NEXT: s_mov_b32 s10, s13
10014 ; GFX6-NEXT: s_mov_b32 s11, s12
10015 ; GFX6-NEXT: s_mov_b32 s14, s5
10016 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10017 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10018 ; GFX6-NEXT: s_mov_b32 s5, s14
10019 ; GFX6-NEXT: s_mov_b32 s6, s13
10020 ; GFX6-NEXT: s_mov_b32 s7, s12
10021 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
10022 ; GFX6-NEXT: s_waitcnt vmcnt(0)
10023 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10024 ; GFX6-NEXT: s_endpgm
10025 ;
10026 ; GFX7-LABEL: global_singlethread_one_as_unordered_load:
10027 ; GFX7: ; %bb.0: ; %entry
10028 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
10029 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10030 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
10031 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10032 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10033 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10034 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
10035 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
10036 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
10037 ; GFX7-NEXT: s_waitcnt vmcnt(0)
10038 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10039 ; GFX7-NEXT: s_endpgm
10040 ;
10041 ; GFX10-WGP-LABEL: global_singlethread_one_as_unordered_load:
10042 ; GFX10-WGP: ; %bb.0: ; %entry
10043 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10044 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10045 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10046 ; GFX10-WGP-NEXT: s_nop 0
10047 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10048 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10049 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
10050 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
10051 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10052 ; GFX10-WGP-NEXT: s_endpgm
10053 ;
10054 ; GFX10-CU-LABEL: global_singlethread_one_as_unordered_load:
10055 ; GFX10-CU: ; %bb.0: ; %entry
10056 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10057 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
10058 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10059 ; GFX10-CU-NEXT: s_nop 0
10060 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10061 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
10062 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
10063 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
10064 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
10065 ; GFX10-CU-NEXT: s_endpgm
10066 ;
10067 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_unordered_load:
10068 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
10069 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
10070 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
10071 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
10072 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
10073 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10074 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
10075 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
10076 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10077 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
10078 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
10079 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
10080 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
10081 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
10082 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
10083 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
10084 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
10085 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
10086 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
10087 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
10088 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
10089 ; SKIP-CACHE-INV-NEXT: s_endpgm
10090 ;
10091 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_unordered_load:
10092 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
10093 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10094 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10095 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10096 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
10097 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10098 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10099 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10100 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10101 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10102 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
10103 ;
10104 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_unordered_load:
10105 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
10106 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10107 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10108 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10109 ; GFX90A-TGSPLIT-NEXT: s_nop 0
10110 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10111 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10112 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10113 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10114 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10115 ; GFX90A-TGSPLIT-NEXT: s_endpgm
10116 ;
10117 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_unordered_load:
10118 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
10119 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10120 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10121 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10122 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
10123 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10124 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10125 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10126 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10127 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10128 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
10129 ;
10130 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_unordered_load:
10131 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
10132 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10133 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10134 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10135 ; GFX940-TGSPLIT-NEXT: s_nop 0
10136 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10137 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10138 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10139 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10140 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10141 ; GFX940-TGSPLIT-NEXT: s_endpgm
10142 ;
10143 ; GFX11-WGP-LABEL: global_singlethread_one_as_unordered_load:
10144 ; GFX11-WGP: ; %bb.0: ; %entry
10145 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10146 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
10147 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10148 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10149 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
10150 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10151 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
10152 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10153 ; GFX11-WGP-NEXT: s_endpgm
10154 ;
10155 ; GFX11-CU-LABEL: global_singlethread_one_as_unordered_load:
10156 ; GFX11-CU: ; %bb.0: ; %entry
10157 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10158 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
10159 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10160 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10161 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
10162 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10163 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
10164 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10165 ; GFX11-CU-NEXT: s_endpgm
10166 ;
10167 ; GFX12-WGP-LABEL: global_singlethread_one_as_unordered_load:
10168 ; GFX12-WGP: ; %bb.0: ; %entry
10169 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10170 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
10171 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10172 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10173 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
10174 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10175 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
10176 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10177 ; GFX12-WGP-NEXT: s_endpgm
10178 ;
10179 ; GFX12-CU-LABEL: global_singlethread_one_as_unordered_load:
10180 ; GFX12-CU: ; %bb.0: ; %entry
10181 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10182 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
10183 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10184 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10185 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
10186 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10187 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
10188 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10189 ; GFX12-CU-NEXT: s_endpgm
10190 ptr addrspace(1) %in, ptr addrspace(1) %out) {
10191 entry:
10192 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread-one-as") unordered, align 4
10193 store i32 %val, ptr addrspace(1) %out
10194 ret void
10195 }
10197 define amdgpu_kernel void @global_singlethread_one_as_monotonic_load(
10198 ; GFX6-LABEL: global_singlethread_one_as_monotonic_load:
10199 ; GFX6: ; %bb.0: ; %entry
10200 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
10201 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10202 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10203 ; GFX6-NEXT: s_mov_b32 s6, s9
10204 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
10205 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
10206 ; GFX6-NEXT: s_mov_b32 s13, -1
10207 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
10208 ; GFX6-NEXT: s_mov_b32 s9, s6
10209 ; GFX6-NEXT: s_mov_b32 s10, s13
10210 ; GFX6-NEXT: s_mov_b32 s11, s12
10211 ; GFX6-NEXT: s_mov_b32 s14, s5
10212 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10213 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10214 ; GFX6-NEXT: s_mov_b32 s5, s14
10215 ; GFX6-NEXT: s_mov_b32 s6, s13
10216 ; GFX6-NEXT: s_mov_b32 s7, s12
10217 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
10218 ; GFX6-NEXT: s_waitcnt vmcnt(0)
10219 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10220 ; GFX6-NEXT: s_endpgm
10221 ;
10222 ; GFX7-LABEL: global_singlethread_one_as_monotonic_load:
10223 ; GFX7: ; %bb.0: ; %entry
10224 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
10225 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10226 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
10227 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10228 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10229 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10230 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
10231 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
10232 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
10233 ; GFX7-NEXT: s_waitcnt vmcnt(0)
10234 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10235 ; GFX7-NEXT: s_endpgm
10236 ;
10237 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_load:
10238 ; GFX10-WGP: ; %bb.0: ; %entry
10239 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10240 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10241 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10242 ; GFX10-WGP-NEXT: s_nop 0
10243 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10244 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10245 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
10246 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
10247 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10248 ; GFX10-WGP-NEXT: s_endpgm
10249 ;
10250 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_load:
10251 ; GFX10-CU: ; %bb.0: ; %entry
10252 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10253 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
10254 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10255 ; GFX10-CU-NEXT: s_nop 0
10256 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10257 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
10258 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
10259 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
10260 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
10261 ; GFX10-CU-NEXT: s_endpgm
10262 ;
10263 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_load:
10264 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
10265 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
10266 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
10267 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
10268 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
10269 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10270 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
10271 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
10272 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10273 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
10274 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
10275 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
10276 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
10277 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
10278 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
10279 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
10280 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
10281 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
10282 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
10283 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
10284 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
10285 ; SKIP-CACHE-INV-NEXT: s_endpgm
10286 ;
10287 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_load:
10288 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
10289 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10290 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10291 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10292 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
10293 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10294 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10295 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10296 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10297 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10298 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
10299 ;
10300 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_load:
10301 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
10302 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10303 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10304 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10305 ; GFX90A-TGSPLIT-NEXT: s_nop 0
10306 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10307 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10308 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10309 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10310 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10311 ; GFX90A-TGSPLIT-NEXT: s_endpgm
10312 ;
10313 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_load:
10314 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
10315 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10316 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10317 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10318 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
10319 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10320 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10321 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10322 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10323 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10324 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
10325 ;
10326 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_load:
10327 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
10328 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10329 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10330 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10331 ; GFX940-TGSPLIT-NEXT: s_nop 0
10332 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10333 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10334 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10335 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10336 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10337 ; GFX940-TGSPLIT-NEXT: s_endpgm
10338 ;
10339 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_load:
10340 ; GFX11-WGP: ; %bb.0: ; %entry
10341 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10342 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
10343 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10344 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10345 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
10346 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10347 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
10348 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10349 ; GFX11-WGP-NEXT: s_endpgm
10350 ;
10351 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_load:
10352 ; GFX11-CU: ; %bb.0: ; %entry
10353 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10354 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
10355 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10356 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10357 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
10358 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10359 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
10360 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10361 ; GFX11-CU-NEXT: s_endpgm
10362 ;
10363 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_load:
10364 ; GFX12-WGP: ; %bb.0: ; %entry
10365 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10366 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
10367 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10368 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10369 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
10370 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10371 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
10372 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10373 ; GFX12-WGP-NEXT: s_endpgm
10374 ;
10375 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_load:
10376 ; GFX12-CU: ; %bb.0: ; %entry
10377 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10378 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
10379 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10380 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10381 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
10382 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10383 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
10384 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10385 ; GFX12-CU-NEXT: s_endpgm
10386 ptr addrspace(1) %in, ptr addrspace(1) %out) {
10387 entry:
10388 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread-one-as") monotonic, align 4
10389 store i32 %val, ptr addrspace(1) %out
10390 ret void
10391 }
10393 define amdgpu_kernel void @global_singlethread_one_as_acquire_load(
10394 ; GFX6-LABEL: global_singlethread_one_as_acquire_load:
10395 ; GFX6: ; %bb.0: ; %entry
10396 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
10397 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10398 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10399 ; GFX6-NEXT: s_mov_b32 s6, s9
10400 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
10401 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
10402 ; GFX6-NEXT: s_mov_b32 s13, -1
10403 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
10404 ; GFX6-NEXT: s_mov_b32 s9, s6
10405 ; GFX6-NEXT: s_mov_b32 s10, s13
10406 ; GFX6-NEXT: s_mov_b32 s11, s12
10407 ; GFX6-NEXT: s_mov_b32 s14, s5
10408 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10409 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10410 ; GFX6-NEXT: s_mov_b32 s5, s14
10411 ; GFX6-NEXT: s_mov_b32 s6, s13
10412 ; GFX6-NEXT: s_mov_b32 s7, s12
10413 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
10414 ; GFX6-NEXT: s_waitcnt vmcnt(0)
10415 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10416 ; GFX6-NEXT: s_endpgm
10417 ;
10418 ; GFX7-LABEL: global_singlethread_one_as_acquire_load:
10419 ; GFX7: ; %bb.0: ; %entry
10420 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
10421 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10422 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
10423 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10424 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10425 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10426 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
10427 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
10428 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
10429 ; GFX7-NEXT: s_waitcnt vmcnt(0)
10430 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10431 ; GFX7-NEXT: s_endpgm
10432 ;
10433 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_load:
10434 ; GFX10-WGP: ; %bb.0: ; %entry
10435 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10436 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10437 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10438 ; GFX10-WGP-NEXT: s_nop 0
10439 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10440 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10441 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
10442 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
10443 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10444 ; GFX10-WGP-NEXT: s_endpgm
10445 ;
10446 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_load:
10447 ; GFX10-CU: ; %bb.0: ; %entry
10448 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10449 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
10450 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10451 ; GFX10-CU-NEXT: s_nop 0
10452 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10453 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
10454 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
10455 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
10456 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
10457 ; GFX10-CU-NEXT: s_endpgm
10458 ;
10459 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_load:
10460 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
10461 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
10462 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
10463 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
10464 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
10465 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10466 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
10467 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
10468 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10469 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
10470 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
10471 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
10472 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
10473 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
10474 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
10475 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
10476 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
10477 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
10478 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
10479 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
10480 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
10481 ; SKIP-CACHE-INV-NEXT: s_endpgm
10482 ;
10483 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_load:
10484 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
10485 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10486 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10487 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10488 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
10489 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10490 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10491 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10492 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10493 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10494 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
10495 ;
10496 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_load:
10497 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
10498 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10499 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10500 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10501 ; GFX90A-TGSPLIT-NEXT: s_nop 0
10502 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10503 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10504 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10505 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10506 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10507 ; GFX90A-TGSPLIT-NEXT: s_endpgm
10508 ;
10509 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_load:
10510 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
10511 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10512 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10513 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10514 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
10515 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10516 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10517 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10518 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10519 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10520 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
10521 ;
10522 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_load:
10523 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
10524 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10525 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10526 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10527 ; GFX940-TGSPLIT-NEXT: s_nop 0
10528 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10529 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10530 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10531 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10532 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10533 ; GFX940-TGSPLIT-NEXT: s_endpgm
10534 ;
10535 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_load:
10536 ; GFX11-WGP: ; %bb.0: ; %entry
10537 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10538 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
10539 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10540 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10541 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
10542 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10543 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
10544 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10545 ; GFX11-WGP-NEXT: s_endpgm
10546 ;
10547 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_load:
10548 ; GFX11-CU: ; %bb.0: ; %entry
10549 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10550 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
10551 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10552 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10553 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
10554 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10555 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
10556 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10557 ; GFX11-CU-NEXT: s_endpgm
10558 ;
10559 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_load:
10560 ; GFX12-WGP: ; %bb.0: ; %entry
10561 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10562 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
10563 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10564 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10565 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
10566 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10567 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
10568 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10569 ; GFX12-WGP-NEXT: s_endpgm
10570 ;
10571 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_load:
10572 ; GFX12-CU: ; %bb.0: ; %entry
10573 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10574 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
10575 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10576 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10577 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
10578 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10579 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
10580 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10581 ; GFX12-CU-NEXT: s_endpgm
10582 ptr addrspace(1) %in, ptr addrspace(1) %out) {
10583 entry:
10584 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread-one-as") acquire, align 4
10585 store i32 %val, ptr addrspace(1) %out
10586 ret void
10587 }
10589 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_load(
10590 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_load:
10591 ; GFX6: ; %bb.0: ; %entry
10592 ; GFX6-NEXT: s_load_dwordx2 s[8:9], s[6:7], 0x0
10593 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10594 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10595 ; GFX6-NEXT: s_mov_b32 s6, s9
10596 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 killed $sgpr8_sgpr9
10597 ; GFX6-NEXT: s_mov_b32 s12, 0x100f000
10598 ; GFX6-NEXT: s_mov_b32 s13, -1
10599 ; GFX6-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
10600 ; GFX6-NEXT: s_mov_b32 s9, s6
10601 ; GFX6-NEXT: s_mov_b32 s10, s13
10602 ; GFX6-NEXT: s_mov_b32 s11, s12
10603 ; GFX6-NEXT: s_mov_b32 s14, s5
10604 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10605 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10606 ; GFX6-NEXT: s_mov_b32 s5, s14
10607 ; GFX6-NEXT: s_mov_b32 s6, s13
10608 ; GFX6-NEXT: s_mov_b32 s7, s12
10609 ; GFX6-NEXT: buffer_load_dword v0, off, s[8:11], 0
10610 ; GFX6-NEXT: s_waitcnt vmcnt(0)
10611 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10612 ; GFX6-NEXT: s_endpgm
10613 ;
10614 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_load:
10615 ; GFX7: ; %bb.0: ; %entry
10616 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
10617 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10618 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x2
10619 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10620 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10621 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10622 ; GFX7-NEXT: flat_load_dword v2, v[0:1]
10623 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
10624 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
10625 ; GFX7-NEXT: s_waitcnt vmcnt(0)
10626 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10627 ; GFX7-NEXT: s_endpgm
10628 ;
10629 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_load:
10630 ; GFX10-WGP: ; %bb.0: ; %entry
10631 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10632 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10633 ; GFX10-WGP-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10634 ; GFX10-WGP-NEXT: s_nop 0
10635 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10636 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10637 ; GFX10-WGP-NEXT: global_load_dword v1, v0, s[6:7]
10638 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
10639 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10640 ; GFX10-WGP-NEXT: s_endpgm
10641 ;
10642 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_load:
10643 ; GFX10-CU: ; %bb.0: ; %entry
10644 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10645 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
10646 ; GFX10-CU-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10647 ; GFX10-CU-NEXT: s_nop 0
10648 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10649 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
10650 ; GFX10-CU-NEXT: global_load_dword v1, v0, s[6:7]
10651 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
10652 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
10653 ; GFX10-CU-NEXT: s_endpgm
10654 ;
10655 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_load:
10656 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
10657 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
10658 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
10659 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
10660 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s5
10661 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10662 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, 0xf000
10663 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s9, -1
10664 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10665 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, s2
10666 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, s9
10667 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s8
10668 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s10, s1
10669 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
10670 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
10671 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s10
10672 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s9
10673 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s8
10674 ; SKIP-CACHE-INV-NEXT: buffer_load_dword v0, off, s[4:7], 0
10675 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
10676 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
10677 ; SKIP-CACHE-INV-NEXT: s_endpgm
10678 ;
10679 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_load:
10680 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
10681 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10682 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10683 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10684 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
10685 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10686 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10687 ; GFX90A-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10688 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10689 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10690 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
10691 ;
10692 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_load:
10693 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
10694 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10695 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10696 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
10697 ; GFX90A-TGSPLIT-NEXT: s_nop 0
10698 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10699 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10700 ; GFX90A-TGSPLIT-NEXT: global_load_dword v1, v0, s[6:7]
10701 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10702 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10703 ; GFX90A-TGSPLIT-NEXT: s_endpgm
10704 ;
10705 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_load:
10706 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
10707 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10708 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10709 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10710 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
10711 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10712 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10713 ; GFX940-NOTTGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10714 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
10715 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10716 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
10717 ;
10718 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_load:
10719 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
10720 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10721 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10722 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
10723 ; GFX940-TGSPLIT-NEXT: s_nop 0
10724 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10725 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10726 ; GFX940-TGSPLIT-NEXT: global_load_dword v1, v0, s[2:3]
10727 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
10728 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10729 ; GFX940-TGSPLIT-NEXT: s_endpgm
10730 ;
10731 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_load:
10732 ; GFX11-WGP: ; %bb.0: ; %entry
10733 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10734 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
10735 ; GFX11-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10736 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10737 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
10738 ; GFX11-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10739 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
10740 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10741 ; GFX11-WGP-NEXT: s_endpgm
10742 ;
10743 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_load:
10744 ; GFX11-CU: ; %bb.0: ; %entry
10745 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10746 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
10747 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10748 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10749 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
10750 ; GFX11-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10751 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
10752 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10753 ; GFX11-CU-NEXT: s_endpgm
10754 ;
10755 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_load:
10756 ; GFX12-WGP: ; %bb.0: ; %entry
10757 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10758 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
10759 ; GFX12-WGP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10760 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10761 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
10762 ; GFX12-WGP-NEXT: global_load_b32 v1, v0, s[2:3]
10763 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
10764 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10765 ; GFX12-WGP-NEXT: s_endpgm
10766 ;
10767 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_load:
10768 ; GFX12-CU: ; %bb.0: ; %entry
10769 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10770 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
10771 ; GFX12-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
10772 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10773 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
10774 ; GFX12-CU-NEXT: global_load_b32 v1, v0, s[2:3]
10775 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
10776 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10777 ; GFX12-CU-NEXT: s_endpgm
10778 ptr addrspace(1) %in, ptr addrspace(1) %out) {
10779 entry:
10780 %val = load atomic i32, ptr addrspace(1) %in syncscope("singlethread-one-as") seq_cst, align 4
10781 store i32 %val, ptr addrspace(1) %out
10782 ret void
10783 }
10785 define amdgpu_kernel void @global_singlethread_one_as_unordered_store(
10786 ; GFX6-LABEL: global_singlethread_one_as_unordered_store:
10787 ; GFX6: ; %bb.0: ; %entry
10788 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
10789 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10790 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10791 ; GFX6-NEXT: s_mov_b32 s11, s5
10792 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10793 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
10794 ; GFX6-NEXT: s_mov_b32 s10, -1
10795 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10796 ; GFX6-NEXT: s_mov_b32 s5, s11
10797 ; GFX6-NEXT: s_mov_b32 s6, s10
10798 ; GFX6-NEXT: s_mov_b32 s7, s9
10799 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
10800 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10801 ; GFX6-NEXT: s_endpgm
10802 ;
10803 ; GFX7-LABEL: global_singlethread_one_as_unordered_store:
10804 ; GFX7: ; %bb.0: ; %entry
10805 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
10806 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
10807 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10808 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10809 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10810 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
10811 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10812 ; GFX7-NEXT: s_endpgm
10813 ;
10814 ; GFX10-WGP-LABEL: global_singlethread_one_as_unordered_store:
10815 ; GFX10-WGP: ; %bb.0: ; %entry
10816 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10817 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
10818 ; GFX10-WGP-NEXT: s_nop 0
10819 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10820 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10821 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10822 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
10823 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10824 ; GFX10-WGP-NEXT: s_endpgm
10825 ;
10826 ; GFX10-CU-LABEL: global_singlethread_one_as_unordered_store:
10827 ; GFX10-CU: ; %bb.0: ; %entry
10828 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10829 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
10830 ; GFX10-CU-NEXT: s_nop 0
10831 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10832 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
10833 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
10834 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
10835 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
10836 ; GFX10-CU-NEXT: s_endpgm
10837 ;
10838 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_unordered_store:
10839 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
10840 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
10841 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
10842 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
10843 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
10844 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
10845 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
10846 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
10847 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
10848 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
10849 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
10850 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
10851 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
10852 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
10853 ; SKIP-CACHE-INV-NEXT: s_endpgm
10854 ;
10855 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_unordered_store:
10856 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
10857 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10858 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
10859 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
10860 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10861 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10862 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10863 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
10864 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10865 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
10866 ;
10867 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_unordered_store:
10868 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
10869 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
10870 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
10871 ; GFX90A-TGSPLIT-NEXT: s_nop 0
10872 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10873 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10874 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10875 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
10876 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
10877 ; GFX90A-TGSPLIT-NEXT: s_endpgm
10878 ;
10879 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_unordered_store:
10880 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
10881 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10882 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
10883 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
10884 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10885 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10886 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10887 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
10888 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10889 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
10890 ;
10891 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_unordered_store:
10892 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
10893 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
10894 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
10895 ; GFX940-TGSPLIT-NEXT: s_nop 0
10896 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
10897 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
10898 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
10899 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
10900 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
10901 ; GFX940-TGSPLIT-NEXT: s_endpgm
10902 ;
10903 ; GFX11-WGP-LABEL: global_singlethread_one_as_unordered_store:
10904 ; GFX11-WGP: ; %bb.0: ; %entry
10905 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10906 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
10907 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10908 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
10909 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
10910 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
10911 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10912 ; GFX11-WGP-NEXT: s_endpgm
10913 ;
10914 ; GFX11-CU-LABEL: global_singlethread_one_as_unordered_store:
10915 ; GFX11-CU: ; %bb.0: ; %entry
10916 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10917 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
10918 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10919 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
10920 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
10921 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
10922 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10923 ; GFX11-CU-NEXT: s_endpgm
10924 ;
10925 ; GFX12-WGP-LABEL: global_singlethread_one_as_unordered_store:
10926 ; GFX12-WGP: ; %bb.0: ; %entry
10927 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
10928 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
10929 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10930 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
10931 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
10932 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
10933 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
10934 ; GFX12-WGP-NEXT: s_endpgm
10935 ;
10936 ; GFX12-CU-LABEL: global_singlethread_one_as_unordered_store:
10937 ; GFX12-CU: ; %bb.0: ; %entry
10938 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
10939 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
10940 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
10941 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
10942 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
10943 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
10944 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
10945 ; GFX12-CU-NEXT: s_endpgm
10946 i32 %in, ptr addrspace(1) %out) {
10947 entry:
10948 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread-one-as") unordered, align 4
10949 ret void
10950 }
10952 define amdgpu_kernel void @global_singlethread_one_as_monotonic_store(
10953 ; GFX6-LABEL: global_singlethread_one_as_monotonic_store:
10954 ; GFX6: ; %bb.0: ; %entry
10955 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
10956 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
10957 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
10958 ; GFX6-NEXT: s_mov_b32 s11, s5
10959 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
10960 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
10961 ; GFX6-NEXT: s_mov_b32 s10, -1
10962 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
10963 ; GFX6-NEXT: s_mov_b32 s5, s11
10964 ; GFX6-NEXT: s_mov_b32 s6, s10
10965 ; GFX6-NEXT: s_mov_b32 s7, s9
10966 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
10967 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
10968 ; GFX6-NEXT: s_endpgm
10969 ;
10970 ; GFX7-LABEL: global_singlethread_one_as_monotonic_store:
10971 ; GFX7: ; %bb.0: ; %entry
10972 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
10973 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
10974 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
10975 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
10976 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
10977 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
10978 ; GFX7-NEXT: flat_store_dword v[0:1], v2
10979 ; GFX7-NEXT: s_endpgm
10980 ;
10981 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_store:
10982 ; GFX10-WGP: ; %bb.0: ; %entry
10983 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
10984 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
10985 ; GFX10-WGP-NEXT: s_nop 0
10986 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10987 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
10988 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
10989 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
10990 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
10991 ; GFX10-WGP-NEXT: s_endpgm
10992 ;
10993 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_store:
10994 ; GFX10-CU: ; %bb.0: ; %entry
10995 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
10996 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
10997 ; GFX10-CU-NEXT: s_nop 0
10998 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
10999 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11000 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11001 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11002 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
11003 ; GFX10-CU-NEXT: s_endpgm
11004 ;
11005 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_store:
11006 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11007 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
11008 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
11009 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11010 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11011 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11012 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11013 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11014 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11015 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11016 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11017 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11018 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11019 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
11020 ; SKIP-CACHE-INV-NEXT: s_endpgm
11021 ;
11022 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_store:
11023 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11024 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11025 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11026 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11027 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11028 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11029 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11030 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11031 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11032 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11033 ;
11034 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_store:
11035 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11036 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11037 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11038 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11039 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11040 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11041 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11042 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11043 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11044 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11045 ;
11046 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_store:
11047 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11048 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11049 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11050 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11051 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11052 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11053 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11054 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11055 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11056 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11057 ;
11058 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_store:
11059 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11060 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11061 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11062 ; GFX940-TGSPLIT-NEXT: s_nop 0
11063 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11064 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11065 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11066 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11067 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11068 ; GFX940-TGSPLIT-NEXT: s_endpgm
11069 ;
11070 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_store:
11071 ; GFX11-WGP: ; %bb.0: ; %entry
11072 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11073 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11074 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11075 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11076 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11077 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11078 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11079 ; GFX11-WGP-NEXT: s_endpgm
11080 ;
11081 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_store:
11082 ; GFX11-CU: ; %bb.0: ; %entry
11083 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11084 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11085 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11086 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11087 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11088 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11089 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11090 ; GFX11-CU-NEXT: s_endpgm
11091 ;
11092 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_store:
11093 ; GFX12-WGP: ; %bb.0: ; %entry
11094 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11095 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11096 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11097 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11098 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11099 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11100 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11101 ; GFX12-WGP-NEXT: s_endpgm
11102 ;
11103 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_store:
11104 ; GFX12-CU: ; %bb.0: ; %entry
11105 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11106 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11107 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11108 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11109 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11110 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11111 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11112 ; GFX12-CU-NEXT: s_endpgm
11113 i32 %in, ptr addrspace(1) %out) {
11114 entry:
11115 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread-one-as") monotonic, align 4
11116 ret void
11117 }
11119 define amdgpu_kernel void @global_singlethread_one_as_release_store(
11120 ; GFX6-LABEL: global_singlethread_one_as_release_store:
11121 ; GFX6: ; %bb.0: ; %entry
11122 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
11123 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
11124 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11125 ; GFX6-NEXT: s_mov_b32 s11, s5
11126 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11127 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11128 ; GFX6-NEXT: s_mov_b32 s10, -1
11129 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11130 ; GFX6-NEXT: s_mov_b32 s5, s11
11131 ; GFX6-NEXT: s_mov_b32 s6, s10
11132 ; GFX6-NEXT: s_mov_b32 s7, s9
11133 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11134 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
11135 ; GFX6-NEXT: s_endpgm
11136 ;
11137 ; GFX7-LABEL: global_singlethread_one_as_release_store:
11138 ; GFX7: ; %bb.0: ; %entry
11139 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
11140 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
11141 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11142 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11143 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11144 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11145 ; GFX7-NEXT: flat_store_dword v[0:1], v2
11146 ; GFX7-NEXT: s_endpgm
11147 ;
11148 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_store:
11149 ; GFX10-WGP: ; %bb.0: ; %entry
11150 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
11151 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
11152 ; GFX10-WGP-NEXT: s_nop 0
11153 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11154 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11155 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11156 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11157 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
11158 ; GFX10-WGP-NEXT: s_endpgm
11159 ;
11160 ; GFX10-CU-LABEL: global_singlethread_one_as_release_store:
11161 ; GFX10-CU: ; %bb.0: ; %entry
11162 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
11163 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
11164 ; GFX10-CU-NEXT: s_nop 0
11165 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11166 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11167 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11168 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11169 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
11170 ; GFX10-CU-NEXT: s_endpgm
11171 ;
11172 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_store:
11173 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11174 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
11175 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
11176 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11177 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11178 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11179 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11180 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11181 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11182 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11183 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11184 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11185 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11186 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
11187 ; SKIP-CACHE-INV-NEXT: s_endpgm
11188 ;
11189 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_store:
11190 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11191 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11192 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11193 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11194 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11195 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11196 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11197 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11198 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11199 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11200 ;
11201 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_store:
11202 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11203 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11204 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11205 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11206 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11207 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11208 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11209 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11210 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11211 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11212 ;
11213 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_store:
11214 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11215 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11216 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11217 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11218 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11219 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11220 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11221 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11222 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11223 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11224 ;
11225 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_store:
11226 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11227 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11228 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11229 ; GFX940-TGSPLIT-NEXT: s_nop 0
11230 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11231 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11232 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11233 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11234 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11235 ; GFX940-TGSPLIT-NEXT: s_endpgm
11236 ;
11237 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_store:
11238 ; GFX11-WGP: ; %bb.0: ; %entry
11239 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11240 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11241 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11242 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11243 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11244 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11245 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11246 ; GFX11-WGP-NEXT: s_endpgm
11247 ;
11248 ; GFX11-CU-LABEL: global_singlethread_one_as_release_store:
11249 ; GFX11-CU: ; %bb.0: ; %entry
11250 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11251 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11252 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11253 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11254 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11255 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11256 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11257 ; GFX11-CU-NEXT: s_endpgm
11258 ;
11259 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_store:
11260 ; GFX12-WGP: ; %bb.0: ; %entry
11261 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11262 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11263 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11264 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11265 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11266 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11267 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11268 ; GFX12-WGP-NEXT: s_endpgm
11269 ;
11270 ; GFX12-CU-LABEL: global_singlethread_one_as_release_store:
11271 ; GFX12-CU: ; %bb.0: ; %entry
11272 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11273 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11274 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11275 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11276 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11277 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11278 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11279 ; GFX12-CU-NEXT: s_endpgm
