| blob | bc2096093ab91b10079bcfa95020625085adda76 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s -check-prefix=SI
3 ; RUN: llc < %s -mtriple=amdgcn -mcpu=hawaii -verify-machineinstrs | FileCheck %s -check-prefix=GFX7
4 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck %s -check-prefix=GFX10
5 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s -check-prefix=GFX1030
6 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck %s -check-prefix=GFX1100
8 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s -check-prefix=G_SI
9 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=hawaii -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX7
10 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX10
11 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX1030
12 ; RUN: llc < %s -global-isel -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck %s -check-prefix=G_GFX1100
14 declare float @llvm.amdgcn.raw.ptr.buffer.atomic.fmin.f32(float, ptr addrspace(8), i32, i32, i32 immarg)
15 declare float @llvm.amdgcn.raw.ptr.buffer.atomic.fmax.f32(float, ptr addrspace(8), i32, i32, i32 immarg)
18 define amdgpu_kernel void @raw_ptr_buffer_atomic_min_noret_f32(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex) {
19 ; SI-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
20 ; SI: ; %bb.0: ; %main_body
21 ; SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
22 ; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
23 ; SI-NEXT: s_waitcnt lgkmcnt(0)
24 ; SI-NEXT: v_mov_b32_e32 v0, s6
25 ; SI-NEXT: v_mov_b32_e32 v1, s7
26 ; SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
27 ; SI-NEXT: s_endpgm
28 ;
29 ; GFX7-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
30 ; GFX7: ; %bb.0: ; %main_body
31 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
32 ; GFX7-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
33 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
34 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
35 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
36 ; GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
37 ; GFX7-NEXT: s_endpgm
38 ;
39 ; GFX10-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
40 ; GFX10: ; %bb.0: ; %main_body
41 ; GFX10-NEXT: s_clause 0x1
42 ; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
43 ; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
44 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
45 ; GFX10-NEXT: v_mov_b32_e32 v0, s6
46 ; GFX10-NEXT: v_mov_b32_e32 v1, s7
47 ; GFX10-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
48 ; GFX10-NEXT: s_endpgm
49 ;
50 ; GFX1030-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
51 ; GFX1030: ; %bb.0: ; %main_body
52 ; GFX1030-NEXT: s_clause 0x1
53 ; GFX1030-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
54 ; GFX1030-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
55 ; GFX1030-NEXT: s_waitcnt lgkmcnt(0)
56 ; GFX1030-NEXT: v_mov_b32_e32 v0, s6
57 ; GFX1030-NEXT: v_mov_b32_e32 v1, s7
58 ; GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
59 ; GFX1030-NEXT: s_endpgm
60 ;
61 ; GFX1100-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
62 ; GFX1100: ; %bb.0: ; %main_body
63 ; GFX1100-NEXT: s_clause 0x1
64 ; GFX1100-NEXT: s_load_b64 s[6:7], s[4:5], 0x34
65 ; GFX1100-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
66 ; GFX1100-NEXT: s_waitcnt lgkmcnt(0)
67 ; GFX1100-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
68 ; GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 0 offen
69 ; GFX1100-NEXT: s_endpgm
70 ;
71 ; G_SI-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
72 ; G_SI: ; %bb.0: ; %main_body
73 ; G_SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
74 ; G_SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
75 ; G_SI-NEXT: s_waitcnt lgkmcnt(0)
76 ; G_SI-NEXT: v_mov_b32_e32 v0, s6
77 ; G_SI-NEXT: v_mov_b32_e32 v1, s7
