1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX7LESS %s
3 ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s
4 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64 -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1064 %s
5 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32 -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1032 %s
6 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1164 %s
7 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize32, -amdgpu-atomic-optimizer-strategy=Iterative -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1132 %s
8 ; RUN: llc -mtriple=amdgcn -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX7LESS-DPP %s
9 ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9-DPP %s
10 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64 -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1064-DPP %s
11 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32 -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1032-DPP %s
12 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1164-DPP %s
13 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize32 -amdgpu-atomic-optimizer-strategy=DPP -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX1132-DPP %s
15 declare float @div.float.value()
16 declare float @div.double.value()
18 define amdgpu_kernel void @global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
19 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
21 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
22 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
23 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
24 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
25 ; GFX7LESS-NEXT: s_cbranch_execz .LBB0_3
26 ; GFX7LESS-NEXT: ; %bb.1:
27 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
28 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
29 ; GFX7LESS-NEXT: s_load_dword s2, s[0:1], 0x0
30 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
31 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
32 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
33 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s2
34 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
35 ; GFX7LESS-NEXT: .LBB0_2: ; %atomicrmw.start
36 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
37 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
38 ; GFX7LESS-NEXT: v_max_f32_e32 v0, 4.0, v0
39 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
40 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v1
41 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v0
42 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
43 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
44 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
45 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
46 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v2
47 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
48 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB0_2
49 ; GFX7LESS-NEXT: .LBB0_3:
50 ; GFX7LESS-NEXT: s_endpgm
52 ; GFX9-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
54 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
55 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
56 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
57 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
58 ; GFX9-NEXT: s_cbranch_execz .LBB0_3
60 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
61 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
62 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
63 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
64 ; GFX9-NEXT: s_load_dword s4, s[0:1], 0x0
65 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
66 ; GFX9-NEXT: v_mov_b32_e32 v1, s4
67 ; GFX9-NEXT: .LBB0_2: ; %atomicrmw.start
68 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
69 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
70 ; GFX9-NEXT: v_max_f32_e32 v0, 4.0, v0
71 ; GFX9-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
72 ; GFX9-NEXT: s_waitcnt vmcnt(0)
73 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
74 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
75 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
76 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
77 ; GFX9-NEXT: s_cbranch_execnz .LBB0_2
81 ; GFX1064-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
83 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
84 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
85 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
86 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
87 ; GFX1064-NEXT: s_cbranch_execz .LBB0_2
88 ; GFX1064-NEXT: ; %bb.1:
89 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
90 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
91 ; GFX1064-NEXT: v_mov_b32_e32 v1, 4.0
92 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
93 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
94 ; GFX1064-NEXT: .LBB0_2:
95 ; GFX1064-NEXT: s_endpgm
97 ; GFX1032-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
99 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
100 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
101 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
102 ; GFX1032-NEXT: s_cbranch_execz .LBB0_2
103 ; GFX1032-NEXT: ; %bb.1:
104 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
105 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
106 ; GFX1032-NEXT: v_mov_b32_e32 v1, 4.0
107 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
108 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
109 ; GFX1032-NEXT: .LBB0_2:
110 ; GFX1032-NEXT: s_endpgm
112 ; GFX1164-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
114 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
115 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
116 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
117 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
118 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
119 ; GFX1164-NEXT: s_cbranch_execz .LBB0_2
120 ; GFX1164-NEXT: ; %bb.1:
121 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
122 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
123 ; GFX1164-NEXT: v_mov_b32_e32 v1, 4.0
124 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
125 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
126 ; GFX1164-NEXT: .LBB0_2:
127 ; GFX1164-NEXT: s_endpgm
129 ; GFX1132-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
131 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
132 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
133 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
134 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
135 ; GFX1132-NEXT: s_cbranch_execz .LBB0_2
136 ; GFX1132-NEXT: ; %bb.1:
137 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
138 ; GFX1132-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
139 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
140 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
141 ; GFX1132-NEXT: .LBB0_2:
142 ; GFX1132-NEXT: s_endpgm
144 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
145 ; GFX7LESS-DPP: ; %bb.0:
146 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
147 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
148 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
149 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
150 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB0_3
151 ; GFX7LESS-DPP-NEXT: ; %bb.1:
152 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
153 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
154 ; GFX7LESS-DPP-NEXT: s_load_dword s2, s[0:1], 0x0
155 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
156 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
157 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
158 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s2
159 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
160 ; GFX7LESS-DPP-NEXT: .LBB0_2: ; %atomicrmw.start
161 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
162 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
163 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
164 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
165 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v1
166 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v0
167 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
168 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
169 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
170 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
171 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v2
172 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
173 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB0_2
174 ; GFX7LESS-DPP-NEXT: .LBB0_3:
175 ; GFX7LESS-DPP-NEXT: s_endpgm
177 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
179 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
180 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
181 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
182 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
183 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB0_3
184 ; GFX9-DPP-NEXT: ; %bb.1:
185 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
186 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
187 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
188 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
189 ; GFX9-DPP-NEXT: s_load_dword s4, s[0:1], 0x0
190 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
191 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s4
192 ; GFX9-DPP-NEXT: .LBB0_2: ; %atomicrmw.start
193 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
194 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
195 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
196 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
197 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
198 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
199 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
200 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
201 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
202 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB0_2
203 ; GFX9-DPP-NEXT: .LBB0_3:
204 ; GFX9-DPP-NEXT: s_endpgm
206 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
207 ; GFX1064-DPP: ; %bb.0:
208 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
209 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
210 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
211 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
212 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB0_2
213 ; GFX1064-DPP-NEXT: ; %bb.1:
214 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
215 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
216 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 4.0
217 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
218 ; GFX1064-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
219 ; GFX1064-DPP-NEXT: .LBB0_2:
220 ; GFX1064-DPP-NEXT: s_endpgm
222 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
223 ; GFX1032-DPP: ; %bb.0:
224 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
225 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
226 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
227 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB0_2
228 ; GFX1032-DPP-NEXT: ; %bb.1:
229 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
230 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
231 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 4.0
232 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
233 ; GFX1032-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
234 ; GFX1032-DPP-NEXT: .LBB0_2:
235 ; GFX1032-DPP-NEXT: s_endpgm
237 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
238 ; GFX1164-DPP: ; %bb.0:
239 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
240 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
241 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
242 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
243 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
244 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB0_2
245 ; GFX1164-DPP-NEXT: ; %bb.1:
246 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
247 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
248 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 4.0
249 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
250 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
251 ; GFX1164-DPP-NEXT: .LBB0_2:
252 ; GFX1164-DPP-NEXT: s_endpgm
254 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_agent_scope_unsafe:
255 ; GFX1132-DPP: ; %bb.0:
256 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
257 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
258 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
259 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
260 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB0_2
261 ; GFX1132-DPP-NEXT: ; %bb.1:
262 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
263 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
264 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
265 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
266 ; GFX1132-DPP-NEXT: .LBB0_2:
267 ; GFX1132-DPP-NEXT: s_endpgm
268 %result = atomicrmw fmax ptr addrspace(1) %ptr, float 4.0 syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
272 define amdgpu_kernel void @global_atomic_fmax_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
273 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
275 ; GFX7LESS-NEXT: s_mov_b32 s32, 0
276 ; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
277 ; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
278 ; GFX7LESS-NEXT: s_mov_b32 s38, -1
279 ; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
280 ; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
281 ; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
282 ; GFX7LESS-NEXT: s_mov_b32 s14, s10
283 ; GFX7LESS-NEXT: s_mov_b32 s13, s9
284 ; GFX7LESS-NEXT: s_mov_b32 s12, s8
285 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
286 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
287 ; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
288 ; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
289 ; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
290 ; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
291 ; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
292 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
293 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
294 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
295 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
296 ; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
297 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
298 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
299 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
300 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
301 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
302 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
303 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
304 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0x7fc00000
305 ; GFX7LESS-NEXT: .LBB1_1: ; %ComputeLoop
306 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
307 ; GFX7LESS-NEXT: s_ff1_i32_b64 s2, s[0:1]
308 ; GFX7LESS-NEXT: v_mul_f32_e32 v1, 1.0, v2
309 ; GFX7LESS-NEXT: v_readlane_b32 s4, v0, s2
310 ; GFX7LESS-NEXT: s_lshl_b64 s[2:3], 1, s2
311 ; GFX7LESS-NEXT: v_mul_f32_e64 v2, 1.0, s4
312 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
313 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
314 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
315 ; GFX7LESS-NEXT: v_max_f32_e32 v2, v1, v2
316 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB1_1
317 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
318 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
319 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
320 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
321 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
322 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
323 ; GFX7LESS-NEXT: s_cbranch_execz .LBB1_5
324 ; GFX7LESS-NEXT: ; %bb.3:
325 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
326 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
327 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
328 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
329 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[0:3], 0
330 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
331 ; GFX7LESS-NEXT: v_mul_f32_e32 v2, 1.0, v2
332 ; GFX7LESS-NEXT: .LBB1_4: ; %atomicrmw.start
333 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
334 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
335 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
336 ; GFX7LESS-NEXT: v_max_f32_e32 v0, v0, v2
337 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
338 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, v1
339 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v0
340 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[3:4], off, s[0:3], 0 glc
341 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
342 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
343 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
344 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v3
345 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
346 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB1_4
347 ; GFX7LESS-NEXT: .LBB1_5:
348 ; GFX7LESS-NEXT: s_endpgm
350 ; GFX9-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
352 ; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
353 ; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
354 ; GFX9-NEXT: s_mov_b32 s38, -1
355 ; GFX9-NEXT: s_mov_b32 s39, 0xe00000
356 ; GFX9-NEXT: s_add_u32 s36, s36, s11
357 ; GFX9-NEXT: s_addc_u32 s37, s37, 0
358 ; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
359 ; GFX9-NEXT: s_mov_b32 s12, s8
360 ; GFX9-NEXT: s_add_u32 s8, s34, 44
361 ; GFX9-NEXT: s_mov_b32 s13, s9
362 ; GFX9-NEXT: s_addc_u32 s9, s35, 0
363 ; GFX9-NEXT: s_getpc_b64 s[4:5]
364 ; GFX9-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
365 ; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
366 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
367 ; GFX9-NEXT: s_mov_b32 s14, s10
368 ; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
369 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
370 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
371 ; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
372 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
373 ; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
374 ; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
375 ; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
376 ; GFX9-NEXT: s_mov_b32 s32, 0
377 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
378 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
379 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
380 ; GFX9-NEXT: v_mov_b32_e32 v2, 0x7fc00000
381 ; GFX9-NEXT: .LBB1_1: ; %ComputeLoop
382 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
383 ; GFX9-NEXT: s_ff1_i32_b64 s2, s[0:1]
384 ; GFX9-NEXT: v_readlane_b32 s4, v0, s2
385 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s2
386 ; GFX9-NEXT: v_max_f32_e32 v1, v2, v2
387 ; GFX9-NEXT: v_max_f32_e64 v2, s4, s4
388 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
389 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
390 ; GFX9-NEXT: v_max_f32_e32 v2, v1, v2
391 ; GFX9-NEXT: s_cbranch_scc1 .LBB1_1
392 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
393 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
394 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
395 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
396 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
397 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
398 ; GFX9-NEXT: s_cbranch_execz .LBB1_5
399 ; GFX9-NEXT: ; %bb.3:
400 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
401 ; GFX9-NEXT: v_mov_b32_e32 v3, 0
402 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
403 ; GFX9-NEXT: v_max_f32_e32 v2, v2, v2
404 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
405 ; GFX9-NEXT: global_load_dword v1, v3, s[0:1]
406 ; GFX9-NEXT: .LBB1_4: ; %atomicrmw.start
407 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
408 ; GFX9-NEXT: s_waitcnt vmcnt(0)
409 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
410 ; GFX9-NEXT: v_max_f32_e32 v0, v0, v2
411 ; GFX9-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
412 ; GFX9-NEXT: s_waitcnt vmcnt(0)
413 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
414 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
415 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
416 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
417 ; GFX9-NEXT: s_cbranch_execnz .LBB1_4
418 ; GFX9-NEXT: .LBB1_5:
419 ; GFX9-NEXT: s_endpgm
421 ; GFX1064-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
423 ; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
424 ; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
425 ; GFX1064-NEXT: s_mov_b32 s38, -1
426 ; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
427 ; GFX1064-NEXT: s_add_u32 s36, s36, s11
428 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
429 ; GFX1064-NEXT: s_addc_u32 s37, s37, 0
430 ; GFX1064-NEXT: s_mov_b32 s12, s8
431 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
432 ; GFX1064-NEXT: s_mov_b32 s13, s9
433 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
434 ; GFX1064-NEXT: s_getpc_b64 s[4:5]
435 ; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
436 ; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
437 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
438 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
439 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
440 ; GFX1064-NEXT: s_mov_b32 s14, s10
441 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
442 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
443 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
444 ; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
445 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
446 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
447 ; GFX1064-NEXT: s_mov_b32 s32, 0
448 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
449 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
450 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0x7fc00000
451 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
452 ; GFX1064-NEXT: .LBB1_1: ; %ComputeLoop
453 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
454 ; GFX1064-NEXT: s_ff1_i32_b64 s2, s[0:1]
455 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v1
456 ; GFX1064-NEXT: v_readlane_b32 s3, v0, s2
457 ; GFX1064-NEXT: v_max_f32_e64 v2, s3, s3
458 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s2
459 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
460 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
461 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v2
462 ; GFX1064-NEXT: s_cbranch_scc1 .LBB1_1
463 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
464 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
465 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
466 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
467 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
468 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
469 ; GFX1064-NEXT: s_cbranch_execz .LBB1_4
470 ; GFX1064-NEXT: ; %bb.3:
471 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
472 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
473 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
474 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
475 ; GFX1064-NEXT: .LBB1_4:
476 ; GFX1064-NEXT: s_endpgm
478 ; GFX1032-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
480 ; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
481 ; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
482 ; GFX1032-NEXT: s_mov_b32 s38, -1
483 ; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
484 ; GFX1032-NEXT: s_add_u32 s36, s36, s11
485 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
486 ; GFX1032-NEXT: s_addc_u32 s37, s37, 0
487 ; GFX1032-NEXT: s_mov_b32 s12, s8
488 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
489 ; GFX1032-NEXT: s_mov_b32 s13, s9
490 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
491 ; GFX1032-NEXT: s_getpc_b64 s[4:5]
492 ; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
493 ; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
494 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
495 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
496 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
497 ; GFX1032-NEXT: s_mov_b32 s14, s10
498 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
499 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
500 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
501 ; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
502 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
503 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
504 ; GFX1032-NEXT: s_mov_b32 s32, 0
505 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
506 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
507 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0x7fc00000
508 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
509 ; GFX1032-NEXT: .LBB1_1: ; %ComputeLoop
510 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
511 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
512 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v1
513 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
514 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
515 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
516 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
517 ; GFX1032-NEXT: v_max_f32_e64 v2, s2, s2
518 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v2
519 ; GFX1032-NEXT: s_cbranch_scc1 .LBB1_1
520 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
521 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
522 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
523 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
524 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
525 ; GFX1032-NEXT: s_cbranch_execz .LBB1_4
526 ; GFX1032-NEXT: ; %bb.3:
527 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
528 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
529 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
530 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
531 ; GFX1032-NEXT: .LBB1_4:
532 ; GFX1032-NEXT: s_endpgm
534 ; GFX1164-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
536 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
537 ; GFX1164-NEXT: s_mov_b32 s12, s8
538 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
539 ; GFX1164-NEXT: s_mov_b32 s13, s9
540 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
541 ; GFX1164-NEXT: s_getpc_b64 s[4:5]
542 ; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
543 ; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
544 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
545 ; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
546 ; GFX1164-NEXT: s_mov_b32 s14, s10
547 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
548 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
549 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
550 ; GFX1164-NEXT: s_mov_b32 s32, 0
551 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
552 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
553 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0x7fc00000
554 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
555 ; GFX1164-NEXT: .LBB1_1: ; %ComputeLoop
556 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
557 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
558 ; GFX1164-NEXT: s_ctz_i32_b64 s2, s[0:1]
559 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v1
560 ; GFX1164-NEXT: v_readlane_b32 s3, v0, s2
561 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
562 ; GFX1164-NEXT: v_max_f32_e64 v2, s3, s3
563 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s2
564 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
565 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
566 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
567 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v2
568 ; GFX1164-NEXT: s_cbranch_scc1 .LBB1_1
569 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
570 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
571 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
572 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
573 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
574 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
575 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
576 ; GFX1164-NEXT: s_cbranch_execz .LBB1_4
577 ; GFX1164-NEXT: ; %bb.3:
578 ; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
579 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
580 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
581 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
582 ; GFX1164-NEXT: .LBB1_4:
583 ; GFX1164-NEXT: s_endpgm
585 ; GFX1132-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
587 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
588 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
589 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
590 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
591 ; GFX1132-NEXT: s_getpc_b64 s[4:5]
592 ; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
593 ; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
594 ; GFX1132-NEXT: s_mov_b32 s12, s13
595 ; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
596 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
597 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
598 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
599 ; GFX1132-NEXT: s_mov_b32 s13, s14
600 ; GFX1132-NEXT: s_mov_b32 s14, s15
601 ; GFX1132-NEXT: s_mov_b32 s32, 0
602 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
603 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
604 ; GFX1132-NEXT: v_mov_b32_e32 v1, 0x7fc00000
605 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
606 ; GFX1132-NEXT: .LBB1_1: ; %ComputeLoop
607 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
608 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
609 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
610 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v1
611 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
612 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
613 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
614 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
615 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
616 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
617 ; GFX1132-NEXT: v_max_f32_e64 v2, s2, s2
618 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v2
619 ; GFX1132-NEXT: s_cbranch_scc1 .LBB1_1
620 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
621 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
622 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
623 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
624 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
625 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
626 ; GFX1132-NEXT: s_cbranch_execz .LBB1_4
627 ; GFX1132-NEXT: ; %bb.3:
628 ; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
629 ; GFX1132-NEXT: v_mov_b32_e32 v0, 0
630 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
631 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
632 ; GFX1132-NEXT: .LBB1_4:
633 ; GFX1132-NEXT: s_endpgm
635 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
636 ; GFX7LESS-DPP: ; %bb.0:
637 ; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
638 ; GFX7LESS-DPP-NEXT: s_mov_b32 s40, SCRATCH_RSRC_DWORD0
639 ; GFX7LESS-DPP-NEXT: s_mov_b32 s41, SCRATCH_RSRC_DWORD1
640 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, -1
641 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, 0xe8f000
642 ; GFX7LESS-DPP-NEXT: s_add_u32 s40, s40, s11
643 ; GFX7LESS-DPP-NEXT: s_addc_u32 s41, s41, 0
644 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
645 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
646 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
647 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
648 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
649 ; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
650 ; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
651 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
652 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
653 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
654 ; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
655 ; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
656 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
657 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
658 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
659 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
660 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
661 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
662 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
663 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[40:41]
664 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[42:43]
665 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
666 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
667 ; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[36:39], 0
668 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
669 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v2, 1.0, v0
670 ; GFX7LESS-DPP-NEXT: .LBB1_1: ; %atomicrmw.start
671 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
672 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
673 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
674 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, v0, v2
675 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
676 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v1
677 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v0
678 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[3:4], off, s[36:39], 0 glc
679 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
680 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
681 ; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
682 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v3
683 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
684 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB1_1
685 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
686 ; GFX7LESS-DPP-NEXT: s_endpgm
688 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
690 ; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
691 ; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
692 ; GFX9-DPP-NEXT: s_mov_b32 s38, -1
693 ; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
694 ; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
695 ; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
696 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
697 ; GFX9-DPP-NEXT: s_mov_b32 s12, s8
698 ; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
699 ; GFX9-DPP-NEXT: s_mov_b32 s13, s9
700 ; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
701 ; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
702 ; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
703 ; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
704 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
705 ; GFX9-DPP-NEXT: s_mov_b32 s14, s10
706 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
707 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
708 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
709 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
710 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
711 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
712 ; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
713 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
714 ; GFX9-DPP-NEXT: s_mov_b32 s32, 0
715 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
716 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
717 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
718 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
719 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
720 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
721 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v4, v3, v0, s[0:1]
722 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
723 ; GFX9-DPP-NEXT: s_nop 1
724 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:1 row_mask:0xf bank_mask:0xf
725 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v4
726 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
727 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
728 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
729 ; GFX9-DPP-NEXT: s_nop 1
730 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:2 row_mask:0xf bank_mask:0xf
731 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
732 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
733 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
734 ; GFX9-DPP-NEXT: s_nop 1
735 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:4 row_mask:0xf bank_mask:0xf
736 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
737 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
738 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
739 ; GFX9-DPP-NEXT: s_nop 1
740 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:8 row_mask:0xf bank_mask:0xf
741 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
742 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
743 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
744 ; GFX9-DPP-NEXT: s_nop 1
745 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_bcast:15 row_mask:0xa bank_mask:0xf
746 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
747 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
748 ; GFX9-DPP-NEXT: s_nop 1
749 ; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v4 row_bcast:31 row_mask:0xc bank_mask:0xf
750 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v3, v3
751 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v4, v3
752 ; GFX9-DPP-NEXT: v_readlane_b32 s4, v3, 63
753 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
754 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
755 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
756 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB1_3
757 ; GFX9-DPP-NEXT: ; %bb.1:
758 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
759 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
760 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
761 ; GFX9-DPP-NEXT: v_max_f32_e64 v6, s4, s4
762 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
763 ; GFX9-DPP-NEXT: global_load_dword v1, v2, s[0:1]
764 ; GFX9-DPP-NEXT: .LBB1_2: ; %atomicrmw.start
765 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
766 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
767 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
768 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v0, v6
769 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
770 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
771 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
772 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
773 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
774 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
775 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB1_2
776 ; GFX9-DPP-NEXT: .LBB1_3:
777 ; GFX9-DPP-NEXT: s_endpgm
779 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
780 ; GFX1064-DPP: ; %bb.0:
781 ; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
782 ; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
783 ; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
784 ; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
785 ; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
786 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
787 ; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
788 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
789 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
790 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
791 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
792 ; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
793 ; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
794 ; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
795 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
796 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
797 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
798 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
799 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
800 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
801 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
802 ; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
803 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
804 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
805 ; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
806 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
807 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
808 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
809 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
810 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s[0:1]
811 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
812 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
813 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
814 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v3
815 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
816 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
817 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
818 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
819 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
820 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
821 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
822 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
823 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
824 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
825 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
826 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
827 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
828 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
829 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
830 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 32
831 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v3, 0
832 ; GFX1064-DPP-NEXT: v_max_f32_e64 v3, s2, s2
833 ; GFX1064-DPP-NEXT: v_max_f32_e64 v4, s3, s3
834 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
835 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
836 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
837 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
838 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
839 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
840 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
841 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
842 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
843 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB1_2
844 ; GFX1064-DPP-NEXT: ; %bb.1:
845 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
846 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
847 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
848 ; GFX1064-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
849 ; GFX1064-DPP-NEXT: .LBB1_2:
850 ; GFX1064-DPP-NEXT: s_endpgm
852 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
853 ; GFX1032-DPP: ; %bb.0:
854 ; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
855 ; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
856 ; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
857 ; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
858 ; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
859 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
860 ; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
861 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
862 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
863 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
864 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
865 ; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
866 ; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
867 ; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
868 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
869 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
870 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
871 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
872 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
873 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
874 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
875 ; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
876 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
877 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
878 ; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
879 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
880 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
881 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
882 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
883 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s0
884 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
885 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
886 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
887 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v3
888 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v4, v3
889 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
890 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
891 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
892 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
893 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
894 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
895 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
896 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
897 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
898 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
899 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
900 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
901 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
902 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
903 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
904 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
905 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
906 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
907 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
908 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB1_2
909 ; GFX1032-DPP-NEXT: ; %bb.1:
910 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
911 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
912 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
913 ; GFX1032-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
914 ; GFX1032-DPP-NEXT: .LBB1_2:
915 ; GFX1032-DPP-NEXT: s_endpgm
917 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
918 ; GFX1164-DPP: ; %bb.0:
919 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
920 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
921 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
922 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
923 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
924 ; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
925 ; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
926 ; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
927 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
928 ; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
929 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
930 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
931 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
932 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
933 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
934 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
935 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
936 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
937 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
938 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s[0:1]
939 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
940 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
941 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
942 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
943 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v1
944 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
945 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v2, v1
946 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
947 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
948 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
949 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
950 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
951 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
952 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
953 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
954 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
955 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
956 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
957 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
958 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
959 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
960 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
961 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
962 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
963 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
964 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
965 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
966 ; GFX1164-DPP-NEXT: v_permlane64_b32 v2, v1
967 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
968 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
969 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
970 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
971 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
972 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
973 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
974 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
975 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
976 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v4, exec_hi, v0
977 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
978 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v1
979 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
980 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
981 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB1_2
982 ; GFX1164-DPP-NEXT: ; %bb.1:
983 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
984 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
985 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
986 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
987 ; GFX1164-DPP-NEXT: .LBB1_2:
988 ; GFX1164-DPP-NEXT: s_endpgm
990 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_div_value_agent_scope_unsafe:
991 ; GFX1132-DPP: ; %bb.0:
992 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
993 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
994 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
995 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
996 ; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
997 ; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
998 ; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
999 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
1000 ; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
1001 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1002 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1003 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1004 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
1005 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
1006 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
1007 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
1008 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1009 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
1010 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
1011 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s0
1012 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1013 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1014 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
1015 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1016 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v2, v2 :: v_dual_max_f32 v1, v1, v1
1017 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v2, v1
1018 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1019 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
1020 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
1021 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1022 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1023 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
1024 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1025 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
1026 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1027 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1028 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
1029 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v3, v3
1030 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1031 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1032 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
1033 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1034 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v2, v2
1035 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1036 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
1037 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
1038 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v4, exec_lo, 0
1039 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v1
1040 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
1041 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
1042 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
1043 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB1_2
1044 ; GFX1132-DPP-NEXT: ; %bb.1:
1045 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
1046 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
1047 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
1048 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
1049 ; GFX1132-DPP-NEXT: .LBB1_2:
1050 ; GFX1132-DPP-NEXT: s_endpgm
1051 %divValue = call float @div.float.value()
1052 %result = atomicrmw fmax ptr addrspace(1) %ptr, float %divValue syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
1056 define amdgpu_kernel void @global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe(ptr addrspace(1) %ptr) #0 {
1057 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1058 ; GFX7LESS: ; %bb.0:
1059 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
1060 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
1061 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1062 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
1063 ; GFX7LESS-NEXT: s_cbranch_execz .LBB2_3
1064 ; GFX7LESS-NEXT: ; %bb.1:
1065 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
1066 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
1067 ; GFX7LESS-NEXT: s_load_dword s2, s[0:1], 0x0
1068 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
1069 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
1070 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
1071 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s2
1072 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
1073 ; GFX7LESS-NEXT: .LBB2_2: ; %atomicrmw.start
1074 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
1075 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
1076 ; GFX7LESS-NEXT: v_max_f32_e32 v0, 4.0, v0
1077 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
1078 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v1
1079 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v0
1080 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
1081 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
1082 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
1083 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1084 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v2
1085 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
1086 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB2_2
1087 ; GFX7LESS-NEXT: .LBB2_3:
1088 ; GFX7LESS-NEXT: s_endpgm
1090 ; GFX9-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1092 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1093 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1094 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1095 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
1096 ; GFX9-NEXT: s_cbranch_execz .LBB2_3
1097 ; GFX9-NEXT: ; %bb.1:
1098 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1099 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
1100 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
1101 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
1102 ; GFX9-NEXT: s_load_dword s4, s[0:1], 0x0
1103 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
1104 ; GFX9-NEXT: v_mov_b32_e32 v1, s4
1105 ; GFX9-NEXT: .LBB2_2: ; %atomicrmw.start
1106 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
1107 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
1108 ; GFX9-NEXT: v_max_f32_e32 v0, 4.0, v0
1109 ; GFX9-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
1110 ; GFX9-NEXT: s_waitcnt vmcnt(0)
1111 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
1112 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
1113 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
1114 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
1115 ; GFX9-NEXT: s_cbranch_execnz .LBB2_2
1116 ; GFX9-NEXT: .LBB2_3:
1117 ; GFX9-NEXT: s_endpgm
1119 ; GFX1064-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1121 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1122 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1123 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1124 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
1125 ; GFX1064-NEXT: s_cbranch_execz .LBB2_2
1126 ; GFX1064-NEXT: ; %bb.1:
1127 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1128 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
1129 ; GFX1064-NEXT: v_mov_b32_e32 v1, 4.0
1130 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
1131 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
1132 ; GFX1064-NEXT: .LBB2_2:
1133 ; GFX1064-NEXT: s_endpgm
1135 ; GFX1032-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1137 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1138 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
1139 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
1140 ; GFX1032-NEXT: s_cbranch_execz .LBB2_2
1141 ; GFX1032-NEXT: ; %bb.1:
1142 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1143 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
1144 ; GFX1032-NEXT: v_mov_b32_e32 v1, 4.0
1145 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
1146 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
1147 ; GFX1032-NEXT: .LBB2_2:
1148 ; GFX1032-NEXT: s_endpgm
1150 ; GFX1164-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1152 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1153 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
1154 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1155 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1156 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
1157 ; GFX1164-NEXT: s_cbranch_execz .LBB2_2
1158 ; GFX1164-NEXT: ; %bb.1:
1159 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
1160 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
1161 ; GFX1164-NEXT: v_mov_b32_e32 v1, 4.0
1162 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
1163 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1164 ; GFX1164-NEXT: .LBB2_2:
1165 ; GFX1164-NEXT: s_endpgm
1167 ; GFX1132-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1169 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1170 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
1171 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
1172 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
1173 ; GFX1132-NEXT: s_cbranch_execz .LBB2_2
1174 ; GFX1132-NEXT: ; %bb.1:
1175 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
1176 ; GFX1132-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
1177 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
1178 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1179 ; GFX1132-NEXT: .LBB2_2:
1180 ; GFX1132-NEXT: s_endpgm
1182 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1183 ; GFX7LESS-DPP: ; %bb.0:
1184 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
1185 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
1186 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1187 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
1188 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB2_3
1189 ; GFX7LESS-DPP-NEXT: ; %bb.1:
1190 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
1191 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
1192 ; GFX7LESS-DPP-NEXT: s_load_dword s2, s[0:1], 0x0
1193 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
1194 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
1195 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
1196 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s2
1197 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
1198 ; GFX7LESS-DPP-NEXT: .LBB2_2: ; %atomicrmw.start
1199 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
1200 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
1201 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
1202 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
1203 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v1
1204 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v0
1205 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
1206 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
1207 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
1208 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1209 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v2
1210 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
1211 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB2_2
1212 ; GFX7LESS-DPP-NEXT: .LBB2_3:
1213 ; GFX7LESS-DPP-NEXT: s_endpgm
1215 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1216 ; GFX9-DPP: ; %bb.0:
1217 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1218 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1219 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1220 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