11280 i32 %in, ptr addrspace(1) %out) {
11281 entry:
11282 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread-one-as") release, align 4
11283 ret void
11284 }
11286 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_store(
11287 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_store:
11288 ; GFX6: ; %bb.0: ; %entry
11289 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x0
11290 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x2
11291 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11292 ; GFX6-NEXT: s_mov_b32 s11, s5
11293 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11294 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11295 ; GFX6-NEXT: s_mov_b32 s10, -1
11296 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11297 ; GFX6-NEXT: s_mov_b32 s5, s11
11298 ; GFX6-NEXT: s_mov_b32 s6, s10
11299 ; GFX6-NEXT: s_mov_b32 s7, s9
11300 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11301 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
11302 ; GFX6-NEXT: s_endpgm
11303 ;
11304 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_store:
11305 ; GFX7: ; %bb.0: ; %entry
11306 ; GFX7-NEXT: s_load_dword s4, s[6:7], 0x0
11307 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x2
11308 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11309 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11310 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11311 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11312 ; GFX7-NEXT: flat_store_dword v[0:1], v2
11313 ; GFX7-NEXT: s_endpgm
11314 ;
11315 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_store:
11316 ; GFX10-WGP: ; %bb.0: ; %entry
11317 ; GFX10-WGP-NEXT: s_mov_b64 s[4:5], s[6:7]
11318 ; GFX10-WGP-NEXT: s_load_dword s6, s[4:5], 0x0
11319 ; GFX10-WGP-NEXT: s_nop 0
11320 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11321 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11322 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11323 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11324 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
11325 ; GFX10-WGP-NEXT: s_endpgm
11326 ;
11327 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_store:
11328 ; GFX10-CU: ; %bb.0: ; %entry
11329 ; GFX10-CU-NEXT: s_mov_b64 s[4:5], s[6:7]
11330 ; GFX10-CU-NEXT: s_load_dword s6, s[4:5], 0x0
11331 ; GFX10-CU-NEXT: s_nop 0
11332 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11333 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11334 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11335 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11336 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
11337 ; GFX10-CU-NEXT: s_endpgm
11338 ;
11339 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_store:
11340 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11341 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x0
11342 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x2
11343 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11344 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11345 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11346 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11347 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11348 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11349 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11350 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11351 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11352 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11353 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
11354 ; SKIP-CACHE-INV-NEXT: s_endpgm
11355 ;
11356 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_store:
11357 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11358 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11359 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11360 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11361 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11362 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11363 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11364 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11365 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11366 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11367 ;
11368 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_store:
11369 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11370 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[6:7]
11371 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[4:5], 0x0
11372 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11373 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x8
11374 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11375 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11376 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11377 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
11378 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11379 ;
11380 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_store:
11381 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11382 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11383 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11384 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11385 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11386 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11387 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11388 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11389 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11390 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11391 ;
11392 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_store:
11393 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11394 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[0:1], s[2:3]
11395 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[0:1], 0x0
11396 ; GFX940-TGSPLIT-NEXT: s_nop 0
11397 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x8
11398 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11399 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11400 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11401 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
11402 ; GFX940-TGSPLIT-NEXT: s_endpgm
11403 ;
11404 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_store:
11405 ; GFX11-WGP: ; %bb.0: ; %entry
11406 ; GFX11-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11407 ; GFX11-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11408 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11409 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11410 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11411 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11412 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11413 ; GFX11-WGP-NEXT: s_endpgm
11414 ;
11415 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_store:
11416 ; GFX11-CU: ; %bb.0: ; %entry
11417 ; GFX11-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11418 ; GFX11-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11419 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11420 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11421 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11422 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11423 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11424 ; GFX11-CU-NEXT: s_endpgm
11425 ;
11426 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_store:
11427 ; GFX12-WGP: ; %bb.0: ; %entry
11428 ; GFX12-WGP-NEXT: s_mov_b64 s[0:1], s[2:3]
11429 ; GFX12-WGP-NEXT: s_load_b32 s2, s[0:1], 0x0
11430 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11431 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11432 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11433 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11434 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
11435 ; GFX12-WGP-NEXT: s_endpgm
11436 ;
11437 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_store:
11438 ; GFX12-CU: ; %bb.0: ; %entry
11439 ; GFX12-CU-NEXT: s_mov_b64 s[0:1], s[2:3]
11440 ; GFX12-CU-NEXT: s_load_b32 s2, s[0:1], 0x0
11441 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[0:1], 0x8
11442 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11443 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11444 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11445 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
11446 ; GFX12-CU-NEXT: s_endpgm
11447 i32 %in, ptr addrspace(1) %out) {
11448 entry:
11449 store atomic i32 %in, ptr addrspace(1) %out syncscope("singlethread-one-as") seq_cst, align 4
11450 ret void
11451 }
11453 define amdgpu_kernel void @global_singlethread_one_as_monotonic_atomicrmw(
11454 ; GFX6-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11455 ; GFX6: ; %bb.0: ; %entry
11456 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11457 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
11458 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11459 ; GFX6-NEXT: s_mov_b32 s11, s5
11460 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11461 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11462 ; GFX6-NEXT: s_mov_b32 s10, -1
11463 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11464 ; GFX6-NEXT: s_mov_b32 s5, s11
11465 ; GFX6-NEXT: s_mov_b32 s6, s10
11466 ; GFX6-NEXT: s_mov_b32 s7, s9
11467 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11468 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
11469 ; GFX6-NEXT: s_endpgm
11470 ;
11471 ; GFX7-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11472 ; GFX7: ; %bb.0: ; %entry
11473 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
11474 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
11475 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
11476 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11477 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11478 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11479 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11480 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
11481 ; GFX7-NEXT: s_endpgm
11482 ;
11483 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11484 ; GFX10-WGP: ; %bb.0: ; %entry
11485 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11486 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11487 ; GFX10-WGP-NEXT: s_nop 0
11488 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
11489 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11490 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11491 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
11492 ; GFX10-WGP-NEXT: s_endpgm
11493 ;
11494 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11495 ; GFX10-CU: ; %bb.0: ; %entry
11496 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11497 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11498 ; GFX10-CU-NEXT: s_nop 0
11499 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
11500 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11501 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11502 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
11503 ; GFX10-CU-NEXT: s_endpgm
11504 ;
11505 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11506 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11507 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11508 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
11509 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11510 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11511 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11512 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11513 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11514 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11515 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11516 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11517 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11518 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11519 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
11520 ; SKIP-CACHE-INV-NEXT: s_endpgm
11521 ;
11522 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11523 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11524 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11525 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11526 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11527 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11528 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11529 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11530 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11531 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11532 ;
11533 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11534 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11535 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11536 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11537 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11538 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11539 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11540 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11541 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11542 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11543 ;
11544 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11545 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11546 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11547 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11548 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11549 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11550 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11551 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11552 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11553 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11554 ;
11555 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11556 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11557 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11558 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11559 ; GFX940-TGSPLIT-NEXT: s_nop 0
11560 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11561 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11562 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11563 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11564 ; GFX940-TGSPLIT-NEXT: s_endpgm
11565 ;
11566 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11567 ; GFX11-WGP: ; %bb.0: ; %entry
11568 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11569 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11570 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11571 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11572 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11573 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11574 ; GFX11-WGP-NEXT: s_endpgm
11575 ;
11576 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11577 ; GFX11-CU: ; %bb.0: ; %entry
11578 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11579 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11580 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11581 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11582 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11583 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11584 ; GFX11-CU-NEXT: s_endpgm
11585 ;
11586 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11587 ; GFX12-WGP: ; %bb.0: ; %entry
11588 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11589 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11590 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11591 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11592 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11593 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11594 ; GFX12-WGP-NEXT: s_endpgm
11595 ;
11596 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
11597 ; GFX12-CU: ; %bb.0: ; %entry
11598 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11599 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11600 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11601 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11602 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11603 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11604 ; GFX12-CU-NEXT: s_endpgm
11605 ptr addrspace(1) %out, i32 %in) {
11606 entry:
11607 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") monotonic
11608 ret void
11609 }
11611 define amdgpu_kernel void @global_singlethread_one_as_acquire_atomicrmw(
11612 ; GFX6-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11613 ; GFX6: ; %bb.0: ; %entry
11614 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11615 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
11616 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11617 ; GFX6-NEXT: s_mov_b32 s11, s5
11618 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11619 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11620 ; GFX6-NEXT: s_mov_b32 s10, -1
11621 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11622 ; GFX6-NEXT: s_mov_b32 s5, s11
11623 ; GFX6-NEXT: s_mov_b32 s6, s10
11624 ; GFX6-NEXT: s_mov_b32 s7, s9
11625 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11626 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
11627 ; GFX6-NEXT: s_endpgm
11628 ;
11629 ; GFX7-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11630 ; GFX7: ; %bb.0: ; %entry
11631 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
11632 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
11633 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
11634 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11635 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11636 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11637 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11638 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
11639 ; GFX7-NEXT: s_endpgm
11640 ;
11641 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11642 ; GFX10-WGP: ; %bb.0: ; %entry
11643 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11644 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11645 ; GFX10-WGP-NEXT: s_nop 0
11646 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
11647 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11648 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11649 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
11650 ; GFX10-WGP-NEXT: s_endpgm
11651 ;
11652 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11653 ; GFX10-CU: ; %bb.0: ; %entry
11654 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11655 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11656 ; GFX10-CU-NEXT: s_nop 0
11657 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
11658 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11659 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11660 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
11661 ; GFX10-CU-NEXT: s_endpgm
11662 ;
11663 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11664 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11665 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11666 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
11667 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11668 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11669 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11670 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11671 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11672 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11673 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11674 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11675 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11676 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11677 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
11678 ; SKIP-CACHE-INV-NEXT: s_endpgm
11679 ;
11680 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11681 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11682 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11683 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11684 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11685 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11686 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11687 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11688 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11689 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11690 ;
11691 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11692 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11693 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11694 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11695 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11696 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11697 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11698 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11699 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11700 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11701 ;
11702 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11703 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11704 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11705 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11706 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11707 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11708 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11709 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11710 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11711 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11712 ;
11713 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11714 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11715 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11716 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11717 ; GFX940-TGSPLIT-NEXT: s_nop 0
11718 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11719 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11720 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11721 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11722 ; GFX940-TGSPLIT-NEXT: s_endpgm
11723 ;
11724 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11725 ; GFX11-WGP: ; %bb.0: ; %entry
11726 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11727 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11728 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11729 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11730 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11731 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11732 ; GFX11-WGP-NEXT: s_endpgm
11733 ;
11734 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11735 ; GFX11-CU: ; %bb.0: ; %entry
11736 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11737 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11738 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11739 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11740 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11741 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11742 ; GFX11-CU-NEXT: s_endpgm
11743 ;
11744 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11745 ; GFX12-WGP: ; %bb.0: ; %entry
11746 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11747 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11748 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11749 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11750 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11751 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11752 ; GFX12-WGP-NEXT: s_endpgm
11753 ;
11754 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_atomicrmw:
11755 ; GFX12-CU: ; %bb.0: ; %entry
11756 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11757 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11758 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11759 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11760 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11761 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11762 ; GFX12-CU-NEXT: s_endpgm
11763 ptr addrspace(1) %out, i32 %in) {
11764 entry:
11765 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") acquire
11766 ret void
11767 }
11769 define amdgpu_kernel void @global_singlethread_one_as_release_atomicrmw(
11770 ; GFX6-LABEL: global_singlethread_one_as_release_atomicrmw:
11771 ; GFX6: ; %bb.0: ; %entry
11772 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11773 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
11774 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11775 ; GFX6-NEXT: s_mov_b32 s11, s5
11776 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11777 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11778 ; GFX6-NEXT: s_mov_b32 s10, -1
11779 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11780 ; GFX6-NEXT: s_mov_b32 s5, s11
11781 ; GFX6-NEXT: s_mov_b32 s6, s10
11782 ; GFX6-NEXT: s_mov_b32 s7, s9
11783 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11784 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
11785 ; GFX6-NEXT: s_endpgm
11786 ;
11787 ; GFX7-LABEL: global_singlethread_one_as_release_atomicrmw:
11788 ; GFX7: ; %bb.0: ; %entry
11789 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
11790 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
11791 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
11792 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11793 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11794 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11795 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11796 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
11797 ; GFX7-NEXT: s_endpgm
11798 ;
11799 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_atomicrmw:
11800 ; GFX10-WGP: ; %bb.0: ; %entry
11801 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11802 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11803 ; GFX10-WGP-NEXT: s_nop 0
11804 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
11805 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11806 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11807 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
11808 ; GFX10-WGP-NEXT: s_endpgm
11809 ;
11810 ; GFX10-CU-LABEL: global_singlethread_one_as_release_atomicrmw:
11811 ; GFX10-CU: ; %bb.0: ; %entry
11812 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11813 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11814 ; GFX10-CU-NEXT: s_nop 0
11815 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
11816 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11817 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11818 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
11819 ; GFX10-CU-NEXT: s_endpgm
11820 ;
11821 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_atomicrmw:
11822 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11823 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11824 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
11825 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11826 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11827 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11828 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11829 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11830 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11831 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11832 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11833 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11834 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11835 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
11836 ; SKIP-CACHE-INV-NEXT: s_endpgm
11837 ;
11838 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_atomicrmw:
11839 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11840 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11841 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11842 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
11843 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11844 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11845 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11846 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11847 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
11848 ;
11849 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_atomicrmw:
11850 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
11851 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11852 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11853 ; GFX90A-TGSPLIT-NEXT: s_nop 0
11854 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
11855 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11856 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
11857 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
11858 ; GFX90A-TGSPLIT-NEXT: s_endpgm
11859 ;
11860 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_atomicrmw:
11861 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
11862 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11863 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11864 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
11865 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11866 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11867 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11868 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11869 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
11870 ;
11871 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_atomicrmw:
11872 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
11873 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11874 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11875 ; GFX940-TGSPLIT-NEXT: s_nop 0
11876 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
11877 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
11878 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
11879 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
11880 ; GFX940-TGSPLIT-NEXT: s_endpgm
11881 ;
11882 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_atomicrmw:
11883 ; GFX11-WGP: ; %bb.0: ; %entry
11884 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
11885 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11886 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11887 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
11888 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
11889 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11890 ; GFX11-WGP-NEXT: s_endpgm
11891 ;
11892 ; GFX11-CU-LABEL: global_singlethread_one_as_release_atomicrmw:
11893 ; GFX11-CU: ; %bb.0: ; %entry
11894 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
11895 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11896 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11897 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
11898 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
11899 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11900 ; GFX11-CU-NEXT: s_endpgm
11901 ;
11902 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_atomicrmw:
11903 ; GFX12-WGP: ; %bb.0: ; %entry
11904 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
11905 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11906 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
11907 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
11908 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
11909 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11910 ; GFX12-WGP-NEXT: s_endpgm
11911 ;
11912 ; GFX12-CU-LABEL: global_singlethread_one_as_release_atomicrmw:
11913 ; GFX12-CU: ; %bb.0: ; %entry
11914 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
11915 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
11916 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
11917 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
11918 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
11919 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
11920 ; GFX12-CU-NEXT: s_endpgm
11921 ptr addrspace(1) %out, i32 %in) {
11922 entry:
11923 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") release
11924 ret void
11925 }
11927 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_atomicrmw(
11928 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11929 ; GFX6: ; %bb.0: ; %entry
11930 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11931 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
11932 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
11933 ; GFX6-NEXT: s_mov_b32 s11, s5
11934 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
11935 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
11936 ; GFX6-NEXT: s_mov_b32 s10, -1
11937 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
11938 ; GFX6-NEXT: s_mov_b32 s5, s11
11939 ; GFX6-NEXT: s_mov_b32 s6, s10
11940 ; GFX6-NEXT: s_mov_b32 s7, s9
11941 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
11942 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
11943 ; GFX6-NEXT: s_endpgm
11944 ;
11945 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11946 ; GFX7: ; %bb.0: ; %entry
11947 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
11948 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
11949 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
11950 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
11951 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
11952 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
11953 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
11954 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
11955 ; GFX7-NEXT: s_endpgm
11956 ;
11957 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11958 ; GFX10-WGP: ; %bb.0: ; %entry
11959 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
11960 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11961 ; GFX10-WGP-NEXT: s_nop 0
11962 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
11963 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
11964 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
11965 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
11966 ; GFX10-WGP-NEXT: s_endpgm
11967 ;
11968 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11969 ; GFX10-CU: ; %bb.0: ; %entry
11970 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
11971 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
11972 ; GFX10-CU-NEXT: s_nop 0
11973 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
11974 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
11975 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
11976 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
11977 ; GFX10-CU-NEXT: s_endpgm
11978 ;
11979 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11980 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
11981 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
11982 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
11983 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
11984 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
11985 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
11986 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
11987 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
11988 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
11989 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
11990 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
11991 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
11992 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
11993 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
11994 ; SKIP-CACHE-INV-NEXT: s_endpgm
11995 ;
11996 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
11997 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
11998 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
11999 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12000 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
12001 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12002 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12003 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12004 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
12005 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12006 ;
12007 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12008 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12009 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12010 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12011 ; GFX90A-TGSPLIT-NEXT: s_nop 0
12012 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12013 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12014 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12015 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
12016 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12017 ;
12018 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12019 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12020 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12021 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12022 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
12023 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12024 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12025 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12026 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
12027 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12028 ;
12029 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12030 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12031 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12032 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12033 ; GFX940-TGSPLIT-NEXT: s_nop 0
12034 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12035 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12036 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12037 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
12038 ; GFX940-TGSPLIT-NEXT: s_endpgm
12039 ;
12040 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12041 ; GFX11-WGP: ; %bb.0: ; %entry
12042 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12043 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12044 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12045 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12046 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
12047 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12048 ; GFX11-WGP-NEXT: s_endpgm
12049 ;
12050 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12051 ; GFX11-CU: ; %bb.0: ; %entry
12052 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12053 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12054 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12055 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12056 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
12057 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12058 ; GFX11-CU-NEXT: s_endpgm
12059 ;
12060 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12061 ; GFX12-WGP: ; %bb.0: ; %entry
12062 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12063 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12064 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12065 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12066 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
12067 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12068 ; GFX12-WGP-NEXT: s_endpgm
12069 ;
12070 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
12071 ; GFX12-CU: ; %bb.0: ; %entry
12072 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
12073 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12074 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12075 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
12076 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
12077 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12078 ; GFX12-CU-NEXT: s_endpgm
12079 ptr addrspace(1) %out, i32 %in) {
12080 entry:
12081 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") acq_rel
12082 ret void
12083 }
12085 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_atomicrmw(
12086 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12087 ; GFX6: ; %bb.0: ; %entry
12088 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12089 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
12090 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
12091 ; GFX6-NEXT: s_mov_b32 s11, s5
12092 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
12093 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
12094 ; GFX6-NEXT: s_mov_b32 s10, -1
12095 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
12096 ; GFX6-NEXT: s_mov_b32 s5, s11
12097 ; GFX6-NEXT: s_mov_b32 s6, s10
12098 ; GFX6-NEXT: s_mov_b32 s7, s9
12099 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
12100 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0
12101 ; GFX6-NEXT: s_endpgm
12102 ;
12103 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12104 ; GFX7: ; %bb.0: ; %entry
12105 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
12106 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x0
12107 ; GFX7-NEXT: s_load_dword s4, s[4:5], 0x2
12108 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
12109 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
12110 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
12111 ; GFX7-NEXT: v_mov_b32_e32 v2, s4
12112 ; GFX7-NEXT: flat_atomic_swap v[0:1], v2
12113 ; GFX7-NEXT: s_endpgm
12114 ;
12115 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12116 ; GFX10-WGP: ; %bb.0: ; %entry
12117 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
12118 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12119 ; GFX10-WGP-NEXT: s_nop 0
12120 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
12121 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
12122 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
12123 ; GFX10-WGP-NEXT: global_atomic_swap v0, v1, s[4:5]
12124 ; GFX10-WGP-NEXT: s_endpgm
12125 ;
12126 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12127 ; GFX10-CU: ; %bb.0: ; %entry
12128 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
12129 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12130 ; GFX10-CU-NEXT: s_nop 0
12131 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
12132 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
12133 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
12134 ; GFX10-CU-NEXT: global_atomic_swap v0, v1, s[4:5]
12135 ; GFX10-CU-NEXT: s_endpgm
12136 ;
12137 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12138 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
12139 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12140 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
12141 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
12142 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
12143 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
12144 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
12145 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
12146 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
12147 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
12148 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
12149 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
12150 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
12151 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0
12152 ; SKIP-CACHE-INV-NEXT: s_endpgm
12153 ;
12154 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12155 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
12156 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12157 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12158 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
12159 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12160 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12161 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12162 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
12163 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12164 ;
12165 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12166 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12167 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12168 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12169 ; GFX90A-TGSPLIT-NEXT: s_nop 0
12170 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12171 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12172 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12173 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[4:5]
12174 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12175 ;
12176 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12177 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12178 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12179 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12180 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
12181 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12182 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12183 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12184 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
12185 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12186 ;
12187 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12188 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12189 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12190 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12191 ; GFX940-TGSPLIT-NEXT: s_nop 0
12192 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12193 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12194 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12195 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v0, v1, s[0:1]
12196 ; GFX940-TGSPLIT-NEXT: s_endpgm
12197 ;
12198 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12199 ; GFX11-WGP: ; %bb.0: ; %entry
12200 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12201 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12202 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12203 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12204 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
12205 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12206 ; GFX11-WGP-NEXT: s_endpgm
12207 ;
12208 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12209 ; GFX11-CU: ; %bb.0: ; %entry
12210 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12211 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12212 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12213 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12214 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
12215 ; GFX11-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12216 ; GFX11-CU-NEXT: s_endpgm
12217 ;
12218 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12219 ; GFX12-WGP: ; %bb.0: ; %entry
12220 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12221 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12222 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12223 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12224 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
12225 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12226 ; GFX12-WGP-NEXT: s_endpgm
12227 ;
12228 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
12229 ; GFX12-CU: ; %bb.0: ; %entry
12230 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
12231 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12232 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12233 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
12234 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
12235 ; GFX12-CU-NEXT: global_atomic_swap_b32 v0, v1, s[0:1]
12236 ; GFX12-CU-NEXT: s_endpgm
12237 ptr addrspace(1) %out, i32 %in) {
12238 entry:
12239 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") seq_cst
12240 ret void
12241 }
12243 define amdgpu_kernel void @global_singlethread_one_as_acquire_ret_atomicrmw(
12244 ; GFX6-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12245 ; GFX6: ; %bb.0: ; %entry
12246 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12247 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
12248 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
12249 ; GFX6-NEXT: s_mov_b32 s11, s5
12250 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
12251 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
12252 ; GFX6-NEXT: s_mov_b32 s10, -1
12253 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
12254 ; GFX6-NEXT: s_mov_b32 s5, s11
12255 ; GFX6-NEXT: s_mov_b32 s6, s10
12256 ; GFX6-NEXT: s_mov_b32 s7, s9
12257 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
12258 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
12259 ; GFX6-NEXT: s_waitcnt vmcnt(0)
12260 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
12261 ; GFX6-NEXT: s_endpgm
12262 ;
12263 ; GFX7-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12264 ; GFX7: ; %bb.0: ; %entry
12265 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12266 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
12267 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
12268 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12269 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12270 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
12271 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
12272 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12273 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12274 ; GFX7-NEXT: s_waitcnt vmcnt(0)
12275 ; GFX7-NEXT: flat_store_dword v[0:1], v2
12276 ; GFX7-NEXT: s_endpgm
12277 ;
12278 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12279 ; GFX10-WGP: ; %bb.0: ; %entry
12280 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
12281 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12282 ; GFX10-WGP-NEXT: s_nop 0
12283 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
12284 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
12285 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
12286 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12287 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
12288 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
12289 ; GFX10-WGP-NEXT: s_endpgm
12290 ;
12291 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12292 ; GFX10-CU: ; %bb.0: ; %entry
12293 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
12294 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12295 ; GFX10-CU-NEXT: s_nop 0
12296 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
12297 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
12298 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
12299 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12300 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
12301 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
12302 ; GFX10-CU-NEXT: s_endpgm
12303 ;
12304 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12305 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
12306 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12307 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
12308 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
12309 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
12310 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
12311 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
12312 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
12313 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
12314 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
12315 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
12316 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
12317 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
12318 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
12319 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
12320 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
12321 ; SKIP-CACHE-INV-NEXT: s_endpgm
12322 ;
12323 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12324 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
12325 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12326 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12327 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
12328 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12329 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12330 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12331 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12332 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12333 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12334 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12335 ;
12336 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12337 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12338 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12339 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12340 ; GFX90A-TGSPLIT-NEXT: s_nop 0
12341 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12342 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12343 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12344 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12345 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12346 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12347 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12348 ;
12349 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12350 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12351 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12352 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12353 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
12354 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12355 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12356 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12357 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12358 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12359 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12360 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12361 ;
12362 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12363 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12364 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12365 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12366 ; GFX940-TGSPLIT-NEXT: s_nop 0
12367 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12368 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12369 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12370 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12371 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12372 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12373 ; GFX940-TGSPLIT-NEXT: s_endpgm
12374 ;
12375 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12376 ; GFX11-WGP: ; %bb.0: ; %entry
12377 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12378 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12379 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12380 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12381 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
12382 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12383 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
12384 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12385 ; GFX11-WGP-NEXT: s_endpgm
12386 ;
12387 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12388 ; GFX11-CU: ; %bb.0: ; %entry
12389 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12390 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12391 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12392 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12393 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
12394 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12395 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
12396 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12397 ; GFX11-CU-NEXT: s_endpgm
12398 ;
12399 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12400 ; GFX12-WGP: ; %bb.0: ; %entry
12401 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12402 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12403 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12404 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12405 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
12406 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12407 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
12408 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12409 ; GFX12-WGP-NEXT: s_endpgm
12410 ;
12411 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
12412 ; GFX12-CU: ; %bb.0: ; %entry
12413 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
12414 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12415 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12416 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
12417 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
12418 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12419 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
12420 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12421 ; GFX12-CU-NEXT: s_endpgm
12422 ptr addrspace(1) %out, i32 %in) {
12423 entry:
12424 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") acquire
12425 store i32 %val, ptr addrspace(1) %out, align 4
12426 ret void
12427 }
12429 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_ret_atomicrmw(
12430 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12431 ; GFX6: ; %bb.0: ; %entry
12432 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12433 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
12434 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
12435 ; GFX6-NEXT: s_mov_b32 s11, s5
12436 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
12437 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
12438 ; GFX6-NEXT: s_mov_b32 s10, -1
12439 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
12440 ; GFX6-NEXT: s_mov_b32 s5, s11
12441 ; GFX6-NEXT: s_mov_b32 s6, s10
12442 ; GFX6-NEXT: s_mov_b32 s7, s9
12443 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
12444 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
12445 ; GFX6-NEXT: s_waitcnt vmcnt(0)
12446 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
12447 ; GFX6-NEXT: s_endpgm
12448 ;
12449 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12450 ; GFX7: ; %bb.0: ; %entry
12451 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12452 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
12453 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
12454 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12455 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12456 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
12457 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
12458 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12459 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12460 ; GFX7-NEXT: s_waitcnt vmcnt(0)
12461 ; GFX7-NEXT: flat_store_dword v[0:1], v2
12462 ; GFX7-NEXT: s_endpgm
12463 ;
12464 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12465 ; GFX10-WGP: ; %bb.0: ; %entry
12466 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
12467 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12468 ; GFX10-WGP-NEXT: s_nop 0
12469 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
12470 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
12471 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
12472 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12473 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
12474 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
12475 ; GFX10-WGP-NEXT: s_endpgm
12476 ;
12477 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12478 ; GFX10-CU: ; %bb.0: ; %entry
12479 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
12480 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12481 ; GFX10-CU-NEXT: s_nop 0
12482 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
12483 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
12484 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
12485 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12486 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
12487 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
12488 ; GFX10-CU-NEXT: s_endpgm
12489 ;
12490 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12491 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
12492 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12493 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
12494 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
12495 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
12496 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
12497 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
12498 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
12499 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
12500 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
12501 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
12502 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
12503 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
12504 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
12505 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
12506 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
12507 ; SKIP-CACHE-INV-NEXT: s_endpgm
12508 ;
12509 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12510 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
12511 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12512 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12513 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
12514 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12515 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12516 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12517 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12518 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12519 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12520 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12521 ;
12522 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12523 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12524 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12525 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12526 ; GFX90A-TGSPLIT-NEXT: s_nop 0
12527 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12528 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12529 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12530 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12531 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12532 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12533 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12534 ;
12535 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12536 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12537 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12538 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12539 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
12540 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12541 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12542 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12543 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12544 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12545 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12546 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12547 ;
12548 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12549 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12550 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12551 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12552 ; GFX940-TGSPLIT-NEXT: s_nop 0
12553 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12554 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12555 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12556 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12557 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12558 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12559 ; GFX940-TGSPLIT-NEXT: s_endpgm
12560 ;
12561 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12562 ; GFX11-WGP: ; %bb.0: ; %entry
12563 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12564 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12565 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12566 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12567 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
12568 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12569 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
12570 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12571 ; GFX11-WGP-NEXT: s_endpgm
12572 ;
12573 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12574 ; GFX11-CU: ; %bb.0: ; %entry