78 ; G_SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
79 ; G_SI-NEXT: s_endpgm
80 ;
81 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
82 ; G_GFX7: ; %bb.0: ; %main_body
83 ; G_GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
84 ; G_GFX7-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
85 ; G_GFX7-NEXT: s_waitcnt lgkmcnt(0)
86 ; G_GFX7-NEXT: v_mov_b32_e32 v0, s6
87 ; G_GFX7-NEXT: v_mov_b32_e32 v1, s7
88 ; G_GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
89 ; G_GFX7-NEXT: s_endpgm
90 ;
91 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
92 ; G_GFX10: ; %bb.0: ; %main_body
93 ; G_GFX10-NEXT: s_clause 0x1
94 ; G_GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
95 ; G_GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
96 ; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
97 ; G_GFX10-NEXT: v_mov_b32_e32 v0, s6
98 ; G_GFX10-NEXT: v_mov_b32_e32 v1, s7
99 ; G_GFX10-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
100 ; G_GFX10-NEXT: s_endpgm
101 ;
102 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
103 ; G_GFX1030: ; %bb.0: ; %main_body
104 ; G_GFX1030-NEXT: s_clause 0x1
105 ; G_GFX1030-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
106 ; G_GFX1030-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
107 ; G_GFX1030-NEXT: s_waitcnt lgkmcnt(0)
108 ; G_GFX1030-NEXT: v_mov_b32_e32 v0, s6
109 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, s7
110 ; G_GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen
111 ; G_GFX1030-NEXT: s_endpgm
112 ;
113 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_min_noret_f32:
114 ; G_GFX1100: ; %bb.0: ; %main_body
115 ; G_GFX1100-NEXT: s_clause 0x1
116 ; G_GFX1100-NEXT: s_load_b64 s[6:7], s[4:5], 0x34
117 ; G_GFX1100-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
118 ; G_GFX1100-NEXT: s_waitcnt lgkmcnt(0)
119 ; G_GFX1100-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
120 ; G_GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 0 offen
121 ; G_GFX1100-NEXT: s_endpgm
122 main_body:
123 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmin.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0)
124 ret void
125 }
127 define amdgpu_ps void @raw_ptr_buffer_atomic_min_rtn_f32(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex) {
128 ; SI-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
129 ; SI: ; %bb.0: ; %main_body
130 ; SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
131 ; SI-NEXT: s_mov_b32 s3, 0xf000
132 ; SI-NEXT: s_mov_b32 s2, -1
133 ; SI-NEXT: s_waitcnt vmcnt(0)
134 ; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
135 ; SI-NEXT: s_endpgm
136 ;
137 ; GFX7-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
138 ; GFX7: ; %bb.0: ; %main_body
139 ; GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
140 ; GFX7-NEXT: s_mov_b32 s3, 0xf000
141 ; GFX7-NEXT: s_mov_b32 s2, -1
142 ; GFX7-NEXT: s_waitcnt vmcnt(0)
143 ; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
144 ; GFX7-NEXT: s_endpgm
145 ;
146 ; GFX10-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
147 ; GFX10: ; %bb.0: ; %main_body
148 ; GFX10-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
149 ; GFX10-NEXT: s_waitcnt vmcnt(0)
150 ; GFX10-NEXT: global_store_dword v[0:1], v0, off
151 ; GFX10-NEXT: s_endpgm
152 ;
153 ; GFX1030-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
154 ; GFX1030: ; %bb.0: ; %main_body
155 ; GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
156 ; GFX1030-NEXT: s_waitcnt vmcnt(0)
157 ; GFX1030-NEXT: global_store_dword v[0:1], v0, off
158 ; GFX1030-NEXT: s_endpgm
159 ;
160 ; GFX1100-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
161 ; GFX1100: ; %bb.0: ; %main_body
162 ; GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 0 offen glc
163 ; GFX1100-NEXT: s_waitcnt vmcnt(0)
164 ; GFX1100-NEXT: global_store_b32 v[0:1], v0, off