1221 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB2_3
1222 ; GFX9-DPP-NEXT: ; %bb.1:
1223 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1224 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
1225 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
1226 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
1227 ; GFX9-DPP-NEXT: s_load_dword s4, s[0:1], 0x0
1228 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
1229 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s4
1230 ; GFX9-DPP-NEXT: .LBB2_2: ; %atomicrmw.start
1231 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
1232 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
1233 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
1234 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
1235 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
1236 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
1237 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
1238 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
1239 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
1240 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB2_2
1241 ; GFX9-DPP-NEXT: .LBB2_3:
1242 ; GFX9-DPP-NEXT: s_endpgm
1244 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1245 ; GFX1064-DPP: ; %bb.0:
1246 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1247 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1248 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1249 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
1250 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB2_2
1251 ; GFX1064-DPP-NEXT: ; %bb.1:
1252 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1253 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
1254 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 4.0
1255 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
1256 ; GFX1064-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
1257 ; GFX1064-DPP-NEXT: .LBB2_2:
1258 ; GFX1064-DPP-NEXT: s_endpgm
1260 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1261 ; GFX1032-DPP: ; %bb.0:
1262 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1263 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
1264 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
1265 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB2_2
1266 ; GFX1032-DPP-NEXT: ; %bb.1:
1267 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
1268 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
1269 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 4.0
1270 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
1271 ; GFX1032-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
1272 ; GFX1032-DPP-NEXT: .LBB2_2:
1273 ; GFX1032-DPP-NEXT: s_endpgm
1275 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1276 ; GFX1164-DPP: ; %bb.0:
1277 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1278 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
1279 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1280 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1281 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
1282 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB2_2
1283 ; GFX1164-DPP-NEXT: ; %bb.1:
1284 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
1285 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
1286 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 4.0
1287 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
1288 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1289 ; GFX1164-DPP-NEXT: .LBB2_2:
1290 ; GFX1164-DPP-NEXT: s_endpgm
1292 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_one_as_scope_unsafe:
1293 ; GFX1132-DPP: ; %bb.0:
1294 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1295 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
1296 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
1297 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
1298 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB2_2
1299 ; GFX1132-DPP-NEXT: ; %bb.1:
1300 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
1301 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
1302 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
1303 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1304 ; GFX1132-DPP-NEXT: .LBB2_2:
1305 ; GFX1132-DPP-NEXT: s_endpgm
1306 %result = atomicrmw fmax ptr addrspace(1) %ptr, float 4.0 syncscope("one-as") monotonic, !amdgpu.no.fine.grained.memory !1
1311 define amdgpu_kernel void @global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe(ptr addrspace(1) %ptr) #0 {
1312 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1313 ; GFX7LESS: ; %bb.0:
1314 ; GFX7LESS-NEXT: s_mov_b32 s32, 0
1315 ; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1316 ; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1317 ; GFX7LESS-NEXT: s_mov_b32 s38, -1
1318 ; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
1319 ; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
1320 ; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
1321 ; GFX7LESS-NEXT: s_mov_b32 s14, s10
1322 ; GFX7LESS-NEXT: s_mov_b32 s13, s9
1323 ; GFX7LESS-NEXT: s_mov_b32 s12, s8
1324 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
1325 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
1326 ; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
1327 ; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
1328 ; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
1329 ; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1330 ; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1331 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1332 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1333 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1334 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
1335 ; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
1336 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
1337 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
1338 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
1339 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
1340 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
1341 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
1342 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
1343 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0x7fc00000
1344 ; GFX7LESS-NEXT: .LBB3_1: ; %ComputeLoop
1345 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
1346 ; GFX7LESS-NEXT: s_ff1_i32_b64 s2, s[0:1]
1347 ; GFX7LESS-NEXT: v_mul_f32_e32 v1, 1.0, v2
1348 ; GFX7LESS-NEXT: v_readlane_b32 s4, v0, s2
1349 ; GFX7LESS-NEXT: s_lshl_b64 s[2:3], 1, s2
1350 ; GFX7LESS-NEXT: v_mul_f32_e64 v2, 1.0, s4
1351 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
1352 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
1353 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
1354 ; GFX7LESS-NEXT: v_max_f32_e32 v2, v1, v2
1355 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB3_1
1356 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
1357 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
1358 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
1359 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1360 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
1361 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
1362 ; GFX7LESS-NEXT: s_cbranch_execz .LBB3_5
1363 ; GFX7LESS-NEXT: ; %bb.3:
1364 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
1365 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
1366 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
1367 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
1368 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[0:3], 0
1369 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
1370 ; GFX7LESS-NEXT: v_mul_f32_e32 v2, 1.0, v2
1371 ; GFX7LESS-NEXT: .LBB3_4: ; %atomicrmw.start
1372 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
1373 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
1374 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
1375 ; GFX7LESS-NEXT: v_max_f32_e32 v0, v0, v2
1376 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
1377 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, v1
1378 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v0
1379 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[3:4], off, s[0:3], 0 glc
1380 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
1381 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
1382 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
1383 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v3
1384 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
1385 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB3_4
1386 ; GFX7LESS-NEXT: .LBB3_5:
1387 ; GFX7LESS-NEXT: s_endpgm
1389 ; GFX9-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1391 ; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1392 ; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1393 ; GFX9-NEXT: s_mov_b32 s38, -1
1394 ; GFX9-NEXT: s_mov_b32 s39, 0xe00000
1395 ; GFX9-NEXT: s_add_u32 s36, s36, s11
1396 ; GFX9-NEXT: s_addc_u32 s37, s37, 0
1397 ; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
1398 ; GFX9-NEXT: s_mov_b32 s12, s8
1399 ; GFX9-NEXT: s_add_u32 s8, s34, 44
1400 ; GFX9-NEXT: s_mov_b32 s13, s9
1401 ; GFX9-NEXT: s_addc_u32 s9, s35, 0
1402 ; GFX9-NEXT: s_getpc_b64 s[4:5]
1403 ; GFX9-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1404 ; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1405 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1406 ; GFX9-NEXT: s_mov_b32 s14, s10
1407 ; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
1408 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1409 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1410 ; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
1411 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
1412 ; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
1413 ; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
1414 ; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
1415 ; GFX9-NEXT: s_mov_b32 s32, 0
1416 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
1417 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
1418 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
1419 ; GFX9-NEXT: v_mov_b32_e32 v2, 0x7fc00000
1420 ; GFX9-NEXT: .LBB3_1: ; %ComputeLoop
1421 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
1422 ; GFX9-NEXT: s_ff1_i32_b64 s2, s[0:1]
1423 ; GFX9-NEXT: v_readlane_b32 s4, v0, s2
1424 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s2
1425 ; GFX9-NEXT: v_max_f32_e32 v1, v2, v2
1426 ; GFX9-NEXT: v_max_f32_e64 v2, s4, s4
1427 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
1428 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
1429 ; GFX9-NEXT: v_max_f32_e32 v2, v1, v2
1430 ; GFX9-NEXT: s_cbranch_scc1 .LBB3_1
1431 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
1432 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1433 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1434 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1435 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
1436 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
1437 ; GFX9-NEXT: s_cbranch_execz .LBB3_5
1438 ; GFX9-NEXT: ; %bb.3:
1439 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1440 ; GFX9-NEXT: v_mov_b32_e32 v3, 0
1441 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
1442 ; GFX9-NEXT: v_max_f32_e32 v2, v2, v2
1443 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
1444 ; GFX9-NEXT: global_load_dword v1, v3, s[0:1]
1445 ; GFX9-NEXT: .LBB3_4: ; %atomicrmw.start
1446 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
1447 ; GFX9-NEXT: s_waitcnt vmcnt(0)
1448 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
1449 ; GFX9-NEXT: v_max_f32_e32 v0, v0, v2
1450 ; GFX9-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
1451 ; GFX9-NEXT: s_waitcnt vmcnt(0)
1452 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
1453 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
1454 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
1455 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
1456 ; GFX9-NEXT: s_cbranch_execnz .LBB3_4
1457 ; GFX9-NEXT: .LBB3_5:
1458 ; GFX9-NEXT: s_endpgm
1460 ; GFX1064-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1462 ; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1463 ; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1464 ; GFX1064-NEXT: s_mov_b32 s38, -1
1465 ; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
1466 ; GFX1064-NEXT: s_add_u32 s36, s36, s11
1467 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
1468 ; GFX1064-NEXT: s_addc_u32 s37, s37, 0
1469 ; GFX1064-NEXT: s_mov_b32 s12, s8
1470 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
1471 ; GFX1064-NEXT: s_mov_b32 s13, s9
1472 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
1473 ; GFX1064-NEXT: s_getpc_b64 s[4:5]
1474 ; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1475 ; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1476 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1477 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1478 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1479 ; GFX1064-NEXT: s_mov_b32 s14, s10
1480 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
1481 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
1482 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
1483 ; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
1484 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
1485 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
1486 ; GFX1064-NEXT: s_mov_b32 s32, 0
1487 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
1488 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
1489 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1490 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
1491 ; GFX1064-NEXT: .LBB3_1: ; %ComputeLoop
1492 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
1493 ; GFX1064-NEXT: s_ff1_i32_b64 s2, s[0:1]
1494 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v1
1495 ; GFX1064-NEXT: v_readlane_b32 s3, v0, s2
1496 ; GFX1064-NEXT: v_max_f32_e64 v2, s3, s3
1497 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s2
1498 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
1499 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
1500 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v2
1501 ; GFX1064-NEXT: s_cbranch_scc1 .LBB3_1
1502 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
1503 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1504 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1505 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1506 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
1507 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
1508 ; GFX1064-NEXT: s_cbranch_execz .LBB3_4
1509 ; GFX1064-NEXT: ; %bb.3:
1510 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1511 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
1512 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
1513 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
1514 ; GFX1064-NEXT: .LBB3_4:
1515 ; GFX1064-NEXT: s_endpgm
1517 ; GFX1032-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1519 ; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1520 ; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1521 ; GFX1032-NEXT: s_mov_b32 s38, -1
1522 ; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
1523 ; GFX1032-NEXT: s_add_u32 s36, s36, s11
1524 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
1525 ; GFX1032-NEXT: s_addc_u32 s37, s37, 0
1526 ; GFX1032-NEXT: s_mov_b32 s12, s8
1527 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
1528 ; GFX1032-NEXT: s_mov_b32 s13, s9
1529 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
1530 ; GFX1032-NEXT: s_getpc_b64 s[4:5]
1531 ; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1532 ; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1533 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1534 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1535 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1536 ; GFX1032-NEXT: s_mov_b32 s14, s10
1537 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
1538 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
1539 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
1540 ; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
1541 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
1542 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
1543 ; GFX1032-NEXT: s_mov_b32 s32, 0
1544 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
1545 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
1546 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1547 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
1548 ; GFX1032-NEXT: .LBB3_1: ; %ComputeLoop
1549 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
1550 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
1551 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v1
1552 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
1553 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
1554 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
1555 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
1556 ; GFX1032-NEXT: v_max_f32_e64 v2, s2, s2
1557 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v2
1558 ; GFX1032-NEXT: s_cbranch_scc1 .LBB3_1
1559 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
1560 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1561 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
1562 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
1563 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
1564 ; GFX1032-NEXT: s_cbranch_execz .LBB3_4
1565 ; GFX1032-NEXT: ; %bb.3:
1566 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1567 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
1568 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
1569 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
1570 ; GFX1032-NEXT: .LBB3_4:
1571 ; GFX1032-NEXT: s_endpgm
1573 ; GFX1164-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1575 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
1576 ; GFX1164-NEXT: s_mov_b32 s12, s8
1577 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
1578 ; GFX1164-NEXT: s_mov_b32 s13, s9
1579 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
1580 ; GFX1164-NEXT: s_getpc_b64 s[4:5]
1581 ; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1582 ; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1583 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
1584 ; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
1585 ; GFX1164-NEXT: s_mov_b32 s14, s10
1586 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
1587 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
1588 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
1589 ; GFX1164-NEXT: s_mov_b32 s32, 0
1590 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
1591 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
1592 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1593 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
1594 ; GFX1164-NEXT: .LBB3_1: ; %ComputeLoop
1595 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
1596 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1597 ; GFX1164-NEXT: s_ctz_i32_b64 s2, s[0:1]
1598 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v1
1599 ; GFX1164-NEXT: v_readlane_b32 s3, v0, s2
1600 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
1601 ; GFX1164-NEXT: v_max_f32_e64 v2, s3, s3
1602 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s2
1603 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
1604 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1605 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
1606 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v2
1607 ; GFX1164-NEXT: s_cbranch_scc1 .LBB3_1
1608 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
1609 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1610 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
1611 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1612 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
1613 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
1614 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
1615 ; GFX1164-NEXT: s_cbranch_execz .LBB3_4
1616 ; GFX1164-NEXT: ; %bb.3:
1617 ; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
1618 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
1619 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
1620 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1621 ; GFX1164-NEXT: .LBB3_4:
1622 ; GFX1164-NEXT: s_endpgm
1624 ; GFX1132-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1626 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
1627 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
1628 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
1629 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
1630 ; GFX1132-NEXT: s_getpc_b64 s[4:5]
1631 ; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1632 ; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1633 ; GFX1132-NEXT: s_mov_b32 s12, s13
1634 ; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
1635 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
1636 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
1637 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
1638 ; GFX1132-NEXT: s_mov_b32 s13, s14
1639 ; GFX1132-NEXT: s_mov_b32 s14, s15
1640 ; GFX1132-NEXT: s_mov_b32 s32, 0
1641 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
1642 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
1643 ; GFX1132-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1644 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
1645 ; GFX1132-NEXT: .LBB3_1: ; %ComputeLoop
1646 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
1647 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1648 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
1649 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v1
1650 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
1651 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
1652 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
1653 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
1654 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
1655 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1656 ; GFX1132-NEXT: v_max_f32_e64 v2, s2, s2
1657 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v2
1658 ; GFX1132-NEXT: s_cbranch_scc1 .LBB3_1
1659 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
1660 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1661 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
1662 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
1663 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
1664 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
1665 ; GFX1132-NEXT: s_cbranch_execz .LBB3_4
1666 ; GFX1132-NEXT: ; %bb.3:
1667 ; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
1668 ; GFX1132-NEXT: v_mov_b32_e32 v0, 0
1669 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
1670 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
1671 ; GFX1132-NEXT: .LBB3_4:
1672 ; GFX1132-NEXT: s_endpgm
1674 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1675 ; GFX7LESS-DPP: ; %bb.0:
1676 ; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
1677 ; GFX7LESS-DPP-NEXT: s_mov_b32 s40, SCRATCH_RSRC_DWORD0
1678 ; GFX7LESS-DPP-NEXT: s_mov_b32 s41, SCRATCH_RSRC_DWORD1
1679 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, -1
1680 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, 0xe8f000
1681 ; GFX7LESS-DPP-NEXT: s_add_u32 s40, s40, s11
1682 ; GFX7LESS-DPP-NEXT: s_addc_u32 s41, s41, 0
1683 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
1684 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
1685 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
1686 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1687 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
1688 ; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
1689 ; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
1690 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
1691 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
1692 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
1693 ; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1694 ; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1695 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1696 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1697 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1698 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
1699 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
1700 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1701 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1702 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[40:41]
1703 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[42:43]
1704 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
1705 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1706 ; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[36:39], 0
1707 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
1708 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v2, 1.0, v0
1709 ; GFX7LESS-DPP-NEXT: .LBB3_1: ; %atomicrmw.start
1710 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
1711 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
1712 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
1713 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, v0, v2
1714 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
1715 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v1
1716 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v0
1717 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[3:4], off, s[36:39], 0 glc
1718 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
1719 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
1720 ; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
1721 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v3
1722 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
1723 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB3_1
1724 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
1725 ; GFX7LESS-DPP-NEXT: s_endpgm
1727 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1728 ; GFX9-DPP: ; %bb.0:
1729 ; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1730 ; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1731 ; GFX9-DPP-NEXT: s_mov_b32 s38, -1
1732 ; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
1733 ; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
1734 ; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
1735 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
1736 ; GFX9-DPP-NEXT: s_mov_b32 s12, s8
1737 ; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
1738 ; GFX9-DPP-NEXT: s_mov_b32 s13, s9
1739 ; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
1740 ; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
1741 ; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1742 ; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1743 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1744 ; GFX9-DPP-NEXT: s_mov_b32 s14, s10
1745 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1746 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1747 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1748 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1749 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1750 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
1751 ; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
1752 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
1753 ; GFX9-DPP-NEXT: s_mov_b32 s32, 0
1754 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
1755 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1756 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
1757 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
1758 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
1759 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1760 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v4, v3, v0, s[0:1]
1761 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1762 ; GFX9-DPP-NEXT: s_nop 1
1763 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:1 row_mask:0xf bank_mask:0xf
1764 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v4
1765 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
1766 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
1767 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1768 ; GFX9-DPP-NEXT: s_nop 1
1769 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:2 row_mask:0xf bank_mask:0xf
1770 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
1771 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
1772 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1773 ; GFX9-DPP-NEXT: s_nop 1
1774 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:4 row_mask:0xf bank_mask:0xf
1775 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
1776 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
1777 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1778 ; GFX9-DPP-NEXT: s_nop 1
1779 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:8 row_mask:0xf bank_mask:0xf
1780 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
1781 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
1782 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1783 ; GFX9-DPP-NEXT: s_nop 1
1784 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_bcast:15 row_mask:0xa bank_mask:0xf
1785 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
1786 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
1787 ; GFX9-DPP-NEXT: s_nop 1
1788 ; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v4 row_bcast:31 row_mask:0xc bank_mask:0xf
1789 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v3, v3
1790 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v4, v3
1791 ; GFX9-DPP-NEXT: v_readlane_b32 s4, v3, 63
1792 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
1793 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
1794 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
1795 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB3_3
1796 ; GFX9-DPP-NEXT: ; %bb.1:
1797 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1798 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
1799 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
1800 ; GFX9-DPP-NEXT: v_max_f32_e64 v6, s4, s4
1801 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
1802 ; GFX9-DPP-NEXT: global_load_dword v1, v2, s[0:1]
1803 ; GFX9-DPP-NEXT: .LBB3_2: ; %atomicrmw.start
1804 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
1805 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
1806 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
1807 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v0, v6
1808 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
1809 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
1810 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
1811 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
1812 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
1813 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
1814 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB3_2
1815 ; GFX9-DPP-NEXT: .LBB3_3:
1816 ; GFX9-DPP-NEXT: s_endpgm
1818 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1819 ; GFX1064-DPP: ; %bb.0:
1820 ; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1821 ; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1822 ; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
1823 ; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
1824 ; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
1825 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
1826 ; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
1827 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
1828 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
1829 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
1830 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
1831 ; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
1832 ; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1833 ; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1834 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1835 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1836 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1837 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
1838 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1839 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1840 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1841 ; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
1842 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
1843 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
1844 ; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
1845 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
1846 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1847 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
1848 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1849 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s[0:1]
1850 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1851 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
1852 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
1853 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v3
1854 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
1855 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
1856 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1857 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1858 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1859 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
1860 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1861 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1862 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1863 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
1864 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1865 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1866 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
1867 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
1868 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1869 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 32
1870 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v3, 0
1871 ; GFX1064-DPP-NEXT: v_max_f32_e64 v3, s2, s2
1872 ; GFX1064-DPP-NEXT: v_max_f32_e64 v4, s3, s3
1873 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
1874 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
1875 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
1876 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
1877 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
1878 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
1879 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
1880 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
1881 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
1882 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB3_2
1883 ; GFX1064-DPP-NEXT: ; %bb.1:
1884 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1885 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
1886 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
1887 ; GFX1064-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
1888 ; GFX1064-DPP-NEXT: .LBB3_2:
1889 ; GFX1064-DPP-NEXT: s_endpgm
1891 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1892 ; GFX1032-DPP: ; %bb.0:
1893 ; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
1894 ; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
1895 ; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
1896 ; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
1897 ; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
1898 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
1899 ; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
1900 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
1901 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
1902 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
1903 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
1904 ; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
1905 ; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1906 ; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1907 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
1908 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
1909 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
1910 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
1911 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1912 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1913 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1914 ; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
1915 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
1916 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
1917 ; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
1918 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
1919 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1920 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
1921 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1922 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s0
1923 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1924 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
1925 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
1926 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v3
1927 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v4, v3
1928 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
1929 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1930 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1931 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1932 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
1933 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1934 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
1935 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1936 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
1937 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
1938 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1939 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
1940 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
1941 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
1942 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
1943 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
1944 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
1945 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
1946 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
1947 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB3_2
1948 ; GFX1032-DPP-NEXT: ; %bb.1:
1949 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
1950 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
1951 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
1952 ; GFX1032-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
1953 ; GFX1032-DPP-NEXT: .LBB3_2:
1954 ; GFX1032-DPP-NEXT: s_endpgm
1956 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
1957 ; GFX1164-DPP: ; %bb.0:
1958 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
1959 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
1960 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
1961 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
1962 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
1963 ; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
1964 ; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
1965 ; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
1966 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
1967 ; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
1968 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
1969 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
1970 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
1971 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
1972 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
1973 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
1974 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
1975 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
1976 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
1977 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s[0:1]
1978 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1979 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
1980 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
1981 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
1982 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v1
1983 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1984 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v2, v1
1985 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
1986 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
1987 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
1988 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1989 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1990 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1991 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
1992 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
1993 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
1994 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
1995 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
1996 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
1997 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
1998 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
1999 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2000 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2001 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
2002 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
2003 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2004 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2005 ; GFX1164-DPP-NEXT: v_permlane64_b32 v2, v1
2006 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
2007 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
2008 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
2009 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2010 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
2011 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
2012 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2013 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
2014 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
2015 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v4, exec_hi, v0
2016 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
2017 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v1
2018 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
2019 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
2020 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB3_2
2021 ; GFX1164-DPP-NEXT: ; %bb.1:
2022 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
2023 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
2024 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
2025 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
2026 ; GFX1164-DPP-NEXT: .LBB3_2:
2027 ; GFX1164-DPP-NEXT: s_endpgm
2029 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_div_value_one_as_scope_unsafe:
2030 ; GFX1132-DPP: ; %bb.0:
2031 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
2032 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
2033 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
2034 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
2035 ; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
2036 ; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2037 ; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2038 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
2039 ; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
2040 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
2041 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
2042 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
2043 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
2044 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
2045 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
2046 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
2047 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
2048 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
2049 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
2050 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s0
2051 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
2052 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
2053 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
2054 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2055 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v2, v2 :: v_dual_max_f32 v1, v1, v1
2056 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v2, v1
2057 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2058 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
2059 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
2060 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2061 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2062 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
2063 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2064 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
2065 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2066 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2067 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
2068 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v3, v3
2069 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2070 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2071 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
2072 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2073 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v2, v2
2074 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
2075 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
2076 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
2077 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v4, exec_lo, 0
2078 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v1
2079 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
2080 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
2081 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
2082 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB3_2
2083 ; GFX1132-DPP-NEXT: ; %bb.1:
2084 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
2085 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
2086 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
2087 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
2088 ; GFX1132-DPP-NEXT: .LBB3_2:
2089 ; GFX1132-DPP-NEXT: s_endpgm
2090 %divValue = call float @div.float.value()
2091 %result = atomicrmw fmax ptr addrspace(1) %ptr, float %divValue syncscope("one-as") monotonic, !amdgpu.no.fine.grained.memory !1
2096 define amdgpu_kernel void @global_atomic_fmax_uni_address_uni_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
2097 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2098 ; GFX7LESS: ; %bb.0:
2099 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
2100 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
2101 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2102 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
2103 ; GFX7LESS-NEXT: s_cbranch_execz .LBB4_3
2104 ; GFX7LESS-NEXT: ; %bb.1:
2105 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
2106 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
2107 ; GFX7LESS-NEXT: s_load_dword s2, s[0:1], 0x0
2108 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
2109 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
2110 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
2111 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s2
2112 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
2113 ; GFX7LESS-NEXT: .LBB4_2: ; %atomicrmw.start
2114 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
2115 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
2116 ; GFX7LESS-NEXT: v_max_f32_e32 v0, 4.0, v0
2117 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
2118 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v1
2119 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v0
2120 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
2121 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
2122 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
2123 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
2124 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v2
2125 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
2126 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB4_2
2127 ; GFX7LESS-NEXT: .LBB4_3:
2128 ; GFX7LESS-NEXT: s_endpgm
2130 ; GFX9-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2132 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2133 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2134 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2135 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
2136 ; GFX9-NEXT: s_cbranch_execz .LBB4_3
2137 ; GFX9-NEXT: ; %bb.1:
2138 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2139 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
2140 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
2141 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
2142 ; GFX9-NEXT: s_load_dword s4, s[0:1], 0x0
2143 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
2144 ; GFX9-NEXT: v_mov_b32_e32 v1, s4
2145 ; GFX9-NEXT: .LBB4_2: ; %atomicrmw.start
2146 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
2147 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
2148 ; GFX9-NEXT: v_max_f32_e32 v0, 4.0, v0
2149 ; GFX9-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
2150 ; GFX9-NEXT: s_waitcnt vmcnt(0)
2151 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
2152 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
2153 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
2154 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
2155 ; GFX9-NEXT: s_cbranch_execnz .LBB4_2
2156 ; GFX9-NEXT: .LBB4_3:
2157 ; GFX9-NEXT: s_endpgm
2159 ; GFX1064-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2161 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2162 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2163 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2164 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
2165 ; GFX1064-NEXT: s_cbranch_execz .LBB4_2
2166 ; GFX1064-NEXT: ; %bb.1:
2167 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2168 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
2169 ; GFX1064-NEXT: v_mov_b32_e32 v1, 4.0
2170 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
2171 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
2172 ; GFX1064-NEXT: .LBB4_2:
2173 ; GFX1064-NEXT: s_endpgm
2175 ; GFX1032-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2177 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2178 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
2179 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
2180 ; GFX1032-NEXT: s_cbranch_execz .LBB4_2
2181 ; GFX1032-NEXT: ; %bb.1:
2182 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2183 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
2184 ; GFX1032-NEXT: v_mov_b32_e32 v1, 4.0
2185 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
2186 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
2187 ; GFX1032-NEXT: .LBB4_2:
2188 ; GFX1032-NEXT: s_endpgm
2190 ; GFX1164-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2192 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2193 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
2194 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2195 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2196 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
2197 ; GFX1164-NEXT: s_cbranch_execz .LBB4_2
2198 ; GFX1164-NEXT: ; %bb.1:
2199 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
2200 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
2201 ; GFX1164-NEXT: v_mov_b32_e32 v1, 4.0
2202 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
2203 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2204 ; GFX1164-NEXT: .LBB4_2:
2205 ; GFX1164-NEXT: s_endpgm
2207 ; GFX1132-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2209 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2210 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
2211 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
2212 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
2213 ; GFX1132-NEXT: s_cbranch_execz .LBB4_2
2214 ; GFX1132-NEXT: ; %bb.1:
2215 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
2216 ; GFX1132-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
2217 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
2218 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2219 ; GFX1132-NEXT: .LBB4_2:
2220 ; GFX1132-NEXT: s_endpgm
2222 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2223 ; GFX7LESS-DPP: ; %bb.0:
2224 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
2225 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
2226 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2227 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
2228 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB4_3
2229 ; GFX7LESS-DPP-NEXT: ; %bb.1:
2230 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
2231 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
2232 ; GFX7LESS-DPP-NEXT: s_load_dword s2, s[0:1], 0x0
2233 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
2234 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
2235 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
2236 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s2
2237 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
2238 ; GFX7LESS-DPP-NEXT: .LBB4_2: ; %atomicrmw.start
2239 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
2240 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
2241 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
2242 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
2243 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v1
2244 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v0
2245 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
2246 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
2247 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
2248 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
2249 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v2
2250 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
2251 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB4_2
2252 ; GFX7LESS-DPP-NEXT: .LBB4_3:
2253 ; GFX7LESS-DPP-NEXT: s_endpgm
2255 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2256 ; GFX9-DPP: ; %bb.0:
2257 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2258 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2259 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2260 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
2261 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB4_3
2262 ; GFX9-DPP-NEXT: ; %bb.1:
2263 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2264 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
2265 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
2266 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
2267 ; GFX9-DPP-NEXT: s_load_dword s4, s[0:1], 0x0
2268 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
2269 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s4
2270 ; GFX9-DPP-NEXT: .LBB4_2: ; %atomicrmw.start
2271 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
2272 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
2273 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
2274 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
2275 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
2276 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
2277 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
2278 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
2279 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
2280 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB4_2
2281 ; GFX9-DPP-NEXT: .LBB4_3:
2282 ; GFX9-DPP-NEXT: s_endpgm
2284 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2285 ; GFX1064-DPP: ; %bb.0:
2286 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2287 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2288 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2289 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
2290 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB4_2
2291 ; GFX1064-DPP-NEXT: ; %bb.1:
2292 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2293 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
2294 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 4.0
2295 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
2296 ; GFX1064-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
2297 ; GFX1064-DPP-NEXT: .LBB4_2:
2298 ; GFX1064-DPP-NEXT: s_endpgm
2300 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2301 ; GFX1032-DPP: ; %bb.0:
2302 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2303 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
2304 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
2305 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB4_2
2306 ; GFX1032-DPP-NEXT: ; %bb.1:
2307 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
2308 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
2309 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 4.0
2310 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
2311 ; GFX1032-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
2312 ; GFX1032-DPP-NEXT: .LBB4_2:
2313 ; GFX1032-DPP-NEXT: s_endpgm
2315 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2316 ; GFX1164-DPP: ; %bb.0:
2317 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2318 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
2319 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2320 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2321 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
2322 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB4_2
2323 ; GFX1164-DPP-NEXT: ; %bb.1:
2324 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
2325 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
2326 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 4.0
2327 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
2328 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2329 ; GFX1164-DPP-NEXT: .LBB4_2:
2330 ; GFX1164-DPP-NEXT: s_endpgm
2332 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_default_scope_unsafe:
2333 ; GFX1132-DPP: ; %bb.0:
2334 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2335 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
2336 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
2337 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
2338 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB4_2
2339 ; GFX1132-DPP-NEXT: ; %bb.1:
2340 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
2341 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
2342 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
2343 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2344 ; GFX1132-DPP-NEXT: .LBB4_2:
2345 ; GFX1132-DPP-NEXT: s_endpgm
2346 %result = atomicrmw fmax ptr addrspace(1) %ptr, float 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1
2350 define amdgpu_kernel void @global_atomic_fmax_uni_address_div_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
2351 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2352 ; GFX7LESS: ; %bb.0:
2353 ; GFX7LESS-NEXT: s_mov_b32 s32, 0
2354 ; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2355 ; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2356 ; GFX7LESS-NEXT: s_mov_b32 s38, -1
2357 ; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
2358 ; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
2359 ; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
2360 ; GFX7LESS-NEXT: s_mov_b32 s14, s10
2361 ; GFX7LESS-NEXT: s_mov_b32 s13, s9
2362 ; GFX7LESS-NEXT: s_mov_b32 s12, s8
2363 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
2364 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
2365 ; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
2366 ; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
2367 ; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
2368 ; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2369 ; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2370 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2371 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2372 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2373 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
2374 ; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
2375 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
2376 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
2377 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
2378 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
2379 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
2380 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
2381 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
2382 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0x7fc00000
2383 ; GFX7LESS-NEXT: .LBB5_1: ; %ComputeLoop
2384 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
2385 ; GFX7LESS-NEXT: s_ff1_i32_b64 s2, s[0:1]
2386 ; GFX7LESS-NEXT: v_mul_f32_e32 v1, 1.0, v2
2387 ; GFX7LESS-NEXT: v_readlane_b32 s4, v0, s2
2388 ; GFX7LESS-NEXT: s_lshl_b64 s[2:3], 1, s2
2389 ; GFX7LESS-NEXT: v_mul_f32_e64 v2, 1.0, s4
2390 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
2391 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
2392 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
2393 ; GFX7LESS-NEXT: v_max_f32_e32 v2, v1, v2
2394 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB5_1
2395 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
2396 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
2397 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
2398 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2399 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
2400 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
2401 ; GFX7LESS-NEXT: s_cbranch_execz .LBB5_5
2402 ; GFX7LESS-NEXT: ; %bb.3:
2403 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
2404 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
2405 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
2406 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
2407 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[0:3], 0
2408 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
2409 ; GFX7LESS-NEXT: v_mul_f32_e32 v2, 1.0, v2
2410 ; GFX7LESS-NEXT: .LBB5_4: ; %atomicrmw.start
2411 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
2412 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
2413 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
2414 ; GFX7LESS-NEXT: v_max_f32_e32 v0, v0, v2
2415 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
2416 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, v1
2417 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v0
2418 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[3:4], off, s[0:3], 0 glc