12575 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12576 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12577 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12578 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12579 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
12580 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12581 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
12582 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12583 ; GFX11-CU-NEXT: s_endpgm
12584 ;
12585 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12586 ; GFX12-WGP: ; %bb.0: ; %entry
12587 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12588 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12589 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12590 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12591 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
12592 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12593 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
12594 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12595 ; GFX12-WGP-NEXT: s_endpgm
12596 ;
12597 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
12598 ; GFX12-CU: ; %bb.0: ; %entry
12599 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
12600 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12601 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12602 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
12603 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
12604 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12605 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
12606 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12607 ; GFX12-CU-NEXT: s_endpgm
12608 ptr addrspace(1) %out, i32 %in) {
12609 entry:
12610 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") acq_rel
12611 store i32 %val, ptr addrspace(1) %out, align 4
12612 ret void
12613 }
12615 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_ret_atomicrmw(
12616 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12617 ; GFX6: ; %bb.0: ; %entry
12618 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12619 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x2
12620 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
12621 ; GFX6-NEXT: s_mov_b32 s11, s5
12622 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
12623 ; GFX6-NEXT: s_mov_b32 s9, 0x100f000
12624 ; GFX6-NEXT: s_mov_b32 s10, -1
12625 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
12626 ; GFX6-NEXT: s_mov_b32 s5, s11
12627 ; GFX6-NEXT: s_mov_b32 s6, s10
12628 ; GFX6-NEXT: s_mov_b32 s7, s9
12629 ; GFX6-NEXT: v_mov_b32_e32 v0, s8
12630 ; GFX6-NEXT: buffer_atomic_swap v0, off, s[4:7], 0 glc
12631 ; GFX6-NEXT: s_waitcnt vmcnt(0)
12632 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
12633 ; GFX6-NEXT: s_endpgm
12634 ;
12635 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12636 ; GFX7: ; %bb.0: ; %entry
12637 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12638 ; GFX7-NEXT: s_load_dword s6, s[6:7], 0x2
12639 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
12640 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12641 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12642 ; GFX7-NEXT: v_mov_b32_e32 v2, s6
12643 ; GFX7-NEXT: flat_atomic_swap v2, v[0:1], v2 glc
12644 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12645 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12646 ; GFX7-NEXT: s_waitcnt vmcnt(0)
12647 ; GFX7-NEXT: flat_store_dword v[0:1], v2
12648 ; GFX7-NEXT: s_endpgm
12649 ;
12650 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12651 ; GFX10-WGP: ; %bb.0: ; %entry
12652 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
12653 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12654 ; GFX10-WGP-NEXT: s_nop 0
12655 ; GFX10-WGP-NEXT: s_load_dword s6, s[6:7], 0x8
12656 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
12657 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s6
12658 ; GFX10-WGP-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12659 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
12660 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
12661 ; GFX10-WGP-NEXT: s_endpgm
12662 ;
12663 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12664 ; GFX10-CU: ; %bb.0: ; %entry
12665 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
12666 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12667 ; GFX10-CU-NEXT: s_nop 0
12668 ; GFX10-CU-NEXT: s_load_dword s6, s[6:7], 0x8
12669 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
12670 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s6
12671 ; GFX10-CU-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12672 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
12673 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
12674 ; GFX10-CU-NEXT: s_endpgm
12675 ;
12676 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12677 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
12678 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12679 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x2
12680 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
12681 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, s1
12682 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
12683 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s5, 0xf000
12684 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, -1
12685 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
12686 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s7
12687 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s6
12688 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s5
12689 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s4
12690 ; SKIP-CACHE-INV-NEXT: buffer_atomic_swap v0, off, s[0:3], 0 glc
12691 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
12692 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
12693 ; SKIP-CACHE-INV-NEXT: s_endpgm
12694 ;
12695 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12696 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
12697 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12698 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12699 ; GFX90A-NOTTGSPLIT-NEXT: s_nop 0
12700 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12701 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12702 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12703 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12704 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12705 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12706 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12707 ;
12708 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12709 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12710 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12711 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12712 ; GFX90A-TGSPLIT-NEXT: s_nop 0
12713 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[6:7], 0x8
12714 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12715 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12716 ; GFX90A-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[4:5] glc
12717 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12718 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
12719 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12720 ;
12721 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12722 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12723 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12724 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12725 ; GFX940-NOTTGSPLIT-NEXT: s_nop 0
12726 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12727 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12728 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12729 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12730 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
12731 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12732 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12733 ;
12734 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12735 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12736 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12737 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12738 ; GFX940-TGSPLIT-NEXT: s_nop 0
12739 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[2:3], 0x8
12740 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12741 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12742 ; GFX940-TGSPLIT-NEXT: global_atomic_swap v1, v0, v1, s[0:1] sc0
12743 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
12744 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
12745 ; GFX940-TGSPLIT-NEXT: s_endpgm
12746 ;
12747 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12748 ; GFX11-WGP: ; %bb.0: ; %entry
12749 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12750 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12751 ; GFX11-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12752 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12753 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s2
12754 ; GFX11-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12755 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
12756 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12757 ; GFX11-WGP-NEXT: s_endpgm
12758 ;
12759 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12760 ; GFX11-CU: ; %bb.0: ; %entry
12761 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12762 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12763 ; GFX11-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12764 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12765 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s2
12766 ; GFX11-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
12767 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
12768 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12769 ; GFX11-CU-NEXT: s_endpgm
12770 ;
12771 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12772 ; GFX12-WGP: ; %bb.0: ; %entry
12773 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12774 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12775 ; GFX12-WGP-NEXT: s_load_b32 s2, s[2:3], 0x8
12776 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12777 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s2
12778 ; GFX12-WGP-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12779 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
12780 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
12781 ; GFX12-WGP-NEXT: s_endpgm
12782 ;
12783 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
12784 ; GFX12-CU: ; %bb.0: ; %entry
12785 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
12786 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[2:3], 0x0
12787 ; GFX12-CU-NEXT: s_load_b32 s2, s[2:3], 0x8
12788 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
12789 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s2
12790 ; GFX12-CU-NEXT: global_atomic_swap_b32 v1, v0, v1, s[0:1] th:TH_ATOMIC_RETURN
12791 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
12792 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
12793 ; GFX12-CU-NEXT: s_endpgm
12794 ptr addrspace(1) %out, i32 %in) {
12795 entry:
12796 %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("singlethread-one-as") seq_cst
12797 store i32 %val, ptr addrspace(1) %out, align 4
12798 ret void
12799 }
12801 define amdgpu_kernel void @global_singlethread_one_as_monotonic_monotonic_cmpxchg(
12802 ; GFX6-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12803 ; GFX6: ; %bb.0: ; %entry
12804 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
12805 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
12806 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
12807 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
12808 ; GFX6-NEXT: s_mov_b32 s12, s5
12809 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
12810 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
12811 ; GFX6-NEXT: s_mov_b32 s11, -1
12812 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
12813 ; GFX6-NEXT: s_mov_b32 s5, s12
12814 ; GFX6-NEXT: s_mov_b32 s6, s11
12815 ; GFX6-NEXT: s_mov_b32 s7, s10
12816 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
12817 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
12818 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
12819 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
12820 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
12821 ; GFX6-NEXT: s_endpgm
12822 ;
12823 ; GFX7-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12824 ; GFX7: ; %bb.0: ; %entry
12825 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
12826 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
12827 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
12828 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
12829 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
12830 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
12831 ; GFX7-NEXT: s_mov_b32 s4, s8
12832 ; GFX7-NEXT: s_mov_b32 s5, s9
12833 ; GFX7-NEXT: s_mov_b32 s9, s10
12834 ; GFX7-NEXT: s_mov_b32 s8, s11
12835 ; GFX7-NEXT: s_add_u32 s4, s4, s9
12836 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
12837 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
12838 ; GFX7-NEXT: s_mov_b32 s5, s8
12839 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
12840 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
12841 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
12842 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
12843 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
12844 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
12845 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
12846 ; GFX7-NEXT: s_endpgm
12847 ;
12848 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12849 ; GFX10-WGP: ; %bb.0: ; %entry
12850 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
12851 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
12852 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
12853 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
12854 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
12855 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
12856 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
12857 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
12858 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
12859 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
12860 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
12861 ; GFX10-WGP-NEXT: s_endpgm
12862 ;
12863 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12864 ; GFX10-CU: ; %bb.0: ; %entry
12865 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
12866 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
12867 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
12868 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
12869 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
12870 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
12871 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
12872 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
12873 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
12874 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
12875 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
12876 ; GFX10-CU-NEXT: s_endpgm
12877 ;
12878 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12879 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
12880 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
12881 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
12882 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
12883 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
12884 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
12885 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
12886 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
12887 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
12888 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
12889 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
12890 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
12891 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
12892 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
12893 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
12894 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
12895 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
12896 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
12897 ; SKIP-CACHE-INV-NEXT: s_endpgm
12898 ;
12899 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12900 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
12901 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
12902 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12903 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
12904 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
12905 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
12906 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12907 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
12908 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12909 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
12910 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
12911 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
12912 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
12913 ;
12914 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12915 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
12916 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
12917 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12918 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
12919 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
12920 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
12921 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12922 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
12923 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
12924 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
12925 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
12926 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
12927 ; GFX90A-TGSPLIT-NEXT: s_endpgm
12928 ;
12929 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12930 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
12931 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
12932 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12933 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
12934 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
12935 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
12936 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12937 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
12938 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12939 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
12940 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
12941 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
12942 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
12943 ;
12944 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12945 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
12946 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
12947 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
12948 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
12949 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
12950 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
12951 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
12952 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
12953 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
12954 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
12955 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
12956 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
12957 ; GFX940-TGSPLIT-NEXT: s_endpgm
12958 ;
12959 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12960 ; GFX11-WGP: ; %bb.0: ; %entry
12961 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
12962 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
12963 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
12964 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
12965 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
12966 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
12967 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
12968 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
12969 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
12970 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
12971 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
12972 ; GFX11-WGP-NEXT: s_endpgm
12973 ;
12974 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12975 ; GFX11-CU: ; %bb.0: ; %entry
12976 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
12977 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
12978 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
12979 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
12980 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
12981 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
12982 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
12983 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
12984 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
12985 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
12986 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
12987 ; GFX11-CU-NEXT: s_endpgm
12988 ;
12989 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
12990 ; GFX12-WGP: ; %bb.0: ; %entry
12991 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
12992 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
12993 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
12994 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
12995 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
12996 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
12997 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
12998 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
12999 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13000 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
13001 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13002 ; GFX12-WGP-NEXT: s_endpgm
13003 ;
13004 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
13005 ; GFX12-CU: ; %bb.0: ; %entry
13006 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13007 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
13008 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13009 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13010 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13011 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
13012 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
13013 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
13014 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13015 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
13016 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13017 ; GFX12-CU-NEXT: s_endpgm
13018 ptr addrspace(1) %out, i32 %in, i32 %old) {
13019 entry:
13020 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
13021 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
13022 ret void
13023 }
13025 define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_cmpxchg(
13026 ; GFX6-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13027 ; GFX6: ; %bb.0: ; %entry
13028 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
13029 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
13030 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
13031 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
13032 ; GFX6-NEXT: s_mov_b32 s12, s5
13033 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
13034 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
13035 ; GFX6-NEXT: s_mov_b32 s11, -1
13036 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
13037 ; GFX6-NEXT: s_mov_b32 s5, s12
13038 ; GFX6-NEXT: s_mov_b32 s6, s11
13039 ; GFX6-NEXT: s_mov_b32 s7, s10
13040 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
13041 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
13042 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13043 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
13044 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
13045 ; GFX6-NEXT: s_endpgm
13046 ;
13047 ; GFX7-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13048 ; GFX7: ; %bb.0: ; %entry
13049 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
13050 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
13051 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
13052 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
13053 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
13054 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
13055 ; GFX7-NEXT: s_mov_b32 s4, s8
13056 ; GFX7-NEXT: s_mov_b32 s5, s9
13057 ; GFX7-NEXT: s_mov_b32 s9, s10
13058 ; GFX7-NEXT: s_mov_b32 s8, s11
13059 ; GFX7-NEXT: s_add_u32 s4, s4, s9
13060 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
13061 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
13062 ; GFX7-NEXT: s_mov_b32 s5, s8
13063 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
13064 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
13065 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13066 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
13067 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
13068 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
13069 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
13070 ; GFX7-NEXT: s_endpgm
13071 ;
13072 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13073 ; GFX10-WGP: ; %bb.0: ; %entry
13074 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
13075 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
13076 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13077 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
13078 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
13079 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
13080 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
13081 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
13082 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13083 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
13084 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13085 ; GFX10-WGP-NEXT: s_endpgm
13086 ;
13087 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13088 ; GFX10-CU: ; %bb.0: ; %entry
13089 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
13090 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
13091 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13092 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
13093 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
13094 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
13095 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
13096 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
13097 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13098 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
13099 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13100 ; GFX10-CU-NEXT: s_endpgm
13101 ;
13102 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13103 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
13104 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
13105 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
13106 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
13107 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
13108 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
13109 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
13110 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
13111 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
13112 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
13113 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
13114 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
13115 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
13116 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
13117 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
13118 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13119 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
13120 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
13121 ; SKIP-CACHE-INV-NEXT: s_endpgm
13122 ;
13123 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13124 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
13125 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13126 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13127 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13128 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13129 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13130 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13131 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13132 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13133 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13134 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13135 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13136 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
13137 ;
13138 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13139 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
13140 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13141 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13142 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13143 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13144 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13145 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13146 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13147 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13148 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13149 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13150 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13151 ; GFX90A-TGSPLIT-NEXT: s_endpgm
13152 ;
13153 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13154 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
13155 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13156 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13157 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13158 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13159 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13160 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13161 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13162 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13163 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13164 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13165 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13166 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
13167 ;
13168 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13169 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
13170 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13171 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13172 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13173 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13174 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13175 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13176 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13177 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13178 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13179 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13180 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13181 ; GFX940-TGSPLIT-NEXT: s_endpgm
13182 ;
13183 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13184 ; GFX11-WGP: ; %bb.0: ; %entry
13185 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13186 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
13187 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13188 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13189 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13190 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
13191 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
13192 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
13193 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13194 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
13195 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13196 ; GFX11-WGP-NEXT: s_endpgm
13197 ;
13198 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13199 ; GFX11-CU: ; %bb.0: ; %entry
13200 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13201 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
13202 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13203 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13204 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13205 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
13206 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
13207 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
13208 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13209 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
13210 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13211 ; GFX11-CU-NEXT: s_endpgm
13212 ;
13213 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13214 ; GFX12-WGP: ; %bb.0: ; %entry
13215 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13216 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
13217 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13218 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13219 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13220 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
13221 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
13222 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
13223 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13224 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
13225 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13226 ; GFX12-WGP-NEXT: s_endpgm
13227 ;
13228 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
13229 ; GFX12-CU: ; %bb.0: ; %entry
13230 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13231 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
13232 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13233 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13234 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13235 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
13236 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
13237 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
13238 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13239 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
13240 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13241 ; GFX12-CU-NEXT: s_endpgm
13242 ptr addrspace(1) %out, i32 %in, i32 %old) {
13243 entry:
13244 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
13245 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
13246 ret void
13247 }
13249 define amdgpu_kernel void @global_singlethread_one_as_release_monotonic_cmpxchg(
13250 ; GFX6-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13251 ; GFX6: ; %bb.0: ; %entry
13252 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
13253 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
13254 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
13255 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
13256 ; GFX6-NEXT: s_mov_b32 s12, s5
13257 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
13258 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
13259 ; GFX6-NEXT: s_mov_b32 s11, -1
13260 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
13261 ; GFX6-NEXT: s_mov_b32 s5, s12
13262 ; GFX6-NEXT: s_mov_b32 s6, s11
13263 ; GFX6-NEXT: s_mov_b32 s7, s10
13264 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
13265 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
13266 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13267 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
13268 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
13269 ; GFX6-NEXT: s_endpgm
13270 ;
13271 ; GFX7-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13272 ; GFX7: ; %bb.0: ; %entry
13273 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
13274 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
13275 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
13276 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
13277 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
13278 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
13279 ; GFX7-NEXT: s_mov_b32 s4, s8
13280 ; GFX7-NEXT: s_mov_b32 s5, s9
13281 ; GFX7-NEXT: s_mov_b32 s9, s10
13282 ; GFX7-NEXT: s_mov_b32 s8, s11
13283 ; GFX7-NEXT: s_add_u32 s4, s4, s9
13284 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
13285 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
13286 ; GFX7-NEXT: s_mov_b32 s5, s8
13287 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
13288 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
13289 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13290 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
13291 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
13292 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
13293 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
13294 ; GFX7-NEXT: s_endpgm
13295 ;
13296 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13297 ; GFX10-WGP: ; %bb.0: ; %entry
13298 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
13299 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
13300 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13301 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
13302 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
13303 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
13304 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
13305 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
13306 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13307 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
13308 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13309 ; GFX10-WGP-NEXT: s_endpgm
13310 ;
13311 ; GFX10-CU-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13312 ; GFX10-CU: ; %bb.0: ; %entry
13313 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
13314 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
13315 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13316 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
13317 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
13318 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
13319 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
13320 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
13321 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13322 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
13323 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13324 ; GFX10-CU-NEXT: s_endpgm
13325 ;
13326 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13327 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
13328 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
13329 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
13330 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
13331 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
13332 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
13333 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
13334 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
13335 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
13336 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
13337 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
13338 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
13339 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
13340 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
13341 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
13342 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13343 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
13344 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
13345 ; SKIP-CACHE-INV-NEXT: s_endpgm
13346 ;
13347 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13348 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
13349 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13350 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13351 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13352 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13353 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13354 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13355 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13356 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13357 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13358 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13359 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13360 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
13361 ;
13362 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13363 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
13364 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13365 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13366 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13367 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13368 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13369 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13370 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13371 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13372 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13373 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13374 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13375 ; GFX90A-TGSPLIT-NEXT: s_endpgm
13376 ;
13377 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13378 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
13379 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13380 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13381 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13382 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13383 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13384 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13385 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13386 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13387 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13388 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13389 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13390 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
13391 ;
13392 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13393 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
13394 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13395 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13396 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13397 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13398 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13399 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13400 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13401 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13402 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13403 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13404 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13405 ; GFX940-TGSPLIT-NEXT: s_endpgm
13406 ;
13407 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13408 ; GFX11-WGP: ; %bb.0: ; %entry
13409 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13410 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
13411 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13412 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13413 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13414 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
13415 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
13416 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
13417 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13418 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
13419 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13420 ; GFX11-WGP-NEXT: s_endpgm
13421 ;
13422 ; GFX11-CU-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13423 ; GFX11-CU: ; %bb.0: ; %entry
13424 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13425 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
13426 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13427 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13428 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13429 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
13430 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
13431 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
13432 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13433 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
13434 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13435 ; GFX11-CU-NEXT: s_endpgm
13436 ;
13437 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13438 ; GFX12-WGP: ; %bb.0: ; %entry
13439 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13440 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
13441 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13442 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13443 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13444 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
13445 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
13446 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
13447 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13448 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
13449 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13450 ; GFX12-WGP-NEXT: s_endpgm
13451 ;
13452 ; GFX12-CU-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
13453 ; GFX12-CU: ; %bb.0: ; %entry
13454 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13455 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
13456 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13457 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13458 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13459 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
13460 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
13461 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
13462 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13463 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
13464 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13465 ; GFX12-CU-NEXT: s_endpgm
13466 ptr addrspace(1) %out, i32 %in, i32 %old) {
13467 entry:
13468 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
13469 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
13470 ret void
13471 }
13473 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_cmpxchg(
13474 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13475 ; GFX6: ; %bb.0: ; %entry
13476 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
13477 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
13478 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
13479 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
13480 ; GFX6-NEXT: s_mov_b32 s12, s5
13481 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
13482 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
13483 ; GFX6-NEXT: s_mov_b32 s11, -1
13484 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
13485 ; GFX6-NEXT: s_mov_b32 s5, s12
13486 ; GFX6-NEXT: s_mov_b32 s6, s11
13487 ; GFX6-NEXT: s_mov_b32 s7, s10
13488 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
13489 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
13490 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13491 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
13492 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
13493 ; GFX6-NEXT: s_endpgm
13494 ;
13495 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13496 ; GFX7: ; %bb.0: ; %entry
13497 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
13498 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
13499 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
13500 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
13501 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
13502 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
13503 ; GFX7-NEXT: s_mov_b32 s4, s8
13504 ; GFX7-NEXT: s_mov_b32 s5, s9
13505 ; GFX7-NEXT: s_mov_b32 s9, s10
13506 ; GFX7-NEXT: s_mov_b32 s8, s11
13507 ; GFX7-NEXT: s_add_u32 s4, s4, s9
13508 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
13509 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
13510 ; GFX7-NEXT: s_mov_b32 s5, s8
13511 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
13512 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
13513 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13514 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
13515 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
13516 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
13517 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
13518 ; GFX7-NEXT: s_endpgm
13519 ;
13520 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13521 ; GFX10-WGP: ; %bb.0: ; %entry
13522 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
13523 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
13524 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13525 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
13526 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
13527 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
13528 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
13529 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
13530 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13531 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
13532 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13533 ; GFX10-WGP-NEXT: s_endpgm
13534 ;
13535 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13536 ; GFX10-CU: ; %bb.0: ; %entry
13537 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
13538 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
13539 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13540 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
13541 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
13542 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
13543 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
13544 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
13545 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13546 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
13547 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13548 ; GFX10-CU-NEXT: s_endpgm
13549 ;
13550 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13551 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
13552 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
13553 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
13554 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
13555 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
13556 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
13557 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
13558 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
13559 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
13560 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
13561 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
13562 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
13563 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
13564 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
13565 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
13566 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13567 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
13568 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
13569 ; SKIP-CACHE-INV-NEXT: s_endpgm
13570 ;
13571 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13572 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
13573 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13574 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13575 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13576 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13577 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13578 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13579 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13580 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13581 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13582 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13583 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13584 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
13585 ;
13586 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13587 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
13588 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13589 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13590 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13591 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13592 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13593 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13594 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13595 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13596 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13597 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13598 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13599 ; GFX90A-TGSPLIT-NEXT: s_endpgm
13600 ;
13601 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13602 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
13603 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13604 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13605 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13606 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13607 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13608 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13609 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13610 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13611 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13612 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13613 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13614 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
13615 ;
13616 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13617 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
13618 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13619 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13620 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13621 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13622 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13623 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13624 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13625 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13626 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13627 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13628 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13629 ; GFX940-TGSPLIT-NEXT: s_endpgm
13630 ;
13631 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13632 ; GFX11-WGP: ; %bb.0: ; %entry
13633 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13634 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
13635 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13636 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13637 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13638 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
13639 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
13640 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
13641 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13642 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
13643 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13644 ; GFX11-WGP-NEXT: s_endpgm
13645 ;
13646 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13647 ; GFX11-CU: ; %bb.0: ; %entry
13648 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13649 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
13650 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13651 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13652 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13653 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
13654 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
13655 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
13656 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13657 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
13658 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13659 ; GFX11-CU-NEXT: s_endpgm
13660 ;
13661 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13662 ; GFX12-WGP: ; %bb.0: ; %entry
13663 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13664 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
13665 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13666 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13667 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13668 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
13669 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
13670 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
13671 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13672 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
13673 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13674 ; GFX12-WGP-NEXT: s_endpgm
13675 ;
13676 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
13677 ; GFX12-CU: ; %bb.0: ; %entry
13678 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13679 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
13680 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13681 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13682 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13683 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
13684 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
13685 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
13686 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13687 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
13688 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13689 ; GFX12-CU-NEXT: s_endpgm
13690 ptr addrspace(1) %out, i32 %in, i32 %old) {
13691 entry:
13692 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
13693 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
13694 ret void
13695 }
13697 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_cmpxchg(
13698 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13699 ; GFX6: ; %bb.0: ; %entry
13700 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
13701 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
13702 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
13703 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
13704 ; GFX6-NEXT: s_mov_b32 s12, s5
13705 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
13706 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
13707 ; GFX6-NEXT: s_mov_b32 s11, -1
13708 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
13709 ; GFX6-NEXT: s_mov_b32 s5, s12
13710 ; GFX6-NEXT: s_mov_b32 s6, s11
13711 ; GFX6-NEXT: s_mov_b32 s7, s10
13712 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
13713 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
13714 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13715 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
13716 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
13717 ; GFX6-NEXT: s_endpgm
13718 ;
13719 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13720 ; GFX7: ; %bb.0: ; %entry
13721 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
13722 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
13723 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
13724 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
13725 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
13726 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
13727 ; GFX7-NEXT: s_mov_b32 s4, s8
13728 ; GFX7-NEXT: s_mov_b32 s5, s9
13729 ; GFX7-NEXT: s_mov_b32 s9, s10
13730 ; GFX7-NEXT: s_mov_b32 s8, s11
13731 ; GFX7-NEXT: s_add_u32 s4, s4, s9
13732 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
13733 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
13734 ; GFX7-NEXT: s_mov_b32 s5, s8
13735 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
13736 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
13737 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13738 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
13739 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
13740 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
13741 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
13742 ; GFX7-NEXT: s_endpgm
13743 ;
13744 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13745 ; GFX10-WGP: ; %bb.0: ; %entry
13746 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
13747 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
13748 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13749 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
13750 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
13751 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
13752 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
13753 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
13754 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13755 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
13756 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13757 ; GFX10-WGP-NEXT: s_endpgm
13758 ;
13759 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13760 ; GFX10-CU: ; %bb.0: ; %entry
13761 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
13762 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
13763 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13764 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
13765 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
13766 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
13767 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
13768 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
13769 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13770 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
13771 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13772 ; GFX10-CU-NEXT: s_endpgm
13773 ;
13774 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13775 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
13776 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
13777 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
13778 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
13779 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
13780 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
13781 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
13782 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
13783 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
13784 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
13785 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
13786 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
13787 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
13788 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
13789 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
13790 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13791 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
13792 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
13793 ; SKIP-CACHE-INV-NEXT: s_endpgm
13794 ;
13795 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13796 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
13797 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13798 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13799 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13800 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13801 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13802 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13803 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13804 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13805 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13806 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13807 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13808 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
13809 ;
13810 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13811 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
13812 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
13813 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13814 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13815 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
13816 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
13817 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13818 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
13819 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
