| blob | fbf8c203dcb3904a49b1d0950619d12b7b831c14 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11,GISEL11 %s
3 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11,DAGISEL11 %s
4 ; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10,GISEL10 %s
5 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10,DAGISEL10 %s
6 ; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11_W64,GISEL11_W64 %s
7 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11_W64,DAGISEL11_W64 %s
8 ; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1030 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10_W64,GISEL10_W64 %s
9 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1030 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10_W64,DAGISEL10_W64 %s
11 define amdgpu_cs_chain void @set_inactive_chain_arg(ptr addrspace(1) %out, i32 %inactive, i32 %active) {
12 ; GFX11-LABEL: set_inactive_chain_arg:
13 ; GFX11: ; %bb.0:
14 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
15 ; GFX11-NEXT: s_or_saveexec_b32 s0, -1
16 ; GFX11-NEXT: v_mov_b32_e32 v0, v10
17 ; GFX11-NEXT: s_mov_b32 exec_lo, s0
18 ; GFX11-NEXT: s_or_saveexec_b32 s0, -1
19 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
20 ; GFX11-NEXT: v_cndmask_b32_e64 v0, v0, v11, s0
21 ; GFX11-NEXT: s_mov_b32 exec_lo, s0
22 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
23 ; GFX11-NEXT: v_mov_b32_e32 v1, v0
24 ; GFX11-NEXT: global_store_b32 v[8:9], v1, off
25 ; GFX11-NEXT: s_endpgm
26 ;
27 ; GFX10-LABEL: set_inactive_chain_arg:
28 ; GFX10: ; %bb.0:
29 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
30 ; GFX10-NEXT: s_or_saveexec_b32 s0, -1
31 ; GFX10-NEXT: v_mov_b32_e32 v0, v10
32 ; GFX10-NEXT: s_mov_b32 exec_lo, s0
33 ; GFX10-NEXT: s_or_saveexec_b32 s0, -1
34 ; GFX10-NEXT: v_cndmask_b32_e64 v0, v0, v11, s0
35 ; GFX10-NEXT: s_mov_b32 exec_lo, s0
36 ; GFX10-NEXT: v_mov_b32_e32 v1, v0
37 ; GFX10-NEXT: global_store_dword v[8:9], v1, off
38 ; GFX10-NEXT: s_endpgm
39 ;
40 ; GFX11_W64-LABEL: set_inactive_chain_arg:
41 ; GFX11_W64: ; %bb.0:
42 ; GFX11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
43 ; GFX11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
44 ; GFX11_W64-NEXT: v_mov_b32_e32 v0, v10
45 ; GFX11_W64-NEXT: s_mov_b64 exec, s[0:1]
46 ; GFX11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
47 ; GFX11_W64-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
48 ; GFX11_W64-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
49 ; GFX11_W64-NEXT: s_mov_b64 exec, s[0:1]
50 ; GFX11_W64-NEXT: s_delay_alu instid0(VALU_DEP_1)
51 ; GFX11_W64-NEXT: v_mov_b32_e32 v1, v0
52 ; GFX11_W64-NEXT: global_store_b32 v[8:9], v1, off
53 ; GFX11_W64-NEXT: s_endpgm
54 ;
55 ; GFX10_W64-LABEL: set_inactive_chain_arg:
56 ; GFX10_W64: ; %bb.0:
57 ; GFX10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
58 ; GFX10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
59 ; GFX10_W64-NEXT: v_mov_b32_e32 v0, v10
60 ; GFX10_W64-NEXT: s_mov_b64 exec, s[0:1]
61 ; GFX10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
62 ; GFX10_W64-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
63 ; GFX10_W64-NEXT: s_mov_b64 exec, s[0:1]
64 ; GFX10_W64-NEXT: v_mov_b32_e32 v1, v0
65 ; GFX10_W64-NEXT: global_store_dword v[8:9], v1, off
66 ; GFX10_W64-NEXT: s_endpgm
67 %tmp = call i32 @llvm.amdgcn.set.inactive.chain.arg.i32(i32 %active, i32 %inactive) #0
68 %wwm = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp)
69 store i32 %wwm, ptr addrspace(1) %out
70 ret void
71 }