2419 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
2420 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
2421 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
2422 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v3
2423 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
2424 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB5_4
2425 ; GFX7LESS-NEXT: .LBB5_5:
2426 ; GFX7LESS-NEXT: s_endpgm
2428 ; GFX9-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2430 ; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2431 ; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2432 ; GFX9-NEXT: s_mov_b32 s38, -1
2433 ; GFX9-NEXT: s_mov_b32 s39, 0xe00000
2434 ; GFX9-NEXT: s_add_u32 s36, s36, s11
2435 ; GFX9-NEXT: s_addc_u32 s37, s37, 0
2436 ; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
2437 ; GFX9-NEXT: s_mov_b32 s12, s8
2438 ; GFX9-NEXT: s_add_u32 s8, s34, 44
2439 ; GFX9-NEXT: s_mov_b32 s13, s9
2440 ; GFX9-NEXT: s_addc_u32 s9, s35, 0
2441 ; GFX9-NEXT: s_getpc_b64 s[4:5]
2442 ; GFX9-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2443 ; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2444 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2445 ; GFX9-NEXT: s_mov_b32 s14, s10
2446 ; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
2447 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2448 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2449 ; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
2450 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
2451 ; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
2452 ; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
2453 ; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
2454 ; GFX9-NEXT: s_mov_b32 s32, 0
2455 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
2456 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
2457 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
2458 ; GFX9-NEXT: v_mov_b32_e32 v2, 0x7fc00000
2459 ; GFX9-NEXT: .LBB5_1: ; %ComputeLoop
2460 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
2461 ; GFX9-NEXT: s_ff1_i32_b64 s2, s[0:1]
2462 ; GFX9-NEXT: v_readlane_b32 s4, v0, s2
2463 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s2
2464 ; GFX9-NEXT: v_max_f32_e32 v1, v2, v2
2465 ; GFX9-NEXT: v_max_f32_e64 v2, s4, s4
2466 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
2467 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
2468 ; GFX9-NEXT: v_max_f32_e32 v2, v1, v2
2469 ; GFX9-NEXT: s_cbranch_scc1 .LBB5_1
2470 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
2471 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2472 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2473 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2474 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
2475 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
2476 ; GFX9-NEXT: s_cbranch_execz .LBB5_5
2477 ; GFX9-NEXT: ; %bb.3:
2478 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2479 ; GFX9-NEXT: v_mov_b32_e32 v3, 0
2480 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
2481 ; GFX9-NEXT: v_max_f32_e32 v2, v2, v2
2482 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
2483 ; GFX9-NEXT: global_load_dword v1, v3, s[0:1]
2484 ; GFX9-NEXT: .LBB5_4: ; %atomicrmw.start
2485 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
2486 ; GFX9-NEXT: s_waitcnt vmcnt(0)
2487 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
2488 ; GFX9-NEXT: v_max_f32_e32 v0, v0, v2
2489 ; GFX9-NEXT: global_atomic_cmpswap v0, v3, v[0:1], s[0:1] glc
2490 ; GFX9-NEXT: s_waitcnt vmcnt(0)
2491 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
2492 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
2493 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
2494 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
2495 ; GFX9-NEXT: s_cbranch_execnz .LBB5_4
2496 ; GFX9-NEXT: .LBB5_5:
2497 ; GFX9-NEXT: s_endpgm
2499 ; GFX1064-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2501 ; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2502 ; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2503 ; GFX1064-NEXT: s_mov_b32 s38, -1
2504 ; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
2505 ; GFX1064-NEXT: s_add_u32 s36, s36, s11
2506 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
2507 ; GFX1064-NEXT: s_addc_u32 s37, s37, 0
2508 ; GFX1064-NEXT: s_mov_b32 s12, s8
2509 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
2510 ; GFX1064-NEXT: s_mov_b32 s13, s9
2511 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
2512 ; GFX1064-NEXT: s_getpc_b64 s[4:5]
2513 ; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2514 ; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2515 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2516 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2517 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2518 ; GFX1064-NEXT: s_mov_b32 s14, s10
2519 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
2520 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
2521 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
2522 ; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
2523 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
2524 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
2525 ; GFX1064-NEXT: s_mov_b32 s32, 0
2526 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
2527 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
2528 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0x7fc00000
2529 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
2530 ; GFX1064-NEXT: .LBB5_1: ; %ComputeLoop
2531 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
2532 ; GFX1064-NEXT: s_ff1_i32_b64 s2, s[0:1]
2533 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v1
2534 ; GFX1064-NEXT: v_readlane_b32 s3, v0, s2
2535 ; GFX1064-NEXT: v_max_f32_e64 v2, s3, s3
2536 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s2
2537 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
2538 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
2539 ; GFX1064-NEXT: v_max_f32_e32 v1, v1, v2
2540 ; GFX1064-NEXT: s_cbranch_scc1 .LBB5_1
2541 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
2542 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2543 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2544 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
2545 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
2546 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
2547 ; GFX1064-NEXT: s_cbranch_execz .LBB5_4
2548 ; GFX1064-NEXT: ; %bb.3:
2549 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2550 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
2551 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
2552 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
2553 ; GFX1064-NEXT: .LBB5_4:
2554 ; GFX1064-NEXT: s_endpgm
2556 ; GFX1032-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2558 ; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2559 ; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2560 ; GFX1032-NEXT: s_mov_b32 s38, -1
2561 ; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
2562 ; GFX1032-NEXT: s_add_u32 s36, s36, s11
2563 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
2564 ; GFX1032-NEXT: s_addc_u32 s37, s37, 0
2565 ; GFX1032-NEXT: s_mov_b32 s12, s8
2566 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
2567 ; GFX1032-NEXT: s_mov_b32 s13, s9
2568 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
2569 ; GFX1032-NEXT: s_getpc_b64 s[4:5]
2570 ; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2571 ; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2572 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2573 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2574 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2575 ; GFX1032-NEXT: s_mov_b32 s14, s10
2576 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
2577 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
2578 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
2579 ; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
2580 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
2581 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
2582 ; GFX1032-NEXT: s_mov_b32 s32, 0
2583 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
2584 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
2585 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0x7fc00000
2586 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
2587 ; GFX1032-NEXT: .LBB5_1: ; %ComputeLoop
2588 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
2589 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
2590 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v1
2591 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
2592 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
2593 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
2594 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
2595 ; GFX1032-NEXT: v_max_f32_e64 v2, s2, s2
2596 ; GFX1032-NEXT: v_max_f32_e32 v1, v1, v2
2597 ; GFX1032-NEXT: s_cbranch_scc1 .LBB5_1
2598 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
2599 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2600 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
2601 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
2602 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
2603 ; GFX1032-NEXT: s_cbranch_execz .LBB5_4
2604 ; GFX1032-NEXT: ; %bb.3:
2605 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2606 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
2607 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
2608 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
2609 ; GFX1032-NEXT: .LBB5_4:
2610 ; GFX1032-NEXT: s_endpgm
2612 ; GFX1164-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2614 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
2615 ; GFX1164-NEXT: s_mov_b32 s12, s8
2616 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
2617 ; GFX1164-NEXT: s_mov_b32 s13, s9
2618 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
2619 ; GFX1164-NEXT: s_getpc_b64 s[4:5]
2620 ; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2621 ; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2622 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
2623 ; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
2624 ; GFX1164-NEXT: s_mov_b32 s14, s10
2625 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
2626 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
2627 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
2628 ; GFX1164-NEXT: s_mov_b32 s32, 0
2629 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
2630 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
2631 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0x7fc00000
2632 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
2633 ; GFX1164-NEXT: .LBB5_1: ; %ComputeLoop
2634 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
2635 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2636 ; GFX1164-NEXT: s_ctz_i32_b64 s2, s[0:1]
2637 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v1
2638 ; GFX1164-NEXT: v_readlane_b32 s3, v0, s2
2639 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
2640 ; GFX1164-NEXT: v_max_f32_e64 v2, s3, s3
2641 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s2
2642 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
2643 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2644 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
2645 ; GFX1164-NEXT: v_max_f32_e32 v1, v1, v2
2646 ; GFX1164-NEXT: s_cbranch_scc1 .LBB5_1
2647 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
2648 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2649 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
2650 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2651 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
2652 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
2653 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
2654 ; GFX1164-NEXT: s_cbranch_execz .LBB5_4
2655 ; GFX1164-NEXT: ; %bb.3:
2656 ; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
2657 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
2658 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
2659 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2660 ; GFX1164-NEXT: .LBB5_4:
2661 ; GFX1164-NEXT: s_endpgm
2663 ; GFX1132-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2665 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
2666 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
2667 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
2668 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
2669 ; GFX1132-NEXT: s_getpc_b64 s[4:5]
2670 ; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2671 ; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2672 ; GFX1132-NEXT: s_mov_b32 s12, s13
2673 ; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
2674 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
2675 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
2676 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
2677 ; GFX1132-NEXT: s_mov_b32 s13, s14
2678 ; GFX1132-NEXT: s_mov_b32 s14, s15
2679 ; GFX1132-NEXT: s_mov_b32 s32, 0
2680 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
2681 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
2682 ; GFX1132-NEXT: v_mov_b32_e32 v1, 0x7fc00000
2683 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
2684 ; GFX1132-NEXT: .LBB5_1: ; %ComputeLoop
2685 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
2686 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2687 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
2688 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v1
2689 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
2690 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
2691 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
2692 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
2693 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
2694 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
2695 ; GFX1132-NEXT: v_max_f32_e64 v2, s2, s2
2696 ; GFX1132-NEXT: v_max_f32_e32 v1, v1, v2
2697 ; GFX1132-NEXT: s_cbranch_scc1 .LBB5_1
2698 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
2699 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2700 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
2701 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
2702 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
2703 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
2704 ; GFX1132-NEXT: s_cbranch_execz .LBB5_4
2705 ; GFX1132-NEXT: ; %bb.3:
2706 ; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
2707 ; GFX1132-NEXT: v_mov_b32_e32 v0, 0
2708 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
2709 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
2710 ; GFX1132-NEXT: .LBB5_4:
2711 ; GFX1132-NEXT: s_endpgm
2713 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2714 ; GFX7LESS-DPP: ; %bb.0:
2715 ; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
2716 ; GFX7LESS-DPP-NEXT: s_mov_b32 s40, SCRATCH_RSRC_DWORD0
2717 ; GFX7LESS-DPP-NEXT: s_mov_b32 s41, SCRATCH_RSRC_DWORD1
2718 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, -1
2719 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, 0xe8f000
2720 ; GFX7LESS-DPP-NEXT: s_add_u32 s40, s40, s11
2721 ; GFX7LESS-DPP-NEXT: s_addc_u32 s41, s41, 0
2722 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
2723 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
2724 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
2725 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
2726 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
2727 ; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
2728 ; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
2729 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
2730 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
2731 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
2732 ; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2733 ; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2734 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2735 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2736 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2737 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
2738 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
2739 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
2740 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
2741 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[40:41]
2742 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[42:43]
2743 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
2744 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
2745 ; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[36:39], 0
2746 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
2747 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v2, 1.0, v0
2748 ; GFX7LESS-DPP-NEXT: .LBB5_1: ; %atomicrmw.start
2749 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
2750 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
2751 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
2752 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, v0, v2
2753 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
2754 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v1
2755 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v0
2756 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[3:4], off, s[36:39], 0 glc
2757 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
2758 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v3, v1
2759 ; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
2760 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v3
2761 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
2762 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB5_1
2763 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
2764 ; GFX7LESS-DPP-NEXT: s_endpgm
2766 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2767 ; GFX9-DPP: ; %bb.0:
2768 ; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2769 ; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2770 ; GFX9-DPP-NEXT: s_mov_b32 s38, -1
2771 ; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
2772 ; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
2773 ; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
2774 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
2775 ; GFX9-DPP-NEXT: s_mov_b32 s12, s8
2776 ; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
2777 ; GFX9-DPP-NEXT: s_mov_b32 s13, s9
2778 ; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
2779 ; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
2780 ; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2781 ; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2782 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2783 ; GFX9-DPP-NEXT: s_mov_b32 s14, s10
2784 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
2785 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2786 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2787 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
2788 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
2789 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
2790 ; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
2791 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
2792 ; GFX9-DPP-NEXT: s_mov_b32 s32, 0
2793 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
2794 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
2795 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
2796 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
2797 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
2798 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
2799 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v4, v3, v0, s[0:1]
2800 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2801 ; GFX9-DPP-NEXT: s_nop 1
2802 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:1 row_mask:0xf bank_mask:0xf
2803 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v4
2804 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
2805 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
2806 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2807 ; GFX9-DPP-NEXT: s_nop 1
2808 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:2 row_mask:0xf bank_mask:0xf
2809 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
2810 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
2811 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2812 ; GFX9-DPP-NEXT: s_nop 1
2813 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:4 row_mask:0xf bank_mask:0xf
2814 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
2815 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
2816 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2817 ; GFX9-DPP-NEXT: s_nop 1
2818 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_shr:8 row_mask:0xf bank_mask:0xf
2819 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
2820 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
2821 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2822 ; GFX9-DPP-NEXT: s_nop 1
2823 ; GFX9-DPP-NEXT: v_mov_b32_dpp v5, v4 row_bcast:15 row_mask:0xa bank_mask:0xf
2824 ; GFX9-DPP-NEXT: v_max_f32_e32 v5, v5, v5
2825 ; GFX9-DPP-NEXT: v_max_f32_e32 v4, v4, v5
2826 ; GFX9-DPP-NEXT: s_nop 1
2827 ; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v4 row_bcast:31 row_mask:0xc bank_mask:0xf
2828 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v3, v3
2829 ; GFX9-DPP-NEXT: v_max_f32_e32 v3, v4, v3
2830 ; GFX9-DPP-NEXT: v_readlane_b32 s4, v3, 63
2831 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
2832 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
2833 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
2834 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB5_3
2835 ; GFX9-DPP-NEXT: ; %bb.1:
2836 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2837 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
2838 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
2839 ; GFX9-DPP-NEXT: v_max_f32_e64 v6, s4, s4
2840 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
2841 ; GFX9-DPP-NEXT: global_load_dword v1, v2, s[0:1]
2842 ; GFX9-DPP-NEXT: .LBB5_2: ; %atomicrmw.start
2843 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
2844 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
2845 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
2846 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v0, v6
2847 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
2848 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
2849 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
2850 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
2851 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
2852 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
2853 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB5_2
2854 ; GFX9-DPP-NEXT: .LBB5_3:
2855 ; GFX9-DPP-NEXT: s_endpgm
2857 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2858 ; GFX1064-DPP: ; %bb.0:
2859 ; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2860 ; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2861 ; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
2862 ; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
2863 ; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
2864 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
2865 ; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
2866 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
2867 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
2868 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
2869 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
2870 ; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
2871 ; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2872 ; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2873 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2874 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2875 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2876 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
2877 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
2878 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
2879 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
2880 ; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
2881 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
2882 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
2883 ; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
2884 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
2885 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
2886 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
2887 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
2888 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s[0:1]
2889 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2890 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
2891 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
2892 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v3
2893 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
2894 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
2895 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2896 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2897 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2898 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
2899 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2900 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2901 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2902 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
2903 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2904 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2905 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
2906 ; GFX1064-DPP-NEXT: v_max_f32_e32 v4, v4, v4
2907 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2908 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 32
2909 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v3, 0
2910 ; GFX1064-DPP-NEXT: v_max_f32_e64 v3, s2, s2
2911 ; GFX1064-DPP-NEXT: v_max_f32_e64 v4, s3, s3
2912 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
2913 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
2914 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
2915 ; GFX1064-DPP-NEXT: v_max_f32_e32 v3, v4, v3
2916 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
2917 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
2918 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
2919 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
2920 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
2921 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB5_2
2922 ; GFX1064-DPP-NEXT: ; %bb.1:
2923 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2924 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
2925 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
2926 ; GFX1064-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
2927 ; GFX1064-DPP-NEXT: .LBB5_2:
2928 ; GFX1064-DPP-NEXT: s_endpgm
2930 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2931 ; GFX1032-DPP: ; %bb.0:
2932 ; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
2933 ; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
2934 ; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
2935 ; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
2936 ; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
2937 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
2938 ; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
2939 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
2940 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
2941 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
2942 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
2943 ; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
2944 ; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
2945 ; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
2946 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
2947 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
2948 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
2949 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
2950 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
2951 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
2952 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
2953 ; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
2954 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
2955 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
2956 ; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
2957 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
2958 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
2959 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
2960 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
2961 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v4, 0x7fc00000, v0, s0
2962 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2963 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
2964 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
2965 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v3
2966 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v4, v3
2967 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
2968 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2969 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2970 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2971 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
2972 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2973 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0x7fc00000
2974 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2975 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
2976 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v5, v5
2977 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2978 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v4, v3, 0, 0
2979 ; GFX1032-DPP-NEXT: v_max_f32_e32 v4, v4, v4
2980 ; GFX1032-DPP-NEXT: v_max_f32_e32 v3, v3, v4
2981 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
2982 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
2983 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
2984 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
2985 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
2986 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB5_2
2987 ; GFX1032-DPP-NEXT: ; %bb.1:
2988 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
2989 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
2990 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
2991 ; GFX1032-DPP-NEXT: global_atomic_fmax v1, v0, s[0:1]
2992 ; GFX1032-DPP-NEXT: .LBB5_2:
2993 ; GFX1032-DPP-NEXT: s_endpgm
2995 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
2996 ; GFX1164-DPP: ; %bb.0:
2997 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
2998 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
2999 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
3000 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
3001 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
3002 ; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
3003 ; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
3004 ; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
3005 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
3006 ; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
3007 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
3008 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
3009 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
3010 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
3011 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
3012 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
3013 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3014 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
3015 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
3016 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s[0:1]
3017 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
3018 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
3019 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
3020 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
3021 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v1
3022 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3023 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v2, v1
3024 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
3025 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
3026 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
3027 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
3028 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3029 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3030 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
3031 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
3032 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
3033 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
3034 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3035 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
3036 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3037 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v3, v3
3038 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3039 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3040 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
3041 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
3042 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3043 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3044 ; GFX1164-DPP-NEXT: v_permlane64_b32 v2, v1
3045 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
3046 ; GFX1164-DPP-NEXT: v_max_f32_e32 v2, v2, v2
3047 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
3048 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
3049 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
3050 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
3051 ; GFX1164-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3052 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
3053 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
3054 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v4, exec_hi, v0
3055 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
3056 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v1
3057 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
3058 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
3059 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB5_2
3060 ; GFX1164-DPP-NEXT: ; %bb.1:
3061 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
3062 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
3063 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
3064 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
3065 ; GFX1164-DPP-NEXT: .LBB5_2:
3066 ; GFX1164-DPP-NEXT: s_endpgm
3068 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_div_value_default_scope_unsafe:
3069 ; GFX1132-DPP: ; %bb.0:
3070 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
3071 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
3072 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
3073 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
3074 ; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
3075 ; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value@gotpcrel32@lo+4
3076 ; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value@gotpcrel32@hi+12
3077 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
3078 ; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
3079 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
3080 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
3081 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
3082 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
3083 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
3084 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
3085 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
3086 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3087 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
3088 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
3089 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v2, 0x7fc00000, v0, s0
3090 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v1, 0x7fc00000
3091 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v3, 0x7fc00000
3092 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v1, v2 row_xmask:1 row_mask:0xf bank_mask:0xf
3093 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3094 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v2, v2 :: v_dual_max_f32 v1, v1, v1
3095 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v2, v1
3096 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3097 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:2 row_mask:0xf bank_mask:0xf
3098 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
3099 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3100 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3101 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:4 row_mask:0xf bank_mask:0xf
3102 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3103 ; GFX1132-DPP-NEXT: v_dual_max_f32 v2, v3, v3 :: v_dual_mov_b32 v3, 0x7fc00000
3104 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3105 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3106 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v1 row_xmask:8 row_mask:0xf bank_mask:0xf
3107 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v3, v3
3108 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3109 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3110 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v2, v1, 0, 0
3111 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3112 ; GFX1132-DPP-NEXT: v_max_f32_e32 v2, v2, v2
3113 ; GFX1132-DPP-NEXT: v_max_f32_e32 v1, v1, v2
3114 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
3115 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
3116 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v4, exec_lo, 0
3117 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v1
3118 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
3119 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
3120 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v4
3121 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB5_2
3122 ; GFX1132-DPP-NEXT: ; %bb.1:
3123 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
3124 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
3125 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
3126 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v4, v0, s[0:1]
3127 ; GFX1132-DPP-NEXT: .LBB5_2:
3128 ; GFX1132-DPP-NEXT: s_endpgm
3129 %divValue = call float @div.float.value()
3130 %result = atomicrmw fmax ptr addrspace(1) %ptr, float %divValue monotonic, align 4, !amdgpu.no.fine.grained.memory !1
3134 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
3135 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3136 ; GFX7LESS: ; %bb.0:
3137 ; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
3138 ; GFX7LESS-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3139 ; GFX7LESS-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3140 ; GFX7LESS-NEXT: s_mov_b32 s50, -1
3141 ; GFX7LESS-NEXT: s_mov_b32 s51, 0xe8f000
3142 ; GFX7LESS-NEXT: s_add_u32 s48, s48, s11
3143 ; GFX7LESS-NEXT: s_addc_u32 s49, s49, 0
3144 ; GFX7LESS-NEXT: s_mov_b64 s[40:41], s[0:1]
3145 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
3146 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
3147 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3148 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
3149 ; GFX7LESS-NEXT: s_cbranch_execz .LBB6_3
3150 ; GFX7LESS-NEXT: ; %bb.1:
3151 ; GFX7LESS-NEXT: s_mov_b32 s33, s10
3152 ; GFX7LESS-NEXT: s_mov_b32 s42, s9
3153 ; GFX7LESS-NEXT: s_mov_b32 s43, s8
3154 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
3155 ; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
3156 ; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
3157 ; GFX7LESS-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
3158 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
3159 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3160 ; GFX7LESS-NEXT: s_mov_b64 s[46:47], 0
3161 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
3162 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
3163 ; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
3164 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
3165 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
3166 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
3167 ; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
3168 ; GFX7LESS-NEXT: .LBB6_2: ; %atomicrmw.start
3169 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
3170 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
3171 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
3172 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
3173 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0
3174 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
3175 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
3176 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
3177 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
3178 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
3179 ; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3180 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3181 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
3182 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
3183 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3184 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
3185 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
3186 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
3187 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
3188 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
3189 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
3190 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
3191 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
3192 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
3193 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
3194 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
3195 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
3196 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
3197 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
3198 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
3199 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
3200 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, s44
3201 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, s45
3202 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
3203 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
3204 ; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
3205 ; GFX7LESS-NEXT: buffer_load_dword v0, off, s[48:51], 0
3206 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
3207 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
3208 ; GFX7LESS-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3209 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[46:47]
3210 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB6_2
3211 ; GFX7LESS-NEXT: .LBB6_3:
3212 ; GFX7LESS-NEXT: s_endpgm
3214 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3216 ; GFX9-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3217 ; GFX9-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3218 ; GFX9-NEXT: s_mov_b32 s50, -1
3219 ; GFX9-NEXT: s_mov_b32 s51, 0xe00000
3220 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3221 ; GFX9-NEXT: s_add_u32 s48, s48, s11
3222 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
3223 ; GFX9-NEXT: s_addc_u32 s49, s49, 0
3224 ; GFX9-NEXT: s_mov_b64 s[40:41], s[0:1]
3225 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3226 ; GFX9-NEXT: s_movk_i32 s32, 0x800
3227 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
3228 ; GFX9-NEXT: s_cbranch_execz .LBB6_3
3229 ; GFX9-NEXT: ; %bb.1:
3230 ; GFX9-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3231 ; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3232 ; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3233 ; GFX9-NEXT: s_mov_b32 s33, s10
3234 ; GFX9-NEXT: s_mov_b32 s42, s9
3235 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
3236 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3237 ; GFX9-NEXT: s_mov_b32 s43, s8
3238 ; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
3239 ; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
3240 ; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
3241 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
3242 ; GFX9-NEXT: v_mov_b32_e32 v2, s1
3243 ; GFX9-NEXT: s_mov_b64 s[46:47], 0
3244 ; GFX9-NEXT: v_mov_b32_e32 v1, s0
3245 ; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
3246 ; GFX9-NEXT: .LBB6_2: ; %atomicrmw.start
3247 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
3248 ; GFX9-NEXT: s_waitcnt vmcnt(0)
3249 ; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3250 ; GFX9-NEXT: s_add_u32 s8, s36, 44
3251 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
3252 ; GFX9-NEXT: s_getpc_b64 s[0:1]
3253 ; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3254 ; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3255 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3256 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
3257 ; GFX9-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3258 ; GFX9-NEXT: buffer_store_dword v1, off, s[48:51], 0
3259 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
3260 ; GFX9-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3261 ; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
3262 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
3263 ; GFX9-NEXT: s_mov_b32 s12, s43
3264 ; GFX9-NEXT: s_mov_b32 s13, s42
3265 ; GFX9-NEXT: s_mov_b32 s14, s33
3266 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
3267 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
3268 ; GFX9-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3269 ; GFX9-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3270 ; GFX9-NEXT: v_mov_b32_e32 v0, 8
3271 ; GFX9-NEXT: v_mov_b32_e32 v1, 0
3272 ; GFX9-NEXT: v_mov_b32_e32 v2, s44
3273 ; GFX9-NEXT: v_mov_b32_e32 v3, s45
3274 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
3275 ; GFX9-NEXT: v_mov_b32_e32 v5, 8
3276 ; GFX9-NEXT: v_mov_b32_e32 v6, 0
3277 ; GFX9-NEXT: v_mov_b32_e32 v7, 0
3278 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
3279 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
3280 ; GFX9-NEXT: buffer_load_dword v1, off, s[48:51], 0
3281 ; GFX9-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3282 ; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
3283 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3284 ; GFX9-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3285 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[46:47]
3286 ; GFX9-NEXT: s_cbranch_execnz .LBB6_2
3287 ; GFX9-NEXT: .LBB6_3:
3288 ; GFX9-NEXT: s_endpgm
3290 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3292 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3293 ; GFX1064-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3294 ; GFX1064-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3295 ; GFX1064-NEXT: s_mov_b32 s50, -1
3296 ; GFX1064-NEXT: s_mov_b32 s51, 0x31e16000
3297 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
3298 ; GFX1064-NEXT: s_add_u32 s48, s48, s11
3299 ; GFX1064-NEXT: s_addc_u32 s49, s49, 0
3300 ; GFX1064-NEXT: s_mov_b64 s[40:41], s[0:1]
3301 ; GFX1064-NEXT: s_movk_i32 s32, 0x800
3302 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3303 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
3304 ; GFX1064-NEXT: s_cbranch_execz .LBB6_3
3305 ; GFX1064-NEXT: ; %bb.1:
3306 ; GFX1064-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3307 ; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3308 ; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3309 ; GFX1064-NEXT: s_mov_b32 s33, s10
3310 ; GFX1064-NEXT: s_mov_b32 s42, s9
3311 ; GFX1064-NEXT: s_mov_b32 s43, s8
3312 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
3313 ; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
3314 ; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
3315 ; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
3316 ; GFX1064-NEXT: s_mov_b64 s[46:47], 0
3317 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
3318 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3319 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
3320 ; GFX1064-NEXT: v_mov_b32_e32 v2, s1
3321 ; GFX1064-NEXT: v_mov_b32_e32 v1, s0
3322 ; GFX1064-NEXT: .LBB6_2: ; %atomicrmw.start
3323 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
3324 ; GFX1064-NEXT: s_waitcnt vmcnt(0)
3325 ; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3326 ; GFX1064-NEXT: s_add_u32 s8, s36, 44
3327 ; GFX1064-NEXT: s_addc_u32 s9, s37, 0
3328 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
3329 ; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3330 ; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3331 ; GFX1064-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3332 ; GFX1064-NEXT: buffer_store_dword v1, off, s[48:51], 0
3333 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3334 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
3335 ; GFX1064-NEXT: v_mov_b32_e32 v0, 8
3336 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0
3337 ; GFX1064-NEXT: v_mov_b32_e32 v2, s44
3338 ; GFX1064-NEXT: v_mov_b32_e32 v5, 8
3339 ; GFX1064-NEXT: v_mov_b32_e32 v6, 0
3340 ; GFX1064-NEXT: v_mov_b32_e32 v7, 0
3341 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
3342 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
3343 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
3344 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
3345 ; GFX1064-NEXT: s_mov_b32 s12, s43
3346 ; GFX1064-NEXT: s_mov_b32 s13, s42
3347 ; GFX1064-NEXT: s_mov_b32 s14, s33
3348 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
3349 ; GFX1064-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3350 ; GFX1064-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3351 ; GFX1064-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3352 ; GFX1064-NEXT: v_mov_b32_e32 v3, s45
3353 ; GFX1064-NEXT: v_mov_b32_e32 v4, 0
3354 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
3355 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
3356 ; GFX1064-NEXT: s_clause 0x1
3357 ; GFX1064-NEXT: buffer_load_dword v1, off, s[48:51], 0
3358 ; GFX1064-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3359 ; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
3360 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3361 ; GFX1064-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3362 ; GFX1064-NEXT: s_andn2_b64 exec, exec, s[46:47]
3363 ; GFX1064-NEXT: s_cbranch_execnz .LBB6_2
3364 ; GFX1064-NEXT: .LBB6_3:
3365 ; GFX1064-NEXT: s_endpgm
3367 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3369 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3370 ; GFX1032-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3371 ; GFX1032-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3372 ; GFX1032-NEXT: s_mov_b32 s50, -1
3373 ; GFX1032-NEXT: s_mov_b32 s51, 0x31c16000
3374 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
3375 ; GFX1032-NEXT: s_add_u32 s48, s48, s11
3376 ; GFX1032-NEXT: s_addc_u32 s49, s49, 0
3377 ; GFX1032-NEXT: s_mov_b64 s[40:41], s[0:1]
3378 ; GFX1032-NEXT: s_mov_b32 s46, 0
3379 ; GFX1032-NEXT: s_movk_i32 s32, 0x400
3380 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
3381 ; GFX1032-NEXT: s_cbranch_execz .LBB6_3
3382 ; GFX1032-NEXT: ; %bb.1:
3383 ; GFX1032-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3384 ; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3385 ; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3386 ; GFX1032-NEXT: s_mov_b32 s33, s10
3387 ; GFX1032-NEXT: s_mov_b32 s42, s9
3388 ; GFX1032-NEXT: s_mov_b32 s43, s8
3389 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
3390 ; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
3391 ; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
3392 ; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
3393 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
3394 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3395 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
3396 ; GFX1032-NEXT: v_mov_b32_e32 v2, s1
3397 ; GFX1032-NEXT: v_mov_b32_e32 v1, s0
3398 ; GFX1032-NEXT: .LBB6_2: ; %atomicrmw.start
3399 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
3400 ; GFX1032-NEXT: s_waitcnt vmcnt(0)
3401 ; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3402 ; GFX1032-NEXT: s_add_u32 s8, s36, 44
3403 ; GFX1032-NEXT: s_addc_u32 s9, s37, 0
3404 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
3405 ; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3406 ; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3407 ; GFX1032-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3408 ; GFX1032-NEXT: buffer_store_dword v1, off, s[48:51], 0
3409 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3410 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
3411 ; GFX1032-NEXT: v_mov_b32_e32 v0, 8
3412 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0
3413 ; GFX1032-NEXT: v_mov_b32_e32 v2, s44
3414 ; GFX1032-NEXT: v_mov_b32_e32 v5, 8
3415 ; GFX1032-NEXT: v_mov_b32_e32 v6, 0
3416 ; GFX1032-NEXT: v_mov_b32_e32 v7, 0
3417 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
3418 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
3419 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
3420 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
3421 ; GFX1032-NEXT: s_mov_b32 s12, s43
3422 ; GFX1032-NEXT: s_mov_b32 s13, s42
3423 ; GFX1032-NEXT: s_mov_b32 s14, s33
3424 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
3425 ; GFX1032-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3426 ; GFX1032-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3427 ; GFX1032-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3428 ; GFX1032-NEXT: v_mov_b32_e32 v3, s45
3429 ; GFX1032-NEXT: v_mov_b32_e32 v4, 0
3430 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
3431 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
3432 ; GFX1032-NEXT: s_clause 0x1
3433 ; GFX1032-NEXT: buffer_load_dword v1, off, s[48:51], 0
3434 ; GFX1032-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3435 ; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
3436 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
3437 ; GFX1032-NEXT: s_or_b32 s46, vcc_lo, s46
3438 ; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
3439 ; GFX1032-NEXT: s_cbranch_execnz .LBB6_2
3440 ; GFX1032-NEXT: .LBB6_3:
3441 ; GFX1032-NEXT: s_endpgm
3443 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3445 ; GFX1164-NEXT: v_mov_b32_e32 v40, v0
3446 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
3447 ; GFX1164-NEXT: s_mov_b64 s[40:41], s[0:1]
3448 ; GFX1164-NEXT: s_mov_b32 s32, 32
3449 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
3450 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3451 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
3452 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
3453 ; GFX1164-NEXT: s_cbranch_execz .LBB6_3
3454 ; GFX1164-NEXT: ; %bb.1:
3455 ; GFX1164-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
3456 ; GFX1164-NEXT: s_mov_b32 s33, s10
3457 ; GFX1164-NEXT: s_mov_b32 s42, s9
3458 ; GFX1164-NEXT: s_mov_b32 s43, s8
3459 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
3460 ; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
3461 ; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
3462 ; GFX1164-NEXT: s_mov_b64 s[46:47], 0
3463 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
3464 ; GFX1164-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
3465 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
3466 ; GFX1164-NEXT: v_mov_b32_e32 v2, s1
3467 ; GFX1164-NEXT: v_mov_b32_e32 v1, s0
3468 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
3469 ; GFX1164-NEXT: .p2align 6
3470 ; GFX1164-NEXT: .LBB6_2: ; %atomicrmw.start
3471 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
3472 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
3473 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
3474 ; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3475 ; GFX1164-NEXT: s_add_u32 s8, s36, 44
3476 ; GFX1164-NEXT: s_addc_u32 s9, s37, 0
3477 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
3478 ; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3479 ; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3480 ; GFX1164-NEXT: v_mov_b32_e32 v31, v40
3481 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
3482 ; GFX1164-NEXT: v_mov_b32_e32 v0, 8
3483 ; GFX1164-NEXT: v_mov_b32_e32 v5, 8
3484 ; GFX1164-NEXT: v_mov_b32_e32 v6, 0
3485 ; GFX1164-NEXT: v_mov_b32_e32 v7, 0
3486 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
3487 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
3488 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
3489 ; GFX1164-NEXT: s_mov_b32 s12, s43
3490 ; GFX1164-NEXT: s_mov_b32 s13, s42
3491 ; GFX1164-NEXT: s_mov_b32 s14, s33
3492 ; GFX1164-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3493 ; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
3494 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0
3495 ; GFX1164-NEXT: v_mov_b32_e32 v2, s44
3496 ; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
3497 ; GFX1164-NEXT: v_mov_b32_e32 v3, s45
3498 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
3499 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
3500 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
3501 ; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
3502 ; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
3503 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
3504 ; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3505 ; GFX1164-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3506 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[46:47]
3507 ; GFX1164-NEXT: s_cbranch_execnz .LBB6_2
3508 ; GFX1164-NEXT: .LBB6_3:
3509 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
3510 ; GFX1164-NEXT: s_endpgm
3512 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3514 ; GFX1132-NEXT: v_mov_b32_e32 v40, v0
3515 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
3516 ; GFX1132-NEXT: s_mov_b64 s[40:41], s[0:1]
3517 ; GFX1132-NEXT: s_mov_b32 s46, 0
3518 ; GFX1132-NEXT: s_mov_b32 s32, 32
3519 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
3520 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
3521 ; GFX1132-NEXT: s_cbranch_execz .LBB6_3
3522 ; GFX1132-NEXT: ; %bb.1:
3523 ; GFX1132-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
3524 ; GFX1132-NEXT: s_mov_b32 s33, s15
3525 ; GFX1132-NEXT: s_mov_b32 s42, s14
3526 ; GFX1132-NEXT: s_mov_b32 s43, s13
3527 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
3528 ; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
3529 ; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
3530 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
3531 ; GFX1132-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
3532 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
3533 ; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
3534 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
3535 ; GFX1132-NEXT: .p2align 6
3536 ; GFX1132-NEXT: .LBB6_2: ; %atomicrmw.start
3537 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
3538 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
3539 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
3540 ; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3541 ; GFX1132-NEXT: s_add_u32 s8, s36, 44
3542 ; GFX1132-NEXT: s_addc_u32 s9, s37, 0
3543 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
3544 ; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3545 ; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3546 ; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
3547 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
3548 ; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
3549 ; GFX1132-NEXT: v_mov_b32_e32 v7, 0
3550 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
3551 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
3552 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
3553 ; GFX1132-NEXT: s_mov_b32 s12, s43
3554 ; GFX1132-NEXT: s_mov_b32 s13, s42
3555 ; GFX1132-NEXT: s_mov_b32 s14, s33
3556 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
3557 ; GFX1132-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3558 ; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
3559 ; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
3560 ; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
3561 ; GFX1132-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
3562 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
3563 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
3564 ; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
3565 ; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
3566 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
3567 ; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
3568 ; GFX1132-NEXT: s_or_b32 s46, vcc_lo, s46
3569 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
3570 ; GFX1132-NEXT: s_cbranch_execnz .LBB6_2
3571 ; GFX1132-NEXT: .LBB6_3:
3572 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
3573 ; GFX1132-NEXT: s_endpgm
3575 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3576 ; GFX7LESS-DPP: ; %bb.0:
3577 ; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
3578 ; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3579 ; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3580 ; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
3581 ; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
3582 ; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
3583 ; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
3584 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3585 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
3586 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
3587 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3588 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
3589 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB6_3
3590 ; GFX7LESS-DPP-NEXT: ; %bb.1:
3591 ; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
3592 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, s9
3593 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, s8
3594 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3595 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3596 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3597 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
3598 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
3599 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3600 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[46:47], 0
3601 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
3602 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
3603 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
3604 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
3605 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
3606 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
3607 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
3608 ; GFX7LESS-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3609 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3610 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