165 ; GFX1100-NEXT: s_endpgm
166 ;
167 ; G_SI-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
168 ; G_SI: ; %bb.0: ; %main_body
169 ; G_SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
170 ; G_SI-NEXT: s_mov_b32 s2, -1
171 ; G_SI-NEXT: s_mov_b32 s3, 0xf000
172 ; G_SI-NEXT: s_waitcnt vmcnt(0)
173 ; G_SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
174 ; G_SI-NEXT: s_endpgm
175 ;
176 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
177 ; G_GFX7: ; %bb.0: ; %main_body
178 ; G_GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
179 ; G_GFX7-NEXT: s_mov_b32 s2, -1
180 ; G_GFX7-NEXT: s_mov_b32 s3, 0xf000
181 ; G_GFX7-NEXT: s_waitcnt vmcnt(0)
182 ; G_GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
183 ; G_GFX7-NEXT: s_endpgm
184 ;
185 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
186 ; G_GFX10: ; %bb.0: ; %main_body
187 ; G_GFX10-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
188 ; G_GFX10-NEXT: s_waitcnt vmcnt(0)
189 ; G_GFX10-NEXT: global_store_dword v[0:1], v0, off
190 ; G_GFX10-NEXT: s_endpgm
191 ;
192 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
193 ; G_GFX1030: ; %bb.0: ; %main_body
194 ; G_GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 0 offen glc
195 ; G_GFX1030-NEXT: s_waitcnt vmcnt(0)
196 ; G_GFX1030-NEXT: global_store_dword v[0:1], v0, off
197 ; G_GFX1030-NEXT: s_endpgm
198 ;
199 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_min_rtn_f32:
200 ; G_GFX1100: ; %bb.0: ; %main_body
201 ; G_GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 0 offen glc
202 ; G_GFX1100-NEXT: s_waitcnt vmcnt(0)
203 ; G_GFX1100-NEXT: global_store_b32 v[0:1], v0, off
204 ; G_GFX1100-NEXT: s_endpgm
205 main_body:
206 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmin.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0)
207 store float %ret, ptr addrspace(1) undef
208 ret void
209 }
211 define amdgpu_kernel void @raw_ptr_buffer_atomic_min_rtn_f32_off4_slc(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex, ptr addrspace(3) %out) {
212 ; SI-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
213 ; SI: ; %bb.0: ; %main_body
214 ; SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
215 ; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
216 ; SI-NEXT: s_mov_b32 m0, -1
217 ; SI-NEXT: s_waitcnt lgkmcnt(0)
218 ; SI-NEXT: v_mov_b32_e32 v0, s6
219 ; SI-NEXT: v_mov_b32_e32 v1, s7
220 ; SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
221 ; SI-NEXT: s_load_dword s0, s[4:5], 0xf
222 ; SI-NEXT: s_waitcnt lgkmcnt(0)
223 ; SI-NEXT: v_mov_b32_e32 v1, s0
224 ; SI-NEXT: s_waitcnt vmcnt(0)
225 ; SI-NEXT: ds_write_b32 v1, v0
226 ; SI-NEXT: s_endpgm
227 ;
228 ; GFX7-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
229 ; GFX7: ; %bb.0: ; %main_body
230 ; GFX7-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
231 ; GFX7-NEXT: s_mov_b32 m0, -1
232 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
233 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
234 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
235 ; GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
236 ; GFX7-NEXT: v_mov_b32_e32 v1, s6
237 ; GFX7-NEXT: s_waitcnt vmcnt(0)
238 ; GFX7-NEXT: ds_write_b32 v1, v0
239 ; GFX7-NEXT: s_endpgm
240 ;
241 ; GFX10-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
242 ; GFX10: ; %bb.0: ; %main_body
243 ; GFX10-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
244 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
245 ; GFX10-NEXT: v_mov_b32_e32 v0, s12
246 ; GFX10-NEXT: v_mov_b32_e32 v1, s13
247 ; GFX10-NEXT: buffer_atomic_fmin v0, v1, s[8:11], 4 offen glc slc
248 ; GFX10-NEXT: v_mov_b32_e32 v1, s14
249 ; GFX10-NEXT: s_waitcnt vmcnt(0)
250 ; GFX10-NEXT: ds_write_b32 v1, v0
251 ; GFX10-NEXT: s_endpgm
252 ;
253 ; GFX1030-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
254 ; GFX1030: ; %bb.0: ; %main_body