73 define amdgpu_cs_chain void @set_inactive_chain_arg_64(ptr addrspace(1) %out, i64 %inactive, i64 %active) {
74 ; GISEL11-LABEL: set_inactive_chain_arg_64:
75 ; GISEL11: ; %bb.0:
76 ; GISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
77 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
78 ; GISEL11-NEXT: v_dual_mov_b32 v0, v10 :: v_dual_mov_b32 v1, v11
79 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
80 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
81 ; GISEL11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
82 ; GISEL11-NEXT: v_cndmask_b32_e64 v0, v0, v12, s0
83 ; GISEL11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
84 ; GISEL11-NEXT: v_cndmask_b32_e64 v1, v1, v13, s0
85 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
86 ; GISEL11-NEXT: v_mov_b32_e32 v2, v0
87 ; GISEL11-NEXT: s_delay_alu instid0(VALU_DEP_2)
88 ; GISEL11-NEXT: v_mov_b32_e32 v3, v1
89 ; GISEL11-NEXT: global_store_b64 v[8:9], v[2:3], off
90 ; GISEL11-NEXT: s_endpgm
91 ;
92 ; DAGISEL11-LABEL: set_inactive_chain_arg_64:
93 ; DAGISEL11: ; %bb.0:
94 ; DAGISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
95 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
96 ; DAGISEL11-NEXT: v_dual_mov_b32 v0, v11 :: v_dual_mov_b32 v1, v10
97 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
98 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
99 ; DAGISEL11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
100 ; DAGISEL11-NEXT: v_cndmask_b32_e64 v2, v0, v13, s0
101 ; DAGISEL11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
102 ; DAGISEL11-NEXT: v_cndmask_b32_e64 v1, v1, v12, s0
103 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
104 ; DAGISEL11-NEXT: v_mov_b32_e32 v3, v1
105 ; DAGISEL11-NEXT: s_delay_alu instid0(VALU_DEP_3)
106 ; DAGISEL11-NEXT: v_mov_b32_e32 v4, v2
107 ; DAGISEL11-NEXT: global_store_b64 v[8:9], v[3:4], off
108 ; DAGISEL11-NEXT: s_endpgm
109 ;
110 ; GISEL10-LABEL: set_inactive_chain_arg_64:
111 ; GISEL10: ; %bb.0:
112 ; GISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
113 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
114 ; GISEL10-NEXT: v_mov_b32_e32 v0, v10
115 ; GISEL10-NEXT: v_mov_b32_e32 v1, v11
116 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
117 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
118 ; GISEL10-NEXT: v_cndmask_b32_e64 v0, v0, v12, s0
119 ; GISEL10-NEXT: v_cndmask_b32_e64 v1, v1, v13, s0
120 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
121 ; GISEL10-NEXT: v_mov_b32_e32 v2, v0
122 ; GISEL10-NEXT: v_mov_b32_e32 v3, v1
123 ; GISEL10-NEXT: global_store_dwordx2 v[8:9], v[2:3], off
124 ; GISEL10-NEXT: s_endpgm
125 ;
126 ; DAGISEL10-LABEL: set_inactive_chain_arg_64:
127 ; DAGISEL10: ; %bb.0:
128 ; DAGISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
129 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
130 ; DAGISEL10-NEXT: v_mov_b32_e32 v0, v11
131 ; DAGISEL10-NEXT: v_mov_b32_e32 v1, v10
132 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
133 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
134 ; DAGISEL10-NEXT: v_cndmask_b32_e64 v2, v0, v13, s0
135 ; DAGISEL10-NEXT: v_cndmask_b32_e64 v1, v1, v12, s0
136 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
137 ; DAGISEL10-NEXT: v_mov_b32_e32 v3, v1
138 ; DAGISEL10-NEXT: v_mov_b32_e32 v4, v2
139 ; DAGISEL10-NEXT: global_store_dwordx2 v[8:9], v[3:4], off
140 ; DAGISEL10-NEXT: s_endpgm
141 ;
142 ; GISEL11_W64-LABEL: set_inactive_chain_arg_64:
143 ; GISEL11_W64: ; %bb.0:
144 ; GISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