3611 ; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
3612 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
3613 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0
3614 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
3615 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
3616 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
3617 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
3618 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
3619 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3620 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3621 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
3622 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
3623 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3624 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
3625 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
3626 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
3627 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
3628 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
3629 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
3630 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
3631 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3632 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3633 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3634 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
3635 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
3636 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
3637 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
3638 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
3639 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
3640 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s44
3641 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s45
3642 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
3643 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3644 ; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
3645 ; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[48:51], 0
3646 ; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
3647 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
3648 ; GFX7LESS-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3649 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
3650 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB6_2
3651 ; GFX7LESS-DPP-NEXT: .LBB6_3:
3652 ; GFX7LESS-DPP-NEXT: s_endpgm
3654 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3655 ; GFX9-DPP: ; %bb.0:
3656 ; GFX9-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3657 ; GFX9-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3658 ; GFX9-DPP-NEXT: s_mov_b32 s50, -1
3659 ; GFX9-DPP-NEXT: s_mov_b32 s51, 0xe00000
3660 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3661 ; GFX9-DPP-NEXT: s_add_u32 s48, s48, s11
3662 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
3663 ; GFX9-DPP-NEXT: s_addc_u32 s49, s49, 0
3664 ; GFX9-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3665 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3666 ; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
3667 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
3668 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB6_3
3669 ; GFX9-DPP-NEXT: ; %bb.1:
3670 ; GFX9-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3671 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3672 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3673 ; GFX9-DPP-NEXT: s_mov_b32 s33, s10
3674 ; GFX9-DPP-NEXT: s_mov_b32 s42, s9
3675 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
3676 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3677 ; GFX9-DPP-NEXT: s_mov_b32 s43, s8
3678 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3679 ; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3680 ; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3681 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
3682 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
3683 ; GFX9-DPP-NEXT: s_mov_b64 s[46:47], 0
3684 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
3685 ; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
3686 ; GFX9-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3687 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3688 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
3689 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3690 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
3691 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
3692 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
3693 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3694 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3695 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3696 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
3697 ; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3698 ; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
3699 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3700 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3701 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3702 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3703 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
3704 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
3705 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
3706 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
3707 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
3708 ; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3709 ; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3710 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
3711 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
3712 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s44
3713 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s45
3714 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
3715 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
3716 ; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
3717 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
3718 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
3719 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3720 ; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
3721 ; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3722 ; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
3723 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3724 ; GFX9-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3725 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
3726 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB6_2
3727 ; GFX9-DPP-NEXT: .LBB6_3:
3728 ; GFX9-DPP-NEXT: s_endpgm
3730 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3731 ; GFX1064-DPP: ; %bb.0:
3732 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3733 ; GFX1064-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3734 ; GFX1064-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3735 ; GFX1064-DPP-NEXT: s_mov_b32 s50, -1
3736 ; GFX1064-DPP-NEXT: s_mov_b32 s51, 0x31e16000
3737 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
3738 ; GFX1064-DPP-NEXT: s_add_u32 s48, s48, s11
3739 ; GFX1064-DPP-NEXT: s_addc_u32 s49, s49, 0
3740 ; GFX1064-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3741 ; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
3742 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
3743 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
3744 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB6_3
3745 ; GFX1064-DPP-NEXT: ; %bb.1:
3746 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3747 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3748 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3749 ; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
3750 ; GFX1064-DPP-NEXT: s_mov_b32 s42, s9
3751 ; GFX1064-DPP-NEXT: s_mov_b32 s43, s8
3752 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3753 ; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
3754 ; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3755 ; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3756 ; GFX1064-DPP-NEXT: s_mov_b64 s[46:47], 0
3757 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
3758 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3759 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
3760 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
3761 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
3762 ; GFX1064-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3763 ; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3764 ; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
3765 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3766 ; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
3767 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
3768 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
3769 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3770 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3771 ; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3772 ; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
3773 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3774 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
3775 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
3776 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
3777 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s44
3778 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
3779 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
3780 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
3781 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
3782 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3783 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3784 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3785 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
3786 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
3787 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
3788 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
3789 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3790 ; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3791 ; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3792 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s45
3793 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
3794 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
3795 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3796 ; GFX1064-DPP-NEXT: s_clause 0x1
3797 ; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
3798 ; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3799 ; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
3800 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3801 ; GFX1064-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3802 ; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
3803 ; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB6_2
3804 ; GFX1064-DPP-NEXT: .LBB6_3:
3805 ; GFX1064-DPP-NEXT: s_endpgm
3807 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3808 ; GFX1032-DPP: ; %bb.0:
3809 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
3810 ; GFX1032-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
3811 ; GFX1032-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
3812 ; GFX1032-DPP-NEXT: s_mov_b32 s50, -1
3813 ; GFX1032-DPP-NEXT: s_mov_b32 s51, 0x31c16000
3814 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
3815 ; GFX1032-DPP-NEXT: s_add_u32 s48, s48, s11
3816 ; GFX1032-DPP-NEXT: s_addc_u32 s49, s49, 0
3817 ; GFX1032-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3818 ; GFX1032-DPP-NEXT: s_mov_b32 s46, 0
3819 ; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
3820 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
3821 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB6_3
3822 ; GFX1032-DPP-NEXT: ; %bb.1:
3823 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
3824 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
3825 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
3826 ; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
3827 ; GFX1032-DPP-NEXT: s_mov_b32 s42, s9
3828 ; GFX1032-DPP-NEXT: s_mov_b32 s43, s8
3829 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3830 ; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
3831 ; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3832 ; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3833 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
3834 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
3835 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
3836 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
3837 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
3838 ; GFX1032-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3839 ; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3840 ; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
3841 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3842 ; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
3843 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
3844 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
3845 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3846 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3847 ; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
3848 ; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
3849 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
3850 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
3851 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
3852 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
3853 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s44
3854 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
3855 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
3856 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
3857 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
3858 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3859 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3860 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3861 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
3862 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
3863 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
3864 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
3865 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3866 ; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
3867 ; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
3868 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s45
3869 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
3870 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
3871 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
3872 ; GFX1032-DPP-NEXT: s_clause 0x1
3873 ; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
3874 ; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
3875 ; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
3876 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
3877 ; GFX1032-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
3878 ; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
3879 ; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB6_2
3880 ; GFX1032-DPP-NEXT: .LBB6_3:
3881 ; GFX1032-DPP-NEXT: s_endpgm
3883 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3884 ; GFX1164-DPP: ; %bb.0:
3885 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
3886 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
3887 ; GFX1164-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3888 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
3889 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
3890 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
3891 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
3892 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
3893 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB6_3
3894 ; GFX1164-DPP-NEXT: ; %bb.1:
3895 ; GFX1164-DPP-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
3896 ; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
3897 ; GFX1164-DPP-NEXT: s_mov_b32 s42, s9
3898 ; GFX1164-DPP-NEXT: s_mov_b32 s43, s8
3899 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3900 ; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3901 ; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3902 ; GFX1164-DPP-NEXT: s_mov_b64 s[46:47], 0
3903 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
3904 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
3905 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
3906 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
3907 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
3908 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
3909 ; GFX1164-DPP-NEXT: .p2align 6
3910 ; GFX1164-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3911 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3912 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
3913 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
3914 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3915 ; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
3916 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
3917 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
3918 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3919 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3920 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
3921 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
3922 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
3923 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
3924 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
3925 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
3926 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3927 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3928 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3929 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
3930 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
3931 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
3932 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3933 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
3934 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
3935 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s44
3936 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
3937 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s45
3938 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
3939 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
3940 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
3941 ; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
3942 ; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
3943 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
3944 ; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
3945 ; GFX1164-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
3946 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[46:47]
3947 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB6_2
3948 ; GFX1164-DPP-NEXT: .LBB6_3:
3949 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
3950 ; GFX1164-DPP-NEXT: s_endpgm
3952 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
3953 ; GFX1132-DPP: ; %bb.0:
3954 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
3955 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
3956 ; GFX1132-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
3957 ; GFX1132-DPP-NEXT: s_mov_b32 s46, 0
3958 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
3959 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
3960 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
3961 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB6_3
3962 ; GFX1132-DPP-NEXT: ; %bb.1:
3963 ; GFX1132-DPP-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
3964 ; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
3965 ; GFX1132-DPP-NEXT: s_mov_b32 s42, s14
3966 ; GFX1132-DPP-NEXT: s_mov_b32 s43, s13
3967 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
3968 ; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
3969 ; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
3970 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
3971 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
3972 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
3973 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
3974 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
3975 ; GFX1132-DPP-NEXT: .p2align 6
3976 ; GFX1132-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
3977 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
3978 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
3979 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
3980 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
3981 ; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
3982 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
3983 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
3984 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
3985 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
3986 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
3987 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
3988 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
3989 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
3990 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
3991 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
3992 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
3993 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s43
3994 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s42
3995 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
3996 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
3997 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
3998 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
3999 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
4000 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
4001 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
4002 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
4003 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
4004 ; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
4005 ; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
4006 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
4007 ; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
4008 ; GFX1132-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
4009 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
4010 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB6_2
4011 ; GFX1132-DPP-NEXT: .LBB6_3:
4012 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
4013 ; GFX1132-DPP-NEXT: s_endpgm
4014 %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
4018 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
4019 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4020 ; GFX7LESS: ; %bb.0:
4021 ; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
4022 ; GFX7LESS-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4023 ; GFX7LESS-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4024 ; GFX7LESS-NEXT: s_mov_b32 s50, -1
4025 ; GFX7LESS-NEXT: s_mov_b32 s51, 0xe8f000
4026 ; GFX7LESS-NEXT: s_add_u32 s48, s48, s11
4027 ; GFX7LESS-NEXT: s_addc_u32 s49, s49, 0
4028 ; GFX7LESS-NEXT: s_mov_b32 s33, s10
4029 ; GFX7LESS-NEXT: s_mov_b32 s42, s9
4030 ; GFX7LESS-NEXT: s_mov_b32 s43, s8
4031 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
4032 ; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
4033 ; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
4034 ; GFX7LESS-NEXT: s_mov_b64 s[40:41], s[0:1]
4035 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
4036 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
4037 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
4038 ; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4039 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4040 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4041 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4042 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4043 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
4044 ; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
4045 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
4046 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
4047 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
4048 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
4049 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
4050 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
4051 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
4052 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
4053 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
4054 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
4055 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
4056 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
4057 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
4058 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4059 ; GFX7LESS-NEXT: .LBB7_1: ; %ComputeLoop
4060 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
4061 ; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
4062 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4063 ; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
4064 ; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
4065 ; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
4066 ; GFX7LESS-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4067 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
4068 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
4069 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
4070 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4071 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB7_1
4072 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
4073 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
4074 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
4075 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
4076 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
4077 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
4078 ; GFX7LESS-NEXT: s_cbranch_execz .LBB7_5
4079 ; GFX7LESS-NEXT: ; %bb.3:
4080 ; GFX7LESS-NEXT: s_load_dwordx2 s[44:45], s[36:37], 0x9
4081 ; GFX7LESS-NEXT: s_mov_b32 s47, 0xf000
4082 ; GFX7LESS-NEXT: s_mov_b32 s46, -1
4083 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
4084 ; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[44:47], 0
4085 ; GFX7LESS-NEXT: s_mov_b64 s[46:47], 0
4086 ; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4087 ; GFX7LESS-NEXT: .LBB7_4: ; %atomicrmw.start
4088 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
4089 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
4090 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
4091 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
4092 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0
4093 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
4094 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
4095 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[41:42]
4096 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
4097 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
4098 ; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4099 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4100 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
4101 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
4102 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4103 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
4104 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
4105 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
4106 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
4107 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
4108 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
4109 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
4110 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
4111 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
4112 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
4113 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
4114 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
4115 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
4116 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
4117 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
4118 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
4119 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, s44
4120 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, s45
4121 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
4122 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
4123 ; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
4124 ; GFX7LESS-NEXT: buffer_load_dword v0, off, s[48:51], 0
4125 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
4126 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
4127 ; GFX7LESS-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
4128 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[46:47]
4129 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB7_4
4130 ; GFX7LESS-NEXT: .LBB7_5:
4131 ; GFX7LESS-NEXT: s_endpgm
4133 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4135 ; GFX9-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4136 ; GFX9-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4137 ; GFX9-NEXT: s_mov_b32 s50, -1
4138 ; GFX9-NEXT: s_mov_b32 s51, 0xe00000
4139 ; GFX9-NEXT: s_add_u32 s48, s48, s11
4140 ; GFX9-NEXT: s_addc_u32 s49, s49, 0
4141 ; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
4142 ; GFX9-NEXT: s_mov_b32 s43, s8
4143 ; GFX9-NEXT: s_add_u32 s8, s36, 44
4144 ; GFX9-NEXT: s_mov_b32 s42, s9
4145 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
4146 ; GFX9-NEXT: s_mov_b64 s[40:41], s[0:1]
4147 ; GFX9-NEXT: s_getpc_b64 s[0:1]
4148 ; GFX9-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4149 ; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4150 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4151 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4152 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4153 ; GFX9-NEXT: s_mov_b32 s33, s10
4154 ; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
4155 ; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
4156 ; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
4157 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
4158 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
4159 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
4160 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
4161 ; GFX9-NEXT: s_mov_b32 s12, s43
4162 ; GFX9-NEXT: s_mov_b32 s13, s42
4163 ; GFX9-NEXT: s_mov_b32 s14, s33
4164 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
4165 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
4166 ; GFX9-NEXT: s_movk_i32 s32, 0x800
4167 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
4168 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
4169 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
4170 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
4171 ; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4172 ; GFX9-NEXT: .LBB7_1: ; %ComputeLoop
4173 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
4174 ; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
4175 ; GFX9-NEXT: v_readlane_b32 s3, v1, s4
4176 ; GFX9-NEXT: v_readlane_b32 s2, v0, s4
4177 ; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4178 ; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4179 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
4180 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
4181 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
4182 ; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4183 ; GFX9-NEXT: s_cbranch_scc1 .LBB7_1
4184 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
4185 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4186 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
4187 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
4188 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
4189 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
4190 ; GFX9-NEXT: s_cbranch_execz .LBB7_5
4191 ; GFX9-NEXT: ; %bb.3:
4192 ; GFX9-NEXT: s_load_dwordx2 s[44:45], s[36:37], 0x24
4193 ; GFX9-NEXT: v_mov_b32_e32 v0, 0
4194 ; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4195 ; GFX9-NEXT: s_mov_b64 s[46:47], 0
4196 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
4197 ; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
4198 ; GFX9-NEXT: .LBB7_4: ; %atomicrmw.start
4199 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
4200 ; GFX9-NEXT: s_waitcnt vmcnt(0)
4201 ; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
4202 ; GFX9-NEXT: s_add_u32 s8, s36, 44
4203 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
4204 ; GFX9-NEXT: s_getpc_b64 s[0:1]
4205 ; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4206 ; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4207 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4208 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
4209 ; GFX9-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
4210 ; GFX9-NEXT: buffer_store_dword v4, off, s[48:51], 0
4211 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
4212 ; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4213 ; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
4214 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
4215 ; GFX9-NEXT: s_mov_b32 s12, s43
4216 ; GFX9-NEXT: s_mov_b32 s13, s42
4217 ; GFX9-NEXT: s_mov_b32 s14, s33
4218 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
4219 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
4220 ; GFX9-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
4221 ; GFX9-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
4222 ; GFX9-NEXT: v_mov_b32_e32 v0, 8
4223 ; GFX9-NEXT: v_mov_b32_e32 v1, 0
4224 ; GFX9-NEXT: v_mov_b32_e32 v2, s44
4225 ; GFX9-NEXT: v_mov_b32_e32 v3, s45
4226 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
4227 ; GFX9-NEXT: v_mov_b32_e32 v5, 8
4228 ; GFX9-NEXT: v_mov_b32_e32 v6, 0
4229 ; GFX9-NEXT: v_mov_b32_e32 v7, 0
4230 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
4231 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
4232 ; GFX9-NEXT: buffer_load_dword v4, off, s[48:51], 0
4233 ; GFX9-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
4234 ; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
4235 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
4236 ; GFX9-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
4237 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[46:47]
4238 ; GFX9-NEXT: s_cbranch_execnz .LBB7_4
4239 ; GFX9-NEXT: .LBB7_5:
4240 ; GFX9-NEXT: s_endpgm
4242 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4244 ; GFX1064-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4245 ; GFX1064-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4246 ; GFX1064-NEXT: s_mov_b32 s50, -1
4247 ; GFX1064-NEXT: s_mov_b32 s51, 0x31e16000
4248 ; GFX1064-NEXT: s_add_u32 s48, s48, s11
4249 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
4250 ; GFX1064-NEXT: s_addc_u32 s49, s49, 0
4251 ; GFX1064-NEXT: s_mov_b32 s43, s8
4252 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
4253 ; GFX1064-NEXT: s_mov_b32 s42, s9
4254 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
4255 ; GFX1064-NEXT: s_mov_b64 s[40:41], s[0:1]
4256 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
4257 ; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4258 ; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4259 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4260 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4261 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4262 ; GFX1064-NEXT: s_mov_b32 s33, s10
4263 ; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
4264 ; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
4265 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
4266 ; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
4267 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
4268 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
4269 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
4270 ; GFX1064-NEXT: s_mov_b32 s12, s43
4271 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
4272 ; GFX1064-NEXT: s_mov_b32 s13, s42
4273 ; GFX1064-NEXT: s_mov_b32 s14, s33
4274 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
4275 ; GFX1064-NEXT: s_movk_i32 s32, 0x800
4276 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
4277 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
4278 ; GFX1064-NEXT: v_mov_b32_e32 v2, 0
4279 ; GFX1064-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4280 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
4281 ; GFX1064-NEXT: .LBB7_1: ; %ComputeLoop
4282 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
4283 ; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
4284 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4285 ; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
4286 ; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
4287 ; GFX1064-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4288 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
4289 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
4290 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
4291 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4292 ; GFX1064-NEXT: s_cbranch_scc1 .LBB7_1
4293 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
4294 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4295 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
4296 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
4297 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
4298 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
4299 ; GFX1064-NEXT: s_cbranch_execz .LBB7_5
4300 ; GFX1064-NEXT: ; %bb.3:
4301 ; GFX1064-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
4302 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
4303 ; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4304 ; GFX1064-NEXT: s_mov_b64 s[46:47], 0
4305 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
4306 ; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
4307 ; GFX1064-NEXT: .LBB7_4: ; %atomicrmw.start
4308 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
4309 ; GFX1064-NEXT: s_waitcnt vmcnt(0)
4310 ; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
4311 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
4312 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
4313 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
4314 ; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4315 ; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4316 ; GFX1064-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
4317 ; GFX1064-NEXT: buffer_store_dword v4, off, s[48:51], 0
4318 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4319 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
4320 ; GFX1064-NEXT: v_mov_b32_e32 v2, s44
4321 ; GFX1064-NEXT: v_mov_b32_e32 v3, s45
4322 ; GFX1064-NEXT: v_mov_b32_e32 v4, 0
4323 ; GFX1064-NEXT: v_mov_b32_e32 v5, 8
4324 ; GFX1064-NEXT: v_mov_b32_e32 v6, 0
4325 ; GFX1064-NEXT: v_mov_b32_e32 v7, 0
4326 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
4327 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
4328 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
4329 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
4330 ; GFX1064-NEXT: s_mov_b32 s12, s43
4331 ; GFX1064-NEXT: s_mov_b32 s13, s42
4332 ; GFX1064-NEXT: s_mov_b32 s14, s33
4333 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
4334 ; GFX1064-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4335 ; GFX1064-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
4336 ; GFX1064-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
4337 ; GFX1064-NEXT: v_mov_b32_e32 v0, 8
4338 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0
4339 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
4340 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
4341 ; GFX1064-NEXT: s_clause 0x1
4342 ; GFX1064-NEXT: buffer_load_dword v4, off, s[48:51], 0
4343 ; GFX1064-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
4344 ; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
4345 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
4346 ; GFX1064-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
4347 ; GFX1064-NEXT: s_andn2_b64 exec, exec, s[46:47]
4348 ; GFX1064-NEXT: s_cbranch_execnz .LBB7_4
4349 ; GFX1064-NEXT: .LBB7_5:
4350 ; GFX1064-NEXT: s_endpgm
4352 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4354 ; GFX1032-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4355 ; GFX1032-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4356 ; GFX1032-NEXT: s_mov_b32 s50, -1
4357 ; GFX1032-NEXT: s_mov_b32 s51, 0x31c16000
4358 ; GFX1032-NEXT: s_add_u32 s48, s48, s11
4359 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
4360 ; GFX1032-NEXT: s_addc_u32 s49, s49, 0
4361 ; GFX1032-NEXT: s_mov_b32 s43, s8
4362 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
4363 ; GFX1032-NEXT: s_mov_b32 s42, s9
4364 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
4365 ; GFX1032-NEXT: s_mov_b64 s[40:41], s[0:1]
4366 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
4367 ; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4368 ; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4369 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4370 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4371 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4372 ; GFX1032-NEXT: s_mov_b32 s33, s10
4373 ; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
4374 ; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
4375 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
4376 ; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
4377 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
4378 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
4379 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
4380 ; GFX1032-NEXT: s_mov_b32 s12, s43
4381 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
4382 ; GFX1032-NEXT: s_mov_b32 s13, s42
4383 ; GFX1032-NEXT: s_mov_b32 s14, s33
4384 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
4385 ; GFX1032-NEXT: s_movk_i32 s32, 0x400
4386 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
4387 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
4388 ; GFX1032-NEXT: v_mov_b32_e32 v2, 0
4389 ; GFX1032-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4390 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
4391 ; GFX1032-NEXT: .LBB7_1: ; %ComputeLoop
4392 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
4393 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
4394 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4395 ; GFX1032-NEXT: v_readlane_b32 s3, v1, s1
4396 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
4397 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
4398 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
4399 ; GFX1032-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4400 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
4401 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4402 ; GFX1032-NEXT: s_cbranch_scc1 .LBB7_1
4403 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
4404 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4405 ; GFX1032-NEXT: s_mov_b32 s46, 0
4406 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
4407 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
4408 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
4409 ; GFX1032-NEXT: s_cbranch_execz .LBB7_5
4410 ; GFX1032-NEXT: ; %bb.3:
4411 ; GFX1032-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
4412 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
4413 ; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4414 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
4415 ; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
4416 ; GFX1032-NEXT: .LBB7_4: ; %atomicrmw.start
4417 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
4418 ; GFX1032-NEXT: s_waitcnt vmcnt(0)
4419 ; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
4420 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
4421 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
4422 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
4423 ; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4424 ; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4425 ; GFX1032-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
4426 ; GFX1032-NEXT: buffer_store_dword v4, off, s[48:51], 0
4427 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4428 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
4429 ; GFX1032-NEXT: v_mov_b32_e32 v2, s44
4430 ; GFX1032-NEXT: v_mov_b32_e32 v3, s45
4431 ; GFX1032-NEXT: v_mov_b32_e32 v4, 0
4432 ; GFX1032-NEXT: v_mov_b32_e32 v5, 8
4433 ; GFX1032-NEXT: v_mov_b32_e32 v6, 0
4434 ; GFX1032-NEXT: v_mov_b32_e32 v7, 0
4435 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
4436 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
4437 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
4438 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
4439 ; GFX1032-NEXT: s_mov_b32 s12, s43
4440 ; GFX1032-NEXT: s_mov_b32 s13, s42
4441 ; GFX1032-NEXT: s_mov_b32 s14, s33
4442 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
4443 ; GFX1032-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4444 ; GFX1032-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
4445 ; GFX1032-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
4446 ; GFX1032-NEXT: v_mov_b32_e32 v0, 8
4447 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0
4448 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
4449 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
4450 ; GFX1032-NEXT: s_clause 0x1
4451 ; GFX1032-NEXT: buffer_load_dword v4, off, s[48:51], 0
4452 ; GFX1032-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
4453 ; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
4454 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
4455 ; GFX1032-NEXT: s_or_b32 s46, vcc_lo, s46
4456 ; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
4457 ; GFX1032-NEXT: s_cbranch_execnz .LBB7_4
4458 ; GFX1032-NEXT: .LBB7_5:
4459 ; GFX1032-NEXT: s_endpgm
4461 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4463 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
4464 ; GFX1164-NEXT: s_mov_b32 s43, s8
4465 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
4466 ; GFX1164-NEXT: s_mov_b32 s42, s9
4467 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
4468 ; GFX1164-NEXT: s_mov_b64 s[40:41], s[0:1]
4469 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
4470 ; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4471 ; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4472 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
4473 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
4474 ; GFX1164-NEXT: s_mov_b32 s33, s10
4475 ; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
4476 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
4477 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
4478 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
4479 ; GFX1164-NEXT: s_mov_b32 s12, s43
4480 ; GFX1164-NEXT: s_mov_b32 s13, s42
4481 ; GFX1164-NEXT: s_mov_b32 s14, s33
4482 ; GFX1164-NEXT: s_mov_b32 s32, 32
4483 ; GFX1164-NEXT: v_mov_b32_e32 v40, v0
4484 ; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
4485 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
4486 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
4487 ; GFX1164-NEXT: v_mov_b32_e32 v2, 0
4488 ; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4489 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
4490 ; GFX1164-NEXT: .LBB7_1: ; %ComputeLoop
4491 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
4492 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
4493 ; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
4494 ; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4495 ; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
4496 ; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
4497 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
4498 ; GFX1164-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4499 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
4500 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
4501 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
4502 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
4503 ; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4504 ; GFX1164-NEXT: s_cbranch_scc1 .LBB7_1
4505 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
4506 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4507 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
4508 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
4509 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
4510 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
4511 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
4512 ; GFX1164-NEXT: s_cbranch_execz .LBB7_5
4513 ; GFX1164-NEXT: ; %bb.3:
4514 ; GFX1164-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
4515 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
4516 ; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4517 ; GFX1164-NEXT: s_mov_b64 s[46:47], 0
4518 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
4519 ; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[44:45]
4520 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
4521 ; GFX1164-NEXT: .p2align 6
4522 ; GFX1164-NEXT: .LBB7_4: ; %atomicrmw.start
4523 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
4524 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
4525 ; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
4526 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
4527 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
4528 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
4529 ; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4530 ; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4531 ; GFX1164-NEXT: v_mov_b32_e32 v31, v40
4532 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
4533 ; GFX1164-NEXT: v_mov_b32_e32 v2, s44
4534 ; GFX1164-NEXT: v_mov_b32_e32 v3, s45
4535 ; GFX1164-NEXT: v_mov_b32_e32 v6, 0
4536 ; GFX1164-NEXT: v_mov_b32_e32 v7, 0
4537 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
4538 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
4539 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
4540 ; GFX1164-NEXT: s_mov_b32 s12, s43
4541 ; GFX1164-NEXT: s_mov_b32 s13, s42
4542 ; GFX1164-NEXT: s_mov_b32 s14, s33
4543 ; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4544 ; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
4545 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
4546 ; GFX1164-NEXT: v_mov_b32_e32 v5, 8
4547 ; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
4548 ; GFX1164-NEXT: v_mov_b32_e32 v0, 8
4549 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0
4550 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
4551 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
4552 ; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
4553 ; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
4554 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
4555 ; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
4556 ; GFX1164-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
4557 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[46:47]
4558 ; GFX1164-NEXT: s_cbranch_execnz .LBB7_4
4559 ; GFX1164-NEXT: .LBB7_5:
4560 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
4561 ; GFX1164-NEXT: s_endpgm
4563 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4565 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
4566 ; GFX1132-NEXT: s_mov_b64 s[40:41], s[0:1]
4567 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
4568 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
4569 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
4570 ; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4571 ; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4572 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
4573 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
4574 ; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
4575 ; GFX1132-NEXT: s_mov_b32 s42, s14
4576 ; GFX1132-NEXT: s_mov_b32 s43, s13
4577 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
4578 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
4579 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
4580 ; GFX1132-NEXT: s_mov_b32 s12, s13
4581 ; GFX1132-NEXT: s_mov_b32 s13, s14
4582 ; GFX1132-NEXT: s_mov_b32 s14, s15
4583 ; GFX1132-NEXT: s_mov_b32 s32, 32
4584 ; GFX1132-NEXT: s_mov_b32 s33, s15
4585 ; GFX1132-NEXT: v_mov_b32_e32 v40, v0
4586 ; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
4587 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
4588 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
4589 ; GFX1132-NEXT: v_mov_b32_e32 v2, 0
4590 ; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
4591 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
4592 ; GFX1132-NEXT: .LBB7_1: ; %ComputeLoop
4593 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
4594 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
4595 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
4596 ; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
4597 ; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
4598 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
4599 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
4600 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
4601 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
4602 ; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
4603 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
4604 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
4605 ; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
4606 ; GFX1132-NEXT: s_cbranch_scc1 .LBB7_1
4607 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
4608 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4609 ; GFX1132-NEXT: s_mov_b32 s46, 0
4610 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
4611 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
4612 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
4613 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
4614 ; GFX1132-NEXT: s_cbranch_execz .LBB7_5
4615 ; GFX1132-NEXT: ; %bb.3:
4616 ; GFX1132-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
4617 ; GFX1132-NEXT: v_mov_b32_e32 v0, 0
4618 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
4619 ; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
4620 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
4621 ; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[44:45]
4622 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
4623 ; GFX1132-NEXT: .p2align 6
4624 ; GFX1132-NEXT: .LBB7_4: ; %atomicrmw.start
4625 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
4626 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
4627 ; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
4628 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
4629 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
4630 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
4631 ; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4632 ; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4633 ; GFX1132-NEXT: v_mov_b32_e32 v31, v40
4634 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
4635 ; GFX1132-NEXT: v_mov_b32_e32 v3, s45
4636 ; GFX1132-NEXT: v_mov_b32_e32 v7, 0
4637 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
4638 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
4639 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
4640 ; GFX1132-NEXT: s_mov_b32 s12, s43
4641 ; GFX1132-NEXT: s_mov_b32 s13, s42
4642 ; GFX1132-NEXT: s_mov_b32 s14, s33
4643 ; GFX1132-NEXT: v_mov_b32_e32 v6, 0
4644 ; GFX1132-NEXT: v_mov_b32_e32 v2, s44
4645 ; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4646 ; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
4647 ; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
4648 ; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
4649 ; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
4650 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
4651 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
4652 ; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
4653 ; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
4654 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
4655 ; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
4656 ; GFX1132-NEXT: s_or_b32 s46, vcc_lo, s46
4657 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
4658 ; GFX1132-NEXT: s_cbranch_execnz .LBB7_4
4659 ; GFX1132-NEXT: .LBB7_5:
4660 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
4661 ; GFX1132-NEXT: s_endpgm
4663 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4664 ; GFX7LESS-DPP: ; %bb.0:
4665 ; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
4666 ; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4667 ; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4668 ; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
4669 ; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
4670 ; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
4671 ; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
4672 ; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
4673 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, s9
4674 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, s8
4675 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
4676 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
4677 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
4678 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
4679 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
4680 ; GFX7LESS-DPP-NEXT: s_mov_b32 s47, 0xf000
4681 ; GFX7LESS-DPP-NEXT: s_mov_b32 s46, -1
4682 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
4683 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
4684 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
4685 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4686 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4687 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4688 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4689 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4690 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
4691 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
4692 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
4693 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
4694 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
4695 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
4696 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
4697 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
4698 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
4699 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
4700 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
4701 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
4702 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
4703 ; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[44:47], 0
4704 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[46:47], 0
4705 ; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
4706 ; GFX7LESS-DPP-NEXT: .LBB7_1: ; %atomicrmw.start
4707 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
4708 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
4709 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
4710 ; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:4
4711 ; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0
4712 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
4713 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
4714 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
4715 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
4716 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4717 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4718 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
4719 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
4720 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4721 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
4722 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
4723 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
4724 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
4725 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
4726 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
4727 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
4728 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
4729 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
4730 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
4731 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
4732 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
4733 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
4734 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
4735 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
4736 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
4737 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s44
4738 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s45
4739 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
4740 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
4741 ; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
4742 ; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0
4743 ; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[48:51], 0 offset:4
4744 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
4745 ; GFX7LESS-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
4746 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
4747 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB7_1
4748 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
4749 ; GFX7LESS-DPP-NEXT: s_endpgm
4751 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4752 ; GFX9-DPP: ; %bb.0:
4753 ; GFX9-DPP-NEXT: s_mov_b32 s52, SCRATCH_RSRC_DWORD0
4754 ; GFX9-DPP-NEXT: s_mov_b32 s53, SCRATCH_RSRC_DWORD1
4755 ; GFX9-DPP-NEXT: s_mov_b32 s54, -1
4756 ; GFX9-DPP-NEXT: s_mov_b32 s55, 0xe00000
4757 ; GFX9-DPP-NEXT: s_add_u32 s52, s52, s11
4758 ; GFX9-DPP-NEXT: s_addc_u32 s53, s53, 0
4759 ; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
4760 ; GFX9-DPP-NEXT: s_mov_b32 s43, s8
4761 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
4762 ; GFX9-DPP-NEXT: s_mov_b32 s42, s9
4763 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
4764 ; GFX9-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
4765 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
4766 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4767 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4768 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4769 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4770 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4771 ; GFX9-DPP-NEXT: s_mov_b32 s33, s10
4772 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
4773 ; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
4774 ; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
4775 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
4776 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[52:53]
4777 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
4778 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
4779 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