255 ; GFX1030-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
256 ; GFX1030-NEXT: s_waitcnt lgkmcnt(0)
257 ; GFX1030-NEXT: v_mov_b32_e32 v0, s4
258 ; GFX1030-NEXT: v_mov_b32_e32 v1, s5
259 ; GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
260 ; GFX1030-NEXT: v_mov_b32_e32 v1, s6
261 ; GFX1030-NEXT: s_waitcnt vmcnt(0)
262 ; GFX1030-NEXT: ds_write_b32 v1, v0
263 ; GFX1030-NEXT: s_endpgm
264 ;
265 ; GFX1100-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
266 ; GFX1100: ; %bb.0: ; %main_body
267 ; GFX1100-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
268 ; GFX1100-NEXT: s_waitcnt lgkmcnt(0)
269 ; GFX1100-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
270 ; GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 4 offen glc slc
271 ; GFX1100-NEXT: v_mov_b32_e32 v1, s6
272 ; GFX1100-NEXT: s_waitcnt vmcnt(0)
273 ; GFX1100-NEXT: ds_store_b32 v1, v0
274 ; GFX1100-NEXT: s_endpgm
275 ;
276 ; G_SI-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
277 ; G_SI: ; %bb.0: ; %main_body
278 ; G_SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
279 ; G_SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
280 ; G_SI-NEXT: s_mov_b32 m0, -1
281 ; G_SI-NEXT: s_waitcnt lgkmcnt(0)
282 ; G_SI-NEXT: v_mov_b32_e32 v0, s6
283 ; G_SI-NEXT: v_mov_b32_e32 v1, s7
284 ; G_SI-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
285 ; G_SI-NEXT: s_load_dword s0, s[4:5], 0xf
286 ; G_SI-NEXT: s_waitcnt lgkmcnt(0)
287 ; G_SI-NEXT: v_mov_b32_e32 v1, s0
288 ; G_SI-NEXT: s_waitcnt vmcnt(0)
289 ; G_SI-NEXT: ds_write_b32 v1, v0
290 ; G_SI-NEXT: s_endpgm
291 ;
292 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
293 ; G_GFX7: ; %bb.0: ; %main_body
294 ; G_GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
295 ; G_GFX7-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
296 ; G_GFX7-NEXT: s_mov_b32 m0, -1
297 ; G_GFX7-NEXT: s_waitcnt lgkmcnt(0)
298 ; G_GFX7-NEXT: v_mov_b32_e32 v0, s6
299 ; G_GFX7-NEXT: v_mov_b32_e32 v1, s7
300 ; G_GFX7-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
301 ; G_GFX7-NEXT: s_load_dword s0, s[4:5], 0xf
302 ; G_GFX7-NEXT: s_waitcnt lgkmcnt(0)
303 ; G_GFX7-NEXT: v_mov_b32_e32 v1, s0
304 ; G_GFX7-NEXT: s_waitcnt vmcnt(0)
305 ; G_GFX7-NEXT: ds_write_b32 v1, v0
306 ; G_GFX7-NEXT: s_endpgm
307 ;
308 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
309 ; G_GFX10: ; %bb.0: ; %main_body
310 ; G_GFX10-NEXT: s_clause 0x1
311 ; G_GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
312 ; G_GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
313 ; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
314 ; G_GFX10-NEXT: v_mov_b32_e32 v0, s6
315 ; G_GFX10-NEXT: v_mov_b32_e32 v1, s7
316 ; G_GFX10-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
317 ; G_GFX10-NEXT: s_waitcnt_depctr 0xffe3
318 ; G_GFX10-NEXT: s_load_dword s0, s[4:5], 0x3c
319 ; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
320 ; G_GFX10-NEXT: v_mov_b32_e32 v1, s0
321 ; G_GFX10-NEXT: s_waitcnt vmcnt(0)
322 ; G_GFX10-NEXT: ds_write_b32 v1, v0
323 ; G_GFX10-NEXT: s_endpgm
324 ;
325 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
326 ; G_GFX1030: ; %bb.0: ; %main_body
327 ; G_GFX1030-NEXT: s_clause 0x1
328 ; G_GFX1030-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
329 ; G_GFX1030-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
330 ; G_GFX1030-NEXT: s_waitcnt lgkmcnt(0)
331 ; G_GFX1030-NEXT: v_mov_b32_e32 v0, s6
332 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, s7
333 ; G_GFX1030-NEXT: buffer_atomic_fmin v0, v1, s[0:3], 4 offen glc slc
334 ; G_GFX1030-NEXT: s_load_dword s0, s[4:5], 0x3c
335 ; G_GFX1030-NEXT: s_waitcnt lgkmcnt(0)
336 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, s0
337 ; G_GFX1030-NEXT: s_waitcnt vmcnt(0)
338 ; G_GFX1030-NEXT: ds_write_b32 v1, v0
339 ; G_GFX1030-NEXT: s_endpgm
340 ;
341 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