145 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
146 ; GISEL11_W64-NEXT: v_mov_b32_e32 v0, v10
147 ; GISEL11_W64-NEXT: v_mov_b32_e32 v1, v11
148 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
149 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
150 ; GISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
151 ; GISEL11_W64-NEXT: v_cndmask_b32_e64 v0, v0, v12, s[0:1]
152 ; GISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
153 ; GISEL11_W64-NEXT: v_cndmask_b32_e64 v1, v1, v13, s[0:1]
154 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
155 ; GISEL11_W64-NEXT: v_mov_b32_e32 v2, v0
156 ; GISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_2)
157 ; GISEL11_W64-NEXT: v_mov_b32_e32 v3, v1
158 ; GISEL11_W64-NEXT: global_store_b64 v[8:9], v[2:3], off
159 ; GISEL11_W64-NEXT: s_endpgm
160 ;
161 ; DAGISEL11_W64-LABEL: set_inactive_chain_arg_64:
162 ; DAGISEL11_W64: ; %bb.0:
163 ; DAGISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
164 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
165 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v0, v11
166 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v1, v10
167 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
168 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
169 ; DAGISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
170 ; DAGISEL11_W64-NEXT: v_cndmask_b32_e64 v2, v0, v13, s[0:1]
171 ; DAGISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
172 ; DAGISEL11_W64-NEXT: v_cndmask_b32_e64 v1, v1, v12, s[0:1]
173 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
174 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v3, v1
175 ; DAGISEL11_W64-NEXT: s_delay_alu instid0(VALU_DEP_3)
176 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v4, v2
177 ; DAGISEL11_W64-NEXT: global_store_b64 v[8:9], v[3:4], off
178 ; DAGISEL11_W64-NEXT: s_endpgm
179 ;
180 ; GISEL10_W64-LABEL: set_inactive_chain_arg_64:
181 ; GISEL10_W64: ; %bb.0:
182 ; GISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
183 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
184 ; GISEL10_W64-NEXT: v_mov_b32_e32 v0, v10
185 ; GISEL10_W64-NEXT: v_mov_b32_e32 v1, v11
186 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
187 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
188 ; GISEL10_W64-NEXT: v_cndmask_b32_e64 v0, v0, v12, s[0:1]
189 ; GISEL10_W64-NEXT: v_cndmask_b32_e64 v1, v1, v13, s[0:1]
190 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
191 ; GISEL10_W64-NEXT: v_mov_b32_e32 v2, v0
192 ; GISEL10_W64-NEXT: v_mov_b32_e32 v3, v1
193 ; GISEL10_W64-NEXT: global_store_dwordx2 v[8:9], v[2:3], off
194 ; GISEL10_W64-NEXT: s_endpgm
195 ;
196 ; DAGISEL10_W64-LABEL: set_inactive_chain_arg_64:
197 ; DAGISEL10_W64: ; %bb.0:
198 ; DAGISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
199 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
200 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v0, v11
201 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v1, v10
202 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
203 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
204 ; DAGISEL10_W64-NEXT: v_cndmask_b32_e64 v2, v0, v13, s[0:1]
205 ; DAGISEL10_W64-NEXT: v_cndmask_b32_e64 v1, v1, v12, s[0:1]
206 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
207 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v3, v1
208 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v4, v2
209 ; DAGISEL10_W64-NEXT: global_store_dwordx2 v[8:9], v[3:4], off