4780 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
4781 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
4782 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
4783 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[54:55]
4784 ; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
4785 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
4786 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
4787 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
4788 ; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
4789 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
4790 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
4791 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
4792 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
4793 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
4794 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
4795 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
4796 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4797 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
4798 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
4799 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
4800 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
4801 ; GFX9-DPP-NEXT: s_nop 0
4802 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
4803 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
4804 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
4805 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
4806 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
4807 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
4808 ; GFX9-DPP-NEXT: s_nop 0
4809 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
4810 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
4811 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
4812 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
4813 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
4814 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
4815 ; GFX9-DPP-NEXT: s_nop 0
4816 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
4817 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
4818 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
4819 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
4820 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
4821 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
4822 ; GFX9-DPP-NEXT: s_nop 0
4823 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
4824 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
4825 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
4826 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
4827 ; GFX9-DPP-NEXT: s_nop 1
4828 ; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
4829 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
4830 ; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
4831 ; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
4832 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
4833 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
4834 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
4835 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
4836 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
4837 ; GFX9-DPP-NEXT: v_readlane_b32 s45, v9, 63
4838 ; GFX9-DPP-NEXT: v_readlane_b32 s44, v8, 63
4839 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
4840 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
4841 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
4842 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB7_3
4843 ; GFX9-DPP-NEXT: ; %bb.1:
4844 ; GFX9-DPP-NEXT: s_load_dwordx2 s[46:47], s[36:37], 0x24
4845 ; GFX9-DPP-NEXT: s_mov_b64 s[48:49], 0
4846 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
4847 ; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[46:47]
4848 ; GFX9-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
4849 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
4850 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[44:45], s[44:45]
4851 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
4852 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
4853 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
4854 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
4855 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
4856 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4857 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4858 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4859 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[52:53]
4860 ; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[52:55], 0 offset:4
4861 ; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[52:55], 0
4862 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
4863 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
4864 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
4865 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
4866 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
4867 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
4868 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
4869 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
4870 ; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[52:55], 0 offset:12
4871 ; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[52:55], 0 offset:8
4872 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[54:55]
4873 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
4874 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
4875 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s46
4876 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s47
4877 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
4878 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
4879 ; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
4880 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
4881 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
4882 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
4883 ; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[52:55], 0
4884 ; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[52:55], 0 offset:4
4885 ; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
4886 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
4887 ; GFX9-DPP-NEXT: s_or_b64 s[48:49], vcc, s[48:49]
4888 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[48:49]
4889 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB7_2
4890 ; GFX9-DPP-NEXT: .LBB7_3:
4891 ; GFX9-DPP-NEXT: s_endpgm
4893 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
4894 ; GFX1064-DPP: ; %bb.0:
4895 ; GFX1064-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
4896 ; GFX1064-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
4897 ; GFX1064-DPP-NEXT: s_mov_b32 s50, -1
4898 ; GFX1064-DPP-NEXT: s_mov_b32 s51, 0x31e16000
4899 ; GFX1064-DPP-NEXT: s_add_u32 s48, s48, s11
4900 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
4901 ; GFX1064-DPP-NEXT: s_addc_u32 s49, s49, 0
4902 ; GFX1064-DPP-NEXT: s_mov_b32 s43, s8
4903 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
4904 ; GFX1064-DPP-NEXT: s_mov_b32 s42, s9
4905 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
4906 ; GFX1064-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
4907 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
4908 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
4909 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
4910 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
4911 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4912 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
4913 ; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
4914 ; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
4915 ; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
4916 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
4917 ; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
4918 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
4919 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
4920 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
4921 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
4922 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
4923 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
4924 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
4925 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
4926 ; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
4927 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
4928 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
4929 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
4930 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
4931 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
4932 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
4933 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
4934 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
4935 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
4936 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4937 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
4938 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
4939 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
4940 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
4941 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
4942 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
4943 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4944 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
4945 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
4946 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
4947 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
4948 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
4949 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4950 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
4951 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
4952 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
4953 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
4954 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
4955 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4956 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
4957 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
4958 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
4959 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
4960 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
4961 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
4962 ; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
4963 ; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
4964 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
4965 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
4966 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
4967 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
4968 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
4969 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
4970 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
4971 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
4972 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
4973 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
4974 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
4975 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
4976 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB7_3
4977 ; GFX1064-DPP-NEXT: ; %bb.1:
4978 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
4979 ; GFX1064-DPP-NEXT: s_mov_b64 s[46:47], 0
4980 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
4981 ; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[44:45]
4982 ; GFX1064-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
4983 ; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
4984 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
4985 ; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
4986 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
4987 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
4988 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
4989 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
4990 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
4991 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
4992 ; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
4993 ; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
4994 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
4995 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
4996 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
4997 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
4998 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s44
4999 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
5000 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
5001 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5002 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
5003 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5004 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
5005 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
5006 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
5007 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
5008 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
5009 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
5010 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
5011 ; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
5012 ; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
5013 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s45
5014 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
5015 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
5016 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
5017 ; GFX1064-DPP-NEXT: s_clause 0x1
5018 ; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
5019 ; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
5020 ; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
5021 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
5022 ; GFX1064-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
5023 ; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
5024 ; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB7_2
5025 ; GFX1064-DPP-NEXT: .LBB7_3:
5026 ; GFX1064-DPP-NEXT: s_endpgm
5028 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
5029 ; GFX1032-DPP: ; %bb.0:
5030 ; GFX1032-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
5031 ; GFX1032-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
5032 ; GFX1032-DPP-NEXT: s_mov_b32 s50, -1
5033 ; GFX1032-DPP-NEXT: s_mov_b32 s51, 0x31c16000
5034 ; GFX1032-DPP-NEXT: s_add_u32 s48, s48, s11
5035 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
5036 ; GFX1032-DPP-NEXT: s_addc_u32 s49, s49, 0
5037 ; GFX1032-DPP-NEXT: s_mov_b32 s43, s8
5038 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
5039 ; GFX1032-DPP-NEXT: s_mov_b32 s42, s9
5040 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
5041 ; GFX1032-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
5042 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
5043 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
5044 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
5045 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
5046 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
5047 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
5048 ; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
5049 ; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
5050 ; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
5051 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
5052 ; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
5053 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
5054 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5055 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5056 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
5057 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
5058 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
5059 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
5060 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
5061 ; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
5062 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
5063 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
5064 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
5065 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
5066 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
5067 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
5068 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
5069 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
5070 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
5071 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5072 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
5073 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
5074 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5075 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
5076 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
5077 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
5078 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5079 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5080 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5081 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
5082 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
5083 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
5084 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5085 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5086 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5087 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
5088 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
5089 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
5090 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5091 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5092 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
5093 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
5094 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5095 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5096 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
5097 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
5098 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
5099 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
5100 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
5101 ; GFX1032-DPP-NEXT: s_mov_b32 s46, 0
5102 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
5103 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
5104 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB7_3
5105 ; GFX1032-DPP-NEXT: ; %bb.1:
5106 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
5107 ; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
5108 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
5109 ; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[44:45]
5110 ; GFX1032-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
5111 ; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5112 ; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
5113 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
5114 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
5115 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
5116 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
5117 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
5118 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
5119 ; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
5120 ; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
5121 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
5122 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
5123 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
5124 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
5125 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s44
5126 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
5127 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
5128 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
5129 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
5130 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5131 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
5132 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5133 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
5134 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
5135 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
5136 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
5137 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
5138 ; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
5139 ; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
5140 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s45
5141 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
5142 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
5143 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
5144 ; GFX1032-DPP-NEXT: s_clause 0x1
5145 ; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
5146 ; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
5147 ; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
5148 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
5149 ; GFX1032-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
5150 ; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
5151 ; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB7_2
5152 ; GFX1032-DPP-NEXT: .LBB7_3:
5153 ; GFX1032-DPP-NEXT: s_endpgm
5155 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
5156 ; GFX1164-DPP: ; %bb.0:
5157 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
5158 ; GFX1164-DPP-NEXT: s_mov_b32 s43, s8
5159 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
5160 ; GFX1164-DPP-NEXT: s_mov_b32 s42, s9
5161 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
5162 ; GFX1164-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
5163 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
5164 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
5165 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
5166 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
5167 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
5168 ; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
5169 ; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
5170 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5171 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
5172 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5173 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
5174 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
5175 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
5176 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
5177 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
5178 ; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
5179 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
5180 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
5181 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
5182 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
5183 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
5184 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
5185 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
5186 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
5187 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
5188 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
5189 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5190 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
5191 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
5192 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
5193 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5194 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
5195 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
5196 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
5197 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
5198 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5199 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5200 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5201 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5202 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
5203 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
5204 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
5205 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
5206 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5207 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5208 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5209 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
5210 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
5211 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
5212 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
5213 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
5214 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5215 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5216 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5217 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5218 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
5219 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
5220 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5221 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5222 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5223 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5224 ; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
5225 ; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
5226 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5227 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5228 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5229 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
5230 ; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5231 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5232 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
5233 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
5234 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
5235 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
5236 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
5237 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
5238 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
5239 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
5240 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB7_3
5241 ; GFX1164-DPP-NEXT: ; %bb.1:
5242 ; GFX1164-DPP-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
5243 ; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
5244 ; GFX1164-DPP-NEXT: s_mov_b64 s[46:47], 0
5245 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
5246 ; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[44:45]
5247 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
5248 ; GFX1164-DPP-NEXT: .p2align 6
5249 ; GFX1164-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
5250 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5251 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
5252 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
5253 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
5254 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
5255 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
5256 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
5257 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
5258 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
5259 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
5260 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
5261 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
5262 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
5263 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
5264 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5265 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
5266 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5267 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
5268 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
5269 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
5270 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
5271 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
5272 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
5273 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s44
5274 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
5275 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s45
5276 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
5277 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
5278 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
5279 ; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
5280 ; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
5281 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
5282 ; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
5283 ; GFX1164-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
5284 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[46:47]
5285 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB7_2
5286 ; GFX1164-DPP-NEXT: .LBB7_3:
5287 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
5288 ; GFX1164-DPP-NEXT: s_endpgm
5290 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
5291 ; GFX1132-DPP: ; %bb.0:
5292 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
5293 ; GFX1132-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
5294 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
5295 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
5296 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
5297 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
5298 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
5299 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
5300 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
5301 ; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
5302 ; GFX1132-DPP-NEXT: s_mov_b32 s42, s14
5303 ; GFX1132-DPP-NEXT: s_mov_b32 s43, s13
5304 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5305 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
5306 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5307 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
5308 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
5309 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
5310 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
5311 ; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
5312 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
5313 ; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
5314 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
5315 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
5316 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
5317 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
5318 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
5319 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
5320 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
5321 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
5322 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
5323 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5324 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
5325 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
5326 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
5327 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
5328 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5329 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
5330 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
5331 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5332 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5333 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5334 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
5335 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5336 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
5337 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
5338 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5339 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5340 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5341 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
5342 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5343 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
5344 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
5345 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5346 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5347 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5348 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
5349 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
5350 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
5351 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5352 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
5353 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
5354 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
5355 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
5356 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
5357 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
5358 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
5359 ; GFX1132-DPP-NEXT: s_mov_b32 s46, 0
5360 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
5361 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
5362 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
5363 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB7_3
5364 ; GFX1132-DPP-NEXT: ; %bb.1:
5365 ; GFX1132-DPP-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
5366 ; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
5367 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
5368 ; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[44:45]
5369 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
5370 ; GFX1132-DPP-NEXT: .p2align 6
5371 ; GFX1132-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
5372 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5373 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
5374 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
5375 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
5376 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
5377 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
5378 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
5379 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
5380 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
5381 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
5382 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
5383 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
5384 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
5385 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
5386 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
5387 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s43
5388 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s42
5389 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
5390 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
5391 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
5392 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
5393 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
5394 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
5395 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
5396 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
5397 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
5398 ; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
5399 ; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
5400 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
5401 ; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
5402 ; GFX1132-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
5403 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
5404 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB7_2
5405 ; GFX1132-DPP-NEXT: .LBB7_3:
5406 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
5407 ; GFX1132-DPP-NEXT: s_endpgm
5408 %divValue = call double @div.double.value()
5409 %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
5413 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe(ptr addrspace(1) %ptr) #0 {
5414 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5415 ; GFX7LESS: ; %bb.0:
5416 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
5417 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
5418 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5419 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
5420 ; GFX7LESS-NEXT: s_cbranch_execz .LBB8_3
5421 ; GFX7LESS-NEXT: ; %bb.1:
5422 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
5423 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
5424 ; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
5425 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
5426 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
5427 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
5428 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
5429 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
5430 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
5431 ; GFX7LESS-NEXT: .LBB8_2: ; %atomicrmw.start
5432 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
5433 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5434 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5435 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
5436 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, v3
5437 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, v2
5438 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, v1
5439 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, v0
5440 ; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
5441 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
5442 ; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
5443 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
5444 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v4
5445 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v5
5446 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
5447 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB8_2
5448 ; GFX7LESS-NEXT: .LBB8_3:
5449 ; GFX7LESS-NEXT: s_endpgm
5451 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5453 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5454 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5455 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5456 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
5457 ; GFX9-NEXT: s_cbranch_execz .LBB8_3
5458 ; GFX9-NEXT: ; %bb.1:
5459 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5460 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
5461 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
5462 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
5463 ; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
5464 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
5465 ; GFX9-NEXT: v_mov_b32_e32 v2, s4
5466 ; GFX9-NEXT: v_mov_b32_e32 v3, s5
5467 ; GFX9-NEXT: .LBB8_2: ; %atomicrmw.start
5468 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
5469 ; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5470 ; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5471 ; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
5472 ; GFX9-NEXT: s_waitcnt vmcnt(0)
5473 ; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
5474 ; GFX9-NEXT: v_mov_b32_e32 v3, v1
5475 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
5476 ; GFX9-NEXT: v_mov_b32_e32 v2, v0
5477 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
5478 ; GFX9-NEXT: s_cbranch_execnz .LBB8_2
5479 ; GFX9-NEXT: .LBB8_3:
5480 ; GFX9-NEXT: s_endpgm
5482 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5484 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5485 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5486 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5487 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
5488 ; GFX1064-NEXT: s_cbranch_execz .LBB8_2
5489 ; GFX1064-NEXT: ; %bb.1:
5490 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5491 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
5492 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0x40100000
5493 ; GFX1064-NEXT: v_mov_b32_e32 v2, 0
5494 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
5495 ; GFX1064-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
5496 ; GFX1064-NEXT: .LBB8_2:
5497 ; GFX1064-NEXT: s_endpgm
5499 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5501 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5502 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
5503 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
5504 ; GFX1032-NEXT: s_cbranch_execz .LBB8_2
5505 ; GFX1032-NEXT: ; %bb.1:
5506 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5507 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
5508 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0x40100000
5509 ; GFX1032-NEXT: v_mov_b32_e32 v2, 0
5510 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
5511 ; GFX1032-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
5512 ; GFX1032-NEXT: .LBB8_2:
5513 ; GFX1032-NEXT: s_endpgm
5515 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5517 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5518 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
5519 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5520 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5521 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
5522 ; GFX1164-NEXT: s_cbranch_execz .LBB8_3
5523 ; GFX1164-NEXT: ; %bb.1:
5524 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
5525 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
5526 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
5527 ; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
5528 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
5529 ; GFX1164-NEXT: v_mov_b32_e32 v2, s2
5530 ; GFX1164-NEXT: v_mov_b32_e32 v3, s3
5531 ; GFX1164-NEXT: s_mov_b64 s[2:3], 0
5532 ; GFX1164-NEXT: .LBB8_2: ; %atomicrmw.start
5533 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
5534 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5535 ; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5536 ; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5537 ; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
5538 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
5539 ; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
5540 ; GFX1164-NEXT: v_mov_b32_e32 v3, v1
5541 ; GFX1164-NEXT: v_mov_b32_e32 v2, v0
5542 ; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
5543 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
5544 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
5545 ; GFX1164-NEXT: s_cbranch_execnz .LBB8_2
5546 ; GFX1164-NEXT: .LBB8_3:
5547 ; GFX1164-NEXT: s_endpgm
5549 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5551 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5552 ; GFX1132-NEXT: s_mov_b32 s2, 0
5553 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
5554 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
5555 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
5556 ; GFX1132-NEXT: s_cbranch_execz .LBB8_3
5557 ; GFX1132-NEXT: ; %bb.1:
5558 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
5559 ; GFX1132-NEXT: v_mov_b32_e32 v4, 0
5560 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
5561 ; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
5562 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
5563 ; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
5564 ; GFX1132-NEXT: .LBB8_2: ; %atomicrmw.start
5565 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
5566 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5567 ; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5568 ; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5569 ; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
5570 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
5571 ; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
5572 ; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
5573 ; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
5574 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
5575 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
5576 ; GFX1132-NEXT: s_cbranch_execnz .LBB8_2
5577 ; GFX1132-NEXT: .LBB8_3:
5578 ; GFX1132-NEXT: s_endpgm
5580 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5581 ; GFX7LESS-DPP: ; %bb.0:
5582 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
5583 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
5584 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5585 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
5586 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB8_3
5587 ; GFX7LESS-DPP-NEXT: ; %bb.1:
5588 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
5589 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
5590 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
5591 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
5592 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
5593 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
5594 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
5595 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
5596 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
5597 ; GFX7LESS-DPP-NEXT: .LBB8_2: ; %atomicrmw.start
5598 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5599 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5600 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5601 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
5602 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
5603 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
5604 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v1
5605 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v0
5606 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
5607 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
5608 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
5609 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
5610 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v4
5611 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v5
5612 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
5613 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB8_2
5614 ; GFX7LESS-DPP-NEXT: .LBB8_3:
5615 ; GFX7LESS-DPP-NEXT: s_endpgm
5617 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5618 ; GFX9-DPP: ; %bb.0:
5619 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5620 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5621 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5622 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
5623 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB8_3
5624 ; GFX9-DPP-NEXT: ; %bb.1:
5625 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5626 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
5627 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
5628 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
5629 ; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
5630 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
5631 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
5632 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
5633 ; GFX9-DPP-NEXT: .LBB8_2: ; %atomicrmw.start
5634 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5635 ; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5636 ; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5637 ; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
5638 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
5639 ; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
5640 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
5641 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
5642 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
5643 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
5644 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB8_2
5645 ; GFX9-DPP-NEXT: .LBB8_3:
5646 ; GFX9-DPP-NEXT: s_endpgm
5648 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5649 ; GFX1064-DPP: ; %bb.0:
5650 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5651 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5652 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5653 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
5654 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB8_2
5655 ; GFX1064-DPP-NEXT: ; %bb.1:
5656 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5657 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
5658 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
5659 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
5660 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
5661 ; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
5662 ; GFX1064-DPP-NEXT: .LBB8_2:
5663 ; GFX1064-DPP-NEXT: s_endpgm
5665 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5666 ; GFX1032-DPP: ; %bb.0:
5667 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5668 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
5669 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
5670 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB8_2
5671 ; GFX1032-DPP-NEXT: ; %bb.1:
5672 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
5673 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
5674 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
5675 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
5676 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
5677 ; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
5678 ; GFX1032-DPP-NEXT: .LBB8_2:
5679 ; GFX1032-DPP-NEXT: s_endpgm
5681 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5682 ; GFX1164-DPP: ; %bb.0:
5683 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5684 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
5685 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5686 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5687 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
5688 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB8_3
5689 ; GFX1164-DPP-NEXT: ; %bb.1:
5690 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
5691 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
5692 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
5693 ; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
5694 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
5695 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
5696 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
5697 ; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
5698 ; GFX1164-DPP-NEXT: .LBB8_2: ; %atomicrmw.start
5699 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5700 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5701 ; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5702 ; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5703 ; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
5704 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
5705 ; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
5706 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
5707 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
5708 ; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
5709 ; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
5710 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
5711 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB8_2
5712 ; GFX1164-DPP-NEXT: .LBB8_3:
5713 ; GFX1164-DPP-NEXT: s_endpgm
5715 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_one_as_scope_unsafe:
5716 ; GFX1132-DPP: ; %bb.0:
5717 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5718 ; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
5719 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
5720 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
5721 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
5722 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB8_3
5723 ; GFX1132-DPP-NEXT: ; %bb.1:
5724 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
5725 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
5726 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
5727 ; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
5728 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
5729 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
5730 ; GFX1132-DPP-NEXT: .LBB8_2: ; %atomicrmw.start
5731 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
5732 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
5733 ; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5734 ; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
5735 ; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
5736 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
5737 ; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
5738 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
5739 ; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
5740 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
5741 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
5742 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB8_2
5743 ; GFX1132-DPP-NEXT: .LBB8_3:
5744 ; GFX1132-DPP-NEXT: s_endpgm
5745 %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 syncscope("one-as") monotonic, !amdgpu.no.fine.grained.memory !1
5749 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe(ptr addrspace(1) %ptr) #0 {
5750 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
5751 ; GFX7LESS: ; %bb.0:
5752 ; GFX7LESS-NEXT: s_mov_b32 s32, 0
5753 ; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
5754 ; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
5755 ; GFX7LESS-NEXT: s_mov_b32 s38, -1
5756 ; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
5757 ; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
5758 ; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
5759 ; GFX7LESS-NEXT: s_mov_b32 s14, s10
5760 ; GFX7LESS-NEXT: s_mov_b32 s13, s9
5761 ; GFX7LESS-NEXT: s_mov_b32 s12, s8
5762 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
5763 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
5764 ; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
5765 ; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
5766 ; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
5767 ; GFX7LESS-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
5768 ; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
5769 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
5770 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
5771 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
5772 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
5773 ; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
5774 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
5775 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
5776 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
5777 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
5778 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
5779 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
5780 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
5781 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
5782 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, 0x7ff80000
5783 ; GFX7LESS-NEXT: .LBB9_1: ; %ComputeLoop
5784 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
5785 ; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
5786 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
5787 ; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
5788 ; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
5789 ; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
5790 ; GFX7LESS-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
5791 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
5792 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
5793 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
5794 ; GFX7LESS-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
5795 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB9_1
5796 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
5797 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
5798 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
5799 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5800 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
5801 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
5802 ; GFX7LESS-NEXT: s_cbranch_execz .LBB9_5
5803 ; GFX7LESS-NEXT: ; %bb.3:
5804 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
5805 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
5806 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
5807 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
5808 ; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
5809 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
5810 ; GFX7LESS-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
5811 ; GFX7LESS-NEXT: .LBB9_4: ; %atomicrmw.start
5812 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
5813 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
5814 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5815 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
5816 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
5817 ; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
5818 ; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
5819 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
5820 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
5821 ; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
5822 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
5823 ; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
5824 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
5825 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
5826 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
5827 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
5828 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB9_4
5829 ; GFX7LESS-NEXT: .LBB9_5:
5830 ; GFX7LESS-NEXT: s_endpgm
5832 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
5834 ; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
5835 ; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
5836 ; GFX9-NEXT: s_mov_b32 s38, -1
5837 ; GFX9-NEXT: s_mov_b32 s39, 0xe00000
5838 ; GFX9-NEXT: s_add_u32 s36, s36, s11
5839 ; GFX9-NEXT: s_addc_u32 s37, s37, 0
5840 ; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
5841 ; GFX9-NEXT: s_mov_b32 s12, s8
5842 ; GFX9-NEXT: s_add_u32 s8, s34, 44
5843 ; GFX9-NEXT: s_mov_b32 s13, s9
5844 ; GFX9-NEXT: s_addc_u32 s9, s35, 0
5845 ; GFX9-NEXT: s_getpc_b64 s[4:5]
5846 ; GFX9-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
5847 ; GFX9-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
5848 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
5849 ; GFX9-NEXT: s_mov_b32 s14, s10
5850 ; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
5851 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
5852 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
5853 ; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
5854 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
5855 ; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
5856 ; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
5857 ; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
5858 ; GFX9-NEXT: s_mov_b32 s32, 0
5859 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
5860 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
5861 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
5862 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
5863 ; GFX9-NEXT: v_mov_b32_e32 v5, 0x7ff80000
5864 ; GFX9-NEXT: .LBB9_1: ; %ComputeLoop
5865 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
5866 ; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
5867 ; GFX9-NEXT: v_readlane_b32 s3, v1, s4
5868 ; GFX9-NEXT: v_readlane_b32 s2, v0, s4
5869 ; GFX9-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
5870 ; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
5871 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
5872 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
5873 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
5874 ; GFX9-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
5875 ; GFX9-NEXT: s_cbranch_scc1 .LBB9_1
5876 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
5877 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5878 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5879 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5880 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
5881 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
5882 ; GFX9-NEXT: s_cbranch_execz .LBB9_5
5883 ; GFX9-NEXT: ; %bb.3:
5884 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
5885 ; GFX9-NEXT: v_mov_b32_e32 v6, 0
5886 ; GFX9-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
5887 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
5888 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
5889 ; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
5890 ; GFX9-NEXT: .LBB9_4: ; %atomicrmw.start
5891 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
5892 ; GFX9-NEXT: s_waitcnt vmcnt(0)
5893 ; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
5894 ; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
5895 ; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
5896 ; GFX9-NEXT: s_waitcnt vmcnt(0)
5897 ; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
5898 ; GFX9-NEXT: v_mov_b32_e32 v3, v1
5899 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
5900 ; GFX9-NEXT: v_mov_b32_e32 v2, v0
5901 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
5902 ; GFX9-NEXT: s_cbranch_execnz .LBB9_4
5903 ; GFX9-NEXT: .LBB9_5:
5904 ; GFX9-NEXT: s_endpgm
5906 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
5908 ; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
5909 ; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
5910 ; GFX1064-NEXT: s_mov_b32 s38, -1
5911 ; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
5912 ; GFX1064-NEXT: s_add_u32 s36, s36, s11
5913 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
5914 ; GFX1064-NEXT: s_addc_u32 s37, s37, 0
5915 ; GFX1064-NEXT: s_mov_b32 s12, s8
5916 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
5917 ; GFX1064-NEXT: s_mov_b32 s13, s9
5918 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
5919 ; GFX1064-NEXT: s_getpc_b64 s[4:5]
5920 ; GFX1064-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
5921 ; GFX1064-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
5922 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
5923 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
5924 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
5925 ; GFX1064-NEXT: s_mov_b32 s14, s10
5926 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
5927 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
5928 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
5929 ; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
5930 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
5931 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
5932 ; GFX1064-NEXT: s_mov_b32 s32, 0
5933 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
5934 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
5935 ; GFX1064-NEXT: v_mov_b32_e32 v2, 0
5936 ; GFX1064-NEXT: v_mov_b32_e32 v3, 0x7ff80000
5937 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
5938 ; GFX1064-NEXT: .LBB9_1: ; %ComputeLoop
5939 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
5940 ; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
5941 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