342 ; G_GFX1100: ; %bb.0: ; %main_body
343 ; G_GFX1100-NEXT: s_clause 0x1
344 ; G_GFX1100-NEXT: s_load_b64 s[6:7], s[4:5], 0x34
345 ; G_GFX1100-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
346 ; G_GFX1100-NEXT: s_waitcnt lgkmcnt(0)
347 ; G_GFX1100-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
348 ; G_GFX1100-NEXT: buffer_atomic_min_f32 v0, v1, s[0:3], 4 offen glc slc
349 ; G_GFX1100-NEXT: s_load_b32 s0, s[4:5], 0x3c
350 ; G_GFX1100-NEXT: s_waitcnt lgkmcnt(0)
351 ; G_GFX1100-NEXT: v_mov_b32_e32 v1, s0
352 ; G_GFX1100-NEXT: s_waitcnt vmcnt(0)
353 ; G_GFX1100-NEXT: ds_store_b32 v1, v0
354 ; G_GFX1100-NEXT: s_endpgm
355 ; GFX1010-LABEL: raw_ptr_buffer_atomic_min_rtn_f32_off4_slc:
356 main_body:
357 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmin.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 4, i32 2)
358 store float %ret, ptr addrspace(3) %out, align 8
359 ret void
360 }
362 define amdgpu_kernel void @raw_ptr_buffer_atomic_max_noret_f32(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex) {
363 ; SI-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
364 ; SI: ; %bb.0: ; %main_body
365 ; SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
366 ; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
367 ; SI-NEXT: s_waitcnt lgkmcnt(0)
368 ; SI-NEXT: v_mov_b32_e32 v0, s6
369 ; SI-NEXT: v_mov_b32_e32 v1, s7
370 ; SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
371 ; SI-NEXT: s_endpgm
372 ;
373 ; GFX7-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
374 ; GFX7: ; %bb.0: ; %main_body
375 ; GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
376 ; GFX7-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
377 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
378 ; GFX7-NEXT: v_mov_b32_e32 v0, s6
379 ; GFX7-NEXT: v_mov_b32_e32 v1, s7
380 ; GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
381 ; GFX7-NEXT: s_endpgm
382 ;
383 ; GFX10-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
384 ; GFX10: ; %bb.0: ; %main_body
385 ; GFX10-NEXT: s_clause 0x1
386 ; GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
387 ; GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
388 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
389 ; GFX10-NEXT: v_mov_b32_e32 v0, s6
390 ; GFX10-NEXT: v_mov_b32_e32 v1, s7
391 ; GFX10-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
392 ; GFX10-NEXT: s_endpgm
393 ;
394 ; GFX1030-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
395 ; GFX1030: ; %bb.0: ; %main_body
396 ; GFX1030-NEXT: s_clause 0x1
397 ; GFX1030-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
398 ; GFX1030-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
399 ; GFX1030-NEXT: s_waitcnt lgkmcnt(0)
400 ; GFX1030-NEXT: v_mov_b32_e32 v0, s6
401 ; GFX1030-NEXT: v_mov_b32_e32 v1, s7
402 ; GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
403 ; GFX1030-NEXT: s_endpgm
404 ;
405 ; GFX1100-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
406 ; GFX1100: ; %bb.0: ; %main_body
407 ; GFX1100-NEXT: s_clause 0x1
408 ; GFX1100-NEXT: s_load_b64 s[6:7], s[4:5], 0x34
409 ; GFX1100-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
410 ; GFX1100-NEXT: s_waitcnt lgkmcnt(0)
411 ; GFX1100-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
412 ; GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 0 offen
413 ; GFX1100-NEXT: s_endpgm
414 ;
415 ; G_SI-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
416 ; G_SI: ; %bb.0: ; %main_body
417 ; G_SI-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
418 ; G_SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
419 ; G_SI-NEXT: s_waitcnt lgkmcnt(0)
420 ; G_SI-NEXT: v_mov_b32_e32 v0, s6
421 ; G_SI-NEXT: v_mov_b32_e32 v1, s7
422 ; G_SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
423 ; G_SI-NEXT: s_endpgm
424 ;
425 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
426 ; G_GFX7: ; %bb.0: ; %main_body