13820 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13821 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13822 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
13823 ; GFX90A-TGSPLIT-NEXT: s_endpgm
13824 ;
13825 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13826 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
13827 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13828 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13829 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13830 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13831 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13832 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13833 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13834 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13835 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13836 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13837 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13838 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
13839 ;
13840 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13841 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
13842 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
13843 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
13844 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
13845 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
13846 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
13847 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
13848 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
13849 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
13850 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13851 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
13852 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
13853 ; GFX940-TGSPLIT-NEXT: s_endpgm
13854 ;
13855 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13856 ; GFX11-WGP: ; %bb.0: ; %entry
13857 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13858 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
13859 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13860 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13861 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13862 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
13863 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
13864 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
13865 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13866 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
13867 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13868 ; GFX11-WGP-NEXT: s_endpgm
13869 ;
13870 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13871 ; GFX11-CU: ; %bb.0: ; %entry
13872 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13873 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
13874 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13875 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13876 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13877 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
13878 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
13879 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
13880 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13881 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
13882 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13883 ; GFX11-CU-NEXT: s_endpgm
13884 ;
13885 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13886 ; GFX12-WGP: ; %bb.0: ; %entry
13887 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
13888 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
13889 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13890 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
13891 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
13892 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
13893 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
13894 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
13895 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13896 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
13897 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13898 ; GFX12-WGP-NEXT: s_endpgm
13899 ;
13900 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
13901 ; GFX12-CU: ; %bb.0: ; %entry
13902 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
13903 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
13904 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
13905 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
13906 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
13907 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
13908 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
13909 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
13910 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13911 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
13912 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
13913 ; GFX12-CU-NEXT: s_endpgm
13914 ptr addrspace(1) %out, i32 %in, i32 %old) {
13915 entry:
13916 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
13917 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
13918 ret void
13919 }
13921 define amdgpu_kernel void @global_singlethread_one_as_monotonic_acquire_cmpxchg(
13922 ; GFX6-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
13923 ; GFX6: ; %bb.0: ; %entry
13924 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
13925 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
13926 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
13927 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
13928 ; GFX6-NEXT: s_mov_b32 s12, s5
13929 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
13930 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
13931 ; GFX6-NEXT: s_mov_b32 s11, -1
13932 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
13933 ; GFX6-NEXT: s_mov_b32 s5, s12
13934 ; GFX6-NEXT: s_mov_b32 s6, s11
13935 ; GFX6-NEXT: s_mov_b32 s7, s10
13936 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
13937 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
13938 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
13939 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
13940 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
13941 ; GFX6-NEXT: s_endpgm
13942 ;
13943 ; GFX7-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
13944 ; GFX7: ; %bb.0: ; %entry
13945 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
13946 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
13947 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
13948 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
13949 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
13950 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
13951 ; GFX7-NEXT: s_mov_b32 s4, s8
13952 ; GFX7-NEXT: s_mov_b32 s5, s9
13953 ; GFX7-NEXT: s_mov_b32 s9, s10
13954 ; GFX7-NEXT: s_mov_b32 s8, s11
13955 ; GFX7-NEXT: s_add_u32 s4, s4, s9
13956 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
13957 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
13958 ; GFX7-NEXT: s_mov_b32 s5, s8
13959 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
13960 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
13961 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
13962 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
13963 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
13964 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
13965 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
13966 ; GFX7-NEXT: s_endpgm
13967 ;
13968 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
13969 ; GFX10-WGP: ; %bb.0: ; %entry
13970 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
13971 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
13972 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13973 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
13974 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
13975 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
13976 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
13977 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
13978 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13979 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
13980 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13981 ; GFX10-WGP-NEXT: s_endpgm
13982 ;
13983 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
13984 ; GFX10-CU: ; %bb.0: ; %entry
13985 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
13986 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
13987 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
13988 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
13989 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
13990 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
13991 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
13992 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
13993 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
13994 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
13995 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
13996 ; GFX10-CU-NEXT: s_endpgm
13997 ;
13998 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
13999 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
14000 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
14001 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
14002 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
14003 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
14004 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
14005 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
14006 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
14007 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
14008 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
14009 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
14010 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
14011 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
14012 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
14013 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
14014 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14015 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
14016 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
14017 ; SKIP-CACHE-INV-NEXT: s_endpgm
14018 ;
14019 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14020 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
14021 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14022 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14023 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14024 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14025 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14026 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14027 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14028 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14029 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14030 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14031 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14032 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
14033 ;
14034 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14035 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
14036 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14037 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14038 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14039 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14040 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14041 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14042 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14043 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14044 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14045 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14046 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14047 ; GFX90A-TGSPLIT-NEXT: s_endpgm
14048 ;
14049 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14050 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
14051 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14052 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14053 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14054 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14055 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14056 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14057 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14058 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14059 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14060 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14061 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14062 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
14063 ;
14064 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14065 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
14066 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14067 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14068 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14069 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14070 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14071 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14072 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14073 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14074 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14075 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14076 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14077 ; GFX940-TGSPLIT-NEXT: s_endpgm
14078 ;
14079 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14080 ; GFX11-WGP: ; %bb.0: ; %entry
14081 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14082 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
14083 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14084 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14085 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14086 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
14087 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
14088 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
14089 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14090 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
14091 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14092 ; GFX11-WGP-NEXT: s_endpgm
14093 ;
14094 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14095 ; GFX11-CU: ; %bb.0: ; %entry
14096 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14097 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
14098 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14099 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14100 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14101 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
14102 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
14103 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
14104 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14105 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
14106 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14107 ; GFX11-CU-NEXT: s_endpgm
14108 ;
14109 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14110 ; GFX12-WGP: ; %bb.0: ; %entry
14111 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14112 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
14113 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14114 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14115 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14116 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
14117 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
14118 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
14119 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14120 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
14121 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14122 ; GFX12-WGP-NEXT: s_endpgm
14123 ;
14124 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
14125 ; GFX12-CU: ; %bb.0: ; %entry
14126 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14127 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
14128 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14129 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14130 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14131 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
14132 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
14133 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
14134 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14135 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
14136 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14137 ; GFX12-CU-NEXT: s_endpgm
14138 ptr addrspace(1) %out, i32 %in, i32 %old) {
14139 entry:
14140 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
14141 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic acquire
14142 ret void
14143 }
14145 define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_cmpxchg(
14146 ; GFX6-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14147 ; GFX6: ; %bb.0: ; %entry
14148 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
14149 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
14150 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
14151 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
14152 ; GFX6-NEXT: s_mov_b32 s12, s5
14153 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
14154 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
14155 ; GFX6-NEXT: s_mov_b32 s11, -1
14156 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
14157 ; GFX6-NEXT: s_mov_b32 s5, s12
14158 ; GFX6-NEXT: s_mov_b32 s6, s11
14159 ; GFX6-NEXT: s_mov_b32 s7, s10
14160 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
14161 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
14162 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14163 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
14164 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
14165 ; GFX6-NEXT: s_endpgm
14166 ;
14167 ; GFX7-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14168 ; GFX7: ; %bb.0: ; %entry
14169 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
14170 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
14171 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
14172 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
14173 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
14174 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
14175 ; GFX7-NEXT: s_mov_b32 s4, s8
14176 ; GFX7-NEXT: s_mov_b32 s5, s9
14177 ; GFX7-NEXT: s_mov_b32 s9, s10
14178 ; GFX7-NEXT: s_mov_b32 s8, s11
14179 ; GFX7-NEXT: s_add_u32 s4, s4, s9
14180 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
14181 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
14182 ; GFX7-NEXT: s_mov_b32 s5, s8
14183 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
14184 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
14185 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14186 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
14187 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
14188 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
14189 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
14190 ; GFX7-NEXT: s_endpgm
14191 ;
14192 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14193 ; GFX10-WGP: ; %bb.0: ; %entry
14194 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
14195 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
14196 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14197 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
14198 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
14199 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
14200 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
14201 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
14202 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14203 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
14204 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14205 ; GFX10-WGP-NEXT: s_endpgm
14206 ;
14207 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14208 ; GFX10-CU: ; %bb.0: ; %entry
14209 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
14210 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
14211 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14212 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
14213 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
14214 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
14215 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
14216 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
14217 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14218 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
14219 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14220 ; GFX10-CU-NEXT: s_endpgm
14221 ;
14222 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14223 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
14224 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
14225 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
14226 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
14227 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
14228 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
14229 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
14230 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
14231 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
14232 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
14233 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
14234 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
14235 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
14236 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
14237 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
14238 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14239 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
14240 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
14241 ; SKIP-CACHE-INV-NEXT: s_endpgm
14242 ;
14243 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14244 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
14245 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14246 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14247 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14248 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14249 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14250 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14251 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14252 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14253 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14254 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14255 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14256 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
14257 ;
14258 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14259 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
14260 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14261 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14262 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14263 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14264 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14265 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14266 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14267 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14268 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14269 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14270 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14271 ; GFX90A-TGSPLIT-NEXT: s_endpgm
14272 ;
14273 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14274 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
14275 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14276 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14277 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14278 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14279 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14280 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14281 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14282 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14283 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14284 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14285 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14286 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
14287 ;
14288 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14289 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
14290 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14291 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14292 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14293 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14294 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14295 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14296 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14297 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14298 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14299 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14300 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14301 ; GFX940-TGSPLIT-NEXT: s_endpgm
14302 ;
14303 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14304 ; GFX11-WGP: ; %bb.0: ; %entry
14305 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14306 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
14307 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14308 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14309 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14310 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
14311 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
14312 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
14313 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14314 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
14315 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14316 ; GFX11-WGP-NEXT: s_endpgm
14317 ;
14318 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14319 ; GFX11-CU: ; %bb.0: ; %entry
14320 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14321 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
14322 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14323 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14324 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14325 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
14326 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
14327 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
14328 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14329 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
14330 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14331 ; GFX11-CU-NEXT: s_endpgm
14332 ;
14333 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14334 ; GFX12-WGP: ; %bb.0: ; %entry
14335 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14336 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
14337 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14338 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14339 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14340 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
14341 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
14342 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
14343 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14344 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
14345 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14346 ; GFX12-WGP-NEXT: s_endpgm
14347 ;
14348 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
14349 ; GFX12-CU: ; %bb.0: ; %entry
14350 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14351 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
14352 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14353 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14354 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14355 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
14356 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
14357 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
14358 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14359 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
14360 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14361 ; GFX12-CU-NEXT: s_endpgm
14362 ptr addrspace(1) %out, i32 %in, i32 %old) {
14363 entry:
14364 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
14365 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
14366 ret void
14367 }
14369 define amdgpu_kernel void @global_singlethread_one_as_release_acquire_cmpxchg(
14370 ; GFX6-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14371 ; GFX6: ; %bb.0: ; %entry
14372 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
14373 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
14374 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
14375 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
14376 ; GFX6-NEXT: s_mov_b32 s12, s5
14377 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
14378 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
14379 ; GFX6-NEXT: s_mov_b32 s11, -1
14380 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
14381 ; GFX6-NEXT: s_mov_b32 s5, s12
14382 ; GFX6-NEXT: s_mov_b32 s6, s11
14383 ; GFX6-NEXT: s_mov_b32 s7, s10
14384 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
14385 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
14386 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14387 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
14388 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
14389 ; GFX6-NEXT: s_endpgm
14390 ;
14391 ; GFX7-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14392 ; GFX7: ; %bb.0: ; %entry
14393 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
14394 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
14395 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
14396 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
14397 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
14398 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
14399 ; GFX7-NEXT: s_mov_b32 s4, s8
14400 ; GFX7-NEXT: s_mov_b32 s5, s9
14401 ; GFX7-NEXT: s_mov_b32 s9, s10
14402 ; GFX7-NEXT: s_mov_b32 s8, s11
14403 ; GFX7-NEXT: s_add_u32 s4, s4, s9
14404 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
14405 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
14406 ; GFX7-NEXT: s_mov_b32 s5, s8
14407 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
14408 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
14409 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14410 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
14411 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
14412 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
14413 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
14414 ; GFX7-NEXT: s_endpgm
14415 ;
14416 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14417 ; GFX10-WGP: ; %bb.0: ; %entry
14418 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
14419 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
14420 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14421 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
14422 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
14423 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
14424 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
14425 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
14426 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14427 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
14428 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14429 ; GFX10-WGP-NEXT: s_endpgm
14430 ;
14431 ; GFX10-CU-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14432 ; GFX10-CU: ; %bb.0: ; %entry
14433 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
14434 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
14435 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14436 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
14437 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
14438 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
14439 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
14440 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
14441 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14442 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
14443 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14444 ; GFX10-CU-NEXT: s_endpgm
14445 ;
14446 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14447 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
14448 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
14449 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
14450 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
14451 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
14452 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
14453 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
14454 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
14455 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
14456 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
14457 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
14458 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
14459 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
14460 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
14461 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
14462 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14463 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
14464 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
14465 ; SKIP-CACHE-INV-NEXT: s_endpgm
14466 ;
14467 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14468 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
14469 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14470 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14471 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14472 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14473 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14474 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14475 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14476 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14477 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14478 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14479 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14480 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
14481 ;
14482 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14483 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
14484 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14485 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14486 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14487 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14488 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14489 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14490 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14491 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14492 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14493 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14494 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14495 ; GFX90A-TGSPLIT-NEXT: s_endpgm
14496 ;
14497 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14498 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
14499 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14500 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14501 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14502 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14503 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14504 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14505 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14506 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14507 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14508 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14509 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14510 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
14511 ;
14512 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14513 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
14514 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14515 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14516 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14517 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14518 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14519 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14520 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14521 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14522 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14523 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14524 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14525 ; GFX940-TGSPLIT-NEXT: s_endpgm
14526 ;
14527 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14528 ; GFX11-WGP: ; %bb.0: ; %entry
14529 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14530 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
14531 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14532 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14533 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14534 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
14535 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
14536 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
14537 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14538 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
14539 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14540 ; GFX11-WGP-NEXT: s_endpgm
14541 ;
14542 ; GFX11-CU-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14543 ; GFX11-CU: ; %bb.0: ; %entry
14544 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14545 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
14546 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14547 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14548 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14549 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
14550 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
14551 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
14552 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14553 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
14554 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14555 ; GFX11-CU-NEXT: s_endpgm
14556 ;
14557 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14558 ; GFX12-WGP: ; %bb.0: ; %entry
14559 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14560 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
14561 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14562 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14563 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14564 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
14565 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
14566 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
14567 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14568 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
14569 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14570 ; GFX12-WGP-NEXT: s_endpgm
14571 ;
14572 ; GFX12-CU-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
14573 ; GFX12-CU: ; %bb.0: ; %entry
14574 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14575 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
14576 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14577 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14578 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14579 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
14580 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
14581 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
14582 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14583 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
14584 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14585 ; GFX12-CU-NEXT: s_endpgm
14586 ptr addrspace(1) %out, i32 %in, i32 %old) {
14587 entry:
14588 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
14589 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
14590 ret void
14591 }
14593 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_cmpxchg(
14594 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14595 ; GFX6: ; %bb.0: ; %entry
14596 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
14597 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
14598 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
14599 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
14600 ; GFX6-NEXT: s_mov_b32 s12, s5
14601 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
14602 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
14603 ; GFX6-NEXT: s_mov_b32 s11, -1
14604 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
14605 ; GFX6-NEXT: s_mov_b32 s5, s12
14606 ; GFX6-NEXT: s_mov_b32 s6, s11
14607 ; GFX6-NEXT: s_mov_b32 s7, s10
14608 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
14609 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
14610 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14611 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
14612 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
14613 ; GFX6-NEXT: s_endpgm
14614 ;
14615 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14616 ; GFX7: ; %bb.0: ; %entry
14617 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
14618 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
14619 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
14620 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
14621 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
14622 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
14623 ; GFX7-NEXT: s_mov_b32 s4, s8
14624 ; GFX7-NEXT: s_mov_b32 s5, s9
14625 ; GFX7-NEXT: s_mov_b32 s9, s10
14626 ; GFX7-NEXT: s_mov_b32 s8, s11
14627 ; GFX7-NEXT: s_add_u32 s4, s4, s9
14628 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
14629 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
14630 ; GFX7-NEXT: s_mov_b32 s5, s8
14631 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
14632 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
14633 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14634 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
14635 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
14636 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
14637 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
14638 ; GFX7-NEXT: s_endpgm
14639 ;
14640 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14641 ; GFX10-WGP: ; %bb.0: ; %entry
14642 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
14643 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
14644 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14645 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
14646 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
14647 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
14648 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
14649 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
14650 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14651 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
14652 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14653 ; GFX10-WGP-NEXT: s_endpgm
14654 ;
14655 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14656 ; GFX10-CU: ; %bb.0: ; %entry
14657 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
14658 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
14659 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14660 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
14661 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
14662 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
14663 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
14664 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
14665 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14666 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
14667 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14668 ; GFX10-CU-NEXT: s_endpgm
14669 ;
14670 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14671 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
14672 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
14673 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
14674 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
14675 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
14676 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
14677 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
14678 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
14679 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
14680 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
14681 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
14682 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
14683 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
14684 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
14685 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
14686 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14687 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
14688 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
14689 ; SKIP-CACHE-INV-NEXT: s_endpgm
14690 ;
14691 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14692 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
14693 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14694 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14695 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14696 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14697 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14698 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14699 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14700 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14701 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14702 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14703 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14704 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
14705 ;
14706 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14707 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
14708 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14709 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14710 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14711 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14712 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14713 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14714 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14715 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14716 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14717 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14718 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14719 ; GFX90A-TGSPLIT-NEXT: s_endpgm
14720 ;
14721 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14722 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
14723 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14724 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14725 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14726 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14727 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14728 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14729 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14730 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14731 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14732 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14733 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14734 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
14735 ;
14736 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14737 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
14738 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14739 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14740 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14741 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14742 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14743 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14744 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14745 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14746 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14747 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14748 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14749 ; GFX940-TGSPLIT-NEXT: s_endpgm
14750 ;
14751 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14752 ; GFX11-WGP: ; %bb.0: ; %entry
14753 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14754 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
14755 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14756 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14757 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14758 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
14759 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
14760 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
14761 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14762 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
14763 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14764 ; GFX11-WGP-NEXT: s_endpgm
14765 ;
14766 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14767 ; GFX11-CU: ; %bb.0: ; %entry
14768 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14769 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
14770 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14771 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14772 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14773 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
14774 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
14775 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
14776 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14777 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
14778 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14779 ; GFX11-CU-NEXT: s_endpgm
14780 ;
14781 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14782 ; GFX12-WGP: ; %bb.0: ; %entry
14783 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14784 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
14785 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14786 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14787 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14788 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
14789 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
14790 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
14791 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14792 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
14793 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14794 ; GFX12-WGP-NEXT: s_endpgm
14795 ;
14796 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
14797 ; GFX12-CU: ; %bb.0: ; %entry
14798 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14799 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
14800 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14801 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14802 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14803 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
14804 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
14805 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
14806 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14807 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
14808 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14809 ; GFX12-CU-NEXT: s_endpgm
14810 ptr addrspace(1) %out, i32 %in, i32 %old) {
14811 entry:
14812 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
14813 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
14814 ret void
14815 }
14817 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_cmpxchg(
14818 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14819 ; GFX6: ; %bb.0: ; %entry
14820 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
14821 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
14822 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
14823 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
14824 ; GFX6-NEXT: s_mov_b32 s12, s5
14825 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
14826 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
14827 ; GFX6-NEXT: s_mov_b32 s11, -1
14828 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
14829 ; GFX6-NEXT: s_mov_b32 s5, s12
14830 ; GFX6-NEXT: s_mov_b32 s6, s11
14831 ; GFX6-NEXT: s_mov_b32 s7, s10
14832 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
14833 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
14834 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14835 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
14836 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
14837 ; GFX6-NEXT: s_endpgm
14838 ;
14839 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14840 ; GFX7: ; %bb.0: ; %entry
14841 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
14842 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
14843 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
14844 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
14845 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
14846 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
14847 ; GFX7-NEXT: s_mov_b32 s4, s8
14848 ; GFX7-NEXT: s_mov_b32 s5, s9
14849 ; GFX7-NEXT: s_mov_b32 s9, s10
14850 ; GFX7-NEXT: s_mov_b32 s8, s11
14851 ; GFX7-NEXT: s_add_u32 s4, s4, s9
14852 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
14853 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
14854 ; GFX7-NEXT: s_mov_b32 s5, s8
14855 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
14856 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
14857 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14858 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
14859 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
14860 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
14861 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
14862 ; GFX7-NEXT: s_endpgm
14863 ;
14864 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14865 ; GFX10-WGP: ; %bb.0: ; %entry
14866 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
14867 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
14868 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14869 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
14870 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
14871 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
14872 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
14873 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
14874 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14875 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
14876 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14877 ; GFX10-WGP-NEXT: s_endpgm
14878 ;
14879 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14880 ; GFX10-CU: ; %bb.0: ; %entry
14881 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
14882 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
14883 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14884 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
14885 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
14886 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
14887 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
14888 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
14889 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14890 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
14891 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
14892 ; GFX10-CU-NEXT: s_endpgm
14893 ;
14894 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14895 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
14896 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
14897 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
14898 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
14899 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
14900 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
14901 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
14902 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
14903 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
14904 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
14905 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
14906 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
14907 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
14908 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
14909 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
14910 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
14911 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
14912 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
14913 ; SKIP-CACHE-INV-NEXT: s_endpgm
14914 ;
14915 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14916 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
14917 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14918 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14919 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14920 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14921 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14922 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14923 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14924 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14925 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14926 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14927 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14928 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
14929 ;
14930 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14931 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
14932 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
14933 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14934 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
14935 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
14936 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
14937 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14938 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
14939 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
14940 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14941 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14942 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
14943 ; GFX90A-TGSPLIT-NEXT: s_endpgm
14944 ;
14945 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14946 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
14947 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14948 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14949 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14950 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14951 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14952 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14953 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14954 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14955 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14956 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14957 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14958 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
14959 ;
14960 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14961 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
14962 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
14963 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
14964 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
14965 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
14966 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
14967 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
14968 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
14969 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
14970 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
14971 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
14972 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
14973 ; GFX940-TGSPLIT-NEXT: s_endpgm
14974 ;
14975 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14976 ; GFX11-WGP: ; %bb.0: ; %entry
14977 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
14978 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
14979 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14980 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
14981 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
14982 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
14983 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
14984 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
14985 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
14986 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
14987 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
14988 ; GFX11-WGP-NEXT: s_endpgm
14989 ;
14990 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
14991 ; GFX11-CU: ; %bb.0: ; %entry
14992 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
14993 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
14994 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
14995 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
14996 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
14997 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
14998 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
14999 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
15000 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15001 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
15002 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15003 ; GFX11-CU-NEXT: s_endpgm
15004 ;
15005 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
15006 ; GFX12-WGP: ; %bb.0: ; %entry
15007 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15008 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
15009 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15010 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15011 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15012 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
15013 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
15014 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
15015 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15016 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
15017 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15018 ; GFX12-WGP-NEXT: s_endpgm
15019 ;
15020 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
15021 ; GFX12-CU: ; %bb.0: ; %entry
15022 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15023 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
15024 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15025 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15026 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15027 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
15028 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
15029 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
15030 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15031 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
15032 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15033 ; GFX12-CU-NEXT: s_endpgm
15034 ptr addrspace(1) %out, i32 %in, i32 %old) {
15035 entry:
15036 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
15037 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
15038 ret void
15039 }
15041 define amdgpu_kernel void @global_singlethread_one_as_monotonic_seq_cst_cmpxchg(
15042 ; GFX6-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15043 ; GFX6: ; %bb.0: ; %entry
15044 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
15045 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
15046 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
15047 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
15048 ; GFX6-NEXT: s_mov_b32 s12, s5
15049 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
15050 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
15051 ; GFX6-NEXT: s_mov_b32 s11, -1
15052 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
15053 ; GFX6-NEXT: s_mov_b32 s5, s12
15054 ; GFX6-NEXT: s_mov_b32 s6, s11
15055 ; GFX6-NEXT: s_mov_b32 s7, s10
15056 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
15057 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
15058 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15059 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
15060 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
15061 ; GFX6-NEXT: s_endpgm
15062 ;
15063 ; GFX7-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15064 ; GFX7: ; %bb.0: ; %entry