5942 ; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
5943 ; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
5944 ; GFX1064-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
5945 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
5946 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
5947 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
5948 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
5949 ; GFX1064-NEXT: s_cbranch_scc1 .LBB9_1
5950 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
5951 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
5952 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
5953 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
5954 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
5955 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
5956 ; GFX1064-NEXT: s_cbranch_execz .LBB9_4
5957 ; GFX1064-NEXT: ; %bb.3:
5958 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
5959 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
5960 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
5961 ; GFX1064-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
5962 ; GFX1064-NEXT: .LBB9_4:
5963 ; GFX1064-NEXT: s_endpgm
5965 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
5967 ; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
5968 ; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
5969 ; GFX1032-NEXT: s_mov_b32 s38, -1
5970 ; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
5971 ; GFX1032-NEXT: s_add_u32 s36, s36, s11
5972 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
5973 ; GFX1032-NEXT: s_addc_u32 s37, s37, 0
5974 ; GFX1032-NEXT: s_mov_b32 s12, s8
5975 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
5976 ; GFX1032-NEXT: s_mov_b32 s13, s9
5977 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
5978 ; GFX1032-NEXT: s_getpc_b64 s[4:5]
5979 ; GFX1032-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
5980 ; GFX1032-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
5981 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
5982 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
5983 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
5984 ; GFX1032-NEXT: s_mov_b32 s14, s10
5985 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
5986 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
5987 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
5988 ; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
5989 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
5990 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
5991 ; GFX1032-NEXT: s_mov_b32 s32, 0
5992 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
5993 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
5994 ; GFX1032-NEXT: v_mov_b32_e32 v2, 0
5995 ; GFX1032-NEXT: v_mov_b32_e32 v3, 0x7ff80000
5996 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
5997 ; GFX1032-NEXT: .LBB9_1: ; %ComputeLoop
5998 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
5999 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
6000 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
6001 ; GFX1032-NEXT: v_readlane_b32 s3, v1, s1
6002 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
6003 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
6004 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
6005 ; GFX1032-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
6006 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
6007 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6008 ; GFX1032-NEXT: s_cbranch_scc1 .LBB9_1
6009 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
6010 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6011 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
6012 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
6013 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
6014 ; GFX1032-NEXT: s_cbranch_execz .LBB9_4
6015 ; GFX1032-NEXT: ; %bb.3:
6016 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
6017 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
6018 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
6019 ; GFX1032-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
6020 ; GFX1032-NEXT: .LBB9_4:
6021 ; GFX1032-NEXT: s_endpgm
6023 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6025 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
6026 ; GFX1164-NEXT: s_mov_b32 s12, s8
6027 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
6028 ; GFX1164-NEXT: s_mov_b32 s13, s9
6029 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
6030 ; GFX1164-NEXT: s_getpc_b64 s[4:5]
6031 ; GFX1164-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6032 ; GFX1164-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6033 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
6034 ; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
6035 ; GFX1164-NEXT: s_mov_b32 s14, s10
6036 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
6037 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
6038 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
6039 ; GFX1164-NEXT: s_mov_b32 s32, 0
6040 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
6041 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
6042 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
6043 ; GFX1164-NEXT: v_mov_b32_e32 v5, 0x7ff80000
6044 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
6045 ; GFX1164-NEXT: .LBB9_1: ; %ComputeLoop
6046 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
6047 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6048 ; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
6049 ; GFX1164-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
6050 ; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
6051 ; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
6052 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
6053 ; GFX1164-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
6054 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
6055 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
6056 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6057 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
6058 ; GFX1164-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
6059 ; GFX1164-NEXT: s_cbranch_scc1 .LBB9_1
6060 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
6061 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6062 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
6063 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6064 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
6065 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
6066 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
6067 ; GFX1164-NEXT: s_cbranch_execz .LBB9_5
6068 ; GFX1164-NEXT: ; %bb.3:
6069 ; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
6070 ; GFX1164-NEXT: v_mov_b32_e32 v6, 0
6071 ; GFX1164-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6072 ; GFX1164-NEXT: s_mov_b64 s[2:3], 0
6073 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
6074 ; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
6075 ; GFX1164-NEXT: .LBB9_4: ; %atomicrmw.start
6076 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
6077 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
6078 ; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
6079 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
6080 ; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
6081 ; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
6082 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
6083 ; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
6084 ; GFX1164-NEXT: v_mov_b32_e32 v3, v1
6085 ; GFX1164-NEXT: v_mov_b32_e32 v2, v0
6086 ; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
6087 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
6088 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
6089 ; GFX1164-NEXT: s_cbranch_execnz .LBB9_4
6090 ; GFX1164-NEXT: .LBB9_5:
6091 ; GFX1164-NEXT: s_endpgm
6093 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6095 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
6096 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
6097 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
6098 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
6099 ; GFX1132-NEXT: s_getpc_b64 s[4:5]
6100 ; GFX1132-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6101 ; GFX1132-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6102 ; GFX1132-NEXT: s_mov_b32 s12, s13
6103 ; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
6104 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
6105 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
6106 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
6107 ; GFX1132-NEXT: s_mov_b32 s13, s14
6108 ; GFX1132-NEXT: s_mov_b32 s14, s15
6109 ; GFX1132-NEXT: s_mov_b32 s32, 0
6110 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
6111 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
6112 ; GFX1132-NEXT: v_mov_b32_e32 v4, 0
6113 ; GFX1132-NEXT: v_mov_b32_e32 v5, 0x7ff80000
6114 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
6115 ; GFX1132-NEXT: .LBB9_1: ; %ComputeLoop
6116 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
6117 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6118 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
6119 ; GFX1132-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
6120 ; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
6121 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
6122 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
6123 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6124 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
6125 ; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
6126 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
6127 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
6128 ; GFX1132-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
6129 ; GFX1132-NEXT: s_cbranch_scc1 .LBB9_1
6130 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
6131 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6132 ; GFX1132-NEXT: s_mov_b32 s2, 0
6133 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
6134 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
6135 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
6136 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
6137 ; GFX1132-NEXT: s_cbranch_execz .LBB9_5
6138 ; GFX1132-NEXT: ; %bb.3:
6139 ; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
6140 ; GFX1132-NEXT: v_mov_b32_e32 v6, 0
6141 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
6142 ; GFX1132-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6143 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
6144 ; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
6145 ; GFX1132-NEXT: .LBB9_4: ; %atomicrmw.start
6146 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
6147 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
6148 ; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
6149 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
6150 ; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
6151 ; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
6152 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
6153 ; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
6154 ; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
6155 ; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
6156 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
6157 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
6158 ; GFX1132-NEXT: s_cbranch_execnz .LBB9_4
6159 ; GFX1132-NEXT: .LBB9_5:
6160 ; GFX1132-NEXT: s_endpgm
6162 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6163 ; GFX7LESS-DPP: ; %bb.0:
6164 ; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
6165 ; GFX7LESS-DPP-NEXT: s_mov_b32 s40, SCRATCH_RSRC_DWORD0
6166 ; GFX7LESS-DPP-NEXT: s_mov_b32 s41, SCRATCH_RSRC_DWORD1
6167 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, -1
6168 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, 0xe8f000
6169 ; GFX7LESS-DPP-NEXT: s_add_u32 s40, s40, s11
6170 ; GFX7LESS-DPP-NEXT: s_addc_u32 s41, s41, 0
6171 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
6172 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
6173 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
6174 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6175 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
6176 ; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
6177 ; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
6178 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
6179 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
6180 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
6181 ; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6182 ; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6183 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
6184 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
6185 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
6186 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
6187 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
6188 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6189 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6190 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[40:41]
6191 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[42:43]
6192 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
6193 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6194 ; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[36:39], 0
6195 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
6196 ; GFX7LESS-DPP-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
6197 ; GFX7LESS-DPP-NEXT: .LBB9_1: ; %atomicrmw.start
6198 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
6199 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
6200 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
6201 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
6202 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
6203 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
6204 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
6205 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
6206 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
6207 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
6208 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
6209 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
6210 ; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
6211 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
6212 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
6213 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
6214 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB9_1
6215 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
6216 ; GFX7LESS-DPP-NEXT: s_endpgm
6218 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6219 ; GFX9-DPP: ; %bb.0:
6220 ; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
6221 ; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
6222 ; GFX9-DPP-NEXT: s_mov_b32 s38, -1
6223 ; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
6224 ; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
6225 ; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
6226 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
6227 ; GFX9-DPP-NEXT: s_mov_b32 s12, s8
6228 ; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
6229 ; GFX9-DPP-NEXT: s_mov_b32 s13, s9
6230 ; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
6231 ; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
6232 ; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6233 ; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6234 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
6235 ; GFX9-DPP-NEXT: s_mov_b32 s14, s10
6236 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6237 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
6238 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
6239 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6240 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6241 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
6242 ; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
6243 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
6244 ; GFX9-DPP-NEXT: s_mov_b32 s32, 0
6245 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
6246 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6247 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
6248 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
6249 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
6250 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
6251 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
6252 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
6253 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
6254 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
6255 ; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
6256 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6257 ; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
6258 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
6259 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
6260 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
6261 ; GFX9-DPP-NEXT: s_nop 0
6262 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
6263 ; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
6264 ; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
6265 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
6266 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
6267 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
6268 ; GFX9-DPP-NEXT: s_nop 0
6269 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
6270 ; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
6271 ; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
6272 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
6273 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
6274 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
6275 ; GFX9-DPP-NEXT: s_nop 0
6276 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
6277 ; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
6278 ; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
6279 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
6280 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
6281 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
6282 ; GFX9-DPP-NEXT: s_nop 0
6283 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
6284 ; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
6285 ; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
6286 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
6287 ; GFX9-DPP-NEXT: s_nop 1
6288 ; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
6289 ; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
6290 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
6291 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
6292 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
6293 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
6294 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
6295 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
6296 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
6297 ; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
6298 ; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
6299 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
6300 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
6301 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
6302 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
6303 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB9_3
6304 ; GFX9-DPP-NEXT: ; %bb.1:
6305 ; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
6306 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
6307 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
6308 ; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
6309 ; GFX9-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
6310 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
6311 ; GFX9-DPP-NEXT: v_max_f64 v[1:2], s[0:1], s[0:1]
6312 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
6313 ; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[11:12], v[11:12]
6314 ; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[9:10], v[1:2]
6315 ; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
6316 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
6317 ; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
6318 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
6319 ; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
6320 ; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
6321 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
6322 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB9_2
6323 ; GFX9-DPP-NEXT: .LBB9_3:
6324 ; GFX9-DPP-NEXT: s_endpgm
6326 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6327 ; GFX1064-DPP: ; %bb.0:
6328 ; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
6329 ; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
6330 ; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
6331 ; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
6332 ; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
6333 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
6334 ; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
6335 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
6336 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
6337 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
6338 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
6339 ; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
6340 ; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6341 ; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6342 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
6343 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
6344 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
6345 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
6346 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6347 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6348 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6349 ; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
6350 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
6351 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
6352 ; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
6353 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
6354 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6355 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
6356 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
6357 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
6358 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s[0:1]
6359 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
6360 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
6361 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
6362 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6363 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
6364 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
6365 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6366 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
6367 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
6368 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
6369 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6370 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6371 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6372 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
6373 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
6374 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
6375 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6376 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6377 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6378 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
6379 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
6380 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
6381 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6382 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6383 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
6384 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
6385 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6386 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6387 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
6388 ; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
6389 ; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
6390 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
6391 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], s[4:5], s[4:5]
6392 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], s[2:3], s[2:3]
6393 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
6394 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
6395 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6396 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
6397 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
6398 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
6399 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
6400 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
6401 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
6402 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB9_2
6403 ; GFX1064-DPP-NEXT: ; %bb.1:
6404 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
6405 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
6406 ; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
6407 ; GFX1064-DPP-NEXT: .LBB9_2:
6408 ; GFX1064-DPP-NEXT: s_endpgm
6410 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6411 ; GFX1032-DPP: ; %bb.0:
6412 ; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
6413 ; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
6414 ; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
6415 ; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
6416 ; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
6417 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
6418 ; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
6419 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
6420 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
6421 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
6422 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
6423 ; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
6424 ; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6425 ; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6426 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
6427 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
6428 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
6429 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
6430 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6431 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6432 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6433 ; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
6434 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
6435 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
6436 ; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
6437 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
6438 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6439 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
6440 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
6441 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
6442 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s0
6443 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
6444 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
6445 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
6446 ; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6447 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
6448 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
6449 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6450 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
6451 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
6452 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
6453 ; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6454 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6455 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6456 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
6457 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
6458 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
6459 ; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6460 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6461 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
6462 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
6463 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
6464 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
6465 ; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6466 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6467 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
6468 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
6469 ; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
6470 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
6471 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
6472 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
6473 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
6474 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
6475 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
6476 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
6477 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
6478 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB9_2
6479 ; GFX1032-DPP-NEXT: ; %bb.1:
6480 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
6481 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
6482 ; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
6483 ; GFX1032-DPP-NEXT: .LBB9_2:
6484 ; GFX1032-DPP-NEXT: s_endpgm
6486 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6487 ; GFX1164-DPP: ; %bb.0:
6488 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
6489 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
6490 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
6491 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
6492 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
6493 ; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
6494 ; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6495 ; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6496 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
6497 ; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
6498 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
6499 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6500 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6501 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6502 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
6503 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
6504 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6505 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
6506 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7ff80000
6507 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
6508 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s[0:1]
6509 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
6510 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
6511 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
6512 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
6513 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6514 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
6515 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
6516 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
6517 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
6518 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
6519 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
6520 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
6521 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
6522 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6523 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6524 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6525 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
6526 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
6527 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
6528 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
6529 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
6530 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6531 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6532 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6533 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
6534 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
6535 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
6536 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
6537 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
6538 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6539 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6540 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6541 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6542 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
6543 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
6544 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6545 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6546 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6547 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6548 ; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
6549 ; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
6550 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6551 ; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6552 ; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6553 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
6554 ; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
6555 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6556 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
6557 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
6558 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
6559 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
6560 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
6561 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
6562 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
6563 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB9_3
6564 ; GFX1164-DPP-NEXT: ; %bb.1:
6565 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
6566 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
6567 ; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
6568 ; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
6569 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
6570 ; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
6571 ; GFX1164-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
6572 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
6573 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
6574 ; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
6575 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
6576 ; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
6577 ; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
6578 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
6579 ; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
6580 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
6581 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
6582 ; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
6583 ; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
6584 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
6585 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB9_2
6586 ; GFX1164-DPP-NEXT: .LBB9_3:
6587 ; GFX1164-DPP-NEXT: s_endpgm
6589 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_one_as_scope_unsafe:
6590 ; GFX1132-DPP: ; %bb.0:
6591 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
6592 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
6593 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
6594 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
6595 ; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
6596 ; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.double.value@gotpcrel32@lo+4
6597 ; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.double.value@gotpcrel32@hi+12
6598 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
6599 ; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
6600 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
6601 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
6602 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
6603 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
6604 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
6605 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
6606 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
6607 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
6608 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
6609 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, 0x7ff80000 :: v_dual_mov_b32 v2, 0
6610 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s0
6611 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
6612 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
6613 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
6614 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
6615 ; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6616 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
6617 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
6618 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
6619 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
6620 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6621 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
6622 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
6623 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6624 ; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6625 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6626 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
6627 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6628 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
6629 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
6630 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6631 ; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6632 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6633 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
6634 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6635 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
6636 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
6637 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6638 ; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6639 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6640 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
6641 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
6642 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
6643 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6644 ; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
6645 ; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
6646 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
6647 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
6648 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
6649 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
6650 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
6651 ; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
6652 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
6653 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
6654 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
6655 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB9_3
6656 ; GFX1132-DPP-NEXT: ; %bb.1:
6657 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
6658 ; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
6659 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
6660 ; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
6661 ; GFX1132-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
6662 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
6663 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
6664 ; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
6665 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
6666 ; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
6667 ; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
6668 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
6669 ; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
6670 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
6671 ; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
6672 ; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
6673 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
6674 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB9_2
6675 ; GFX1132-DPP-NEXT: .LBB9_3:
6676 ; GFX1132-DPP-NEXT: s_endpgm
6677 %divValue = call double @div.double.value()
6678 %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue syncscope("one-as") monotonic, !amdgpu.no.fine.grained.memory !1
6682 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
6683 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
6684 ; GFX7LESS: ; %bb.0:
6685 ; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
6686 ; GFX7LESS-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
6687 ; GFX7LESS-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
6688 ; GFX7LESS-NEXT: s_mov_b32 s50, -1
6689 ; GFX7LESS-NEXT: s_mov_b32 s51, 0xe8f000
6690 ; GFX7LESS-NEXT: s_add_u32 s48, s48, s11
6691 ; GFX7LESS-NEXT: s_addc_u32 s49, s49, 0
6692 ; GFX7LESS-NEXT: s_mov_b64 s[40:41], s[0:1]
6693 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
6694 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
6695 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
6696 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
6697 ; GFX7LESS-NEXT: s_cbranch_execz .LBB10_3
6698 ; GFX7LESS-NEXT: ; %bb.1:
6699 ; GFX7LESS-NEXT: s_mov_b32 s33, s10
6700 ; GFX7LESS-NEXT: s_mov_b32 s42, s9
6701 ; GFX7LESS-NEXT: s_mov_b32 s43, s8
6702 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
6703 ; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
6704 ; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
6705 ; GFX7LESS-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
6706 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
6707 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
6708 ; GFX7LESS-NEXT: s_mov_b64 s[46:47], 0
6709 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
6710 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
6711 ; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
6712 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
6713 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
6714 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
6715 ; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
6716 ; GFX7LESS-NEXT: .LBB10_2: ; %atomicrmw.start
6717 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
6718 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
6719 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
6720 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
6721 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0
6722 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
6723 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
6724 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
6725 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
6726 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
6727 ; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
6728 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
6729 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
6730 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
6731 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
6732 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
6733 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
6734 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
6735 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
6736 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
6737 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
6738 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
6739 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
6740 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
6741 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
6742 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
6743 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
6744 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
6745 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
6746 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
6747 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
6748 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, s44
6749 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, s45
6750 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
6751 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
6752 ; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
6753 ; GFX7LESS-NEXT: buffer_load_dword v0, off, s[48:51], 0
6754 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
6755 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
6756 ; GFX7LESS-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
6757 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[46:47]
6758 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB10_2
6759 ; GFX7LESS-NEXT: .LBB10_3:
6760 ; GFX7LESS-NEXT: s_endpgm
6762 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
6764 ; GFX9-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
6765 ; GFX9-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
6766 ; GFX9-NEXT: s_mov_b32 s50, -1
6767 ; GFX9-NEXT: s_mov_b32 s51, 0xe00000
6768 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
6769 ; GFX9-NEXT: s_add_u32 s48, s48, s11
6770 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
6771 ; GFX9-NEXT: s_addc_u32 s49, s49, 0
6772 ; GFX9-NEXT: s_mov_b64 s[40:41], s[0:1]
6773 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
6774 ; GFX9-NEXT: s_movk_i32 s32, 0x800
6775 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
6776 ; GFX9-NEXT: s_cbranch_execz .LBB10_3
6777 ; GFX9-NEXT: ; %bb.1:
6778 ; GFX9-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
6779 ; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
6780 ; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
6781 ; GFX9-NEXT: s_mov_b32 s33, s10
6782 ; GFX9-NEXT: s_mov_b32 s42, s9
6783 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
6784 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
6785 ; GFX9-NEXT: s_mov_b32 s43, s8
6786 ; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
6787 ; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
6788 ; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
6789 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
6790 ; GFX9-NEXT: v_mov_b32_e32 v2, s1
6791 ; GFX9-NEXT: s_mov_b64 s[46:47], 0
6792 ; GFX9-NEXT: v_mov_b32_e32 v1, s0
6793 ; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
6794 ; GFX9-NEXT: .LBB10_2: ; %atomicrmw.start
6795 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
6796 ; GFX9-NEXT: s_waitcnt vmcnt(0)
6797 ; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
6798 ; GFX9-NEXT: s_add_u32 s8, s36, 44
6799 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
6800 ; GFX9-NEXT: s_getpc_b64 s[0:1]
6801 ; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
6802 ; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
6803 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
6804 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
6805 ; GFX9-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
6806 ; GFX9-NEXT: buffer_store_dword v1, off, s[48:51], 0
6807 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
6808 ; GFX9-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
6809 ; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
6810 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
6811 ; GFX9-NEXT: s_mov_b32 s12, s43
6812 ; GFX9-NEXT: s_mov_b32 s13, s42
6813 ; GFX9-NEXT: s_mov_b32 s14, s33
6814 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
6815 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
6816 ; GFX9-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
6817 ; GFX9-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
6818 ; GFX9-NEXT: v_mov_b32_e32 v0, 8
6819 ; GFX9-NEXT: v_mov_b32_e32 v1, 0
6820 ; GFX9-NEXT: v_mov_b32_e32 v2, s44
6821 ; GFX9-NEXT: v_mov_b32_e32 v3, s45
6822 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
6823 ; GFX9-NEXT: v_mov_b32_e32 v5, 8
6824 ; GFX9-NEXT: v_mov_b32_e32 v6, 0
6825 ; GFX9-NEXT: v_mov_b32_e32 v7, 0
6826 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
6827 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
6828 ; GFX9-NEXT: buffer_load_dword v1, off, s[48:51], 0
6829 ; GFX9-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
6830 ; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
6831 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
6832 ; GFX9-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
6833 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[46:47]
6834 ; GFX9-NEXT: s_cbranch_execnz .LBB10_2
6835 ; GFX9-NEXT: .LBB10_3:
6836 ; GFX9-NEXT: s_endpgm
6838 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
6840 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
6841 ; GFX1064-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
6842 ; GFX1064-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
6843 ; GFX1064-NEXT: s_mov_b32 s50, -1
6844 ; GFX1064-NEXT: s_mov_b32 s51, 0x31e16000
6845 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
6846 ; GFX1064-NEXT: s_add_u32 s48, s48, s11
6847 ; GFX1064-NEXT: s_addc_u32 s49, s49, 0
6848 ; GFX1064-NEXT: s_mov_b64 s[40:41], s[0:1]
6849 ; GFX1064-NEXT: s_movk_i32 s32, 0x800
6850 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
6851 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
6852 ; GFX1064-NEXT: s_cbranch_execz .LBB10_3
6853 ; GFX1064-NEXT: ; %bb.1:
6854 ; GFX1064-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
6855 ; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
6856 ; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
6857 ; GFX1064-NEXT: s_mov_b32 s33, s10
6858 ; GFX1064-NEXT: s_mov_b32 s42, s9
6859 ; GFX1064-NEXT: s_mov_b32 s43, s8
6860 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
6861 ; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
6862 ; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
6863 ; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
6864 ; GFX1064-NEXT: s_mov_b64 s[46:47], 0
6865 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
6866 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
6867 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
6868 ; GFX1064-NEXT: v_mov_b32_e32 v2, s1
6869 ; GFX1064-NEXT: v_mov_b32_e32 v1, s0
6870 ; GFX1064-NEXT: .LBB10_2: ; %atomicrmw.start
6871 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
6872 ; GFX1064-NEXT: s_waitcnt vmcnt(0)
6873 ; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
6874 ; GFX1064-NEXT: s_add_u32 s8, s36, 44
6875 ; GFX1064-NEXT: s_addc_u32 s9, s37, 0
6876 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
6877 ; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
6878 ; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
6879 ; GFX1064-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
6880 ; GFX1064-NEXT: buffer_store_dword v1, off, s[48:51], 0
6881 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
6882 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
6883 ; GFX1064-NEXT: v_mov_b32_e32 v0, 8
6884 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0
6885 ; GFX1064-NEXT: v_mov_b32_e32 v2, s44
6886 ; GFX1064-NEXT: v_mov_b32_e32 v5, 8
6887 ; GFX1064-NEXT: v_mov_b32_e32 v6, 0
6888 ; GFX1064-NEXT: v_mov_b32_e32 v7, 0
6889 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
6890 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
6891 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
6892 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
6893 ; GFX1064-NEXT: s_mov_b32 s12, s43
6894 ; GFX1064-NEXT: s_mov_b32 s13, s42
6895 ; GFX1064-NEXT: s_mov_b32 s14, s33
6896 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
6897 ; GFX1064-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
6898 ; GFX1064-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
6899 ; GFX1064-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
6900 ; GFX1064-NEXT: v_mov_b32_e32 v3, s45
6901 ; GFX1064-NEXT: v_mov_b32_e32 v4, 0
6902 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
6903 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
6904 ; GFX1064-NEXT: s_clause 0x1
6905 ; GFX1064-NEXT: buffer_load_dword v1, off, s[48:51], 0
6906 ; GFX1064-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
6907 ; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
6908 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
6909 ; GFX1064-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
6910 ; GFX1064-NEXT: s_andn2_b64 exec, exec, s[46:47]
6911 ; GFX1064-NEXT: s_cbranch_execnz .LBB10_2
6912 ; GFX1064-NEXT: .LBB10_3:
6913 ; GFX1064-NEXT: s_endpgm
6915 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
6917 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
6918 ; GFX1032-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
6919 ; GFX1032-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
6920 ; GFX1032-NEXT: s_mov_b32 s50, -1
6921 ; GFX1032-NEXT: s_mov_b32 s51, 0x31c16000
6922 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
6923 ; GFX1032-NEXT: s_add_u32 s48, s48, s11
6924 ; GFX1032-NEXT: s_addc_u32 s49, s49, 0
6925 ; GFX1032-NEXT: s_mov_b64 s[40:41], s[0:1]
6926 ; GFX1032-NEXT: s_mov_b32 s46, 0
6927 ; GFX1032-NEXT: s_movk_i32 s32, 0x400
6928 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
6929 ; GFX1032-NEXT: s_cbranch_execz .LBB10_3
6930 ; GFX1032-NEXT: ; %bb.1:
6931 ; GFX1032-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
6932 ; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
6933 ; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
6934 ; GFX1032-NEXT: s_mov_b32 s33, s10
6935 ; GFX1032-NEXT: s_mov_b32 s42, s9
6936 ; GFX1032-NEXT: s_mov_b32 s43, s8
6937 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
6938 ; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
6939 ; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
6940 ; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
6941 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
6942 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
6943 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
6944 ; GFX1032-NEXT: v_mov_b32_e32 v2, s1
6945 ; GFX1032-NEXT: v_mov_b32_e32 v1, s0
6946 ; GFX1032-NEXT: .LBB10_2: ; %atomicrmw.start
6947 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
6948 ; GFX1032-NEXT: s_waitcnt vmcnt(0)
6949 ; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
6950 ; GFX1032-NEXT: s_add_u32 s8, s36, 44
6951 ; GFX1032-NEXT: s_addc_u32 s9, s37, 0
6952 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
6953 ; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
6954 ; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
6955 ; GFX1032-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
6956 ; GFX1032-NEXT: buffer_store_dword v1, off, s[48:51], 0
6957 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
6958 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
6959 ; GFX1032-NEXT: v_mov_b32_e32 v0, 8
6960 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0
6961 ; GFX1032-NEXT: v_mov_b32_e32 v2, s44
6962 ; GFX1032-NEXT: v_mov_b32_e32 v5, 8
6963 ; GFX1032-NEXT: v_mov_b32_e32 v6, 0
6964 ; GFX1032-NEXT: v_mov_b32_e32 v7, 0
6965 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
6966 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
6967 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
6968 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
6969 ; GFX1032-NEXT: s_mov_b32 s12, s43
6970 ; GFX1032-NEXT: s_mov_b32 s13, s42
6971 ; GFX1032-NEXT: s_mov_b32 s14, s33
6972 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
6973 ; GFX1032-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
6974 ; GFX1032-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
6975 ; GFX1032-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
6976 ; GFX1032-NEXT: v_mov_b32_e32 v3, s45
6977 ; GFX1032-NEXT: v_mov_b32_e32 v4, 0
6978 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
6979 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
6980 ; GFX1032-NEXT: s_clause 0x1
6981 ; GFX1032-NEXT: buffer_load_dword v1, off, s[48:51], 0
6982 ; GFX1032-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
6983 ; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
6984 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
6985 ; GFX1032-NEXT: s_or_b32 s46, vcc_lo, s46
6986 ; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
6987 ; GFX1032-NEXT: s_cbranch_execnz .LBB10_2
6988 ; GFX1032-NEXT: .LBB10_3:
6989 ; GFX1032-NEXT: s_endpgm
6991 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
6993 ; GFX1164-NEXT: v_mov_b32_e32 v40, v0
6994 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
6995 ; GFX1164-NEXT: s_mov_b64 s[40:41], s[0:1]
6996 ; GFX1164-NEXT: s_mov_b32 s32, 32
6997 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
6998 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
6999 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
7000 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
7001 ; GFX1164-NEXT: s_cbranch_execz .LBB10_3
7002 ; GFX1164-NEXT: ; %bb.1:
7003 ; GFX1164-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
7004 ; GFX1164-NEXT: s_mov_b32 s33, s10
7005 ; GFX1164-NEXT: s_mov_b32 s42, s9
7006 ; GFX1164-NEXT: s_mov_b32 s43, s8
7007 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
7008 ; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
7009 ; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
7010 ; GFX1164-NEXT: s_mov_b64 s[46:47], 0
7011 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
7012 ; GFX1164-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
7013 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
7014 ; GFX1164-NEXT: v_mov_b32_e32 v2, s1
7015 ; GFX1164-NEXT: v_mov_b32_e32 v1, s0
7016 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
7017 ; GFX1164-NEXT: .p2align 6
7018 ; GFX1164-NEXT: .LBB10_2: ; %atomicrmw.start
7019 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
7020 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
7021 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
7022 ; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7023 ; GFX1164-NEXT: s_add_u32 s8, s36, 44
7024 ; GFX1164-NEXT: s_addc_u32 s9, s37, 0
7025 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
7026 ; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7027 ; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7028 ; GFX1164-NEXT: v_mov_b32_e32 v31, v40
7029 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
7030 ; GFX1164-NEXT: v_mov_b32_e32 v0, 8
7031 ; GFX1164-NEXT: v_mov_b32_e32 v5, 8
7032 ; GFX1164-NEXT: v_mov_b32_e32 v6, 0
7033 ; GFX1164-NEXT: v_mov_b32_e32 v7, 0
7034 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
7035 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
7036 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
7037 ; GFX1164-NEXT: s_mov_b32 s12, s43
7038 ; GFX1164-NEXT: s_mov_b32 s13, s42
7039 ; GFX1164-NEXT: s_mov_b32 s14, s33
7040 ; GFX1164-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7041 ; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
7042 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0
7043 ; GFX1164-NEXT: v_mov_b32_e32 v2, s44
7044 ; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
7045 ; GFX1164-NEXT: v_mov_b32_e32 v3, s45
7046 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
7047 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
7048 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
7049 ; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
7050 ; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
7051 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
7052 ; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7053 ; GFX1164-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7054 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[46:47]
7055 ; GFX1164-NEXT: s_cbranch_execnz .LBB10_2
7056 ; GFX1164-NEXT: .LBB10_3:
7057 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
7058 ; GFX1164-NEXT: s_endpgm
7060 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7062 ; GFX1132-NEXT: v_mov_b32_e32 v40, v0
7063 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7064 ; GFX1132-NEXT: s_mov_b64 s[40:41], s[0:1]
7065 ; GFX1132-NEXT: s_mov_b32 s46, 0
7066 ; GFX1132-NEXT: s_mov_b32 s32, 32
7067 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
7068 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
7069 ; GFX1132-NEXT: s_cbranch_execz .LBB10_3
7070 ; GFX1132-NEXT: ; %bb.1:
7071 ; GFX1132-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
7072 ; GFX1132-NEXT: s_mov_b32 s33, s15
7073 ; GFX1132-NEXT: s_mov_b32 s42, s14
7074 ; GFX1132-NEXT: s_mov_b32 s43, s13
7075 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
7076 ; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
7077 ; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
7078 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
7079 ; GFX1132-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
7080 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
7081 ; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
7082 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
7083 ; GFX1132-NEXT: .p2align 6
7084 ; GFX1132-NEXT: .LBB10_2: ; %atomicrmw.start
7085 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
7086 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
7087 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
7088 ; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7089 ; GFX1132-NEXT: s_add_u32 s8, s36, 44
7090 ; GFX1132-NEXT: s_addc_u32 s9, s37, 0
7091 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
7092 ; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7093 ; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7094 ; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
7095 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
7096 ; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
7097 ; GFX1132-NEXT: v_mov_b32_e32 v7, 0
7098 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
7099 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
7100 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
7101 ; GFX1132-NEXT: s_mov_b32 s12, s43
7102 ; GFX1132-NEXT: s_mov_b32 s13, s42
7103 ; GFX1132-NEXT: s_mov_b32 s14, s33
7104 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
7105 ; GFX1132-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7106 ; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
7107 ; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
7108 ; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
7109 ; GFX1132-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
7110 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
7111 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
7112 ; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
7113 ; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