427 ; G_GFX7-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0xd
428 ; G_GFX7-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9
429 ; G_GFX7-NEXT: s_waitcnt lgkmcnt(0)
430 ; G_GFX7-NEXT: v_mov_b32_e32 v0, s6
431 ; G_GFX7-NEXT: v_mov_b32_e32 v1, s7
432 ; G_GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
433 ; G_GFX7-NEXT: s_endpgm
434 ;
435 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
436 ; G_GFX10: ; %bb.0: ; %main_body
437 ; G_GFX10-NEXT: s_clause 0x1
438 ; G_GFX10-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
439 ; G_GFX10-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
440 ; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
441 ; G_GFX10-NEXT: v_mov_b32_e32 v0, s6
442 ; G_GFX10-NEXT: v_mov_b32_e32 v1, s7
443 ; G_GFX10-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
444 ; G_GFX10-NEXT: s_endpgm
445 ;
446 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
447 ; G_GFX1030: ; %bb.0: ; %main_body
448 ; G_GFX1030-NEXT: s_clause 0x1
449 ; G_GFX1030-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x34
450 ; G_GFX1030-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24
451 ; G_GFX1030-NEXT: s_waitcnt lgkmcnt(0)
452 ; G_GFX1030-NEXT: v_mov_b32_e32 v0, s6
453 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, s7
454 ; G_GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen
455 ; G_GFX1030-NEXT: s_endpgm
456 ;
457 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_max_noret_f32:
458 ; G_GFX1100: ; %bb.0: ; %main_body
459 ; G_GFX1100-NEXT: s_clause 0x1
460 ; G_GFX1100-NEXT: s_load_b64 s[6:7], s[4:5], 0x34
461 ; G_GFX1100-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
462 ; G_GFX1100-NEXT: s_waitcnt lgkmcnt(0)
463 ; G_GFX1100-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
464 ; G_GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 0 offen
465 ; G_GFX1100-NEXT: s_endpgm
466 main_body:
467 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmax.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0)
468 ret void
469 }
471 define amdgpu_ps void @raw_ptr_buffer_atomic_max_rtn_f32(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex) {
472 ; SI-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
473 ; SI: ; %bb.0: ; %main_body
474 ; SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
475 ; SI-NEXT: s_mov_b32 s3, 0xf000
476 ; SI-NEXT: s_mov_b32 s2, -1
477 ; SI-NEXT: s_waitcnt vmcnt(0)
478 ; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
479 ; SI-NEXT: s_endpgm
480 ;
481 ; GFX7-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
482 ; GFX7: ; %bb.0: ; %main_body
483 ; GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
484 ; GFX7-NEXT: s_mov_b32 s3, 0xf000
485 ; GFX7-NEXT: s_mov_b32 s2, -1
486 ; GFX7-NEXT: s_waitcnt vmcnt(0)
487 ; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
488 ; GFX7-NEXT: s_endpgm
489 ;
490 ; GFX10-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
491 ; GFX10: ; %bb.0: ; %main_body
492 ; GFX10-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
493 ; GFX10-NEXT: s_waitcnt vmcnt(0)
494 ; GFX10-NEXT: global_store_dword v[0:1], v0, off
495 ; GFX10-NEXT: s_endpgm
496 ;
497 ; GFX1030-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
498 ; GFX1030: ; %bb.0: ; %main_body
499 ; GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
500 ; GFX1030-NEXT: s_waitcnt vmcnt(0)
501 ; GFX1030-NEXT: global_store_dword v[0:1], v0, off
502 ; GFX1030-NEXT: s_endpgm
503 ;
504 ; GFX1100-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
505 ; GFX1100: ; %bb.0: ; %main_body
506 ; GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 0 offen glc
507 ; GFX1100-NEXT: s_waitcnt vmcnt(0)
508 ; GFX1100-NEXT: global_store_b32 v[0:1], v0, off
509 ; GFX1100-NEXT: s_endpgm
510 ;
511 ; G_SI-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
512 ; G_SI: ; %bb.0: ; %main_body