15065 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
15066 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
15067 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
15068 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
15069 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
15070 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
15071 ; GFX7-NEXT: s_mov_b32 s4, s8
15072 ; GFX7-NEXT: s_mov_b32 s5, s9
15073 ; GFX7-NEXT: s_mov_b32 s9, s10
15074 ; GFX7-NEXT: s_mov_b32 s8, s11
15075 ; GFX7-NEXT: s_add_u32 s4, s4, s9
15076 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
15077 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
15078 ; GFX7-NEXT: s_mov_b32 s5, s8
15079 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
15080 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
15081 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15082 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
15083 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
15084 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
15085 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
15086 ; GFX7-NEXT: s_endpgm
15087 ;
15088 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15089 ; GFX10-WGP: ; %bb.0: ; %entry
15090 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
15091 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
15092 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15093 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
15094 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
15095 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
15096 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
15097 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
15098 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15099 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
15100 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15101 ; GFX10-WGP-NEXT: s_endpgm
15102 ;
15103 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15104 ; GFX10-CU: ; %bb.0: ; %entry
15105 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
15106 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
15107 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15108 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
15109 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
15110 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
15111 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
15112 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
15113 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15114 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
15115 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15116 ; GFX10-CU-NEXT: s_endpgm
15117 ;
15118 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15119 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
15120 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
15121 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
15122 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
15123 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
15124 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
15125 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
15126 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
15127 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
15128 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
15129 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
15130 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
15131 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
15132 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
15133 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
15134 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15135 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
15136 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
15137 ; SKIP-CACHE-INV-NEXT: s_endpgm
15138 ;
15139 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15140 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
15141 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15142 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15143 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15144 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15145 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15146 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15147 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15148 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15149 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15150 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15151 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15152 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
15153 ;
15154 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15155 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
15156 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15157 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15158 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15159 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15160 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15161 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15162 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15163 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15164 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15165 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15166 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15167 ; GFX90A-TGSPLIT-NEXT: s_endpgm
15168 ;
15169 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15170 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
15171 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15172 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15173 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15174 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15175 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15176 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15177 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15178 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15179 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15180 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15181 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15182 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
15183 ;
15184 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15185 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
15186 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15187 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15188 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15189 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15190 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15191 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15192 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15193 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15194 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15195 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15196 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15197 ; GFX940-TGSPLIT-NEXT: s_endpgm
15198 ;
15199 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15200 ; GFX11-WGP: ; %bb.0: ; %entry
15201 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15202 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
15203 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15204 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15205 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15206 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
15207 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
15208 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
15209 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15210 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
15211 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15212 ; GFX11-WGP-NEXT: s_endpgm
15213 ;
15214 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15215 ; GFX11-CU: ; %bb.0: ; %entry
15216 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15217 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
15218 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15219 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15220 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15221 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
15222 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
15223 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
15224 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15225 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
15226 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15227 ; GFX11-CU-NEXT: s_endpgm
15228 ;
15229 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15230 ; GFX12-WGP: ; %bb.0: ; %entry
15231 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15232 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
15233 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15234 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15235 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15236 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
15237 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
15238 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
15239 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15240 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
15241 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15242 ; GFX12-WGP-NEXT: s_endpgm
15243 ;
15244 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
15245 ; GFX12-CU: ; %bb.0: ; %entry
15246 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15247 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
15248 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15249 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15250 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15251 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
15252 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
15253 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
15254 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15255 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
15256 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15257 ; GFX12-CU-NEXT: s_endpgm
15258 ptr addrspace(1) %out, i32 %in, i32 %old) {
15259 entry:
15260 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
15261 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic seq_cst
15262 ret void
15263 }
15265 define amdgpu_kernel void @global_singlethread_one_as_acquire_seq_cst_cmpxchg(
15266 ; GFX6-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15267 ; GFX6: ; %bb.0: ; %entry
15268 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
15269 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
15270 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
15271 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
15272 ; GFX6-NEXT: s_mov_b32 s12, s5
15273 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
15274 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
15275 ; GFX6-NEXT: s_mov_b32 s11, -1
15276 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
15277 ; GFX6-NEXT: s_mov_b32 s5, s12
15278 ; GFX6-NEXT: s_mov_b32 s6, s11
15279 ; GFX6-NEXT: s_mov_b32 s7, s10
15280 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
15281 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
15282 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15283 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
15284 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
15285 ; GFX6-NEXT: s_endpgm
15286 ;
15287 ; GFX7-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15288 ; GFX7: ; %bb.0: ; %entry
15289 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
15290 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
15291 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
15292 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
15293 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
15294 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
15295 ; GFX7-NEXT: s_mov_b32 s4, s8
15296 ; GFX7-NEXT: s_mov_b32 s5, s9
15297 ; GFX7-NEXT: s_mov_b32 s9, s10
15298 ; GFX7-NEXT: s_mov_b32 s8, s11
15299 ; GFX7-NEXT: s_add_u32 s4, s4, s9
15300 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
15301 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
15302 ; GFX7-NEXT: s_mov_b32 s5, s8
15303 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
15304 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
15305 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15306 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
15307 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
15308 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
15309 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
15310 ; GFX7-NEXT: s_endpgm
15311 ;
15312 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15313 ; GFX10-WGP: ; %bb.0: ; %entry
15314 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
15315 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
15316 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15317 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
15318 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
15319 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
15320 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
15321 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
15322 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15323 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
15324 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15325 ; GFX10-WGP-NEXT: s_endpgm
15326 ;
15327 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15328 ; GFX10-CU: ; %bb.0: ; %entry
15329 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
15330 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
15331 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15332 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
15333 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
15334 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
15335 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
15336 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
15337 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15338 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
15339 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15340 ; GFX10-CU-NEXT: s_endpgm
15341 ;
15342 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15343 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
15344 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
15345 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
15346 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
15347 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
15348 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
15349 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
15350 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
15351 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
15352 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
15353 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
15354 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
15355 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
15356 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
15357 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
15358 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15359 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
15360 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
15361 ; SKIP-CACHE-INV-NEXT: s_endpgm
15362 ;
15363 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15364 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
15365 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15366 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15367 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15368 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15369 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15370 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15371 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15372 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15373 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15374 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15375 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15376 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
15377 ;
15378 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15379 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
15380 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15381 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15382 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15383 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15384 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15385 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15386 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15387 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15388 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15389 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15390 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15391 ; GFX90A-TGSPLIT-NEXT: s_endpgm
15392 ;
15393 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15394 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
15395 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15396 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15397 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15398 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15399 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15400 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15401 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15402 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15403 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15404 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15405 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15406 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
15407 ;
15408 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15409 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
15410 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15411 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15412 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15413 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15414 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15415 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15416 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15417 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15418 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15419 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15420 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15421 ; GFX940-TGSPLIT-NEXT: s_endpgm
15422 ;
15423 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15424 ; GFX11-WGP: ; %bb.0: ; %entry
15425 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15426 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
15427 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15428 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15429 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15430 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
15431 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
15432 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
15433 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15434 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
15435 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15436 ; GFX11-WGP-NEXT: s_endpgm
15437 ;
15438 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15439 ; GFX11-CU: ; %bb.0: ; %entry
15440 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15441 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
15442 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15443 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15444 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15445 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
15446 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
15447 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
15448 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15449 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
15450 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15451 ; GFX11-CU-NEXT: s_endpgm
15452 ;
15453 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15454 ; GFX12-WGP: ; %bb.0: ; %entry
15455 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15456 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
15457 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15458 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15459 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15460 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
15461 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
15462 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
15463 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15464 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
15465 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15466 ; GFX12-WGP-NEXT: s_endpgm
15467 ;
15468 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
15469 ; GFX12-CU: ; %bb.0: ; %entry
15470 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15471 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
15472 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15473 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15474 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15475 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
15476 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
15477 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
15478 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15479 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
15480 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15481 ; GFX12-CU-NEXT: s_endpgm
15482 ptr addrspace(1) %out, i32 %in, i32 %old) {
15483 entry:
15484 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
15485 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire seq_cst
15486 ret void
15487 }
15489 define amdgpu_kernel void @global_singlethread_one_as_release_seq_cst_cmpxchg(
15490 ; GFX6-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15491 ; GFX6: ; %bb.0: ; %entry
15492 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
15493 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
15494 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
15495 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
15496 ; GFX6-NEXT: s_mov_b32 s12, s5
15497 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
15498 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
15499 ; GFX6-NEXT: s_mov_b32 s11, -1
15500 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
15501 ; GFX6-NEXT: s_mov_b32 s5, s12
15502 ; GFX6-NEXT: s_mov_b32 s6, s11
15503 ; GFX6-NEXT: s_mov_b32 s7, s10
15504 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
15505 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
15506 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15507 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
15508 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
15509 ; GFX6-NEXT: s_endpgm
15510 ;
15511 ; GFX7-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15512 ; GFX7: ; %bb.0: ; %entry
15513 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
15514 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
15515 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
15516 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
15517 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
15518 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
15519 ; GFX7-NEXT: s_mov_b32 s4, s8
15520 ; GFX7-NEXT: s_mov_b32 s5, s9
15521 ; GFX7-NEXT: s_mov_b32 s9, s10
15522 ; GFX7-NEXT: s_mov_b32 s8, s11
15523 ; GFX7-NEXT: s_add_u32 s4, s4, s9
15524 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
15525 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
15526 ; GFX7-NEXT: s_mov_b32 s5, s8
15527 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
15528 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
15529 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15530 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
15531 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
15532 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
15533 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
15534 ; GFX7-NEXT: s_endpgm
15535 ;
15536 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15537 ; GFX10-WGP: ; %bb.0: ; %entry
15538 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
15539 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
15540 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15541 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
15542 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
15543 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
15544 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
15545 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
15546 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15547 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
15548 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15549 ; GFX10-WGP-NEXT: s_endpgm
15550 ;
15551 ; GFX10-CU-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15552 ; GFX10-CU: ; %bb.0: ; %entry
15553 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
15554 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
15555 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15556 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
15557 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
15558 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
15559 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
15560 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
15561 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15562 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
15563 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15564 ; GFX10-CU-NEXT: s_endpgm
15565 ;
15566 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15567 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
15568 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
15569 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
15570 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
15571 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
15572 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
15573 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
15574 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
15575 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
15576 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
15577 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
15578 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
15579 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
15580 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
15581 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
15582 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15583 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
15584 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
15585 ; SKIP-CACHE-INV-NEXT: s_endpgm
15586 ;
15587 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15588 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
15589 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15590 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15591 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15592 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15593 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15594 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15595 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15596 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15597 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15598 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15599 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15600 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
15601 ;
15602 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15603 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
15604 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15605 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15606 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15607 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15608 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15609 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15610 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15611 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15612 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15613 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15614 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15615 ; GFX90A-TGSPLIT-NEXT: s_endpgm
15616 ;
15617 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15618 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
15619 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15620 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15621 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15622 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15623 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15624 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15625 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15626 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15627 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15628 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15629 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15630 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
15631 ;
15632 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15633 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
15634 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15635 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15636 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15637 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15638 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15639 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15640 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15641 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15642 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15643 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15644 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15645 ; GFX940-TGSPLIT-NEXT: s_endpgm
15646 ;
15647 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15648 ; GFX11-WGP: ; %bb.0: ; %entry
15649 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15650 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
15651 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15652 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15653 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15654 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
15655 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
15656 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
15657 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15658 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
15659 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15660 ; GFX11-WGP-NEXT: s_endpgm
15661 ;
15662 ; GFX11-CU-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15663 ; GFX11-CU: ; %bb.0: ; %entry
15664 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15665 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
15666 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15667 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15668 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15669 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
15670 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
15671 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
15672 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15673 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
15674 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15675 ; GFX11-CU-NEXT: s_endpgm
15676 ;
15677 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15678 ; GFX12-WGP: ; %bb.0: ; %entry
15679 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15680 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
15681 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15682 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15683 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15684 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
15685 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
15686 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
15687 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15688 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
15689 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15690 ; GFX12-WGP-NEXT: s_endpgm
15691 ;
15692 ; GFX12-CU-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
15693 ; GFX12-CU: ; %bb.0: ; %entry
15694 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15695 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
15696 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15697 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15698 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15699 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
15700 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
15701 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
15702 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15703 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
15704 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15705 ; GFX12-CU-NEXT: s_endpgm
15706 ptr addrspace(1) %out, i32 %in, i32 %old) {
15707 entry:
15708 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
15709 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release seq_cst
15710 ret void
15711 }
15713 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_seq_cst_cmpxchg(
15714 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15715 ; GFX6: ; %bb.0: ; %entry
15716 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
15717 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
15718 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
15719 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
15720 ; GFX6-NEXT: s_mov_b32 s12, s5
15721 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
15722 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
15723 ; GFX6-NEXT: s_mov_b32 s11, -1
15724 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
15725 ; GFX6-NEXT: s_mov_b32 s5, s12
15726 ; GFX6-NEXT: s_mov_b32 s6, s11
15727 ; GFX6-NEXT: s_mov_b32 s7, s10
15728 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
15729 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
15730 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15731 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
15732 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
15733 ; GFX6-NEXT: s_endpgm
15734 ;
15735 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15736 ; GFX7: ; %bb.0: ; %entry
15737 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
15738 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
15739 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
15740 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
15741 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
15742 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
15743 ; GFX7-NEXT: s_mov_b32 s4, s8
15744 ; GFX7-NEXT: s_mov_b32 s5, s9
15745 ; GFX7-NEXT: s_mov_b32 s9, s10
15746 ; GFX7-NEXT: s_mov_b32 s8, s11
15747 ; GFX7-NEXT: s_add_u32 s4, s4, s9
15748 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
15749 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
15750 ; GFX7-NEXT: s_mov_b32 s5, s8
15751 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
15752 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
15753 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15754 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
15755 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
15756 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
15757 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
15758 ; GFX7-NEXT: s_endpgm
15759 ;
15760 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15761 ; GFX10-WGP: ; %bb.0: ; %entry
15762 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
15763 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
15764 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15765 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
15766 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
15767 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
15768 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
15769 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
15770 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15771 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
15772 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15773 ; GFX10-WGP-NEXT: s_endpgm
15774 ;
15775 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15776 ; GFX10-CU: ; %bb.0: ; %entry
15777 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
15778 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
15779 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15780 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
15781 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
15782 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
15783 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
15784 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
15785 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15786 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
15787 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15788 ; GFX10-CU-NEXT: s_endpgm
15789 ;
15790 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15791 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
15792 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
15793 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
15794 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
15795 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
15796 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
15797 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
15798 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
15799 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
15800 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
15801 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
15802 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
15803 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
15804 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
15805 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
15806 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15807 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
15808 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
15809 ; SKIP-CACHE-INV-NEXT: s_endpgm
15810 ;
15811 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15812 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
15813 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15814 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15815 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15816 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15817 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15818 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15819 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15820 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15821 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15822 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15823 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15824 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
15825 ;
15826 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15827 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
15828 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
15829 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15830 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15831 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
15832 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
15833 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15834 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
15835 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
15836 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15837 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15838 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
15839 ; GFX90A-TGSPLIT-NEXT: s_endpgm
15840 ;
15841 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15842 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
15843 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15844 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15845 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15846 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15847 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15848 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15849 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15850 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15851 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15852 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15853 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15854 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
15855 ;
15856 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15857 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
15858 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
15859 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
15860 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
15861 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
15862 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
15863 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
15864 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
15865 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
15866 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15867 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
15868 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
15869 ; GFX940-TGSPLIT-NEXT: s_endpgm
15870 ;
15871 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15872 ; GFX11-WGP: ; %bb.0: ; %entry
15873 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15874 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
15875 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15876 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15877 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15878 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
15879 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
15880 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
15881 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15882 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
15883 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15884 ; GFX11-WGP-NEXT: s_endpgm
15885 ;
15886 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15887 ; GFX11-CU: ; %bb.0: ; %entry
15888 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15889 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
15890 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15891 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15892 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15893 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
15894 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
15895 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
15896 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15897 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
15898 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15899 ; GFX11-CU-NEXT: s_endpgm
15900 ;
15901 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15902 ; GFX12-WGP: ; %bb.0: ; %entry
15903 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
15904 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
15905 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15906 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
15907 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
15908 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
15909 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
15910 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
15911 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15912 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
15913 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15914 ; GFX12-WGP-NEXT: s_endpgm
15915 ;
15916 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
15917 ; GFX12-CU: ; %bb.0: ; %entry
15918 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
15919 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
15920 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
15921 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
15922 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
15923 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
15924 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
15925 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
15926 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15927 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
15928 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
15929 ; GFX12-CU-NEXT: s_endpgm
15930 ptr addrspace(1) %out, i32 %in, i32 %old) {
15931 entry:
15932 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
15933 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel seq_cst
15934 ret void
15935 }
15937 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_cmpxchg(
15938 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
15939 ; GFX6: ; %bb.0: ; %entry
15940 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
15941 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
15942 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
15943 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
15944 ; GFX6-NEXT: s_mov_b32 s12, s5
15945 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
15946 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
15947 ; GFX6-NEXT: s_mov_b32 s11, -1
15948 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
15949 ; GFX6-NEXT: s_mov_b32 s5, s12
15950 ; GFX6-NEXT: s_mov_b32 s6, s11
15951 ; GFX6-NEXT: s_mov_b32 s7, s10
15952 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
15953 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
15954 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
15955 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
15956 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
15957 ; GFX6-NEXT: s_endpgm
15958 ;
15959 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
15960 ; GFX7: ; %bb.0: ; %entry
15961 ; GFX7-NEXT: s_mov_b64 s[4:5], s[6:7]
15962 ; GFX7-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0x0
15963 ; GFX7-NEXT: s_load_dword s7, s[4:5], 0x2
15964 ; GFX7-NEXT: s_load_dword s6, s[4:5], 0x3
15965 ; GFX7-NEXT: s_mov_b64 s[10:11], 16
15966 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
15967 ; GFX7-NEXT: s_mov_b32 s4, s8
15968 ; GFX7-NEXT: s_mov_b32 s5, s9
15969 ; GFX7-NEXT: s_mov_b32 s9, s10
15970 ; GFX7-NEXT: s_mov_b32 s8, s11
15971 ; GFX7-NEXT: s_add_u32 s4, s4, s9
15972 ; GFX7-NEXT: s_addc_u32 s8, s5, s8
15973 ; GFX7-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
15974 ; GFX7-NEXT: s_mov_b32 s5, s8
15975 ; GFX7-NEXT: v_mov_b32_e32 v2, s7
15976 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
15977 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
15978 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
15979 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
15980 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
15981 ; GFX7-NEXT: flat_atomic_cmpswap v[0:1], v[2:3]
15982 ; GFX7-NEXT: s_endpgm
15983 ;
15984 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
15985 ; GFX10-WGP: ; %bb.0: ; %entry
15986 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
15987 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
15988 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
15989 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
15990 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
15991 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
15992 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
15993 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
15994 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
15995 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
15996 ; GFX10-WGP-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
15997 ; GFX10-WGP-NEXT: s_endpgm
15998 ;
15999 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16000 ; GFX10-CU: ; %bb.0: ; %entry
16001 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
16002 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
16003 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16004 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
16005 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
16006 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
16007 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
16008 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
16009 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16010 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
16011 ; GFX10-CU-NEXT: global_atomic_cmpswap v0, v[1:2], s[4:5] offset:16
16012 ; GFX10-CU-NEXT: s_endpgm
16013 ;
16014 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16015 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
16016 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
16017 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
16018 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
16019 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
16020 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
16021 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
16022 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
16023 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
16024 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
16025 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
16026 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
16027 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
16028 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
16029 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
16030 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16031 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
16032 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16
16033 ; SKIP-CACHE-INV-NEXT: s_endpgm
16034 ;
16035 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16036 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
16037 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16038 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16039 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16040 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16041 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16042 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16043 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16044 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16045 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16046 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16047 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
16048 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
16049 ;
16050 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16051 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
16052 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16053 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16054 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16055 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16056 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16057 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16058 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16059 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16060 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16061 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16062 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[4:5] offset:16
16063 ; GFX90A-TGSPLIT-NEXT: s_endpgm
16064 ;
16065 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16066 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
16067 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16068 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16069 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16070 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16071 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16072 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16073 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16074 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16075 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16076 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16077 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
16078 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
16079 ;
16080 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16081 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
16082 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16083 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16084 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16085 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16086 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16087 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16088 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16089 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16090 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16091 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16092 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v0, v[2:3], s[0:1] offset:16
16093 ; GFX940-TGSPLIT-NEXT: s_endpgm
16094 ;
16095 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16096 ; GFX11-WGP: ; %bb.0: ; %entry
16097 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16098 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
16099 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16100 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16101 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16102 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
16103 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
16104 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
16105 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16106 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
16107 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
16108 ; GFX11-WGP-NEXT: s_endpgm
16109 ;
16110 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16111 ; GFX11-CU: ; %bb.0: ; %entry
16112 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16113 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
16114 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16115 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16116 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16117 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
16118 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
16119 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
16120 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16121 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
16122 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
16123 ; GFX11-CU-NEXT: s_endpgm
16124 ;
16125 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16126 ; GFX12-WGP: ; %bb.0: ; %entry
16127 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16128 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
16129 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16130 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16131 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16132 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
16133 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
16134 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
16135 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16136 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
16137 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
16138 ; GFX12-WGP-NEXT: s_endpgm
16139 ;
16140 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
16141 ; GFX12-CU: ; %bb.0: ; %entry
16142 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16143 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
16144 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16145 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16146 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16147 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
16148 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
16149 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
16150 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16151 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
16152 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v0, v[1:2], s[0:1] offset:16
16153 ; GFX12-CU-NEXT: s_endpgm
16154 ptr addrspace(1) %out, i32 %in, i32 %old) {
16155 entry:
16156 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
16157 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
16158 ret void
16159 }
16161 define amdgpu_kernel void @global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg(
16162 ; GFX6-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16163 ; GFX6: ; %bb.0: ; %entry
16164 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16165 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
16166 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
16167 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
16168 ; GFX6-NEXT: s_mov_b32 s12, s5
16169 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
16170 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
16171 ; GFX6-NEXT: s_mov_b32 s11, -1
16172 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
16173 ; GFX6-NEXT: s_mov_b32 s5, s12
16174 ; GFX6-NEXT: s_mov_b32 s6, s11
16175 ; GFX6-NEXT: s_mov_b32 s7, s10
16176 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
16177 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
16178 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16179 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
16180 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
16181 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16182 ; GFX6-NEXT: s_waitcnt vmcnt(0)
16183 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
16184 ; GFX6-NEXT: s_endpgm
16185 ;
16186 ; GFX7-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16187 ; GFX7: ; %bb.0: ; %entry
16188 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16189 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
16190 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
16191 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
16192 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
16193 ; GFX7-NEXT: s_mov_b32 s6, s4
16194 ; GFX7-NEXT: s_mov_b32 s7, s5
16195 ; GFX7-NEXT: s_mov_b32 s11, s12
16196 ; GFX7-NEXT: s_mov_b32 s10, s13
16197 ; GFX7-NEXT: s_add_u32 s6, s6, s11
16198 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
16199 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
16200 ; GFX7-NEXT: s_mov_b32 s7, s10
16201 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
16202 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
16203 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16204 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
16205 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
16206 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
16207 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
16208 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
16209 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
16210 ; GFX7-NEXT: s_waitcnt vmcnt(0)
16211 ; GFX7-NEXT: flat_store_dword v[0:1], v2
16212 ; GFX7-NEXT: s_endpgm
16213 ;
16214 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16215 ; GFX10-WGP: ; %bb.0: ; %entry
16216 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
16217 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
16218 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16219 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
16220 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
16221 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
16222 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
16223 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
16224 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16225 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
16226 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16227 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
16228 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
16229 ; GFX10-WGP-NEXT: s_endpgm
16230 ;
16231 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16232 ; GFX10-CU: ; %bb.0: ; %entry
16233 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
16234 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
16235 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16236 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
16237 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
16238 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
16239 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
16240 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
16241 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16242 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
16243 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16244 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
16245 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
16246 ; GFX10-CU-NEXT: s_endpgm
16247 ;
16248 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16249 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
16250 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
16251 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
16252 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
16253 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
16254 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
16255 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
16256 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
16257 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
16258 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
16259 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
16260 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
16261 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
16262 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
16263 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
16264 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16265 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
16266 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
16267 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16268 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
16269 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
16270 ; SKIP-CACHE-INV-NEXT: s_endpgm
16271 ;
16272 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16273 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
16274 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16275 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16276 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16277 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16278 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16279 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16280 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16281 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16282 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16283 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16284 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16285 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16286 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16287 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
16288 ;
16289 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16290 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
16291 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16292 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16293 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16294 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16295 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16296 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16297 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16298 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16299 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16300 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16301 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16302 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16303 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16304 ; GFX90A-TGSPLIT-NEXT: s_endpgm