7114 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
7115 ; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
7116 ; GFX1132-NEXT: s_or_b32 s46, vcc_lo, s46
7117 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
7118 ; GFX1132-NEXT: s_cbranch_execnz .LBB10_2
7119 ; GFX1132-NEXT: .LBB10_3:
7120 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
7121 ; GFX1132-NEXT: s_endpgm
7123 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7124 ; GFX7LESS-DPP: ; %bb.0:
7125 ; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
7126 ; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7127 ; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7128 ; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
7129 ; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
7130 ; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
7131 ; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
7132 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7133 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
7134 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
7135 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
7136 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
7137 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB10_3
7138 ; GFX7LESS-DPP-NEXT: ; %bb.1:
7139 ; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
7140 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, s9
7141 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, s8
7142 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7143 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7144 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7145 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
7146 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
7147 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
7148 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[46:47], 0
7149 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
7150 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
7151 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
7152 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
7153 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
7154 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
7155 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
7156 ; GFX7LESS-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7157 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7158 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
7159 ; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
7160 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
7161 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0
7162 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
7163 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
7164 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
7165 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
7166 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
7167 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7168 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7169 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
7170 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
7171 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7172 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
7173 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
7174 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
7175 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
7176 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
7177 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
7178 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
7179 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7180 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7181 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7182 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
7183 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
7184 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
7185 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
7186 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
7187 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
7188 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s44
7189 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s45
7190 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
7191 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
7192 ; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
7193 ; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[48:51], 0
7194 ; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
7195 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
7196 ; GFX7LESS-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7197 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
7198 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_2
7199 ; GFX7LESS-DPP-NEXT: .LBB10_3:
7200 ; GFX7LESS-DPP-NEXT: s_endpgm
7202 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7203 ; GFX9-DPP: ; %bb.0:
7204 ; GFX9-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7205 ; GFX9-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7206 ; GFX9-DPP-NEXT: s_mov_b32 s50, -1
7207 ; GFX9-DPP-NEXT: s_mov_b32 s51, 0xe00000
7208 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
7209 ; GFX9-DPP-NEXT: s_add_u32 s48, s48, s11
7210 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
7211 ; GFX9-DPP-NEXT: s_addc_u32 s49, s49, 0
7212 ; GFX9-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7213 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
7214 ; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
7215 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
7216 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB10_3
7217 ; GFX9-DPP-NEXT: ; %bb.1:
7218 ; GFX9-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
7219 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
7220 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
7221 ; GFX9-DPP-NEXT: s_mov_b32 s33, s10
7222 ; GFX9-DPP-NEXT: s_mov_b32 s42, s9
7223 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
7224 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
7225 ; GFX9-DPP-NEXT: s_mov_b32 s43, s8
7226 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7227 ; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7228 ; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7229 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
7230 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
7231 ; GFX9-DPP-NEXT: s_mov_b64 s[46:47], 0
7232 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
7233 ; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
7234 ; GFX9-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7235 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7236 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
7237 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7238 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
7239 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
7240 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
7241 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7242 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7243 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7244 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
7245 ; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
7246 ; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
7247 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7248 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7249 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7250 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7251 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
7252 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
7253 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
7254 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
7255 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
7256 ; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
7257 ; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
7258 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
7259 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
7260 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s44
7261 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s45
7262 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
7263 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
7264 ; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
7265 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
7266 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
7267 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
7268 ; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
7269 ; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
7270 ; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
7271 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7272 ; GFX9-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7273 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
7274 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB10_2
7275 ; GFX9-DPP-NEXT: .LBB10_3:
7276 ; GFX9-DPP-NEXT: s_endpgm
7278 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7279 ; GFX1064-DPP: ; %bb.0:
7280 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
7281 ; GFX1064-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7282 ; GFX1064-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7283 ; GFX1064-DPP-NEXT: s_mov_b32 s50, -1
7284 ; GFX1064-DPP-NEXT: s_mov_b32 s51, 0x31e16000
7285 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
7286 ; GFX1064-DPP-NEXT: s_add_u32 s48, s48, s11
7287 ; GFX1064-DPP-NEXT: s_addc_u32 s49, s49, 0
7288 ; GFX1064-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7289 ; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
7290 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
7291 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
7292 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_3
7293 ; GFX1064-DPP-NEXT: ; %bb.1:
7294 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
7295 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
7296 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
7297 ; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
7298 ; GFX1064-DPP-NEXT: s_mov_b32 s42, s9
7299 ; GFX1064-DPP-NEXT: s_mov_b32 s43, s8
7300 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7301 ; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
7302 ; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7303 ; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7304 ; GFX1064-DPP-NEXT: s_mov_b64 s[46:47], 0
7305 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
7306 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
7307 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
7308 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
7309 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
7310 ; GFX1064-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7311 ; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7312 ; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
7313 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7314 ; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
7315 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
7316 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
7317 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7318 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7319 ; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
7320 ; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
7321 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7322 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
7323 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
7324 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
7325 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s44
7326 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
7327 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
7328 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
7329 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
7330 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7331 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7332 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7333 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
7334 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
7335 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
7336 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
7337 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7338 ; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
7339 ; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
7340 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s45
7341 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
7342 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
7343 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
7344 ; GFX1064-DPP-NEXT: s_clause 0x1
7345 ; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
7346 ; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
7347 ; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
7348 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7349 ; GFX1064-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7350 ; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
7351 ; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB10_2
7352 ; GFX1064-DPP-NEXT: .LBB10_3:
7353 ; GFX1064-DPP-NEXT: s_endpgm
7355 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7356 ; GFX1032-DPP: ; %bb.0:
7357 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
7358 ; GFX1032-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7359 ; GFX1032-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7360 ; GFX1032-DPP-NEXT: s_mov_b32 s50, -1
7361 ; GFX1032-DPP-NEXT: s_mov_b32 s51, 0x31c16000
7362 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
7363 ; GFX1032-DPP-NEXT: s_add_u32 s48, s48, s11
7364 ; GFX1032-DPP-NEXT: s_addc_u32 s49, s49, 0
7365 ; GFX1032-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7366 ; GFX1032-DPP-NEXT: s_mov_b32 s46, 0
7367 ; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
7368 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
7369 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_3
7370 ; GFX1032-DPP-NEXT: ; %bb.1:
7371 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x24
7372 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
7373 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
7374 ; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
7375 ; GFX1032-DPP-NEXT: s_mov_b32 s42, s9
7376 ; GFX1032-DPP-NEXT: s_mov_b32 s43, s8
7377 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7378 ; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
7379 ; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7380 ; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7381 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
7382 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[44:45], 0x0
7383 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
7384 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
7385 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
7386 ; GFX1032-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7387 ; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7388 ; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
7389 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7390 ; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
7391 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
7392 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
7393 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7394 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7395 ; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
7396 ; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
7397 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7398 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
7399 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
7400 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
7401 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s44
7402 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
7403 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
7404 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
7405 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
7406 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7407 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7408 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7409 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
7410 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
7411 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
7412 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
7413 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7414 ; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
7415 ; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
7416 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s45
7417 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
7418 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
7419 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
7420 ; GFX1032-DPP-NEXT: s_clause 0x1
7421 ; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
7422 ; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
7423 ; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
7424 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
7425 ; GFX1032-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
7426 ; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
7427 ; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB10_2
7428 ; GFX1032-DPP-NEXT: .LBB10_3:
7429 ; GFX1032-DPP-NEXT: s_endpgm
7431 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7432 ; GFX1164-DPP: ; %bb.0:
7433 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
7434 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7435 ; GFX1164-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7436 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
7437 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
7438 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
7439 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
7440 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
7441 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB10_3
7442 ; GFX1164-DPP-NEXT: ; %bb.1:
7443 ; GFX1164-DPP-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
7444 ; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
7445 ; GFX1164-DPP-NEXT: s_mov_b32 s42, s9
7446 ; GFX1164-DPP-NEXT: s_mov_b32 s43, s8
7447 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7448 ; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7449 ; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7450 ; GFX1164-DPP-NEXT: s_mov_b64 s[46:47], 0
7451 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
7452 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
7453 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
7454 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
7455 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
7456 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
7457 ; GFX1164-DPP-NEXT: .p2align 6
7458 ; GFX1164-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7459 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7460 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
7461 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
7462 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7463 ; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
7464 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
7465 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
7466 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7467 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7468 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
7469 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
7470 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
7471 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
7472 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
7473 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
7474 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7475 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7476 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7477 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
7478 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
7479 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
7480 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7481 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
7482 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
7483 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s44
7484 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
7485 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s45
7486 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
7487 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
7488 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
7489 ; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
7490 ; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
7491 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
7492 ; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7493 ; GFX1164-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7494 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[46:47]
7495 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB10_2
7496 ; GFX1164-DPP-NEXT: .LBB10_3:
7497 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
7498 ; GFX1164-DPP-NEXT: s_endpgm
7500 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
7501 ; GFX1132-DPP: ; %bb.0:
7502 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
7503 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7504 ; GFX1132-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
7505 ; GFX1132-DPP-NEXT: s_mov_b32 s46, 0
7506 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
7507 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
7508 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
7509 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB10_3
7510 ; GFX1132-DPP-NEXT: ; %bb.1:
7511 ; GFX1132-DPP-NEXT: s_load_b64 s[44:45], s[4:5], 0x24
7512 ; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
7513 ; GFX1132-DPP-NEXT: s_mov_b32 s42, s14
7514 ; GFX1132-DPP-NEXT: s_mov_b32 s43, s13
7515 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
7516 ; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
7517 ; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
7518 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
7519 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[44:45], 0x0
7520 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
7521 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
7522 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
7523 ; GFX1132-DPP-NEXT: .p2align 6
7524 ; GFX1132-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
7525 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
7526 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
7527 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
7528 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
7529 ; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
7530 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
7531 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
7532 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7533 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7534 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
7535 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
7536 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
7537 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
7538 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
7539 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
7540 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
7541 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s43
7542 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s42
7543 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
7544 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
7545 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
7546 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
7547 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
7548 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
7549 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
7550 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
7551 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
7552 ; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
7553 ; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
7554 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
7555 ; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
7556 ; GFX1132-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
7557 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
7558 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB10_2
7559 ; GFX1132-DPP-NEXT: .LBB10_3:
7560 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
7561 ; GFX1132-DPP-NEXT: s_endpgm
7562 %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1
7566 define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
7567 ; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
7568 ; GFX7LESS: ; %bb.0:
7569 ; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
7570 ; GFX7LESS-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7571 ; GFX7LESS-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7572 ; GFX7LESS-NEXT: s_mov_b32 s50, -1
7573 ; GFX7LESS-NEXT: s_mov_b32 s51, 0xe8f000
7574 ; GFX7LESS-NEXT: s_add_u32 s48, s48, s11
7575 ; GFX7LESS-NEXT: s_addc_u32 s49, s49, 0
7576 ; GFX7LESS-NEXT: s_mov_b32 s33, s10
7577 ; GFX7LESS-NEXT: s_mov_b32 s42, s9
7578 ; GFX7LESS-NEXT: s_mov_b32 s43, s8
7579 ; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
7580 ; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
7581 ; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
7582 ; GFX7LESS-NEXT: s_mov_b64 s[40:41], s[0:1]
7583 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
7584 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
7585 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
7586 ; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
7587 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
7588 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7589 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
7590 ; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
7591 ; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
7592 ; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
7593 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
7594 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
7595 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
7596 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
7597 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
7598 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
7599 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
7600 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
7601 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
7602 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
7603 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
7604 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
7605 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
7606 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
7607 ; GFX7LESS-NEXT: .LBB11_1: ; %ComputeLoop
7608 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
7609 ; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
7610 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
7611 ; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
7612 ; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
7613 ; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
7614 ; GFX7LESS-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
7615 ; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
7616 ; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
7617 ; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
7618 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
7619 ; GFX7LESS-NEXT: s_cbranch_vccnz .LBB11_1
7620 ; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
7621 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
7622 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
7623 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
7624 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
7625 ; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
7626 ; GFX7LESS-NEXT: s_cbranch_execz .LBB11_5
7627 ; GFX7LESS-NEXT: ; %bb.3:
7628 ; GFX7LESS-NEXT: s_load_dwordx2 s[44:45], s[36:37], 0x9
7629 ; GFX7LESS-NEXT: s_mov_b32 s47, 0xf000
7630 ; GFX7LESS-NEXT: s_mov_b32 s46, -1
7631 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
7632 ; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[44:47], 0
7633 ; GFX7LESS-NEXT: s_mov_b64 s[46:47], 0
7634 ; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
7635 ; GFX7LESS-NEXT: .LBB11_4: ; %atomicrmw.start
7636 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
7637 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
7638 ; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
7639 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:4
7640 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0
7641 ; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
7642 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
7643 ; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[41:42]
7644 ; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
7645 ; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
7646 ; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7647 ; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7648 ; GFX7LESS-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
7649 ; GFX7LESS-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
7650 ; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7651 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
7652 ; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
7653 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
7654 ; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
7655 ; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
7656 ; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
7657 ; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
7658 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[40:41]
7659 ; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
7660 ; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
7661 ; GFX7LESS-NEXT: s_mov_b32 s12, s43
7662 ; GFX7LESS-NEXT: s_mov_b32 s13, s42
7663 ; GFX7LESS-NEXT: s_mov_b32 s14, s33
7664 ; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
7665 ; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[48:49]
7666 ; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[50:51]
7667 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, s44
7668 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, s45
7669 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
7670 ; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
7671 ; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
7672 ; GFX7LESS-NEXT: buffer_load_dword v0, off, s[48:51], 0
7673 ; GFX7LESS-NEXT: buffer_load_dword v1, off, s[48:51], 0 offset:4
7674 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
7675 ; GFX7LESS-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7676 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[46:47]
7677 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB11_4
7678 ; GFX7LESS-NEXT: .LBB11_5:
7679 ; GFX7LESS-NEXT: s_endpgm
7681 ; GFX9-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
7683 ; GFX9-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7684 ; GFX9-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7685 ; GFX9-NEXT: s_mov_b32 s50, -1
7686 ; GFX9-NEXT: s_mov_b32 s51, 0xe00000
7687 ; GFX9-NEXT: s_add_u32 s48, s48, s11
7688 ; GFX9-NEXT: s_addc_u32 s49, s49, 0
7689 ; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
7690 ; GFX9-NEXT: s_mov_b32 s43, s8
7691 ; GFX9-NEXT: s_add_u32 s8, s36, 44
7692 ; GFX9-NEXT: s_mov_b32 s42, s9
7693 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
7694 ; GFX9-NEXT: s_mov_b64 s[40:41], s[0:1]
7695 ; GFX9-NEXT: s_getpc_b64 s[0:1]
7696 ; GFX9-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
7697 ; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
7698 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7699 ; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
7700 ; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
7701 ; GFX9-NEXT: s_mov_b32 s33, s10
7702 ; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
7703 ; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
7704 ; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
7705 ; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
7706 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
7707 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
7708 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
7709 ; GFX9-NEXT: s_mov_b32 s12, s43
7710 ; GFX9-NEXT: s_mov_b32 s13, s42
7711 ; GFX9-NEXT: s_mov_b32 s14, s33
7712 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
7713 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
7714 ; GFX9-NEXT: s_movk_i32 s32, 0x800
7715 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
7716 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
7717 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
7718 ; GFX9-NEXT: s_mov_b64 s[0:1], exec
7719 ; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
7720 ; GFX9-NEXT: .LBB11_1: ; %ComputeLoop
7721 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
7722 ; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
7723 ; GFX9-NEXT: v_readlane_b32 s3, v1, s4
7724 ; GFX9-NEXT: v_readlane_b32 s2, v0, s4
7725 ; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
7726 ; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
7727 ; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
7728 ; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
7729 ; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
7730 ; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
7731 ; GFX9-NEXT: s_cbranch_scc1 .LBB11_1
7732 ; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
7733 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7734 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
7735 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
7736 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
7737 ; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
7738 ; GFX9-NEXT: s_cbranch_execz .LBB11_5
7739 ; GFX9-NEXT: ; %bb.3:
7740 ; GFX9-NEXT: s_load_dwordx2 s[44:45], s[36:37], 0x24
7741 ; GFX9-NEXT: v_mov_b32_e32 v0, 0
7742 ; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
7743 ; GFX9-NEXT: s_mov_b64 s[46:47], 0
7744 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
7745 ; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
7746 ; GFX9-NEXT: .LBB11_4: ; %atomicrmw.start
7747 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
7748 ; GFX9-NEXT: s_waitcnt vmcnt(0)
7749 ; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
7750 ; GFX9-NEXT: s_add_u32 s8, s36, 44
7751 ; GFX9-NEXT: s_addc_u32 s9, s37, 0
7752 ; GFX9-NEXT: s_getpc_b64 s[0:1]
7753 ; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7754 ; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7755 ; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7756 ; GFX9-NEXT: s_mov_b64 s[0:1], s[48:49]
7757 ; GFX9-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
7758 ; GFX9-NEXT: buffer_store_dword v4, off, s[48:51], 0
7759 ; GFX9-NEXT: s_mov_b64 s[4:5], s[40:41]
7760 ; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
7761 ; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
7762 ; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
7763 ; GFX9-NEXT: s_mov_b32 s12, s43
7764 ; GFX9-NEXT: s_mov_b32 s13, s42
7765 ; GFX9-NEXT: s_mov_b32 s14, s33
7766 ; GFX9-NEXT: v_mov_b32_e32 v31, v40
7767 ; GFX9-NEXT: s_mov_b64 s[2:3], s[50:51]
7768 ; GFX9-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
7769 ; GFX9-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
7770 ; GFX9-NEXT: v_mov_b32_e32 v0, 8
7771 ; GFX9-NEXT: v_mov_b32_e32 v1, 0
7772 ; GFX9-NEXT: v_mov_b32_e32 v2, s44
7773 ; GFX9-NEXT: v_mov_b32_e32 v3, s45
7774 ; GFX9-NEXT: v_mov_b32_e32 v4, 0
7775 ; GFX9-NEXT: v_mov_b32_e32 v5, 8
7776 ; GFX9-NEXT: v_mov_b32_e32 v6, 0
7777 ; GFX9-NEXT: v_mov_b32_e32 v7, 0
7778 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
7779 ; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
7780 ; GFX9-NEXT: buffer_load_dword v4, off, s[48:51], 0
7781 ; GFX9-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
7782 ; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
7783 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7784 ; GFX9-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7785 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[46:47]
7786 ; GFX9-NEXT: s_cbranch_execnz .LBB11_4
7787 ; GFX9-NEXT: .LBB11_5:
7788 ; GFX9-NEXT: s_endpgm
7790 ; GFX1064-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
7792 ; GFX1064-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7793 ; GFX1064-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7794 ; GFX1064-NEXT: s_mov_b32 s50, -1
7795 ; GFX1064-NEXT: s_mov_b32 s51, 0x31e16000
7796 ; GFX1064-NEXT: s_add_u32 s48, s48, s11
7797 ; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
7798 ; GFX1064-NEXT: s_addc_u32 s49, s49, 0
7799 ; GFX1064-NEXT: s_mov_b32 s43, s8
7800 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
7801 ; GFX1064-NEXT: s_mov_b32 s42, s9
7802 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
7803 ; GFX1064-NEXT: s_mov_b64 s[40:41], s[0:1]
7804 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
7805 ; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
7806 ; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
7807 ; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
7808 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7809 ; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
7810 ; GFX1064-NEXT: s_mov_b32 s33, s10
7811 ; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
7812 ; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
7813 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
7814 ; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
7815 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
7816 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
7817 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
7818 ; GFX1064-NEXT: s_mov_b32 s12, s43
7819 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
7820 ; GFX1064-NEXT: s_mov_b32 s13, s42
7821 ; GFX1064-NEXT: s_mov_b32 s14, s33
7822 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
7823 ; GFX1064-NEXT: s_movk_i32 s32, 0x800
7824 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
7825 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
7826 ; GFX1064-NEXT: v_mov_b32_e32 v2, 0
7827 ; GFX1064-NEXT: v_mov_b32_e32 v3, 0x7ff80000
7828 ; GFX1064-NEXT: s_mov_b64 s[0:1], exec
7829 ; GFX1064-NEXT: .LBB11_1: ; %ComputeLoop
7830 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
7831 ; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
7832 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
7833 ; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
7834 ; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
7835 ; GFX1064-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
7836 ; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
7837 ; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
7838 ; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
7839 ; GFX1064-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
7840 ; GFX1064-NEXT: s_cbranch_scc1 .LBB11_1
7841 ; GFX1064-NEXT: ; %bb.2: ; %ComputeEnd
7842 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7843 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
7844 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
7845 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
7846 ; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
7847 ; GFX1064-NEXT: s_cbranch_execz .LBB11_5
7848 ; GFX1064-NEXT: ; %bb.3:
7849 ; GFX1064-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
7850 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
7851 ; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
7852 ; GFX1064-NEXT: s_mov_b64 s[46:47], 0
7853 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
7854 ; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
7855 ; GFX1064-NEXT: .LBB11_4: ; %atomicrmw.start
7856 ; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
7857 ; GFX1064-NEXT: s_waitcnt vmcnt(0)
7858 ; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
7859 ; GFX1064-NEXT: s_add_u32 s8, s34, 44
7860 ; GFX1064-NEXT: s_addc_u32 s9, s35, 0
7861 ; GFX1064-NEXT: s_getpc_b64 s[0:1]
7862 ; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7863 ; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7864 ; GFX1064-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
7865 ; GFX1064-NEXT: buffer_store_dword v4, off, s[48:51], 0
7866 ; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7867 ; GFX1064-NEXT: v_mov_b32_e32 v31, v40
7868 ; GFX1064-NEXT: v_mov_b32_e32 v2, s44
7869 ; GFX1064-NEXT: v_mov_b32_e32 v3, s45
7870 ; GFX1064-NEXT: v_mov_b32_e32 v4, 0
7871 ; GFX1064-NEXT: v_mov_b32_e32 v5, 8
7872 ; GFX1064-NEXT: v_mov_b32_e32 v6, 0
7873 ; GFX1064-NEXT: v_mov_b32_e32 v7, 0
7874 ; GFX1064-NEXT: s_mov_b64 s[0:1], s[48:49]
7875 ; GFX1064-NEXT: s_mov_b64 s[4:5], s[40:41]
7876 ; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
7877 ; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
7878 ; GFX1064-NEXT: s_mov_b32 s12, s43
7879 ; GFX1064-NEXT: s_mov_b32 s13, s42
7880 ; GFX1064-NEXT: s_mov_b32 s14, s33
7881 ; GFX1064-NEXT: s_mov_b64 s[2:3], s[50:51]
7882 ; GFX1064-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
7883 ; GFX1064-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
7884 ; GFX1064-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
7885 ; GFX1064-NEXT: v_mov_b32_e32 v0, 8
7886 ; GFX1064-NEXT: v_mov_b32_e32 v1, 0
7887 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
7888 ; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
7889 ; GFX1064-NEXT: s_clause 0x1
7890 ; GFX1064-NEXT: buffer_load_dword v4, off, s[48:51], 0
7891 ; GFX1064-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
7892 ; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
7893 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
7894 ; GFX1064-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
7895 ; GFX1064-NEXT: s_andn2_b64 exec, exec, s[46:47]
7896 ; GFX1064-NEXT: s_cbranch_execnz .LBB11_4
7897 ; GFX1064-NEXT: .LBB11_5:
7898 ; GFX1064-NEXT: s_endpgm
7900 ; GFX1032-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
7902 ; GFX1032-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
7903 ; GFX1032-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
7904 ; GFX1032-NEXT: s_mov_b32 s50, -1
7905 ; GFX1032-NEXT: s_mov_b32 s51, 0x31c16000
7906 ; GFX1032-NEXT: s_add_u32 s48, s48, s11
7907 ; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
7908 ; GFX1032-NEXT: s_addc_u32 s49, s49, 0
7909 ; GFX1032-NEXT: s_mov_b32 s43, s8
7910 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
7911 ; GFX1032-NEXT: s_mov_b32 s42, s9
7912 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
7913 ; GFX1032-NEXT: s_mov_b64 s[40:41], s[0:1]
7914 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
7915 ; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
7916 ; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
7917 ; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
7918 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7919 ; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
7920 ; GFX1032-NEXT: s_mov_b32 s33, s10
7921 ; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
7922 ; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
7923 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
7924 ; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
7925 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
7926 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
7927 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
7928 ; GFX1032-NEXT: s_mov_b32 s12, s43
7929 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
7930 ; GFX1032-NEXT: s_mov_b32 s13, s42
7931 ; GFX1032-NEXT: s_mov_b32 s14, s33
7932 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
7933 ; GFX1032-NEXT: s_movk_i32 s32, 0x400
7934 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
7935 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
7936 ; GFX1032-NEXT: v_mov_b32_e32 v2, 0
7937 ; GFX1032-NEXT: v_mov_b32_e32 v3, 0x7ff80000
7938 ; GFX1032-NEXT: s_mov_b32 s0, exec_lo
7939 ; GFX1032-NEXT: .LBB11_1: ; %ComputeLoop
7940 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
7941 ; GFX1032-NEXT: s_ff1_i32_b32 s1, s0
7942 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
7943 ; GFX1032-NEXT: v_readlane_b32 s3, v1, s1
7944 ; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
7945 ; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
7946 ; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
7947 ; GFX1032-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
7948 ; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
7949 ; GFX1032-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
7950 ; GFX1032-NEXT: s_cbranch_scc1 .LBB11_1
7951 ; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
7952 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
7953 ; GFX1032-NEXT: s_mov_b32 s46, 0
7954 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
7955 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
7956 ; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
7957 ; GFX1032-NEXT: s_cbranch_execz .LBB11_5
7958 ; GFX1032-NEXT: ; %bb.3:
7959 ; GFX1032-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
7960 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
7961 ; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
7962 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
7963 ; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[44:45]
7964 ; GFX1032-NEXT: .LBB11_4: ; %atomicrmw.start
7965 ; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
7966 ; GFX1032-NEXT: s_waitcnt vmcnt(0)
7967 ; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
7968 ; GFX1032-NEXT: s_add_u32 s8, s34, 44
7969 ; GFX1032-NEXT: s_addc_u32 s9, s35, 0
7970 ; GFX1032-NEXT: s_getpc_b64 s[0:1]
7971 ; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
7972 ; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
7973 ; GFX1032-NEXT: buffer_store_dword v5, off, s[48:51], 0 offset:4
7974 ; GFX1032-NEXT: buffer_store_dword v4, off, s[48:51], 0
7975 ; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
7976 ; GFX1032-NEXT: v_mov_b32_e32 v31, v40
7977 ; GFX1032-NEXT: v_mov_b32_e32 v2, s44
7978 ; GFX1032-NEXT: v_mov_b32_e32 v3, s45
7979 ; GFX1032-NEXT: v_mov_b32_e32 v4, 0
7980 ; GFX1032-NEXT: v_mov_b32_e32 v5, 8
7981 ; GFX1032-NEXT: v_mov_b32_e32 v6, 0
7982 ; GFX1032-NEXT: v_mov_b32_e32 v7, 0
7983 ; GFX1032-NEXT: s_mov_b64 s[0:1], s[48:49]
7984 ; GFX1032-NEXT: s_mov_b64 s[4:5], s[40:41]
7985 ; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
7986 ; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
7987 ; GFX1032-NEXT: s_mov_b32 s12, s43
7988 ; GFX1032-NEXT: s_mov_b32 s13, s42
7989 ; GFX1032-NEXT: s_mov_b32 s14, s33
7990 ; GFX1032-NEXT: s_mov_b64 s[2:3], s[50:51]
7991 ; GFX1032-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
7992 ; GFX1032-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
7993 ; GFX1032-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
7994 ; GFX1032-NEXT: v_mov_b32_e32 v0, 8
7995 ; GFX1032-NEXT: v_mov_b32_e32 v1, 0
7996 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
7997 ; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
7998 ; GFX1032-NEXT: s_clause 0x1
7999 ; GFX1032-NEXT: buffer_load_dword v4, off, s[48:51], 0
8000 ; GFX1032-NEXT: buffer_load_dword v5, off, s[48:51], 0 offset:4
8001 ; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
8002 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
8003 ; GFX1032-NEXT: s_or_b32 s46, vcc_lo, s46
8004 ; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
8005 ; GFX1032-NEXT: s_cbranch_execnz .LBB11_4
8006 ; GFX1032-NEXT: .LBB11_5:
8007 ; GFX1032-NEXT: s_endpgm
8009 ; GFX1164-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8011 ; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
8012 ; GFX1164-NEXT: s_mov_b32 s43, s8
8013 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
8014 ; GFX1164-NEXT: s_mov_b32 s42, s9
8015 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
8016 ; GFX1164-NEXT: s_mov_b64 s[40:41], s[0:1]
8017 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
8018 ; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8019 ; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8020 ; GFX1164-NEXT: v_mov_b32_e32 v31, v0
8021 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8022 ; GFX1164-NEXT: s_mov_b32 s33, s10
8023 ; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
8024 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
8025 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
8026 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
8027 ; GFX1164-NEXT: s_mov_b32 s12, s43
8028 ; GFX1164-NEXT: s_mov_b32 s13, s42
8029 ; GFX1164-NEXT: s_mov_b32 s14, s33
8030 ; GFX1164-NEXT: s_mov_b32 s32, 32
8031 ; GFX1164-NEXT: v_mov_b32_e32 v40, v0
8032 ; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
8033 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
8034 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
8035 ; GFX1164-NEXT: v_mov_b32_e32 v2, 0
8036 ; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
8037 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
8038 ; GFX1164-NEXT: .LBB11_1: ; %ComputeLoop
8039 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
8040 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8041 ; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
8042 ; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
8043 ; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
8044 ; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
8045 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
8046 ; GFX1164-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
8047 ; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
8048 ; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
8049 ; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8050 ; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
8051 ; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
8052 ; GFX1164-NEXT: s_cbranch_scc1 .LBB11_1
8053 ; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
8054 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
8055 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
8056 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8057 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
8058 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
8059 ; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
8060 ; GFX1164-NEXT: s_cbranch_execz .LBB11_5
8061 ; GFX1164-NEXT: ; %bb.3:
8062 ; GFX1164-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
8063 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
8064 ; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
8065 ; GFX1164-NEXT: s_mov_b64 s[46:47], 0
8066 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
8067 ; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[44:45]
8068 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
8069 ; GFX1164-NEXT: .p2align 6
8070 ; GFX1164-NEXT: .LBB11_4: ; %atomicrmw.start
8071 ; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
8072 ; GFX1164-NEXT: s_waitcnt vmcnt(0)
8073 ; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
8074 ; GFX1164-NEXT: s_add_u32 s8, s34, 44
8075 ; GFX1164-NEXT: s_addc_u32 s9, s35, 0
8076 ; GFX1164-NEXT: s_getpc_b64 s[0:1]
8077 ; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8078 ; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8079 ; GFX1164-NEXT: v_mov_b32_e32 v31, v40
8080 ; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8081 ; GFX1164-NEXT: v_mov_b32_e32 v2, s44
8082 ; GFX1164-NEXT: v_mov_b32_e32 v3, s45
8083 ; GFX1164-NEXT: v_mov_b32_e32 v6, 0
8084 ; GFX1164-NEXT: v_mov_b32_e32 v7, 0
8085 ; GFX1164-NEXT: s_mov_b64 s[4:5], s[40:41]
8086 ; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
8087 ; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
8088 ; GFX1164-NEXT: s_mov_b32 s12, s43
8089 ; GFX1164-NEXT: s_mov_b32 s13, s42
8090 ; GFX1164-NEXT: s_mov_b32 s14, s33
8091 ; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
8092 ; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
8093 ; GFX1164-NEXT: v_mov_b32_e32 v4, 0
8094 ; GFX1164-NEXT: v_mov_b32_e32 v5, 8
8095 ; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
8096 ; GFX1164-NEXT: v_mov_b32_e32 v0, 8
8097 ; GFX1164-NEXT: v_mov_b32_e32 v1, 0
8098 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
8099 ; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
8100 ; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
8101 ; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
8102 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
8103 ; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
8104 ; GFX1164-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
8105 ; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[46:47]
8106 ; GFX1164-NEXT: s_cbranch_execnz .LBB11_4
8107 ; GFX1164-NEXT: .LBB11_5:
8108 ; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
8109 ; GFX1164-NEXT: s_endpgm
8111 ; GFX1132-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8113 ; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
8114 ; GFX1132-NEXT: s_mov_b64 s[40:41], s[0:1]
8115 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
8116 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
8117 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
8118 ; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8119 ; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8120 ; GFX1132-NEXT: v_mov_b32_e32 v31, v0
8121 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8122 ; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
8123 ; GFX1132-NEXT: s_mov_b32 s42, s14
8124 ; GFX1132-NEXT: s_mov_b32 s43, s13
8125 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
8126 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
8127 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
8128 ; GFX1132-NEXT: s_mov_b32 s12, s13
8129 ; GFX1132-NEXT: s_mov_b32 s13, s14
8130 ; GFX1132-NEXT: s_mov_b32 s14, s15
8131 ; GFX1132-NEXT: s_mov_b32 s32, 32
8132 ; GFX1132-NEXT: s_mov_b32 s33, s15
8133 ; GFX1132-NEXT: v_mov_b32_e32 v40, v0
8134 ; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
8135 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
8136 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
8137 ; GFX1132-NEXT: v_mov_b32_e32 v2, 0
8138 ; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
8139 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
8140 ; GFX1132-NEXT: .LBB11_1: ; %ComputeLoop
8141 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
8142 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8143 ; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
8144 ; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
8145 ; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
8146 ; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
8147 ; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
8148 ; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8149 ; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
8150 ; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
8151 ; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
8152 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
8153 ; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
8154 ; GFX1132-NEXT: s_cbranch_scc1 .LBB11_1
8155 ; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
8156 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
8157 ; GFX1132-NEXT: s_mov_b32 s46, 0
8158 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
8159 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
8160 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
8161 ; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
8162 ; GFX1132-NEXT: s_cbranch_execz .LBB11_5
8163 ; GFX1132-NEXT: ; %bb.3:
8164 ; GFX1132-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
8165 ; GFX1132-NEXT: v_mov_b32_e32 v0, 0
8166 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
8167 ; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
8168 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
8169 ; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[44:45]
8170 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
8171 ; GFX1132-NEXT: .p2align 6
8172 ; GFX1132-NEXT: .LBB11_4: ; %atomicrmw.start
8173 ; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
8174 ; GFX1132-NEXT: s_waitcnt vmcnt(0)
8175 ; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
8176 ; GFX1132-NEXT: s_add_u32 s8, s34, 44
8177 ; GFX1132-NEXT: s_addc_u32 s9, s35, 0
8178 ; GFX1132-NEXT: s_getpc_b64 s[0:1]
8179 ; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8180 ; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8181 ; GFX1132-NEXT: v_mov_b32_e32 v31, v40
8182 ; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8183 ; GFX1132-NEXT: v_mov_b32_e32 v3, s45
8184 ; GFX1132-NEXT: v_mov_b32_e32 v7, 0
8185 ; GFX1132-NEXT: s_mov_b64 s[4:5], s[40:41]
8186 ; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
8187 ; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
8188 ; GFX1132-NEXT: s_mov_b32 s12, s43
8189 ; GFX1132-NEXT: s_mov_b32 s13, s42
8190 ; GFX1132-NEXT: s_mov_b32 s14, s33
8191 ; GFX1132-NEXT: v_mov_b32_e32 v6, 0
8192 ; GFX1132-NEXT: v_mov_b32_e32 v2, s44
8193 ; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
8194 ; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
8195 ; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
8196 ; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
8197 ; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
8198 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
8199 ; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
8200 ; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
8201 ; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
8202 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
8203 ; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
8204 ; GFX1132-NEXT: s_or_b32 s46, vcc_lo, s46
8205 ; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
8206 ; GFX1132-NEXT: s_cbranch_execnz .LBB11_4
8207 ; GFX1132-NEXT: .LBB11_5:
8208 ; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
8209 ; GFX1132-NEXT: s_endpgm
8211 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8212 ; GFX7LESS-DPP: ; %bb.0:
8213 ; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
8214 ; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
8215 ; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
8216 ; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
8217 ; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
8218 ; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
8219 ; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
8220 ; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
8221 ; GFX7LESS-DPP-NEXT: s_mov_b32 s42, s9
8222 ; GFX7LESS-DPP-NEXT: s_mov_b32 s43, s8
8223 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
8224 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
8225 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8226 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8227 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[44:45], s[4:5], 0x9
8228 ; GFX7LESS-DPP-NEXT: s_mov_b32 s47, 0xf000
8229 ; GFX7LESS-DPP-NEXT: s_mov_b32 s46, -1
8230 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
8231 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
8232 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
8233 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8234 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8235 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8236 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
8237 ; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
8238 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
8239 ; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
8240 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8241 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8242 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
8243 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
8244 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
8245 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
8246 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
8247 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8248 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8249 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
8250 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8251 ; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[44:47], 0
8252 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[46:47], 0
8253 ; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
8254 ; GFX7LESS-DPP-NEXT: .LBB11_1: ; %atomicrmw.start
8255 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8256 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
8257 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
8258 ; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:4
8259 ; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0
8260 ; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
8261 ; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
8262 ; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
8263 ; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
8264 ; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8265 ; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8266 ; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0 offset:12
8267 ; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[48:51], 0 offset:8
8268 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8269 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
8270 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
8271 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
8272 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
8273 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
8274 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
8275 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
8276 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8277 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8278 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
8279 ; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s43
8280 ; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s42
8281 ; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
8282 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
8283 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8284 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8285 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s44
8286 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s45
8287 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
8288 ; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8289 ; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8290 ; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0