513 ; G_SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
514 ; G_SI-NEXT: s_mov_b32 s2, -1
515 ; G_SI-NEXT: s_mov_b32 s3, 0xf000
516 ; G_SI-NEXT: s_waitcnt vmcnt(0)
517 ; G_SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
518 ; G_SI-NEXT: s_endpgm
519 ;
520 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
521 ; G_GFX7: ; %bb.0: ; %main_body
522 ; G_GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
523 ; G_GFX7-NEXT: s_mov_b32 s2, -1
524 ; G_GFX7-NEXT: s_mov_b32 s3, 0xf000
525 ; G_GFX7-NEXT: s_waitcnt vmcnt(0)
526 ; G_GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
527 ; G_GFX7-NEXT: s_endpgm
528 ;
529 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
530 ; G_GFX10: ; %bb.0: ; %main_body
531 ; G_GFX10-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
532 ; G_GFX10-NEXT: s_waitcnt vmcnt(0)
533 ; G_GFX10-NEXT: global_store_dword v[0:1], v0, off
534 ; G_GFX10-NEXT: s_endpgm
535 ;
536 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
537 ; G_GFX1030: ; %bb.0: ; %main_body
538 ; G_GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 0 offen glc
539 ; G_GFX1030-NEXT: s_waitcnt vmcnt(0)
540 ; G_GFX1030-NEXT: global_store_dword v[0:1], v0, off
541 ; G_GFX1030-NEXT: s_endpgm
542 ;
543 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_max_rtn_f32:
544 ; G_GFX1100: ; %bb.0: ; %main_body
545 ; G_GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 0 offen glc
546 ; G_GFX1100-NEXT: s_waitcnt vmcnt(0)
547 ; G_GFX1100-NEXT: global_store_b32 v[0:1], v0, off
548 ; G_GFX1100-NEXT: s_endpgm
549 main_body:
550 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmax.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 0, i32 0)
551 store float %ret, ptr addrspace(1) undef
552 ret void
553 }
555 define amdgpu_kernel void @raw_ptr_buffer_atomic_max_rtn_f32_off4_slc(ptr addrspace(8) inreg %rsrc, float %data, i32 %vindex, ptr addrspace(1) %out) {
556 ; SI-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
557 ; SI: ; %bb.0: ; %main_body
558 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
559 ; SI-NEXT: s_waitcnt lgkmcnt(0)
560 ; SI-NEXT: v_mov_b32_e32 v0, s4
561 ; SI-NEXT: v_mov_b32_e32 v1, s5
562 ; SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
563 ; SI-NEXT: s_mov_b32 s3, 0xf000
564 ; SI-NEXT: s_mov_b32 s2, -1
565 ; SI-NEXT: s_mov_b32 s0, s6
566 ; SI-NEXT: s_mov_b32 s1, s7
567 ; SI-NEXT: s_waitcnt vmcnt(0)
568 ; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
569 ; SI-NEXT: s_endpgm
570 ;
571 ; GFX7-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
572 ; GFX7: ; %bb.0: ; %main_body
573 ; GFX7-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
574 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
575 ; GFX7-NEXT: v_mov_b32_e32 v0, s4
576 ; GFX7-NEXT: v_mov_b32_e32 v1, s5
577 ; GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
578 ; GFX7-NEXT: s_mov_b32 s3, 0xf000
579 ; GFX7-NEXT: s_mov_b32 s2, -1
580 ; GFX7-NEXT: s_mov_b32 s0, s6
581 ; GFX7-NEXT: s_mov_b32 s1, s7
582 ; GFX7-NEXT: s_waitcnt vmcnt(0)
583 ; GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
584 ; GFX7-NEXT: s_endpgm
585 ;
586 ; GFX10-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
587 ; GFX10: ; %bb.0: ; %main_body
588 ; GFX10-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
589 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
590 ; GFX10-NEXT: v_mov_b32_e32 v0, s12
591 ; GFX10-NEXT: v_mov_b32_e32 v1, s13
592 ; GFX10-NEXT: buffer_atomic_fmax v0, v1, s[8:11], 4 offen glc slc
593 ; GFX10-NEXT: v_mov_b32_e32 v1, 0
594 ; GFX10-NEXT: s_waitcnt vmcnt(0)
595 ; GFX10-NEXT: global_store_dword v1, v0, s[14:15]
596 ; GFX10-NEXT: s_endpgm
597 ;
598 ; GFX1030-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
599 ; GFX1030: ; %bb.0: ; %main_body
600 ; GFX1030-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
601 ; GFX1030-NEXT: s_waitcnt lgkmcnt(0)
602 ; GFX1030-NEXT: v_mov_b32_e32 v0, s4