16305 ;
16306 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16307 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
16308 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16309 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16310 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16311 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16312 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16313 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16314 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16315 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16316 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16317 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16318 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16319 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16320 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16321 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
16322 ;
16323 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16324 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
16325 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16326 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16327 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16328 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16329 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16330 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16331 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16332 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16333 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16334 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16335 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16336 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16337 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16338 ; GFX940-TGSPLIT-NEXT: s_endpgm
16339 ;
16340 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16341 ; GFX11-WGP: ; %bb.0: ; %entry
16342 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16343 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
16344 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16345 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16346 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16347 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
16348 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
16349 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
16350 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16351 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
16352 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16353 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
16354 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16355 ; GFX11-WGP-NEXT: s_endpgm
16356 ;
16357 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16358 ; GFX11-CU: ; %bb.0: ; %entry
16359 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16360 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
16361 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16362 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16363 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16364 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
16365 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
16366 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
16367 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16368 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
16369 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16370 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
16371 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16372 ; GFX11-CU-NEXT: s_endpgm
16373 ;
16374 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16375 ; GFX12-WGP: ; %bb.0: ; %entry
16376 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16377 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
16378 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16379 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16380 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16381 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
16382 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
16383 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
16384 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16385 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
16386 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16387 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
16388 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16389 ; GFX12-WGP-NEXT: s_endpgm
16390 ;
16391 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
16392 ; GFX12-CU: ; %bb.0: ; %entry
16393 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16394 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
16395 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16396 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16397 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16398 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
16399 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
16400 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
16401 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16402 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
16403 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16404 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
16405 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16406 ; GFX12-CU-NEXT: s_endpgm
16407 ptr addrspace(1) %out, i32 %in, i32 %old) {
16408 entry:
16409 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
16410 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
16411 %val0 = extractvalue { i32, i1 } %val, 0
16412 store i32 %val0, ptr addrspace(1) %out, align 4
16413 ret void
16414 }
16416 define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_ret_cmpxchg(
16417 ; GFX6-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16418 ; GFX6: ; %bb.0: ; %entry
16419 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16420 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
16421 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
16422 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
16423 ; GFX6-NEXT: s_mov_b32 s12, s5
16424 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
16425 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
16426 ; GFX6-NEXT: s_mov_b32 s11, -1
16427 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
16428 ; GFX6-NEXT: s_mov_b32 s5, s12
16429 ; GFX6-NEXT: s_mov_b32 s6, s11
16430 ; GFX6-NEXT: s_mov_b32 s7, s10
16431 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
16432 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
16433 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16434 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
16435 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
16436 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16437 ; GFX6-NEXT: s_waitcnt vmcnt(0)
16438 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
16439 ; GFX6-NEXT: s_endpgm
16440 ;
16441 ; GFX7-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16442 ; GFX7: ; %bb.0: ; %entry
16443 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16444 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
16445 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
16446 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
16447 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
16448 ; GFX7-NEXT: s_mov_b32 s6, s4
16449 ; GFX7-NEXT: s_mov_b32 s7, s5
16450 ; GFX7-NEXT: s_mov_b32 s11, s12
16451 ; GFX7-NEXT: s_mov_b32 s10, s13
16452 ; GFX7-NEXT: s_add_u32 s6, s6, s11
16453 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
16454 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
16455 ; GFX7-NEXT: s_mov_b32 s7, s10
16456 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
16457 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
16458 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16459 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
16460 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
16461 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
16462 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
16463 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
16464 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
16465 ; GFX7-NEXT: s_waitcnt vmcnt(0)
16466 ; GFX7-NEXT: flat_store_dword v[0:1], v2
16467 ; GFX7-NEXT: s_endpgm
16468 ;
16469 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16470 ; GFX10-WGP: ; %bb.0: ; %entry
16471 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
16472 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
16473 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16474 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
16475 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
16476 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
16477 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
16478 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
16479 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16480 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
16481 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16482 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
16483 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
16484 ; GFX10-WGP-NEXT: s_endpgm
16485 ;
16486 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16487 ; GFX10-CU: ; %bb.0: ; %entry
16488 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
16489 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
16490 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16491 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
16492 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
16493 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
16494 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
16495 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
16496 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16497 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
16498 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16499 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
16500 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
16501 ; GFX10-CU-NEXT: s_endpgm
16502 ;
16503 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16504 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
16505 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
16506 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
16507 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
16508 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
16509 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
16510 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
16511 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
16512 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
16513 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
16514 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
16515 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
16516 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
16517 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
16518 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
16519 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16520 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
16521 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
16522 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16523 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
16524 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
16525 ; SKIP-CACHE-INV-NEXT: s_endpgm
16526 ;
16527 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16528 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
16529 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16530 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16531 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16532 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16533 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16534 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16535 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16536 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16537 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16538 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16539 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16540 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16541 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16542 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
16543 ;
16544 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16545 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
16546 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16547 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16548 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16549 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16550 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16551 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16552 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16553 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16554 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16555 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16556 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16557 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16558 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16559 ; GFX90A-TGSPLIT-NEXT: s_endpgm
16560 ;
16561 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16562 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
16563 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16564 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16565 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16566 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16567 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16568 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16569 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16570 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16571 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16572 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16573 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16574 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16575 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16576 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
16577 ;
16578 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16579 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
16580 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16581 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16582 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16583 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16584 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16585 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16586 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16587 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16588 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16589 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16590 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16591 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16592 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16593 ; GFX940-TGSPLIT-NEXT: s_endpgm
16594 ;
16595 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16596 ; GFX11-WGP: ; %bb.0: ; %entry
16597 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16598 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
16599 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16600 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16601 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16602 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
16603 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
16604 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
16605 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16606 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
16607 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16608 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
16609 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16610 ; GFX11-WGP-NEXT: s_endpgm
16611 ;
16612 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16613 ; GFX11-CU: ; %bb.0: ; %entry
16614 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16615 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
16616 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16617 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16618 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16619 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
16620 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
16621 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
16622 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16623 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
16624 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16625 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
16626 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16627 ; GFX11-CU-NEXT: s_endpgm
16628 ;
16629 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16630 ; GFX12-WGP: ; %bb.0: ; %entry
16631 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16632 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
16633 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16634 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16635 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16636 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
16637 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
16638 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
16639 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16640 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
16641 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16642 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
16643 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16644 ; GFX12-WGP-NEXT: s_endpgm
16645 ;
16646 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
16647 ; GFX12-CU: ; %bb.0: ; %entry
16648 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16649 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
16650 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16651 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16652 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16653 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
16654 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
16655 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
16656 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16657 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
16658 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16659 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
16660 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16661 ; GFX12-CU-NEXT: s_endpgm
16662 ptr addrspace(1) %out, i32 %in, i32 %old) {
16663 entry:
16664 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
16665 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
16666 %val0 = extractvalue { i32, i1 } %val, 0
16667 store i32 %val0, ptr addrspace(1) %out, align 4
16668 ret void
16669 }
16671 define amdgpu_kernel void @global_singlethread_one_as_release_monotonic_ret_cmpxchg(
16672 ; GFX6-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16673 ; GFX6: ; %bb.0: ; %entry
16674 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16675 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
16676 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
16677 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
16678 ; GFX6-NEXT: s_mov_b32 s12, s5
16679 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
16680 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
16681 ; GFX6-NEXT: s_mov_b32 s11, -1
16682 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
16683 ; GFX6-NEXT: s_mov_b32 s5, s12
16684 ; GFX6-NEXT: s_mov_b32 s6, s11
16685 ; GFX6-NEXT: s_mov_b32 s7, s10
16686 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
16687 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
16688 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16689 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
16690 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
16691 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16692 ; GFX6-NEXT: s_waitcnt vmcnt(0)
16693 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
16694 ; GFX6-NEXT: s_endpgm
16695 ;
16696 ; GFX7-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16697 ; GFX7: ; %bb.0: ; %entry
16698 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16699 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
16700 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
16701 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
16702 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
16703 ; GFX7-NEXT: s_mov_b32 s6, s4
16704 ; GFX7-NEXT: s_mov_b32 s7, s5
16705 ; GFX7-NEXT: s_mov_b32 s11, s12
16706 ; GFX7-NEXT: s_mov_b32 s10, s13
16707 ; GFX7-NEXT: s_add_u32 s6, s6, s11
16708 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
16709 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
16710 ; GFX7-NEXT: s_mov_b32 s7, s10
16711 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
16712 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
16713 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16714 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
16715 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
16716 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
16717 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
16718 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
16719 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
16720 ; GFX7-NEXT: s_waitcnt vmcnt(0)
16721 ; GFX7-NEXT: flat_store_dword v[0:1], v2
16722 ; GFX7-NEXT: s_endpgm
16723 ;
16724 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16725 ; GFX10-WGP: ; %bb.0: ; %entry
16726 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
16727 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
16728 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16729 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
16730 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
16731 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
16732 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
16733 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
16734 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16735 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
16736 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16737 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
16738 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
16739 ; GFX10-WGP-NEXT: s_endpgm
16740 ;
16741 ; GFX10-CU-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16742 ; GFX10-CU: ; %bb.0: ; %entry
16743 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
16744 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
16745 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16746 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
16747 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
16748 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
16749 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
16750 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
16751 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16752 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
16753 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16754 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
16755 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
16756 ; GFX10-CU-NEXT: s_endpgm
16757 ;
16758 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16759 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
16760 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
16761 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
16762 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
16763 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
16764 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
16765 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
16766 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
16767 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
16768 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
16769 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
16770 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
16771 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
16772 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
16773 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
16774 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16775 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
16776 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
16777 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16778 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
16779 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
16780 ; SKIP-CACHE-INV-NEXT: s_endpgm
16781 ;
16782 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16783 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
16784 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16785 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16786 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16787 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16788 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16789 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16790 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16791 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16792 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16793 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16794 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16795 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16796 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16797 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
16798 ;
16799 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16800 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
16801 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
16802 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16803 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16804 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
16805 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
16806 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16807 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
16808 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
16809 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16810 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16811 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
16812 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16813 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
16814 ; GFX90A-TGSPLIT-NEXT: s_endpgm
16815 ;
16816 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16817 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
16818 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16819 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16820 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16821 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16822 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16823 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16824 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16825 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16826 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16827 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16828 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16829 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
16830 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16831 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
16832 ;
16833 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16834 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
16835 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
16836 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
16837 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
16838 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
16839 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
16840 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
16841 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
16842 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
16843 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16844 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
16845 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
16846 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
16847 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
16848 ; GFX940-TGSPLIT-NEXT: s_endpgm
16849 ;
16850 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16851 ; GFX11-WGP: ; %bb.0: ; %entry
16852 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16853 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
16854 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16855 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16856 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16857 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
16858 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
16859 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
16860 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16861 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
16862 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16863 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
16864 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16865 ; GFX11-WGP-NEXT: s_endpgm
16866 ;
16867 ; GFX11-CU-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16868 ; GFX11-CU: ; %bb.0: ; %entry
16869 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16870 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
16871 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16872 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16873 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16874 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
16875 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
16876 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
16877 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16878 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
16879 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
16880 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
16881 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16882 ; GFX11-CU-NEXT: s_endpgm
16883 ;
16884 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16885 ; GFX12-WGP: ; %bb.0: ; %entry
16886 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
16887 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
16888 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16889 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
16890 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
16891 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
16892 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
16893 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
16894 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16895 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
16896 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16897 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
16898 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
16899 ; GFX12-WGP-NEXT: s_endpgm
16900 ;
16901 ; GFX12-CU-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
16902 ; GFX12-CU: ; %bb.0: ; %entry
16903 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
16904 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
16905 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
16906 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
16907 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
16908 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
16909 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
16910 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
16911 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16912 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
16913 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
16914 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
16915 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
16916 ; GFX12-CU-NEXT: s_endpgm
16917 ptr addrspace(1) %out, i32 %in, i32 %old) {
16918 entry:
16919 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
16920 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
16921 %val0 = extractvalue { i32, i1 } %val, 0
16922 store i32 %val0, ptr addrspace(1) %out, align 4
16923 ret void
16924 }
16926 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg(
16927 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
16928 ; GFX6: ; %bb.0: ; %entry
16929 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16930 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
16931 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
16932 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
16933 ; GFX6-NEXT: s_mov_b32 s12, s5
16934 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
16935 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
16936 ; GFX6-NEXT: s_mov_b32 s11, -1
16937 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
16938 ; GFX6-NEXT: s_mov_b32 s5, s12
16939 ; GFX6-NEXT: s_mov_b32 s6, s11
16940 ; GFX6-NEXT: s_mov_b32 s7, s10
16941 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
16942 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
16943 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
16944 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
16945 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
16946 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
16947 ; GFX6-NEXT: s_waitcnt vmcnt(0)
16948 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
16949 ; GFX6-NEXT: s_endpgm
16950 ;
16951 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
16952 ; GFX7: ; %bb.0: ; %entry
16953 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
16954 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
16955 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
16956 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
16957 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
16958 ; GFX7-NEXT: s_mov_b32 s6, s4
16959 ; GFX7-NEXT: s_mov_b32 s7, s5
16960 ; GFX7-NEXT: s_mov_b32 s11, s12
16961 ; GFX7-NEXT: s_mov_b32 s10, s13
16962 ; GFX7-NEXT: s_add_u32 s6, s6, s11
16963 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
16964 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
16965 ; GFX7-NEXT: s_mov_b32 s7, s10
16966 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
16967 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
16968 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
16969 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
16970 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
16971 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
16972 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
16973 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
16974 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
16975 ; GFX7-NEXT: s_waitcnt vmcnt(0)
16976 ; GFX7-NEXT: flat_store_dword v[0:1], v2
16977 ; GFX7-NEXT: s_endpgm
16978 ;
16979 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
16980 ; GFX10-WGP: ; %bb.0: ; %entry
16981 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
16982 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
16983 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
16984 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
16985 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
16986 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
16987 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
16988 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
16989 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
16990 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
16991 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
16992 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
16993 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
16994 ; GFX10-WGP-NEXT: s_endpgm
16995 ;
16996 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
16997 ; GFX10-CU: ; %bb.0: ; %entry
16998 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
16999 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
17000 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17001 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
17002 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
17003 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
17004 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
17005 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
17006 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17007 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
17008 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17009 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
17010 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
17011 ; GFX10-CU-NEXT: s_endpgm
17012 ;
17013 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17014 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
17015 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
17016 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
17017 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
17018 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
17019 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
17020 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
17021 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
17022 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
17023 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
17024 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
17025 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
17026 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
17027 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
17028 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
17029 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17030 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
17031 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
17032 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17033 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
17034 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
17035 ; SKIP-CACHE-INV-NEXT: s_endpgm
17036 ;
17037 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17038 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
17039 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17040 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17041 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17042 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17043 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17044 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17045 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17046 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17047 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17048 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17049 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17050 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17051 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17052 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
17053 ;
17054 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17055 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
17056 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17057 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17058 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17059 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17060 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17061 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17062 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17063 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17064 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17065 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17066 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17067 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17068 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17069 ; GFX90A-TGSPLIT-NEXT: s_endpgm
17070 ;
17071 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17072 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
17073 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17074 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17075 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17076 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17077 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17078 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17079 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17080 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17081 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17082 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17083 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17084 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17085 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17086 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
17087 ;
17088 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17089 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
17090 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17091 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17092 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17093 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17094 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17095 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17096 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17097 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17098 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17099 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17100 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17101 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17102 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17103 ; GFX940-TGSPLIT-NEXT: s_endpgm
17104 ;
17105 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17106 ; GFX11-WGP: ; %bb.0: ; %entry
17107 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17108 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
17109 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17110 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17111 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17112 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
17113 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
17114 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
17115 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17116 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
17117 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17118 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
17119 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17120 ; GFX11-WGP-NEXT: s_endpgm
17121 ;
17122 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17123 ; GFX11-CU: ; %bb.0: ; %entry
17124 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17125 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
17126 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17127 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17128 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17129 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
17130 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
17131 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
17132 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17133 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
17134 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17135 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
17136 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17137 ; GFX11-CU-NEXT: s_endpgm
17138 ;
17139 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17140 ; GFX12-WGP: ; %bb.0: ; %entry
17141 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17142 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
17143 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17144 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17145 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17146 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
17147 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
17148 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
17149 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17150 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
17151 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17152 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
17153 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17154 ; GFX12-WGP-NEXT: s_endpgm
17155 ;
17156 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
17157 ; GFX12-CU: ; %bb.0: ; %entry
17158 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17159 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
17160 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17161 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17162 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17163 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
17164 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
17165 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
17166 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17167 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
17168 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17169 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
17170 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17171 ; GFX12-CU-NEXT: s_endpgm
17172 ptr addrspace(1) %out, i32 %in, i32 %old) {
17173 entry:
17174 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
17175 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
17176 %val0 = extractvalue { i32, i1 } %val, 0
17177 store i32 %val0, ptr addrspace(1) %out, align 4
17178 ret void
17179 }
17181 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg(
17182 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17183 ; GFX6: ; %bb.0: ; %entry
17184 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17185 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
17186 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
17187 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
17188 ; GFX6-NEXT: s_mov_b32 s12, s5
17189 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
17190 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
17191 ; GFX6-NEXT: s_mov_b32 s11, -1
17192 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
17193 ; GFX6-NEXT: s_mov_b32 s5, s12
17194 ; GFX6-NEXT: s_mov_b32 s6, s11
17195 ; GFX6-NEXT: s_mov_b32 s7, s10
17196 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
17197 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
17198 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17199 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
17200 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
17201 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17202 ; GFX6-NEXT: s_waitcnt vmcnt(0)
17203 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
17204 ; GFX6-NEXT: s_endpgm
17205 ;
17206 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17207 ; GFX7: ; %bb.0: ; %entry
17208 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17209 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
17210 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
17211 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
17212 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
17213 ; GFX7-NEXT: s_mov_b32 s6, s4
17214 ; GFX7-NEXT: s_mov_b32 s7, s5
17215 ; GFX7-NEXT: s_mov_b32 s11, s12
17216 ; GFX7-NEXT: s_mov_b32 s10, s13
17217 ; GFX7-NEXT: s_add_u32 s6, s6, s11
17218 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
17219 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
17220 ; GFX7-NEXT: s_mov_b32 s7, s10
17221 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
17222 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
17223 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17224 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
17225 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
17226 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
17227 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
17228 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
17229 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
17230 ; GFX7-NEXT: s_waitcnt vmcnt(0)
17231 ; GFX7-NEXT: flat_store_dword v[0:1], v2
17232 ; GFX7-NEXT: s_endpgm
17233 ;
17234 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17235 ; GFX10-WGP: ; %bb.0: ; %entry
17236 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
17237 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
17238 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17239 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
17240 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
17241 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
17242 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
17243 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
17244 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17245 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
17246 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17247 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
17248 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
17249 ; GFX10-WGP-NEXT: s_endpgm
17250 ;
17251 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17252 ; GFX10-CU: ; %bb.0: ; %entry
17253 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
17254 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
17255 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17256 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
17257 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
17258 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
17259 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
17260 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
17261 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17262 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
17263 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17264 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
17265 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
17266 ; GFX10-CU-NEXT: s_endpgm
17267 ;
17268 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17269 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
17270 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
17271 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
17272 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
17273 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
17274 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
17275 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
17276 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
17277 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
17278 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
17279 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
17280 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
17281 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
17282 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
17283 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
17284 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17285 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
17286 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
17287 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17288 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
17289 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
17290 ; SKIP-CACHE-INV-NEXT: s_endpgm
17291 ;
17292 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17293 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
17294 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17295 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17296 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17297 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17298 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17299 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17300 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17301 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17302 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17303 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17304 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17305 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17306 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17307 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
17308 ;
17309 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17310 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
17311 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17312 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17313 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17314 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17315 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17316 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17317 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17318 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17319 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17320 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17321 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17322 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17323 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17324 ; GFX90A-TGSPLIT-NEXT: s_endpgm
17325 ;
17326 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17327 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
17328 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17329 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17330 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17331 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17332 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17333 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17334 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17335 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17336 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17337 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17338 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17339 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17340 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17341 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
17342 ;
17343 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17344 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
17345 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17346 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17347 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17348 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17349 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17350 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17351 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17352 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17353 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17354 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17355 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17356 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17357 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17358 ; GFX940-TGSPLIT-NEXT: s_endpgm
17359 ;
17360 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17361 ; GFX11-WGP: ; %bb.0: ; %entry
17362 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17363 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
17364 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17365 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17366 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17367 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
17368 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
17369 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
17370 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17371 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
17372 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17373 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
17374 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17375 ; GFX11-WGP-NEXT: s_endpgm
17376 ;
17377 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17378 ; GFX11-CU: ; %bb.0: ; %entry
17379 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17380 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
17381 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17382 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17383 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17384 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
17385 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
17386 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
17387 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17388 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
17389 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17390 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
17391 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17392 ; GFX11-CU-NEXT: s_endpgm
17393 ;
17394 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17395 ; GFX12-WGP: ; %bb.0: ; %entry
17396 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17397 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
17398 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17399 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17400 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17401 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
17402 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
17403 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
17404 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17405 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
17406 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17407 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
17408 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17409 ; GFX12-WGP-NEXT: s_endpgm
17410 ;
17411 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
17412 ; GFX12-CU: ; %bb.0: ; %entry
17413 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17414 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
17415 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17416 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17417 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17418 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
17419 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
17420 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
17421 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17422 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
17423 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17424 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
17425 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17426 ; GFX12-CU-NEXT: s_endpgm
17427 ptr addrspace(1) %out, i32 %in, i32 %old) {
17428 entry:
17429 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
17430 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
17431 %val0 = extractvalue { i32, i1 } %val, 0
17432 store i32 %val0, ptr addrspace(1) %out, align 4
17433 ret void
17434 }
17436 define amdgpu_kernel void @global_singlethread_one_as_monotonic_acquire_ret_cmpxchg(
17437 ; GFX6-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17438 ; GFX6: ; %bb.0: ; %entry
17439 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17440 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
17441 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
17442 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
17443 ; GFX6-NEXT: s_mov_b32 s12, s5
17444 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
17445 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
17446 ; GFX6-NEXT: s_mov_b32 s11, -1
17447 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
17448 ; GFX6-NEXT: s_mov_b32 s5, s12
17449 ; GFX6-NEXT: s_mov_b32 s6, s11
17450 ; GFX6-NEXT: s_mov_b32 s7, s10
17451 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
17452 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
17453 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17454 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
17455 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
17456 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17457 ; GFX6-NEXT: s_waitcnt vmcnt(0)
17458 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
17459 ; GFX6-NEXT: s_endpgm
17460 ;
17461 ; GFX7-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17462 ; GFX7: ; %bb.0: ; %entry
17463 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17464 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
17465 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
17466 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
17467 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
17468 ; GFX7-NEXT: s_mov_b32 s6, s4
17469 ; GFX7-NEXT: s_mov_b32 s7, s5
17470 ; GFX7-NEXT: s_mov_b32 s11, s12
17471 ; GFX7-NEXT: s_mov_b32 s10, s13
17472 ; GFX7-NEXT: s_add_u32 s6, s6, s11
17473 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
17474 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
17475 ; GFX7-NEXT: s_mov_b32 s7, s10
17476 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
17477 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
17478 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17479 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
17480 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
17481 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
17482 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
17483 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
17484 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
17485 ; GFX7-NEXT: s_waitcnt vmcnt(0)
17486 ; GFX7-NEXT: flat_store_dword v[0:1], v2
17487 ; GFX7-NEXT: s_endpgm
17488 ;
17489 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17490 ; GFX10-WGP: ; %bb.0: ; %entry
17491 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
17492 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
17493 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17494 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
17495 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
17496 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
17497 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
17498 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
17499 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17500 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
17501 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17502 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
17503 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
17504 ; GFX10-WGP-NEXT: s_endpgm
17505 ;
17506 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17507 ; GFX10-CU: ; %bb.0: ; %entry
17508 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
17509 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
17510 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17511 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
17512 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
17513 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
17514 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
17515 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
17516 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17517 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
17518 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17519 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
17520 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
17521 ; GFX10-CU-NEXT: s_endpgm
17522 ;
17523 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17524 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
17525 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
17526 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
17527 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
17528 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
17529 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
17530 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
17531 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
17532 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