8291 ; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[48:51], 0 offset:4
8292 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
8293 ; GFX7LESS-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
8294 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
8295 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB11_1
8296 ; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
8297 ; GFX7LESS-DPP-NEXT: s_endpgm
8299 ; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8300 ; GFX9-DPP: ; %bb.0:
8301 ; GFX9-DPP-NEXT: s_mov_b32 s52, SCRATCH_RSRC_DWORD0
8302 ; GFX9-DPP-NEXT: s_mov_b32 s53, SCRATCH_RSRC_DWORD1
8303 ; GFX9-DPP-NEXT: s_mov_b32 s54, -1
8304 ; GFX9-DPP-NEXT: s_mov_b32 s55, 0xe00000
8305 ; GFX9-DPP-NEXT: s_add_u32 s52, s52, s11
8306 ; GFX9-DPP-NEXT: s_addc_u32 s53, s53, 0
8307 ; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
8308 ; GFX9-DPP-NEXT: s_mov_b32 s43, s8
8309 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
8310 ; GFX9-DPP-NEXT: s_mov_b32 s42, s9
8311 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
8312 ; GFX9-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8313 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
8314 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8315 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8316 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8317 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
8318 ; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
8319 ; GFX9-DPP-NEXT: s_mov_b32 s33, s10
8320 ; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
8321 ; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8322 ; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
8323 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8324 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[52:53]
8325 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8326 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
8327 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
8328 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
8329 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
8330 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
8331 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[54:55]
8332 ; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
8333 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
8334 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8335 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
8336 ; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
8337 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
8338 ; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
8339 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
8340 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
8341 ; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
8342 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
8343 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
8344 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8345 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
8346 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
8347 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
8348 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
8349 ; GFX9-DPP-NEXT: s_nop 0
8350 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
8351 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
8352 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
8353 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
8354 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
8355 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
8356 ; GFX9-DPP-NEXT: s_nop 0
8357 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
8358 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
8359 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
8360 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
8361 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
8362 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
8363 ; GFX9-DPP-NEXT: s_nop 0
8364 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
8365 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
8366 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
8367 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
8368 ; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
8369 ; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
8370 ; GFX9-DPP-NEXT: s_nop 0
8371 ; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
8372 ; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
8373 ; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
8374 ; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
8375 ; GFX9-DPP-NEXT: s_nop 1
8376 ; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
8377 ; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
8378 ; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
8379 ; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8380 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
8381 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
8382 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
8383 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
8384 ; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
8385 ; GFX9-DPP-NEXT: v_readlane_b32 s45, v9, 63
8386 ; GFX9-DPP-NEXT: v_readlane_b32 s44, v8, 63
8387 ; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
8388 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
8389 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
8390 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB11_3
8391 ; GFX9-DPP-NEXT: ; %bb.1:
8392 ; GFX9-DPP-NEXT: s_load_dwordx2 s[46:47], s[36:37], 0x24
8393 ; GFX9-DPP-NEXT: s_mov_b64 s[48:49], 0
8394 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
8395 ; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[46:47]
8396 ; GFX9-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
8397 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8398 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[44:45], s[44:45]
8399 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
8400 ; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
8401 ; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
8402 ; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
8403 ; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
8404 ; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8405 ; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8406 ; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8407 ; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[52:53]
8408 ; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[52:55], 0 offset:4
8409 ; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[52:55], 0
8410 ; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8411 ; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
8412 ; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8413 ; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
8414 ; GFX9-DPP-NEXT: s_mov_b32 s12, s43
8415 ; GFX9-DPP-NEXT: s_mov_b32 s13, s42
8416 ; GFX9-DPP-NEXT: s_mov_b32 s14, s33
8417 ; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
8418 ; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[52:55], 0 offset:12
8419 ; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[52:55], 0 offset:8
8420 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[54:55]
8421 ; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
8422 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
8423 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s46
8424 ; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s47
8425 ; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
8426 ; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
8427 ; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
8428 ; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
8429 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
8430 ; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8431 ; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[52:55], 0
8432 ; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[52:55], 0 offset:4
8433 ; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8434 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
8435 ; GFX9-DPP-NEXT: s_or_b64 s[48:49], vcc, s[48:49]
8436 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[48:49]
8437 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB11_2
8438 ; GFX9-DPP-NEXT: .LBB11_3:
8439 ; GFX9-DPP-NEXT: s_endpgm
8441 ; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8442 ; GFX1064-DPP: ; %bb.0:
8443 ; GFX1064-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
8444 ; GFX1064-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
8445 ; GFX1064-DPP-NEXT: s_mov_b32 s50, -1
8446 ; GFX1064-DPP-NEXT: s_mov_b32 s51, 0x31e16000
8447 ; GFX1064-DPP-NEXT: s_add_u32 s48, s48, s11
8448 ; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
8449 ; GFX1064-DPP-NEXT: s_addc_u32 s49, s49, 0
8450 ; GFX1064-DPP-NEXT: s_mov_b32 s43, s8
8451 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
8452 ; GFX1064-DPP-NEXT: s_mov_b32 s42, s9
8453 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
8454 ; GFX1064-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8455 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
8456 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8457 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8458 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
8459 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8460 ; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
8461 ; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
8462 ; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
8463 ; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8464 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8465 ; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
8466 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8467 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8468 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8469 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
8470 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
8471 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
8472 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
8473 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8474 ; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
8475 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
8476 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8477 ; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
8478 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
8479 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
8480 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
8481 ; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
8482 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
8483 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
8484 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8485 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
8486 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8487 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8488 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
8489 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
8490 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
8491 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8492 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8493 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8494 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
8495 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
8496 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
8497 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8498 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8499 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8500 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
8501 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
8502 ; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
8503 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8504 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8505 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
8506 ; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
8507 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8508 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8509 ; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
8510 ; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
8511 ; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
8512 ; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
8513 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
8514 ; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
8515 ; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8516 ; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
8517 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
8518 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
8519 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
8520 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
8521 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
8522 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
8523 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
8524 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB11_3
8525 ; GFX1064-DPP-NEXT: ; %bb.1:
8526 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
8527 ; GFX1064-DPP-NEXT: s_mov_b64 s[46:47], 0
8528 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
8529 ; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[44:45]
8530 ; GFX1064-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
8531 ; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8532 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
8533 ; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
8534 ; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
8535 ; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
8536 ; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
8537 ; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
8538 ; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8539 ; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8540 ; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
8541 ; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
8542 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8543 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
8544 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
8545 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
8546 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s44
8547 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
8548 ; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8549 ; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8550 ; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8551 ; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8552 ; GFX1064-DPP-NEXT: s_mov_b32 s12, s43
8553 ; GFX1064-DPP-NEXT: s_mov_b32 s13, s42
8554 ; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
8555 ; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8556 ; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
8557 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
8558 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
8559 ; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
8560 ; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
8561 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s45
8562 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
8563 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
8564 ; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8565 ; GFX1064-DPP-NEXT: s_clause 0x1
8566 ; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
8567 ; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
8568 ; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8569 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
8570 ; GFX1064-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
8571 ; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[46:47]
8572 ; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB11_2
8573 ; GFX1064-DPP-NEXT: .LBB11_3:
8574 ; GFX1064-DPP-NEXT: s_endpgm
8576 ; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8577 ; GFX1032-DPP: ; %bb.0:
8578 ; GFX1032-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
8579 ; GFX1032-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
8580 ; GFX1032-DPP-NEXT: s_mov_b32 s50, -1
8581 ; GFX1032-DPP-NEXT: s_mov_b32 s51, 0x31c16000
8582 ; GFX1032-DPP-NEXT: s_add_u32 s48, s48, s11
8583 ; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
8584 ; GFX1032-DPP-NEXT: s_addc_u32 s49, s49, 0
8585 ; GFX1032-DPP-NEXT: s_mov_b32 s43, s8
8586 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
8587 ; GFX1032-DPP-NEXT: s_mov_b32 s42, s9
8588 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
8589 ; GFX1032-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8590 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
8591 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8592 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8593 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
8594 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8595 ; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
8596 ; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
8597 ; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
8598 ; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8599 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8600 ; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
8601 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8602 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8603 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8604 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
8605 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
8606 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
8607 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
8608 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8609 ; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
8610 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
8611 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8612 ; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
8613 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
8614 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
8615 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
8616 ; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
8617 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
8618 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
8619 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8620 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
8621 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8622 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8623 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
8624 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
8625 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
8626 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8627 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8628 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8629 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
8630 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
8631 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
8632 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8633 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8634 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8635 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
8636 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
8637 ; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
8638 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8639 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8640 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
8641 ; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
8642 ; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8643 ; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8644 ; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
8645 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
8646 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
8647 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
8648 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
8649 ; GFX1032-DPP-NEXT: s_mov_b32 s46, 0
8650 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
8651 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
8652 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB11_3
8653 ; GFX1032-DPP-NEXT: ; %bb.1:
8654 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[44:45], s[34:35], 0x24
8655 ; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
8656 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
8657 ; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[44:45]
8658 ; GFX1032-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
8659 ; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8660 ; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
8661 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
8662 ; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
8663 ; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
8664 ; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
8665 ; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8666 ; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8667 ; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[48:51], 0 offset:4
8668 ; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[48:51], 0
8669 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
8670 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
8671 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
8672 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
8673 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s44
8674 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
8675 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
8676 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
8677 ; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
8678 ; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8679 ; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8680 ; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8681 ; GFX1032-DPP-NEXT: s_mov_b32 s12, s43
8682 ; GFX1032-DPP-NEXT: s_mov_b32 s13, s42
8683 ; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
8684 ; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
8685 ; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
8686 ; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[48:51], 0 offset:12
8687 ; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[48:51], 0 offset:8
8688 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s45
8689 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
8690 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
8691 ; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
8692 ; GFX1032-DPP-NEXT: s_clause 0x1
8693 ; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[48:51], 0
8694 ; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[48:51], 0 offset:4
8695 ; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8696 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
8697 ; GFX1032-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
8698 ; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s46
8699 ; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB11_2
8700 ; GFX1032-DPP-NEXT: .LBB11_3:
8701 ; GFX1032-DPP-NEXT: s_endpgm
8703 ; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8704 ; GFX1164-DPP: ; %bb.0:
8705 ; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
8706 ; GFX1164-DPP-NEXT: s_mov_b32 s43, s8
8707 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
8708 ; GFX1164-DPP-NEXT: s_mov_b32 s42, s9
8709 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
8710 ; GFX1164-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8711 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
8712 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8713 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8714 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
8715 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8716 ; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
8717 ; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
8718 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8719 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8720 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8721 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
8722 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
8723 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
8724 ; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
8725 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
8726 ; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8727 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
8728 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
8729 ; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
8730 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
8731 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
8732 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
8733 ; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
8734 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
8735 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
8736 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
8737 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8738 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
8739 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
8740 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8741 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8742 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
8743 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
8744 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
8745 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
8746 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8747 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8748 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8749 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8750 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
8751 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
8752 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
8753 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
8754 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8755 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8756 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8757 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
8758 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
8759 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
8760 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
8761 ; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
8762 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8763 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8764 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8765 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8766 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
8767 ; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
8768 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8769 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8770 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8771 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8772 ; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
8773 ; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
8774 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8775 ; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8776 ; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8777 ; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
8778 ; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8779 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
8780 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
8781 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
8782 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
8783 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
8784 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
8785 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
8786 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
8787 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
8788 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB11_3
8789 ; GFX1164-DPP-NEXT: ; %bb.1:
8790 ; GFX1164-DPP-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
8791 ; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
8792 ; GFX1164-DPP-NEXT: s_mov_b64 s[46:47], 0
8793 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
8794 ; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[44:45]
8795 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
8796 ; GFX1164-DPP-NEXT: .p2align 6
8797 ; GFX1164-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
8798 ; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8799 ; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
8800 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
8801 ; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
8802 ; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
8803 ; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
8804 ; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8805 ; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8806 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
8807 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8808 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
8809 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
8810 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
8811 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
8812 ; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8813 ; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8814 ; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8815 ; GFX1164-DPP-NEXT: s_mov_b32 s12, s43
8816 ; GFX1164-DPP-NEXT: s_mov_b32 s13, s42
8817 ; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
8818 ; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
8819 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
8820 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
8821 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s44
8822 ; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
8823 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s45
8824 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
8825 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
8826 ; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
8827 ; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
8828 ; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8829 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
8830 ; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
8831 ; GFX1164-DPP-NEXT: s_or_b64 s[46:47], vcc, s[46:47]
8832 ; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[46:47]
8833 ; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB11_2
8834 ; GFX1164-DPP-NEXT: .LBB11_3:
8835 ; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
8836 ; GFX1164-DPP-NEXT: s_endpgm
8838 ; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
8839 ; GFX1132-DPP: ; %bb.0:
8840 ; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
8841 ; GFX1132-DPP-NEXT: s_mov_b64 s[40:41], s[0:1]
8842 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
8843 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
8844 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
8845 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value@gotpcrel32@lo+4
8846 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value@gotpcrel32@hi+12
8847 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
8848 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8849 ; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
8850 ; GFX1132-DPP-NEXT: s_mov_b32 s42, s14
8851 ; GFX1132-DPP-NEXT: s_mov_b32 s43, s13
8852 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8853 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
8854 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8855 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
8856 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
8857 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
8858 ; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
8859 ; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
8860 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
8861 ; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
8862 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
8863 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
8864 ; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
8865 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
8866 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
8867 ; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
8868 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
8869 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
8870 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
8871 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8872 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
8873 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
8874 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
8875 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
8876 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8877 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
8878 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
8879 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8880 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8881 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8882 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
8883 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8884 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
8885 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
8886 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8887 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8888 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8889 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
8890 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8891 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
8892 ; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
8893 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8894 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8895 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8896 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
8897 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
8898 ; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
8899 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
8900 ; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
8901 ; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
8902 ; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
8903 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
8904 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
8905 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
8906 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
8907 ; GFX1132-DPP-NEXT: s_mov_b32 s46, 0
8908 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
8909 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
8910 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
8911 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB11_3
8912 ; GFX1132-DPP-NEXT: ; %bb.1:
8913 ; GFX1132-DPP-NEXT: s_load_b64 s[44:45], s[34:35], 0x24
8914 ; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
8915 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
8916 ; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[44:45]
8917 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
8918 ; GFX1132-DPP-NEXT: .p2align 6
8919 ; GFX1132-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
8920 ; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
8921 ; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
8922 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
8923 ; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
8924 ; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
8925 ; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
8926 ; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange@gotpcrel32@lo+4
8927 ; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange@gotpcrel32@hi+12
8928 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
8929 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
8930 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
8931 ; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
8932 ; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[40:41]
8933 ; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
8934 ; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
8935 ; GFX1132-DPP-NEXT: s_mov_b32 s12, s43
8936 ; GFX1132-DPP-NEXT: s_mov_b32 s13, s42
8937 ; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
8938 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
8939 ; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
8940 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
8941 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s44
8942 ; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
8943 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v4, 0
8944 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
8945 ; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
8946 ; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
8947 ; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
8948 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
8949 ; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
8950 ; GFX1132-DPP-NEXT: s_or_b32 s46, vcc_lo, s46
8951 ; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s46
8952 ; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB11_2
8953 ; GFX1132-DPP-NEXT: .LBB11_3:
8954 ; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
8955 ; GFX1132-DPP-NEXT: s_endpgm
8956 %divValue = call double @div.double.value()
8957 %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue monotonic, align 4, !amdgpu.no.fine.grained.memory !1
8961 define amdgpu_kernel void @global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr) {
8962 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
8963 ; GFX7LESS: ; %bb.0:
8964 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
8965 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
8966 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
8967 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
8968 ; GFX7LESS-NEXT: s_cbranch_execz .LBB12_3
8969 ; GFX7LESS-NEXT: ; %bb.1:
8970 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
8971 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
8972 ; GFX7LESS-NEXT: s_load_dword s2, s[0:1], 0x0
8973 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
8974 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
8975 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
8976 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s2
8977 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
8978 ; GFX7LESS-NEXT: .LBB12_2: ; %atomicrmw.start
8979 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
8980 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
8981 ; GFX7LESS-NEXT: v_max_f32_e32 v0, 4.0, v0
8982 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
8983 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v1
8984 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v0
8985 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
8986 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
8987 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
8988 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
8989 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v2
8990 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
8991 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB12_2
8992 ; GFX7LESS-NEXT: .LBB12_3:
8993 ; GFX7LESS-NEXT: s_endpgm
8995 ; GFX9-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
8997 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
8998 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
8999 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9000 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
9001 ; GFX9-NEXT: s_cbranch_execz .LBB12_3
9002 ; GFX9-NEXT: ; %bb.1:
9003 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9004 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
9005 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
9006 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
9007 ; GFX9-NEXT: s_load_dword s4, s[0:1], 0x0
9008 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
9009 ; GFX9-NEXT: v_mov_b32_e32 v1, s4
9010 ; GFX9-NEXT: .LBB12_2: ; %atomicrmw.start
9011 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
9012 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
9013 ; GFX9-NEXT: v_max_f32_e32 v0, 4.0, v0
9014 ; GFX9-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
9015 ; GFX9-NEXT: s_waitcnt vmcnt(0)
9016 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
9017 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
9018 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
9019 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
9020 ; GFX9-NEXT: s_cbranch_execnz .LBB12_2
9021 ; GFX9-NEXT: .LBB12_3:
9022 ; GFX9-NEXT: s_endpgm
9024 ; GFX1064-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9026 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9027 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9028 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9029 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
9030 ; GFX1064-NEXT: s_cbranch_execz .LBB12_2
9031 ; GFX1064-NEXT: ; %bb.1:
9032 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9033 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
9034 ; GFX1064-NEXT: v_mov_b32_e32 v1, 4.0
9035 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
9036 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
9037 ; GFX1064-NEXT: .LBB12_2:
9038 ; GFX1064-NEXT: s_endpgm
9040 ; GFX1032-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9042 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9043 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
9044 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
9045 ; GFX1032-NEXT: s_cbranch_execz .LBB12_2
9046 ; GFX1032-NEXT: ; %bb.1:
9047 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9048 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
9049 ; GFX1032-NEXT: v_mov_b32_e32 v1, 4.0
9050 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
9051 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
9052 ; GFX1032-NEXT: .LBB12_2:
9053 ; GFX1032-NEXT: s_endpgm
9055 ; GFX1164-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9057 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9058 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
9059 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
9060 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9061 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
9062 ; GFX1164-NEXT: s_cbranch_execz .LBB12_2
9063 ; GFX1164-NEXT: ; %bb.1:
9064 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9065 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
9066 ; GFX1164-NEXT: v_mov_b32_e32 v1, 4.0
9067 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
9068 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9069 ; GFX1164-NEXT: .LBB12_2:
9070 ; GFX1164-NEXT: s_endpgm
9072 ; GFX1132-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9074 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9075 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
9076 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
9077 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
9078 ; GFX1132-NEXT: s_cbranch_execz .LBB12_2
9079 ; GFX1132-NEXT: ; %bb.1:
9080 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9081 ; GFX1132-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
9082 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
9083 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9084 ; GFX1132-NEXT: .LBB12_2:
9085 ; GFX1132-NEXT: s_endpgm
9087 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9088 ; GFX7LESS-DPP: ; %bb.0:
9089 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
9090 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
9091 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9092 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9093 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB12_3
9094 ; GFX7LESS-DPP-NEXT: ; %bb.1:
9095 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
9096 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
9097 ; GFX7LESS-DPP-NEXT: s_load_dword s2, s[0:1], 0x0
9098 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
9099 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
9100 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
9101 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s2
9102 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
9103 ; GFX7LESS-DPP-NEXT: .LBB12_2: ; %atomicrmw.start
9104 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
9105 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
9106 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
9107 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
9108 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v1
9109 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v0
9110 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
9111 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
9112 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
9113 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
9114 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v2
9115 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
9116 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB12_2
9117 ; GFX7LESS-DPP-NEXT: .LBB12_3:
9118 ; GFX7LESS-DPP-NEXT: s_endpgm
9120 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9121 ; GFX9-DPP: ; %bb.0:
9122 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9123 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9124 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9125 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9126 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB12_3
9127 ; GFX9-DPP-NEXT: ; %bb.1:
9128 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9129 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
9130 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
9131 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
9132 ; GFX9-DPP-NEXT: s_load_dword s4, s[0:1], 0x0
9133 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
9134 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s4
9135 ; GFX9-DPP-NEXT: .LBB12_2: ; %atomicrmw.start
9136 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
9137 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
9138 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
9139 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
9140 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
9141 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
9142 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
9143 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
9144 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
9145 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB12_2
9146 ; GFX9-DPP-NEXT: .LBB12_3:
9147 ; GFX9-DPP-NEXT: s_endpgm
9149 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9150 ; GFX1064-DPP: ; %bb.0:
9151 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9152 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9153 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9154 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9155 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB12_2
9156 ; GFX1064-DPP-NEXT: ; %bb.1:
9157 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9158 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
9159 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9160 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
9161 ; GFX1064-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
9162 ; GFX1064-DPP-NEXT: .LBB12_2:
9163 ; GFX1064-DPP-NEXT: s_endpgm
9165 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9166 ; GFX1032-DPP: ; %bb.0:
9167 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9168 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
9169 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
9170 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB12_2
9171 ; GFX1032-DPP-NEXT: ; %bb.1:
9172 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9173 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
9174 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9175 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
9176 ; GFX1032-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
9177 ; GFX1032-DPP-NEXT: .LBB12_2:
9178 ; GFX1032-DPP-NEXT: s_endpgm
9180 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9181 ; GFX1164-DPP: ; %bb.0:
9182 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9183 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
9184 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
9185 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9186 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
9187 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB12_2
9188 ; GFX1164-DPP-NEXT: ; %bb.1:
9189 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9190 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
9191 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9192 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
9193 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9194 ; GFX1164-DPP-NEXT: .LBB12_2:
9195 ; GFX1164-DPP-NEXT: s_endpgm
9197 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_ignore_denormal_mode__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9198 ; GFX1132-DPP: ; %bb.0:
9199 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9200 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
9201 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
9202 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
9203 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB12_2
9204 ; GFX1132-DPP-NEXT: ; %bb.1:
9205 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9206 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
9207 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
9208 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9209 ; GFX1132-DPP-NEXT: .LBB12_2:
9210 ; GFX1132-DPP-NEXT: s_endpgm
9211 %result = atomicrmw fmax ptr addrspace(1) %ptr, float 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.no.remote.memory !1, !amdgpu.ignore.denormal.mode !1
9215 define amdgpu_kernel void @global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory(ptr addrspace(1) %ptr) {
9216 ; GFX7LESS-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9217 ; GFX7LESS: ; %bb.0:
9218 ; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
9219 ; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
9220 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9221 ; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
9222 ; GFX7LESS-NEXT: s_cbranch_execz .LBB13_3
9223 ; GFX7LESS-NEXT: ; %bb.1:
9224 ; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
9225 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
9226 ; GFX7LESS-NEXT: s_load_dword s2, s[0:1], 0x0
9227 ; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
9228 ; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
9229 ; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
9230 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, s2
9231 ; GFX7LESS-NEXT: s_mov_b32 s2, -1
9232 ; GFX7LESS-NEXT: .LBB13_2: ; %atomicrmw.start
9233 ; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
9234 ; GFX7LESS-NEXT: v_mul_f32_e32 v0, 1.0, v1
9235 ; GFX7LESS-NEXT: v_max_f32_e32 v0, 4.0, v0
9236 ; GFX7LESS-NEXT: s_waitcnt expcnt(0)
9237 ; GFX7LESS-NEXT: v_mov_b32_e32 v3, v1
9238 ; GFX7LESS-NEXT: v_mov_b32_e32 v2, v0
9239 ; GFX7LESS-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
9240 ; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
9241 ; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
9242 ; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
9243 ; GFX7LESS-NEXT: v_mov_b32_e32 v1, v2
9244 ; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
9245 ; GFX7LESS-NEXT: s_cbranch_execnz .LBB13_2
9246 ; GFX7LESS-NEXT: .LBB13_3:
9247 ; GFX7LESS-NEXT: s_endpgm
9249 ; GFX9-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9251 ; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9252 ; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9253 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9254 ; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
9255 ; GFX9-NEXT: s_cbranch_execz .LBB13_3
9256 ; GFX9-NEXT: ; %bb.1:
9257 ; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9258 ; GFX9-NEXT: s_mov_b64 s[2:3], 0
9259 ; GFX9-NEXT: v_mov_b32_e32 v2, 0
9260 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
9261 ; GFX9-NEXT: s_load_dword s4, s[0:1], 0x0
9262 ; GFX9-NEXT: s_waitcnt lgkmcnt(0)
9263 ; GFX9-NEXT: v_mov_b32_e32 v1, s4
9264 ; GFX9-NEXT: .LBB13_2: ; %atomicrmw.start
9265 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
9266 ; GFX9-NEXT: v_max_f32_e32 v0, v1, v1
9267 ; GFX9-NEXT: v_max_f32_e32 v0, 4.0, v0
9268 ; GFX9-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
9269 ; GFX9-NEXT: s_waitcnt vmcnt(0)
9270 ; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
9271 ; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
9272 ; GFX9-NEXT: v_mov_b32_e32 v1, v0
9273 ; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
9274 ; GFX9-NEXT: s_cbranch_execnz .LBB13_2
9275 ; GFX9-NEXT: .LBB13_3:
9276 ; GFX9-NEXT: s_endpgm
9278 ; GFX1064-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9280 ; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9281 ; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9282 ; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9283 ; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
9284 ; GFX1064-NEXT: s_cbranch_execz .LBB13_2
9285 ; GFX1064-NEXT: ; %bb.1:
9286 ; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9287 ; GFX1064-NEXT: v_mov_b32_e32 v0, 0
9288 ; GFX1064-NEXT: v_mov_b32_e32 v1, 4.0
9289 ; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
9290 ; GFX1064-NEXT: global_atomic_fmax v0, v1, s[0:1]
9291 ; GFX1064-NEXT: .LBB13_2:
9292 ; GFX1064-NEXT: s_endpgm
9294 ; GFX1032-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9296 ; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9297 ; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
9298 ; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
9299 ; GFX1032-NEXT: s_cbranch_execz .LBB13_2
9300 ; GFX1032-NEXT: ; %bb.1:
9301 ; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9302 ; GFX1032-NEXT: v_mov_b32_e32 v0, 0
9303 ; GFX1032-NEXT: v_mov_b32_e32 v1, 4.0
9304 ; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
9305 ; GFX1032-NEXT: global_atomic_fmax v0, v1, s[0:1]
9306 ; GFX1032-NEXT: .LBB13_2:
9307 ; GFX1032-NEXT: s_endpgm
9309 ; GFX1164-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9311 ; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9312 ; GFX1164-NEXT: s_mov_b64 s[0:1], exec
9313 ; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
9314 ; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9315 ; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
9316 ; GFX1164-NEXT: s_cbranch_execz .LBB13_2
9317 ; GFX1164-NEXT: ; %bb.1:
9318 ; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9319 ; GFX1164-NEXT: v_mov_b32_e32 v0, 0
9320 ; GFX1164-NEXT: v_mov_b32_e32 v1, 4.0
9321 ; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
9322 ; GFX1164-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9323 ; GFX1164-NEXT: .LBB13_2:
9324 ; GFX1164-NEXT: s_endpgm
9326 ; GFX1132-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9328 ; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9329 ; GFX1132-NEXT: s_mov_b32 s0, exec_lo
9330 ; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
9331 ; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
9332 ; GFX1132-NEXT: s_cbranch_execz .LBB13_2
9333 ; GFX1132-NEXT: ; %bb.1:
9334 ; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9335 ; GFX1132-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
9336 ; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
9337 ; GFX1132-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9338 ; GFX1132-NEXT: .LBB13_2:
9339 ; GFX1132-NEXT: s_endpgm
9341 ; GFX7LESS-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9342 ; GFX7LESS-DPP: ; %bb.0:
9343 ; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
9344 ; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
9345 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9346 ; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9347 ; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB13_3
9348 ; GFX7LESS-DPP-NEXT: ; %bb.1:
9349 ; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
9350 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
9351 ; GFX7LESS-DPP-NEXT: s_load_dword s2, s[0:1], 0x0
9352 ; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
9353 ; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
9354 ; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
9355 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s2
9356 ; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
9357 ; GFX7LESS-DPP-NEXT: .LBB13_2: ; %atomicrmw.start
9358 ; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
9359 ; GFX7LESS-DPP-NEXT: v_mul_f32_e32 v0, 1.0, v1
9360 ; GFX7LESS-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
9361 ; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
9362 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v1
9363 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v0
9364 ; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap v[2:3], off, s[0:3], 0 glc
9365 ; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
9366 ; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1
9367 ; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
9368 ; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, v2
9369 ; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
9370 ; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB13_2
9371 ; GFX7LESS-DPP-NEXT: .LBB13_3:
9372 ; GFX7LESS-DPP-NEXT: s_endpgm
9374 ; GFX9-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9375 ; GFX9-DPP: ; %bb.0:
9376 ; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9377 ; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9378 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9379 ; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9380 ; GFX9-DPP-NEXT: s_cbranch_execz .LBB13_3
9381 ; GFX9-DPP-NEXT: ; %bb.1:
9382 ; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9383 ; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
9384 ; GFX9-DPP-NEXT: v_mov_b32_e32 v2, 0
9385 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
9386 ; GFX9-DPP-NEXT: s_load_dword s4, s[0:1], 0x0
9387 ; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
9388 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s4
9389 ; GFX9-DPP-NEXT: .LBB13_2: ; %atomicrmw.start
9390 ; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
9391 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, v1, v1
9392 ; GFX9-DPP-NEXT: v_max_f32_e32 v0, 4.0, v0
9393 ; GFX9-DPP-NEXT: global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
9394 ; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
9395 ; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, v0, v1
9396 ; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
9397 ; GFX9-DPP-NEXT: v_mov_b32_e32 v1, v0
9398 ; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
9399 ; GFX9-DPP-NEXT: s_cbranch_execnz .LBB13_2
9400 ; GFX9-DPP-NEXT: .LBB13_3:
9401 ; GFX9-DPP-NEXT: s_endpgm
9403 ; GFX1064-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9404 ; GFX1064-DPP: ; %bb.0:
9405 ; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9406 ; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9407 ; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
9408 ; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
9409 ; GFX1064-DPP-NEXT: s_cbranch_execz .LBB13_2
9410 ; GFX1064-DPP-NEXT: ; %bb.1:
9411 ; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9412 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
9413 ; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9414 ; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
9415 ; GFX1064-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
9416 ; GFX1064-DPP-NEXT: .LBB13_2:
9417 ; GFX1064-DPP-NEXT: s_endpgm
9419 ; GFX1032-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9420 ; GFX1032-DPP: ; %bb.0:
9421 ; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9422 ; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
9423 ; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
9424 ; GFX1032-DPP-NEXT: s_cbranch_execz .LBB13_2
9425 ; GFX1032-DPP-NEXT: ; %bb.1:
9426 ; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
9427 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
9428 ; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9429 ; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
9430 ; GFX1032-DPP-NEXT: global_atomic_fmax v0, v1, s[0:1]
9431 ; GFX1032-DPP-NEXT: .LBB13_2:
9432 ; GFX1032-DPP-NEXT: s_endpgm
9434 ; GFX1164-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9435 ; GFX1164-DPP: ; %bb.0:
9436 ; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9437 ; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
9438 ; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
9439 ; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
9440 ; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
9441 ; GFX1164-DPP-NEXT: s_cbranch_execz .LBB13_2
9442 ; GFX1164-DPP-NEXT: ; %bb.1:
9443 ; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9444 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
9445 ; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 4.0
9446 ; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
9447 ; GFX1164-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9448 ; GFX1164-DPP-NEXT: .LBB13_2:
9449 ; GFX1164-DPP-NEXT: s_endpgm
9451 ; GFX1132-DPP-LABEL: global_atomic_fmax_uni_address_uni_value_system_scope__amdgpu_no_fine_grained_memory__amdgpu_no_remote_memory:
9452 ; GFX1132-DPP: ; %bb.0:
9453 ; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
9454 ; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
9455 ; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
9456 ; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
9457 ; GFX1132-DPP-NEXT: s_cbranch_execz .LBB13_2
9458 ; GFX1132-DPP-NEXT: ; %bb.1:
9459 ; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
9460 ; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
9461 ; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
9462 ; GFX1132-DPP-NEXT: global_atomic_max_f32 v0, v1, s[0:1]
9463 ; GFX1132-DPP-NEXT: .LBB13_2:
9464 ; GFX1132-DPP-NEXT: s_endpgm
9465 %result = atomicrmw fmax ptr addrspace(1) %ptr, float 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.no.remote.memory !1
9469 attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
9471 !llvm.module.flags = !{!0}
9472 !0 = !{i32 1, !"amdhsa_code_object_version", i32 500}