603 ; GFX1030-NEXT: v_mov_b32_e32 v1, s5
604 ; GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
605 ; GFX1030-NEXT: v_mov_b32_e32 v1, 0
606 ; GFX1030-NEXT: s_waitcnt vmcnt(0)
607 ; GFX1030-NEXT: global_store_dword v1, v0, s[6:7]
608 ; GFX1030-NEXT: s_endpgm
609 ;
610 ; GFX1100-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
611 ; GFX1100: ; %bb.0: ; %main_body
612 ; GFX1100-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
613 ; GFX1100-NEXT: s_waitcnt lgkmcnt(0)
614 ; GFX1100-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
615 ; GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 4 offen glc slc
616 ; GFX1100-NEXT: v_mov_b32_e32 v1, 0
617 ; GFX1100-NEXT: s_waitcnt vmcnt(0)
618 ; GFX1100-NEXT: global_store_b32 v1, v0, s[6:7]
619 ; GFX1100-NEXT: s_endpgm
620 ;
621 ; G_SI-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
622 ; G_SI: ; %bb.0: ; %main_body
623 ; G_SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
624 ; G_SI-NEXT: s_waitcnt lgkmcnt(0)
625 ; G_SI-NEXT: v_mov_b32_e32 v0, s4
626 ; G_SI-NEXT: v_mov_b32_e32 v1, s5
627 ; G_SI-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
628 ; G_SI-NEXT: s_mov_b32 s2, -1
629 ; G_SI-NEXT: s_mov_b32 s3, 0xf000
630 ; G_SI-NEXT: s_mov_b64 s[0:1], s[6:7]
631 ; G_SI-NEXT: s_waitcnt vmcnt(0)
632 ; G_SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
633 ; G_SI-NEXT: s_endpgm
634 ;
635 ; G_GFX7-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
636 ; G_GFX7: ; %bb.0: ; %main_body
637 ; G_GFX7-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
638 ; G_GFX7-NEXT: s_waitcnt lgkmcnt(0)
639 ; G_GFX7-NEXT: v_mov_b32_e32 v0, s4
640 ; G_GFX7-NEXT: v_mov_b32_e32 v1, s5
641 ; G_GFX7-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
642 ; G_GFX7-NEXT: s_mov_b32 s2, -1
643 ; G_GFX7-NEXT: s_mov_b32 s3, 0xf000
644 ; G_GFX7-NEXT: s_mov_b64 s[0:1], s[6:7]
645 ; G_GFX7-NEXT: s_waitcnt vmcnt(0)
646 ; G_GFX7-NEXT: buffer_store_dword v0, off, s[0:3], 0
647 ; G_GFX7-NEXT: s_endpgm
648 ;
649 ; G_GFX10-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
650 ; G_GFX10: ; %bb.0: ; %main_body
651 ; G_GFX10-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
652 ; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
653 ; G_GFX10-NEXT: v_mov_b32_e32 v0, s12
654 ; G_GFX10-NEXT: v_mov_b32_e32 v1, s13
655 ; G_GFX10-NEXT: buffer_atomic_fmax v0, v1, s[8:11], 4 offen glc slc
656 ; G_GFX10-NEXT: v_mov_b32_e32 v1, 0
657 ; G_GFX10-NEXT: s_waitcnt vmcnt(0)
658 ; G_GFX10-NEXT: global_store_dword v1, v0, s[14:15]
659 ; G_GFX10-NEXT: s_endpgm
660 ;
661 ; G_GFX1030-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
662 ; G_GFX1030: ; %bb.0: ; %main_body
663 ; G_GFX1030-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
664 ; G_GFX1030-NEXT: s_waitcnt lgkmcnt(0)
665 ; G_GFX1030-NEXT: v_mov_b32_e32 v0, s4
666 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, s5
667 ; G_GFX1030-NEXT: buffer_atomic_fmax v0, v1, s[0:3], 4 offen glc slc
668 ; G_GFX1030-NEXT: v_mov_b32_e32 v1, 0
669 ; G_GFX1030-NEXT: s_waitcnt vmcnt(0)
670 ; G_GFX1030-NEXT: global_store_dword v1, v0, s[6:7]
671 ; G_GFX1030-NEXT: s_endpgm
672 ;
673 ; G_GFX1100-LABEL: raw_ptr_buffer_atomic_max_rtn_f32_off4_slc:
674 ; G_GFX1100: ; %bb.0: ; %main_body
675 ; G_GFX1100-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
676 ; G_GFX1100-NEXT: s_waitcnt lgkmcnt(0)
677 ; G_GFX1100-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
678 ; G_GFX1100-NEXT: buffer_atomic_max_f32 v0, v1, s[0:3], 4 offen glc slc
679 ; G_GFX1100-NEXT: v_mov_b32_e32 v1, 0
680 ; G_GFX1100-NEXT: s_waitcnt vmcnt(0)
681 ; G_GFX1100-NEXT: global_store_b32 v1, v0, s[6:7]
682 ; G_GFX1100-NEXT: s_endpgm
683 main_body:
684 %ret = call float @llvm.amdgcn.raw.ptr.buffer.atomic.fmax.f32(float %data, ptr addrspace(8) %rsrc, i32 %vindex, i32 4, i32 2)
685 store float %ret, ptr addrspace(1) %out, align 8
686 ret void
687 }