17533 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
17534 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
17535 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
17536 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
17537 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
17538 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
17539 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17540 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
17541 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
17542 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17543 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
17544 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
17545 ; SKIP-CACHE-INV-NEXT: s_endpgm
17546 ;
17547 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17548 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
17549 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17550 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17551 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17552 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17553 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17554 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17555 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17556 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17557 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17558 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17559 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17560 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17561 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17562 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
17563 ;
17564 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17565 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
17566 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17567 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17568 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17569 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17570 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17571 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17572 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17573 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17574 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17575 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17576 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17577 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17578 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17579 ; GFX90A-TGSPLIT-NEXT: s_endpgm
17580 ;
17581 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17582 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
17583 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17584 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17585 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17586 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17587 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17588 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17589 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17590 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17591 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17592 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17593 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17594 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17595 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17596 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
17597 ;
17598 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17599 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
17600 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17601 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17602 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17603 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17604 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17605 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17606 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17607 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17608 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17609 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17610 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17611 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17612 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17613 ; GFX940-TGSPLIT-NEXT: s_endpgm
17614 ;
17615 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17616 ; GFX11-WGP: ; %bb.0: ; %entry
17617 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17618 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
17619 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17620 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17621 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17622 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
17623 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
17624 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
17625 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17626 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
17627 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17628 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
17629 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17630 ; GFX11-WGP-NEXT: s_endpgm
17631 ;
17632 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17633 ; GFX11-CU: ; %bb.0: ; %entry
17634 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17635 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
17636 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17637 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17638 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17639 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
17640 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
17641 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
17642 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17643 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
17644 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17645 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
17646 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17647 ; GFX11-CU-NEXT: s_endpgm
17648 ;
17649 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17650 ; GFX12-WGP: ; %bb.0: ; %entry
17651 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17652 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
17653 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17654 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17655 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17656 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
17657 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
17658 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
17659 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17660 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
17661 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17662 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
17663 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17664 ; GFX12-WGP-NEXT: s_endpgm
17665 ;
17666 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
17667 ; GFX12-CU: ; %bb.0: ; %entry
17668 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17669 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
17670 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17671 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17672 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17673 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
17674 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
17675 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
17676 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17677 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
17678 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17679 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
17680 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17681 ; GFX12-CU-NEXT: s_endpgm
17682 ptr addrspace(1) %out, i32 %in, i32 %old) {
17683 entry:
17684 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
17685 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic acquire
17686 %val0 = extractvalue { i32, i1 } %val, 0
17687 store i32 %val0, ptr addrspace(1) %out, align 4
17688 ret void
17689 }
17691 define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_ret_cmpxchg(
17692 ; GFX6-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17693 ; GFX6: ; %bb.0: ; %entry
17694 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17695 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
17696 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
17697 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
17698 ; GFX6-NEXT: s_mov_b32 s12, s5
17699 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
17700 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
17701 ; GFX6-NEXT: s_mov_b32 s11, -1
17702 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
17703 ; GFX6-NEXT: s_mov_b32 s5, s12
17704 ; GFX6-NEXT: s_mov_b32 s6, s11
17705 ; GFX6-NEXT: s_mov_b32 s7, s10
17706 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
17707 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
17708 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17709 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
17710 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
17711 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17712 ; GFX6-NEXT: s_waitcnt vmcnt(0)
17713 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
17714 ; GFX6-NEXT: s_endpgm
17715 ;
17716 ; GFX7-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17717 ; GFX7: ; %bb.0: ; %entry
17718 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17719 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
17720 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
17721 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
17722 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
17723 ; GFX7-NEXT: s_mov_b32 s6, s4
17724 ; GFX7-NEXT: s_mov_b32 s7, s5
17725 ; GFX7-NEXT: s_mov_b32 s11, s12
17726 ; GFX7-NEXT: s_mov_b32 s10, s13
17727 ; GFX7-NEXT: s_add_u32 s6, s6, s11
17728 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
17729 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
17730 ; GFX7-NEXT: s_mov_b32 s7, s10
17731 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
17732 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
17733 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17734 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
17735 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
17736 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
17737 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
17738 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
17739 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
17740 ; GFX7-NEXT: s_waitcnt vmcnt(0)
17741 ; GFX7-NEXT: flat_store_dword v[0:1], v2
17742 ; GFX7-NEXT: s_endpgm
17743 ;
17744 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17745 ; GFX10-WGP: ; %bb.0: ; %entry
17746 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
17747 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
17748 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17749 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
17750 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
17751 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
17752 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
17753 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
17754 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17755 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
17756 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17757 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
17758 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
17759 ; GFX10-WGP-NEXT: s_endpgm
17760 ;
17761 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17762 ; GFX10-CU: ; %bb.0: ; %entry
17763 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
17764 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
17765 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17766 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
17767 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
17768 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
17769 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
17770 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
17771 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17772 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
17773 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
17774 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
17775 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
17776 ; GFX10-CU-NEXT: s_endpgm
17777 ;
17778 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17779 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
17780 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
17781 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
17782 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
17783 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
17784 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
17785 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
17786 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
17787 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
17788 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
17789 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
17790 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
17791 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
17792 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
17793 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
17794 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17795 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
17796 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
17797 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17798 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
17799 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
17800 ; SKIP-CACHE-INV-NEXT: s_endpgm
17801 ;
17802 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17803 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
17804 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17805 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17806 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17807 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17808 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17809 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17810 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17811 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17812 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17813 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17814 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17815 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17816 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17817 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
17818 ;
17819 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17820 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
17821 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
17822 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17823 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
17824 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
17825 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
17826 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17827 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
17828 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
17829 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17830 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17831 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
17832 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17833 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
17834 ; GFX90A-TGSPLIT-NEXT: s_endpgm
17835 ;
17836 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17837 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
17838 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17839 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17840 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17841 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17842 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17843 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17844 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17845 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17846 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17847 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17848 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17849 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
17850 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17851 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
17852 ;
17853 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17854 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
17855 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
17856 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
17857 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
17858 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
17859 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
17860 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
17861 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
17862 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
17863 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17864 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
17865 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
17866 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
17867 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
17868 ; GFX940-TGSPLIT-NEXT: s_endpgm
17869 ;
17870 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17871 ; GFX11-WGP: ; %bb.0: ; %entry
17872 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17873 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
17874 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17875 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17876 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17877 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
17878 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
17879 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
17880 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17881 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
17882 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17883 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
17884 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17885 ; GFX11-WGP-NEXT: s_endpgm
17886 ;
17887 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17888 ; GFX11-CU: ; %bb.0: ; %entry
17889 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17890 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
17891 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17892 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17893 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17894 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
17895 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
17896 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
17897 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17898 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
17899 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
17900 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
17901 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17902 ; GFX11-CU-NEXT: s_endpgm
17903 ;
17904 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17905 ; GFX12-WGP: ; %bb.0: ; %entry
17906 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
17907 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
17908 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17909 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
17910 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
17911 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
17912 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
17913 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
17914 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17915 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
17916 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17917 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
17918 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
17919 ; GFX12-WGP-NEXT: s_endpgm
17920 ;
17921 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
17922 ; GFX12-CU: ; %bb.0: ; %entry
17923 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
17924 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
17925 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
17926 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
17927 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
17928 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
17929 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
17930 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
17931 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
17932 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
17933 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
17934 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
17935 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
17936 ; GFX12-CU-NEXT: s_endpgm
17937 ptr addrspace(1) %out, i32 %in, i32 %old) {
17938 entry:
17939 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
17940 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
17941 %val0 = extractvalue { i32, i1 } %val, 0
17942 store i32 %val0, ptr addrspace(1) %out, align 4
17943 ret void
17944 }
17946 define amdgpu_kernel void @global_singlethread_one_as_release_acquire_ret_cmpxchg(
17947 ; GFX6-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
17948 ; GFX6: ; %bb.0: ; %entry
17949 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17950 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
17951 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
17952 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
17953 ; GFX6-NEXT: s_mov_b32 s12, s5
17954 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
17955 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
17956 ; GFX6-NEXT: s_mov_b32 s11, -1
17957 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
17958 ; GFX6-NEXT: s_mov_b32 s5, s12
17959 ; GFX6-NEXT: s_mov_b32 s6, s11
17960 ; GFX6-NEXT: s_mov_b32 s7, s10
17961 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
17962 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
17963 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
17964 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
17965 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
17966 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
17967 ; GFX6-NEXT: s_waitcnt vmcnt(0)
17968 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
17969 ; GFX6-NEXT: s_endpgm
17970 ;
17971 ; GFX7-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
17972 ; GFX7: ; %bb.0: ; %entry
17973 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
17974 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
17975 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
17976 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
17977 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
17978 ; GFX7-NEXT: s_mov_b32 s6, s4
17979 ; GFX7-NEXT: s_mov_b32 s7, s5
17980 ; GFX7-NEXT: s_mov_b32 s11, s12
17981 ; GFX7-NEXT: s_mov_b32 s10, s13
17982 ; GFX7-NEXT: s_add_u32 s6, s6, s11
17983 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
17984 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
17985 ; GFX7-NEXT: s_mov_b32 s7, s10
17986 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
17987 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
17988 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
17989 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
17990 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
17991 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
17992 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
17993 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
17994 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
17995 ; GFX7-NEXT: s_waitcnt vmcnt(0)
17996 ; GFX7-NEXT: flat_store_dword v[0:1], v2
17997 ; GFX7-NEXT: s_endpgm
17998 ;
17999 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18000 ; GFX10-WGP: ; %bb.0: ; %entry
18001 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
18002 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
18003 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18004 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
18005 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
18006 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
18007 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
18008 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
18009 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18010 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
18011 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18012 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
18013 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
18014 ; GFX10-WGP-NEXT: s_endpgm
18015 ;
18016 ; GFX10-CU-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18017 ; GFX10-CU: ; %bb.0: ; %entry
18018 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
18019 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
18020 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18021 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
18022 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
18023 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
18024 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
18025 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
18026 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18027 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
18028 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18029 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
18030 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
18031 ; GFX10-CU-NEXT: s_endpgm
18032 ;
18033 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18034 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
18035 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
18036 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
18037 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
18038 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
18039 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
18040 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
18041 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
18042 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
18043 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
18044 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
18045 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
18046 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
18047 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
18048 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
18049 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18050 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
18051 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
18052 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18053 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
18054 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
18055 ; SKIP-CACHE-INV-NEXT: s_endpgm
18056 ;
18057 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18058 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
18059 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18060 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18061 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18062 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18063 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18064 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18065 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18066 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18067 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18068 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18069 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18070 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18071 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18072 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
18073 ;
18074 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18075 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
18076 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18077 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18078 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18079 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18080 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18081 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18082 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18083 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18084 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18085 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18086 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18087 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18088 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18089 ; GFX90A-TGSPLIT-NEXT: s_endpgm
18090 ;
18091 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18092 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
18093 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18094 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18095 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18096 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18097 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18098 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18099 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18100 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18101 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18102 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18103 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18104 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18105 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18106 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
18107 ;
18108 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18109 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
18110 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18111 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18112 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18113 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18114 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18115 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18116 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18117 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18118 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18119 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18120 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18121 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18122 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18123 ; GFX940-TGSPLIT-NEXT: s_endpgm
18124 ;
18125 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18126 ; GFX11-WGP: ; %bb.0: ; %entry
18127 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18128 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
18129 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18130 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18131 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18132 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
18133 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
18134 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
18135 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18136 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
18137 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18138 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
18139 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18140 ; GFX11-WGP-NEXT: s_endpgm
18141 ;
18142 ; GFX11-CU-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18143 ; GFX11-CU: ; %bb.0: ; %entry
18144 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18145 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
18146 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18147 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18148 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18149 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
18150 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
18151 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
18152 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18153 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
18154 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18155 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
18156 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18157 ; GFX11-CU-NEXT: s_endpgm
18158 ;
18159 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18160 ; GFX12-WGP: ; %bb.0: ; %entry
18161 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18162 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
18163 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18164 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18165 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18166 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
18167 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
18168 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
18169 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18170 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
18171 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18172 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
18173 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18174 ; GFX12-WGP-NEXT: s_endpgm
18175 ;
18176 ; GFX12-CU-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
18177 ; GFX12-CU: ; %bb.0: ; %entry
18178 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18179 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
18180 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18181 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18182 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18183 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
18184 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
18185 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
18186 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18187 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
18188 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18189 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
18190 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18191 ; GFX12-CU-NEXT: s_endpgm
18192 ptr addrspace(1) %out, i32 %in, i32 %old) {
18193 entry:
18194 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
18195 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
18196 %val0 = extractvalue { i32, i1 } %val, 0
18197 store i32 %val0, ptr addrspace(1) %out, align 4
18198 ret void
18199 }
18201 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg(
18202 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18203 ; GFX6: ; %bb.0: ; %entry
18204 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18205 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
18206 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
18207 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
18208 ; GFX6-NEXT: s_mov_b32 s12, s5
18209 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
18210 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
18211 ; GFX6-NEXT: s_mov_b32 s11, -1
18212 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
18213 ; GFX6-NEXT: s_mov_b32 s5, s12
18214 ; GFX6-NEXT: s_mov_b32 s6, s11
18215 ; GFX6-NEXT: s_mov_b32 s7, s10
18216 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
18217 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
18218 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18219 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
18220 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
18221 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18222 ; GFX6-NEXT: s_waitcnt vmcnt(0)
18223 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
18224 ; GFX6-NEXT: s_endpgm
18225 ;
18226 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18227 ; GFX7: ; %bb.0: ; %entry
18228 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18229 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
18230 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
18231 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
18232 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
18233 ; GFX7-NEXT: s_mov_b32 s6, s4
18234 ; GFX7-NEXT: s_mov_b32 s7, s5
18235 ; GFX7-NEXT: s_mov_b32 s11, s12
18236 ; GFX7-NEXT: s_mov_b32 s10, s13
18237 ; GFX7-NEXT: s_add_u32 s6, s6, s11
18238 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
18239 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
18240 ; GFX7-NEXT: s_mov_b32 s7, s10
18241 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
18242 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
18243 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18244 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
18245 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
18246 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
18247 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
18248 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
18249 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
18250 ; GFX7-NEXT: s_waitcnt vmcnt(0)
18251 ; GFX7-NEXT: flat_store_dword v[0:1], v2
18252 ; GFX7-NEXT: s_endpgm
18253 ;
18254 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18255 ; GFX10-WGP: ; %bb.0: ; %entry
18256 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
18257 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
18258 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18259 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
18260 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
18261 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
18262 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
18263 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
18264 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18265 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
18266 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18267 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
18268 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
18269 ; GFX10-WGP-NEXT: s_endpgm
18270 ;
18271 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18272 ; GFX10-CU: ; %bb.0: ; %entry
18273 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
18274 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
18275 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18276 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
18277 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
18278 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
18279 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
18280 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
18281 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18282 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
18283 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18284 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
18285 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
18286 ; GFX10-CU-NEXT: s_endpgm
18287 ;
18288 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18289 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
18290 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
18291 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
18292 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
18293 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
18294 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
18295 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
18296 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
18297 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
18298 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
18299 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
18300 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
18301 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
18302 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
18303 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
18304 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18305 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
18306 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
18307 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18308 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
18309 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
18310 ; SKIP-CACHE-INV-NEXT: s_endpgm
18311 ;
18312 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18313 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
18314 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18315 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18316 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18317 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18318 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18319 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18320 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18321 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18322 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18323 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18324 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18325 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18326 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18327 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
18328 ;
18329 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18330 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
18331 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18332 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18333 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18334 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18335 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18336 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18337 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18338 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18339 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18340 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18341 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18342 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18343 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18344 ; GFX90A-TGSPLIT-NEXT: s_endpgm
18345 ;
18346 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18347 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
18348 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18349 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18350 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18351 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18352 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18353 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18354 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18355 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18356 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18357 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18358 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18359 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18360 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18361 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
18362 ;
18363 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18364 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
18365 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18366 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18367 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18368 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18369 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18370 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18371 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18372 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18373 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18374 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18375 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18376 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18377 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18378 ; GFX940-TGSPLIT-NEXT: s_endpgm
18379 ;
18380 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18381 ; GFX11-WGP: ; %bb.0: ; %entry
18382 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18383 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
18384 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18385 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18386 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18387 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
18388 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
18389 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
18390 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18391 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
18392 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18393 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
18394 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18395 ; GFX11-WGP-NEXT: s_endpgm
18396 ;
18397 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18398 ; GFX11-CU: ; %bb.0: ; %entry
18399 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18400 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
18401 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18402 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18403 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18404 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
18405 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
18406 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
18407 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18408 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
18409 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18410 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
18411 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18412 ; GFX11-CU-NEXT: s_endpgm
18413 ;
18414 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18415 ; GFX12-WGP: ; %bb.0: ; %entry
18416 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18417 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
18418 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18419 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18420 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18421 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
18422 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
18423 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
18424 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18425 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
18426 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18427 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
18428 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18429 ; GFX12-WGP-NEXT: s_endpgm
18430 ;
18431 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
18432 ; GFX12-CU: ; %bb.0: ; %entry
18433 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18434 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
18435 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18436 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18437 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18438 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
18439 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
18440 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
18441 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18442 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
18443 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18444 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
18445 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18446 ; GFX12-CU-NEXT: s_endpgm
18447 ptr addrspace(1) %out, i32 %in, i32 %old) {
18448 entry:
18449 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
18450 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
18451 %val0 = extractvalue { i32, i1 } %val, 0
18452 store i32 %val0, ptr addrspace(1) %out, align 4
18453 ret void
18454 }
18456 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg(
18457 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18458 ; GFX6: ; %bb.0: ; %entry
18459 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18460 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
18461 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
18462 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
18463 ; GFX6-NEXT: s_mov_b32 s12, s5
18464 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
18465 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
18466 ; GFX6-NEXT: s_mov_b32 s11, -1
18467 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
18468 ; GFX6-NEXT: s_mov_b32 s5, s12
18469 ; GFX6-NEXT: s_mov_b32 s6, s11
18470 ; GFX6-NEXT: s_mov_b32 s7, s10
18471 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
18472 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
18473 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18474 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
18475 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
18476 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18477 ; GFX6-NEXT: s_waitcnt vmcnt(0)
18478 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
18479 ; GFX6-NEXT: s_endpgm
18480 ;
18481 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18482 ; GFX7: ; %bb.0: ; %entry
18483 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18484 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
18485 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
18486 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
18487 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
18488 ; GFX7-NEXT: s_mov_b32 s6, s4
18489 ; GFX7-NEXT: s_mov_b32 s7, s5
18490 ; GFX7-NEXT: s_mov_b32 s11, s12
18491 ; GFX7-NEXT: s_mov_b32 s10, s13
18492 ; GFX7-NEXT: s_add_u32 s6, s6, s11
18493 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
18494 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
18495 ; GFX7-NEXT: s_mov_b32 s7, s10
18496 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
18497 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
18498 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18499 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
18500 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
18501 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
18502 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
18503 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
18504 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
18505 ; GFX7-NEXT: s_waitcnt vmcnt(0)
18506 ; GFX7-NEXT: flat_store_dword v[0:1], v2
18507 ; GFX7-NEXT: s_endpgm
18508 ;
18509 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18510 ; GFX10-WGP: ; %bb.0: ; %entry
18511 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
18512 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
18513 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18514 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
18515 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
18516 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
18517 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
18518 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
18519 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18520 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
18521 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18522 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
18523 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
18524 ; GFX10-WGP-NEXT: s_endpgm
18525 ;
18526 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18527 ; GFX10-CU: ; %bb.0: ; %entry
18528 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
18529 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
18530 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18531 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
18532 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
18533 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
18534 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
18535 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
18536 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18537 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
18538 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18539 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
18540 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
18541 ; GFX10-CU-NEXT: s_endpgm
18542 ;
18543 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18544 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
18545 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
18546 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
18547 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
18548 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
18549 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
18550 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
18551 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
18552 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
18553 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
18554 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
18555 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
18556 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
18557 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
18558 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
18559 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18560 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
18561 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
18562 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18563 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
18564 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
18565 ; SKIP-CACHE-INV-NEXT: s_endpgm
18566 ;
18567 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18568 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
18569 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18570 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18571 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18572 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18573 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18574 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18575 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18576 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18577 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18578 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18579 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18580 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18581 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18582 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
18583 ;
18584 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18585 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
18586 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18587 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18588 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18589 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18590 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18591 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18592 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18593 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18594 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18595 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18596 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18597 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18598 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18599 ; GFX90A-TGSPLIT-NEXT: s_endpgm
18600 ;
18601 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18602 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
18603 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18604 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18605 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18606 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18607 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18608 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18609 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18610 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18611 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18612 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18613 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18614 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18615 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18616 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
18617 ;
18618 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18619 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
18620 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18621 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18622 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18623 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18624 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18625 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18626 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18627 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18628 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18629 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18630 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18631 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18632 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18633 ; GFX940-TGSPLIT-NEXT: s_endpgm
18634 ;
18635 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18636 ; GFX11-WGP: ; %bb.0: ; %entry
18637 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18638 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
18639 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18640 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18641 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18642 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
18643 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
18644 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
18645 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18646 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
18647 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18648 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
18649 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18650 ; GFX11-WGP-NEXT: s_endpgm
18651 ;
18652 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18653 ; GFX11-CU: ; %bb.0: ; %entry
18654 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18655 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
18656 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18657 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18658 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18659 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
18660 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
18661 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
18662 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18663 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
18664 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18665 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
18666 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18667 ; GFX11-CU-NEXT: s_endpgm
18668 ;
18669 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18670 ; GFX12-WGP: ; %bb.0: ; %entry
18671 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18672 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
18673 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18674 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18675 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18676 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
18677 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
18678 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
18679 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18680 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
18681 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18682 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
18683 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18684 ; GFX12-WGP-NEXT: s_endpgm
18685 ;
18686 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
18687 ; GFX12-CU: ; %bb.0: ; %entry
18688 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18689 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
18690 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18691 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18692 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18693 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
18694 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
18695 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
18696 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18697 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
18698 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18699 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
18700 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18701 ; GFX12-CU-NEXT: s_endpgm
18702 ptr addrspace(1) %out, i32 %in, i32 %old) {
18703 entry:
18704 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
18705 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
18706 %val0 = extractvalue { i32, i1 } %val, 0
18707 store i32 %val0, ptr addrspace(1) %out, align 4
18708 ret void
18709 }
18711 define amdgpu_kernel void @global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg(
18712 ; GFX6-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18713 ; GFX6: ; %bb.0: ; %entry
18714 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18715 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
18716 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
18717 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
18718 ; GFX6-NEXT: s_mov_b32 s12, s5
18719 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
18720 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
18721 ; GFX6-NEXT: s_mov_b32 s11, -1
18722 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
18723 ; GFX6-NEXT: s_mov_b32 s5, s12
18724 ; GFX6-NEXT: s_mov_b32 s6, s11
18725 ; GFX6-NEXT: s_mov_b32 s7, s10
18726 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
18727 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
18728 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18729 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
18730 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
18731 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18732 ; GFX6-NEXT: s_waitcnt vmcnt(0)
18733 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
18734 ; GFX6-NEXT: s_endpgm
18735 ;
18736 ; GFX7-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18737 ; GFX7: ; %bb.0: ; %entry
18738 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18739 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
18740 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
18741 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
18742 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
18743 ; GFX7-NEXT: s_mov_b32 s6, s4
18744 ; GFX7-NEXT: s_mov_b32 s7, s5
18745 ; GFX7-NEXT: s_mov_b32 s11, s12
18746 ; GFX7-NEXT: s_mov_b32 s10, s13
18747 ; GFX7-NEXT: s_add_u32 s6, s6, s11
18748 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
18749 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
18750 ; GFX7-NEXT: s_mov_b32 s7, s10
18751 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
18752 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
18753 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18754 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
18755 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
18756 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
18757 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
18758 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
18759 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
18760 ; GFX7-NEXT: s_waitcnt vmcnt(0)
18761 ; GFX7-NEXT: flat_store_dword v[0:1], v2
18762 ; GFX7-NEXT: s_endpgm
18763 ;
18764 ; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18765 ; GFX10-WGP: ; %bb.0: ; %entry
18766 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
18767 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
18768 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18769 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
18770 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
18771 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
18772 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
18773 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
18774 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18775 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
18776 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18777 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
18778 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
18779 ; GFX10-WGP-NEXT: s_endpgm
18780 ;
18781 ; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18782 ; GFX10-CU: ; %bb.0: ; %entry
18783 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
18784 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
18785 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18786 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
18787 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
18788 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
18789 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
18790 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
18791 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18792 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
18793 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
18794 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
18795 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
18796 ; GFX10-CU-NEXT: s_endpgm
18797 ;
18798 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18799 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
18800 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
18801 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
18802 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
18803 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
18804 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
18805 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
18806 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
18807 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
18808 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
18809 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
18810 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
18811 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
18812 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
18813 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
18814 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18815 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
18816 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
18817 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18818 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
18819 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
18820 ; SKIP-CACHE-INV-NEXT: s_endpgm
18821 ;
18822 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18823 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
18824 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18825 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18826 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18827 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18828 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18829 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18830 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18831 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18832 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18833 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18834 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18835 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18836 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18837 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
18838 ;
18839 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18840 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
18841 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
18842 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18843 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
18844 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
18845 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
18846 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18847 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
18848 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
18849 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18850 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18851 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
18852 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18853 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
18854 ; GFX90A-TGSPLIT-NEXT: s_endpgm
18855 ;
18856 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18857 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
18858 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18859 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18860 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18861 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18862 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18863 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18864 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18865 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18866 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18867 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18868 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18869 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
18870 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18871 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
18872 ;
18873 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18874 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
18875 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
18876 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
18877 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
18878 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
18879 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
18880 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
18881 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
18882 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
18883 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
18884 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
18885 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
18886 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
18887 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
18888 ; GFX940-TGSPLIT-NEXT: s_endpgm
18889 ;
18890 ; GFX11-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18891 ; GFX11-WGP: ; %bb.0: ; %entry
18892 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18893 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
18894 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18895 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18896 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18897 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
18898 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
18899 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
18900 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18901 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
18902 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18903 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
18904 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18905 ; GFX11-WGP-NEXT: s_endpgm
18906 ;
18907 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18908 ; GFX11-CU: ; %bb.0: ; %entry
18909 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18910 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
18911 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18912 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18913 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18914 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
18915 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
18916 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
18917 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18918 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
18919 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
18920 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
18921 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18922 ; GFX11-CU-NEXT: s_endpgm
18923 ;
18924 ; GFX12-WGP-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18925 ; GFX12-WGP: ; %bb.0: ; %entry
18926 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
18927 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
18928 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18929 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
18930 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
18931 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
18932 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
18933 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
18934 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18935 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
18936 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18937 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
18938 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
18939 ; GFX12-WGP-NEXT: s_endpgm
18940 ;
18941 ; GFX12-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
18942 ; GFX12-CU: ; %bb.0: ; %entry
18943 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
18944 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
18945 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
18946 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
18947 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
18948 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
18949 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
18950 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
18951 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
18952 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
18953 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
18954 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
18955 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
18956 ; GFX12-CU-NEXT: s_endpgm
18957 ptr addrspace(1) %out, i32 %in, i32 %old) {
18958 entry:
18959 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
18960 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic seq_cst
18961 %val0 = extractvalue { i32, i1 } %val, 0
18962 store i32 %val0, ptr addrspace(1) %out, align 4
18963 ret void
18964 }
18966 define amdgpu_kernel void @global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg(
18967 ; GFX6-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
18968 ; GFX6: ; %bb.0: ; %entry
18969 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18970 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
18971 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
18972 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
18973 ; GFX6-NEXT: s_mov_b32 s12, s5
18974 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
18975 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
18976 ; GFX6-NEXT: s_mov_b32 s11, -1
18977 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
18978 ; GFX6-NEXT: s_mov_b32 s5, s12
18979 ; GFX6-NEXT: s_mov_b32 s6, s11
18980 ; GFX6-NEXT: s_mov_b32 s7, s10
18981 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
18982 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
18983 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
18984 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
18985 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
18986 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
18987 ; GFX6-NEXT: s_waitcnt vmcnt(0)
18988 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
18989 ; GFX6-NEXT: s_endpgm
18990 ;
18991 ; GFX7-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
18992 ; GFX7: ; %bb.0: ; %entry
18993 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
18994 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
18995 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
18996 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
18997 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
18998 ; GFX7-NEXT: s_mov_b32 s6, s4
18999 ; GFX7-NEXT: s_mov_b32 s7, s5
19000 ; GFX7-NEXT: s_mov_b32 s11, s12
19001 ; GFX7-NEXT: s_mov_b32 s10, s13
19002 ; GFX7-NEXT: s_add_u32 s6, s6, s11
19003 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
19004 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
19005 ; GFX7-NEXT: s_mov_b32 s7, s10
19006 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
19007 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
19008 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19009 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
19010 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
19011 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
19012 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
19013 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
19014 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
19015 ; GFX7-NEXT: s_waitcnt vmcnt(0)
19016 ; GFX7-NEXT: flat_store_dword v[0:1], v2
19017 ; GFX7-NEXT: s_endpgm
19018 ;
19019 ; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19020 ; GFX10-WGP: ; %bb.0: ; %entry
19021 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
19022 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
19023 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19024 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
19025 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
19026 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
19027 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
19028 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
19029 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19030 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
19031 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19032 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
19033 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
19034 ; GFX10-WGP-NEXT: s_endpgm
19035 ;
19036 ; GFX10-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19037 ; GFX10-CU: ; %bb.0: ; %entry
19038 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
19039 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
19040 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19041 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
19042 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
19043 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
19044 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
19045 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
19046 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19047 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
19048 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19049 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
19050 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
19051 ; GFX10-CU-NEXT: s_endpgm
19052 ;
19053 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19054 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
19055 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
19056 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
19057 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
19058 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
19059 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
19060 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
19061 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
19062 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
19063 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
19064 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
19065 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
19066 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
19067 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
19068 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
19069 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19070 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
19071 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
19072 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19073 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
19074 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
19075 ; SKIP-CACHE-INV-NEXT: s_endpgm
19076 ;
19077 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19078 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
19079 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19080 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19081 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19082 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19083 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19084 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19085 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19086 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19087 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19088 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19089 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19090 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19091 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19092 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
19093 ;
19094 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19095 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
19096 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19097 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19098 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19099 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19100 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19101 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19102 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19103 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19104 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19105 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19106 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19107 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19108 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19109 ; GFX90A-TGSPLIT-NEXT: s_endpgm
19110 ;
19111 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19112 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
19113 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19114 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19115 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19116 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19117 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19118 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19119 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19120 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19121 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19122 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19123 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19124 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19125 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19126 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
19127 ;
19128 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19129 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
19130 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19131 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19132 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19133 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19134 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19135 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19136 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19137 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19138 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19139 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19140 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19141 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19142 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19143 ; GFX940-TGSPLIT-NEXT: s_endpgm
19144 ;
19145 ; GFX11-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19146 ; GFX11-WGP: ; %bb.0: ; %entry
19147 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19148 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
19149 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19150 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19151 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19152 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
19153 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
19154 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
19155 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19156 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
19157 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19158 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
19159 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19160 ; GFX11-WGP-NEXT: s_endpgm
19161 ;
19162 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19163 ; GFX11-CU: ; %bb.0: ; %entry
19164 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19165 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
19166 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19167 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19168 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19169 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
19170 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
19171 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
19172 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19173 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
19174 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19175 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
19176 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19177 ; GFX11-CU-NEXT: s_endpgm
19178 ;
19179 ; GFX12-WGP-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19180 ; GFX12-WGP: ; %bb.0: ; %entry
19181 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19182 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
19183 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19184 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19185 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19186 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
19187 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
19188 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
19189 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19190 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
19191 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19192 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
19193 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19194 ; GFX12-WGP-NEXT: s_endpgm
19195 ;
19196 ; GFX12-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
19197 ; GFX12-CU: ; %bb.0: ; %entry
19198 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19199 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
19200 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19201 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19202 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19203 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
19204 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
19205 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
19206 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19207 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
19208 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19209 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
19210 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19211 ; GFX12-CU-NEXT: s_endpgm
19212 ptr addrspace(1) %out, i32 %in, i32 %old) {
19213 entry:
19214 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
19215 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire seq_cst
19216 %val0 = extractvalue { i32, i1 } %val, 0
19217 store i32 %val0, ptr addrspace(1) %out, align 4
19218 ret void
19219 }
19221 define amdgpu_kernel void @global_singlethread_one_as_release_seq_cst_ret_cmpxchg(
19222 ; GFX6-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19223 ; GFX6: ; %bb.0: ; %entry
19224 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19225 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
19226 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
19227 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
19228 ; GFX6-NEXT: s_mov_b32 s12, s5
19229 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
19230 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
19231 ; GFX6-NEXT: s_mov_b32 s11, -1
19232 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
19233 ; GFX6-NEXT: s_mov_b32 s5, s12
19234 ; GFX6-NEXT: s_mov_b32 s6, s11
19235 ; GFX6-NEXT: s_mov_b32 s7, s10
19236 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
19237 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
19238 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19239 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
19240 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
19241 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19242 ; GFX6-NEXT: s_waitcnt vmcnt(0)
19243 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
19244 ; GFX6-NEXT: s_endpgm
19245 ;
19246 ; GFX7-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19247 ; GFX7: ; %bb.0: ; %entry
19248 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19249 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
19250 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
19251 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
19252 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
19253 ; GFX7-NEXT: s_mov_b32 s6, s4
19254 ; GFX7-NEXT: s_mov_b32 s7, s5
19255 ; GFX7-NEXT: s_mov_b32 s11, s12
19256 ; GFX7-NEXT: s_mov_b32 s10, s13
19257 ; GFX7-NEXT: s_add_u32 s6, s6, s11
19258 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
19259 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
19260 ; GFX7-NEXT: s_mov_b32 s7, s10
19261 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
19262 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
19263 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19264 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
19265 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
19266 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
19267 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
19268 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
19269 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
19270 ; GFX7-NEXT: s_waitcnt vmcnt(0)
19271 ; GFX7-NEXT: flat_store_dword v[0:1], v2
19272 ; GFX7-NEXT: s_endpgm
19273 ;
19274 ; GFX10-WGP-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19275 ; GFX10-WGP: ; %bb.0: ; %entry
19276 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
19277 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
19278 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19279 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
19280 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
19281 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
19282 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
19283 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
19284 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19285 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
19286 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19287 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
19288 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
19289 ; GFX10-WGP-NEXT: s_endpgm
19290 ;
19291 ; GFX10-CU-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19292 ; GFX10-CU: ; %bb.0: ; %entry
19293 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
19294 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
19295 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19296 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
19297 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
19298 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
19299 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
19300 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
19301 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19302 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
19303 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19304 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
19305 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
19306 ; GFX10-CU-NEXT: s_endpgm
19307 ;
19308 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19309 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
19310 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
19311 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
19312 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
19313 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
19314 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
19315 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
19316 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
19317 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
19318 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
19319 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
19320 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
19321 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
19322 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
19323 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
19324 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19325 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
19326 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
19327 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19328 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
19329 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
19330 ; SKIP-CACHE-INV-NEXT: s_endpgm
19331 ;
19332 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19333 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
19334 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19335 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19336 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19337 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19338 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19339 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19340 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19341 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19342 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19343 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19344 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19345 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19346 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19347 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
19348 ;
19349 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19350 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
19351 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19352 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19353 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19354 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19355 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19356 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19357 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19358 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19359 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19360 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19361 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19362 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19363 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19364 ; GFX90A-TGSPLIT-NEXT: s_endpgm
19365 ;
19366 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19367 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
19368 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19369 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19370 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19371 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19372 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19373 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19374 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19375 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19376 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19377 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19378 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19379 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19380 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19381 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
19382 ;
19383 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19384 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
19385 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19386 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19387 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19388 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19389 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19390 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19391 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19392 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19393 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19394 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19395 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19396 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19397 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19398 ; GFX940-TGSPLIT-NEXT: s_endpgm
19399 ;
19400 ; GFX11-WGP-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19401 ; GFX11-WGP: ; %bb.0: ; %entry
19402 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19403 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
19404 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19405 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19406 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19407 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
19408 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
19409 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
19410 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19411 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
19412 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19413 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
19414 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19415 ; GFX11-WGP-NEXT: s_endpgm
19416 ;
19417 ; GFX11-CU-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19418 ; GFX11-CU: ; %bb.0: ; %entry
19419 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19420 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
19421 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19422 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19423 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19424 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
19425 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
19426 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
19427 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19428 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
19429 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19430 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
19431 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19432 ; GFX11-CU-NEXT: s_endpgm
19433 ;
19434 ; GFX12-WGP-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19435 ; GFX12-WGP: ; %bb.0: ; %entry
19436 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19437 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
19438 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19439 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19440 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19441 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
19442 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
19443 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
19444 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19445 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
19446 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19447 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
19448 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19449 ; GFX12-WGP-NEXT: s_endpgm
19450 ;
19451 ; GFX12-CU-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
19452 ; GFX12-CU: ; %bb.0: ; %entry
19453 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19454 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
19455 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19456 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19457 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19458 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
19459 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
19460 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
19461 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19462 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
19463 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19464 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
19465 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19466 ; GFX12-CU-NEXT: s_endpgm
19467 ptr addrspace(1) %out, i32 %in, i32 %old) {
19468 entry:
19469 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
19470 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release seq_cst
19471 %val0 = extractvalue { i32, i1 } %val, 0
19472 store i32 %val0, ptr addrspace(1) %out, align 4
19473 ret void
19474 }
19476 define amdgpu_kernel void @global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg(
19477 ; GFX6-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19478 ; GFX6: ; %bb.0: ; %entry
19479 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19480 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
19481 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
19482 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
19483 ; GFX6-NEXT: s_mov_b32 s12, s5
19484 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
19485 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
19486 ; GFX6-NEXT: s_mov_b32 s11, -1
19487 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
19488 ; GFX6-NEXT: s_mov_b32 s5, s12
19489 ; GFX6-NEXT: s_mov_b32 s6, s11
19490 ; GFX6-NEXT: s_mov_b32 s7, s10
19491 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
19492 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
19493 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19494 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
19495 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
19496 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19497 ; GFX6-NEXT: s_waitcnt vmcnt(0)
19498 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
19499 ; GFX6-NEXT: s_endpgm
19500 ;
19501 ; GFX7-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19502 ; GFX7: ; %bb.0: ; %entry
19503 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19504 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
19505 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
19506 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
19507 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
19508 ; GFX7-NEXT: s_mov_b32 s6, s4
19509 ; GFX7-NEXT: s_mov_b32 s7, s5
19510 ; GFX7-NEXT: s_mov_b32 s11, s12
19511 ; GFX7-NEXT: s_mov_b32 s10, s13
19512 ; GFX7-NEXT: s_add_u32 s6, s6, s11
19513 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
19514 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
19515 ; GFX7-NEXT: s_mov_b32 s7, s10
19516 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
19517 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
19518 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19519 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
19520 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
19521 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
19522 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
19523 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
19524 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
19525 ; GFX7-NEXT: s_waitcnt vmcnt(0)
19526 ; GFX7-NEXT: flat_store_dword v[0:1], v2
19527 ; GFX7-NEXT: s_endpgm
19528 ;
19529 ; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19530 ; GFX10-WGP: ; %bb.0: ; %entry
19531 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
19532 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
19533 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19534 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
19535 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
19536 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
19537 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
19538 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
19539 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19540 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
19541 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19542 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
19543 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
19544 ; GFX10-WGP-NEXT: s_endpgm
19545 ;
19546 ; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19547 ; GFX10-CU: ; %bb.0: ; %entry
19548 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
19549 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
19550 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19551 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
19552 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
19553 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
19554 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
19555 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
19556 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19557 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
19558 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19559 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
19560 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
19561 ; GFX10-CU-NEXT: s_endpgm
19562 ;
19563 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19564 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
19565 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
19566 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
19567 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
19568 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
19569 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
19570 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
19571 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
19572 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
19573 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
19574 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
19575 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
19576 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
19577 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
19578 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
19579 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19580 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
19581 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
19582 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19583 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
19584 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
19585 ; SKIP-CACHE-INV-NEXT: s_endpgm
19586 ;
19587 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19588 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
19589 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19590 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19591 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19592 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19593 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19594 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19595 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19596 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19597 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19598 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19599 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19600 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19601 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19602 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
19603 ;
19604 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19605 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
19606 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19607 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19608 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19609 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19610 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19611 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19612 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19613 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19614 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19615 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19616 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19617 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19618 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19619 ; GFX90A-TGSPLIT-NEXT: s_endpgm
19620 ;
19621 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19622 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
19623 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19624 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19625 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19626 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19627 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19628 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19629 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19630 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19631 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19632 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19633 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19634 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19635 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19636 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
19637 ;
19638 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19639 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
19640 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19641 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19642 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19643 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19644 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19645 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19646 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19647 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19648 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19649 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19650 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19651 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19652 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19653 ; GFX940-TGSPLIT-NEXT: s_endpgm
19654 ;
19655 ; GFX11-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19656 ; GFX11-WGP: ; %bb.0: ; %entry
19657 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19658 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
19659 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19660 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19661 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19662 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
19663 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
19664 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
19665 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19666 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
19667 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19668 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
19669 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19670 ; GFX11-WGP-NEXT: s_endpgm
19671 ;
19672 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19673 ; GFX11-CU: ; %bb.0: ; %entry
19674 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19675 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
19676 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19677 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19678 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19679 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
19680 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
19681 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
19682 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19683 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
19684 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19685 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
19686 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19687 ; GFX11-CU-NEXT: s_endpgm
19688 ;
19689 ; GFX12-WGP-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19690 ; GFX12-WGP: ; %bb.0: ; %entry
19691 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19692 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
19693 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19694 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19695 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19696 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
19697 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
19698 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
19699 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19700 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
19701 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19702 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
19703 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19704 ; GFX12-WGP-NEXT: s_endpgm
19705 ;
19706 ; GFX12-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
19707 ; GFX12-CU: ; %bb.0: ; %entry
19708 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19709 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
19710 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19711 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19712 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19713 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
19714 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
19715 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
19716 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19717 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
19718 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19719 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
19720 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19721 ; GFX12-CU-NEXT: s_endpgm
19722 ptr addrspace(1) %out, i32 %in, i32 %old) {
19723 entry:
19724 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
19725 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel seq_cst
19726 %val0 = extractvalue { i32, i1 } %val, 0
19727 store i32 %val0, ptr addrspace(1) %out, align 4
19728 ret void
19729 }
19731 define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg(
19732 ; GFX6-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19733 ; GFX6: ; %bb.0: ; %entry
19734 ; GFX6-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19735 ; GFX6-NEXT: s_load_dword s9, s[6:7], 0x2
19736 ; GFX6-NEXT: s_load_dword s8, s[6:7], 0x3
19737 ; GFX6-NEXT: s_waitcnt lgkmcnt(0)
19738 ; GFX6-NEXT: s_mov_b32 s12, s5
19739 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 killed $sgpr4_sgpr5
19740 ; GFX6-NEXT: s_mov_b32 s10, 0x100f000
19741 ; GFX6-NEXT: s_mov_b32 s11, -1
19742 ; GFX6-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
19743 ; GFX6-NEXT: s_mov_b32 s5, s12
19744 ; GFX6-NEXT: s_mov_b32 s6, s11
19745 ; GFX6-NEXT: s_mov_b32 s7, s10
19746 ; GFX6-NEXT: v_mov_b32_e32 v0, s9
19747 ; GFX6-NEXT: v_mov_b32_e32 v2, s8
19748 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19749 ; GFX6-NEXT: v_mov_b32_e32 v1, v2
19750 ; GFX6-NEXT: buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
19751 ; GFX6-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19752 ; GFX6-NEXT: s_waitcnt vmcnt(0)
19753 ; GFX6-NEXT: buffer_store_dword v0, off, s[4:7], 0
19754 ; GFX6-NEXT: s_endpgm
19755 ;
19756 ; GFX7-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19757 ; GFX7: ; %bb.0: ; %entry
19758 ; GFX7-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
19759 ; GFX7-NEXT: s_load_dword s9, s[6:7], 0x2
19760 ; GFX7-NEXT: s_load_dword s8, s[6:7], 0x3
19761 ; GFX7-NEXT: s_mov_b64 s[12:13], 16
19762 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
19763 ; GFX7-NEXT: s_mov_b32 s6, s4
19764 ; GFX7-NEXT: s_mov_b32 s7, s5
19765 ; GFX7-NEXT: s_mov_b32 s11, s12
19766 ; GFX7-NEXT: s_mov_b32 s10, s13
19767 ; GFX7-NEXT: s_add_u32 s6, s6, s11
19768 ; GFX7-NEXT: s_addc_u32 s10, s7, s10
19769 ; GFX7-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
19770 ; GFX7-NEXT: s_mov_b32 s7, s10
19771 ; GFX7-NEXT: v_mov_b32_e32 v2, s9
19772 ; GFX7-NEXT: v_mov_b32_e32 v0, s8
19773 ; GFX7-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19774 ; GFX7-NEXT: v_mov_b32_e32 v3, v0
19775 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
19776 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
19777 ; GFX7-NEXT: flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
19778 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
19779 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
19780 ; GFX7-NEXT: s_waitcnt vmcnt(0)
19781 ; GFX7-NEXT: flat_store_dword v[0:1], v2
19782 ; GFX7-NEXT: s_endpgm
19783 ;
19784 ; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19785 ; GFX10-WGP: ; %bb.0: ; %entry
19786 ; GFX10-WGP-NEXT: s_mov_b64 s[8:9], s[6:7]
19787 ; GFX10-WGP-NEXT: v_mov_b32_e32 v0, 0
19788 ; GFX10-WGP-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19789 ; GFX10-WGP-NEXT: s_load_dword s7, s[8:9], 0x8
19790 ; GFX10-WGP-NEXT: s_load_dword s6, s[8:9], 0xc
19791 ; GFX10-WGP-NEXT: s_waitcnt lgkmcnt(0)
19792 ; GFX10-WGP-NEXT: v_mov_b32_e32 v1, s7
19793 ; GFX10-WGP-NEXT: v_mov_b32_e32 v3, s6
19794 ; GFX10-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19795 ; GFX10-WGP-NEXT: v_mov_b32_e32 v2, v3
19796 ; GFX10-WGP-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19797 ; GFX10-WGP-NEXT: s_waitcnt vmcnt(0)
19798 ; GFX10-WGP-NEXT: global_store_dword v0, v1, s[4:5]
19799 ; GFX10-WGP-NEXT: s_endpgm
19800 ;
19801 ; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19802 ; GFX10-CU: ; %bb.0: ; %entry
19803 ; GFX10-CU-NEXT: s_mov_b64 s[8:9], s[6:7]
19804 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, 0
19805 ; GFX10-CU-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19806 ; GFX10-CU-NEXT: s_load_dword s7, s[8:9], 0x8
19807 ; GFX10-CU-NEXT: s_load_dword s6, s[8:9], 0xc
19808 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
19809 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s7
19810 ; GFX10-CU-NEXT: v_mov_b32_e32 v3, s6
19811 ; GFX10-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19812 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, v3
19813 ; GFX10-CU-NEXT: global_atomic_cmpswap v1, v0, v[1:2], s[4:5] offset:16 glc
19814 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
19815 ; GFX10-CU-NEXT: global_store_dword v0, v1, s[4:5]
19816 ; GFX10-CU-NEXT: s_endpgm
19817 ;
19818 ; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19819 ; SKIP-CACHE-INV: ; %bb.0: ; %entry
19820 ; SKIP-CACHE-INV-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x0
19821 ; SKIP-CACHE-INV-NEXT: s_load_dword s5, s[2:3], 0x2
19822 ; SKIP-CACHE-INV-NEXT: s_load_dword s4, s[2:3], 0x3
19823 ; SKIP-CACHE-INV-NEXT: s_waitcnt lgkmcnt(0)
19824 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s8, s1
19825 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
19826 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s6, 0xf000
19827 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s7, -1
19828 ; SKIP-CACHE-INV-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
19829 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s1, s8
19830 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s2, s7
19831 ; SKIP-CACHE-INV-NEXT: s_mov_b32 s3, s6
19832 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v0, s5
19833 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v2, s4
19834 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
19835 ; SKIP-CACHE-INV-NEXT: v_mov_b32_e32 v1, v2
19836 ; SKIP-CACHE-INV-NEXT: buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 offset:16 glc
19837 ; SKIP-CACHE-INV-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
19838 ; SKIP-CACHE-INV-NEXT: s_waitcnt vmcnt(0)
19839 ; SKIP-CACHE-INV-NEXT: buffer_store_dword v0, off, s[0:3], 0
19840 ; SKIP-CACHE-INV-NEXT: s_endpgm
19841 ;
19842 ; GFX90A-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19843 ; GFX90A-NOTTGSPLIT: ; %bb.0: ; %entry
19844 ; GFX90A-NOTTGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19845 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19846 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19847 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19848 ; GFX90A-NOTTGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19849 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19850 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19851 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19852 ; GFX90A-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19853 ; GFX90A-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19854 ; GFX90A-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19855 ; GFX90A-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19856 ; GFX90A-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19857 ; GFX90A-NOTTGSPLIT-NEXT: s_endpgm
19858 ;
19859 ; GFX90A-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19860 ; GFX90A-TGSPLIT: ; %bb.0: ; %entry
19861 ; GFX90A-TGSPLIT-NEXT: s_mov_b64 s[8:9], s[6:7]
19862 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19863 ; GFX90A-TGSPLIT-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
19864 ; GFX90A-TGSPLIT-NEXT: s_load_dword s7, s[8:9], 0x8
19865 ; GFX90A-TGSPLIT-NEXT: s_load_dword s6, s[8:9], 0xc
19866 ; GFX90A-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19867 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v2, s7
19868 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v1, s6
19869 ; GFX90A-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19870 ; GFX90A-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19871 ; GFX90A-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[4:5] offset:16 glc
19872 ; GFX90A-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19873 ; GFX90A-TGSPLIT-NEXT: global_store_dword v0, v1, s[4:5]
19874 ; GFX90A-TGSPLIT-NEXT: s_endpgm
19875 ;
19876 ; GFX940-NOTTGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19877 ; GFX940-NOTTGSPLIT: ; %bb.0: ; %entry
19878 ; GFX940-NOTTGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19879 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19880 ; GFX940-NOTTGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19881 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19882 ; GFX940-NOTTGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19883 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19884 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19885 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19886 ; GFX940-NOTTGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19887 ; GFX940-NOTTGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19888 ; GFX940-NOTTGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19889 ; GFX940-NOTTGSPLIT-NEXT: s_waitcnt vmcnt(0)
19890 ; GFX940-NOTTGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19891 ; GFX940-NOTTGSPLIT-NEXT: s_endpgm
19892 ;
19893 ; GFX940-TGSPLIT-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19894 ; GFX940-TGSPLIT: ; %bb.0: ; %entry
19895 ; GFX940-TGSPLIT-NEXT: s_mov_b64 s[4:5], s[2:3]
19896 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v0, 0
19897 ; GFX940-TGSPLIT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x0
19898 ; GFX940-TGSPLIT-NEXT: s_load_dword s3, s[4:5], 0x8
19899 ; GFX940-TGSPLIT-NEXT: s_load_dword s2, s[4:5], 0xc
19900 ; GFX940-TGSPLIT-NEXT: s_waitcnt lgkmcnt(0)
19901 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v2, s3
19902 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v1, s2
19903 ; GFX940-TGSPLIT-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
19904 ; GFX940-TGSPLIT-NEXT: v_mov_b32_e32 v3, v1
19905 ; GFX940-TGSPLIT-NEXT: global_atomic_cmpswap v1, v0, v[2:3], s[0:1] offset:16 sc0
19906 ; GFX940-TGSPLIT-NEXT: s_waitcnt vmcnt(0)
19907 ; GFX940-TGSPLIT-NEXT: global_store_dword v0, v1, s[0:1] sc0 sc1
19908 ; GFX940-TGSPLIT-NEXT: s_endpgm
19909 ;
19910 ; GFX11-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19911 ; GFX11-WGP: ; %bb.0: ; %entry
19912 ; GFX11-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19913 ; GFX11-WGP-NEXT: v_mov_b32_e32 v0, 0
19914 ; GFX11-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19915 ; GFX11-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19916 ; GFX11-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19917 ; GFX11-WGP-NEXT: s_waitcnt lgkmcnt(0)
19918 ; GFX11-WGP-NEXT: v_mov_b32_e32 v1, s3
19919 ; GFX11-WGP-NEXT: v_mov_b32_e32 v3, s2
19920 ; GFX11-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19921 ; GFX11-WGP-NEXT: v_mov_b32_e32 v2, v3
19922 ; GFX11-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19923 ; GFX11-WGP-NEXT: s_waitcnt vmcnt(0)
19924 ; GFX11-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19925 ; GFX11-WGP-NEXT: s_endpgm
19926 ;
19927 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19928 ; GFX11-CU: ; %bb.0: ; %entry
19929 ; GFX11-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19930 ; GFX11-CU-NEXT: v_mov_b32_e32 v0, 0
19931 ; GFX11-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19932 ; GFX11-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19933 ; GFX11-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19934 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
19935 ; GFX11-CU-NEXT: v_mov_b32_e32 v1, s3
19936 ; GFX11-CU-NEXT: v_mov_b32_e32 v3, s2
19937 ; GFX11-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19938 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, v3
19939 ; GFX11-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 glc
19940 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
19941 ; GFX11-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19942 ; GFX11-CU-NEXT: s_endpgm
19943 ;
19944 ; GFX12-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19945 ; GFX12-WGP: ; %bb.0: ; %entry
19946 ; GFX12-WGP-NEXT: s_mov_b64 s[4:5], s[2:3]
19947 ; GFX12-WGP-NEXT: v_mov_b32_e32 v0, 0
19948 ; GFX12-WGP-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19949 ; GFX12-WGP-NEXT: s_load_b32 s3, s[4:5], 0x8
19950 ; GFX12-WGP-NEXT: s_load_b32 s2, s[4:5], 0xc
19951 ; GFX12-WGP-NEXT: s_wait_kmcnt 0x0
19952 ; GFX12-WGP-NEXT: v_mov_b32_e32 v1, s3
19953 ; GFX12-WGP-NEXT: v_mov_b32_e32 v3, s2
19954 ; GFX12-WGP-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19955 ; GFX12-WGP-NEXT: v_mov_b32_e32 v2, v3
19956 ; GFX12-WGP-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19957 ; GFX12-WGP-NEXT: s_wait_loadcnt 0x0
19958 ; GFX12-WGP-NEXT: global_store_b32 v0, v1, s[0:1]
19959 ; GFX12-WGP-NEXT: s_endpgm
19960 ;
19961 ; GFX12-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
19962 ; GFX12-CU: ; %bb.0: ; %entry
19963 ; GFX12-CU-NEXT: s_mov_b64 s[4:5], s[2:3]
19964 ; GFX12-CU-NEXT: v_mov_b32_e32 v0, 0
19965 ; GFX12-CU-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
19966 ; GFX12-CU-NEXT: s_load_b32 s3, s[4:5], 0x8
19967 ; GFX12-CU-NEXT: s_load_b32 s2, s[4:5], 0xc
19968 ; GFX12-CU-NEXT: s_wait_kmcnt 0x0
19969 ; GFX12-CU-NEXT: v_mov_b32_e32 v1, s3
19970 ; GFX12-CU-NEXT: v_mov_b32_e32 v3, s2
19971 ; GFX12-CU-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
19972 ; GFX12-CU-NEXT: v_mov_b32_e32 v2, v3
19973 ; GFX12-CU-NEXT: global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] offset:16 th:TH_ATOMIC_RETURN
19974 ; GFX12-CU-NEXT: s_wait_loadcnt 0x0
19975 ; GFX12-CU-NEXT: global_store_b32 v0, v1, s[0:1]
19976 ; GFX12-CU-NEXT: s_endpgm
19977 ptr addrspace(1) %out, i32 %in, i32 %old) {
19978 entry:
19979 %gep = getelementptr i32, ptr addrspace(1) %out, i32 4
19980 %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
19981 %val0 = extractvalue { i32, i1 } %val, 0
19982 store i32 %val0, ptr addrspace(1) %out, align 4
19983 ret void
19984 }