210 ; DAGISEL10_W64-NEXT: s_endpgm
211 %tmp = call i64 @llvm.amdgcn.set.inactive.chain.arg.i64(i64 %active, i64 %inactive) #0
212 %wwm = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp)
213 store i64 %wwm, ptr addrspace(1) %out
214 ret void
215 }
217 define amdgpu_cs_chain void @set_inactive_chain_arg_dpp(ptr addrspace(1) %out, i32 %inactive, i32 %active) {
218 ; GFX11-LABEL: set_inactive_chain_arg_dpp:
219 ; GFX11: ; %bb.0:
220 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
221 ; GFX11-NEXT: s_or_saveexec_b32 s0, -1
222 ; GFX11-NEXT: v_mov_b32_e32 v0, v10
223 ; GFX11-NEXT: s_mov_b32 exec_lo, s0
224 ; GFX11-NEXT: s_or_saveexec_b32 s0, -1
225 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
226 ; GFX11-NEXT: v_cndmask_b32_e64 v0, v0, v11, s0
227 ; GFX11-NEXT: v_mov_b32_e32 v1, 0
228 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
229 ; GFX11-NEXT: v_mov_b32_dpp v1, v0 row_xmask:1 row_mask:0xf bank_mask:0xf
230 ; GFX11-NEXT: s_mov_b32 exec_lo, s0
231 ; GFX11-NEXT: v_mov_b32_e32 v2, v1
232 ; GFX11-NEXT: global_store_b32 v[8:9], v2, off
233 ; GFX11-NEXT: s_endpgm
234 ;
235 ; GFX10-LABEL: set_inactive_chain_arg_dpp:
236 ; GFX10: ; %bb.0:
237 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
238 ; GFX10-NEXT: s_or_saveexec_b32 s0, -1
239 ; GFX10-NEXT: v_mov_b32_e32 v0, v10
240 ; GFX10-NEXT: s_mov_b32 exec_lo, s0
241 ; GFX10-NEXT: s_or_saveexec_b32 s0, -1
242 ; GFX10-NEXT: v_cndmask_b32_e64 v0, v0, v11, s0
243 ; GFX10-NEXT: v_mov_b32_e32 v1, 0
244 ; GFX10-NEXT: v_mov_b32_dpp v1, v0 row_xmask:1 row_mask:0xf bank_mask:0xf
245 ; GFX10-NEXT: s_mov_b32 exec_lo, s0
246 ; GFX10-NEXT: v_mov_b32_e32 v2, v1
247 ; GFX10-NEXT: global_store_dword v[8:9], v2, off
248 ; GFX10-NEXT: s_endpgm
249 ;
250 ; GFX11_W64-LABEL: set_inactive_chain_arg_dpp:
251 ; GFX11_W64: ; %bb.0:
252 ; GFX11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
253 ; GFX11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
254 ; GFX11_W64-NEXT: v_mov_b32_e32 v0, v10
255 ; GFX11_W64-NEXT: s_mov_b64 exec, s[0:1]
256 ; GFX11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
257 ; GFX11_W64-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
258 ; GFX11_W64-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
259 ; GFX11_W64-NEXT: v_mov_b32_e32 v1, 0
260 ; GFX11_W64-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
261 ; GFX11_W64-NEXT: v_mov_b32_dpp v1, v0 row_xmask:1 row_mask:0xf bank_mask:0xf
262 ; GFX11_W64-NEXT: s_mov_b64 exec, s[0:1]
263 ; GFX11_W64-NEXT: v_mov_b32_e32 v2, v1
264 ; GFX11_W64-NEXT: global_store_b32 v[8:9], v2, off
265 ; GFX11_W64-NEXT: s_endpgm
266 ;
267 ; GFX10_W64-LABEL: set_inactive_chain_arg_dpp:
268 ; GFX10_W64: ; %bb.0:
269 ; GFX10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
270 ; GFX10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
271 ; GFX10_W64-NEXT: v_mov_b32_e32 v0, v10
272 ; GFX10_W64-NEXT: s_mov_b64 exec, s[0:1]
273 ; GFX10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
274 ; GFX10_W64-NEXT: v_cndmask_b32_e64 v0, v0, v11, s[0:1]
275 ; GFX10_W64-NEXT: v_mov_b32_e32 v1, 0
276 ; GFX10_W64-NEXT: v_mov_b32_dpp v1, v0 row_xmask:1 row_mask:0xf bank_mask:0xf
277 ; GFX10_W64-NEXT: s_mov_b64 exec, s[0:1]
278 ; GFX10_W64-NEXT: v_mov_b32_e32 v2, v1
279 ; GFX10_W64-NEXT: global_store_dword v[8:9], v2, off
280 ; GFX10_W64-NEXT: s_endpgm
281 %tmp = call i32 @llvm.amdgcn.set.inactive.chain.arg.i32(i32 %active, i32 %inactive) #0
282 %dpp = call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp, i32 353, i32 15, i32 15, i1 false)
283 %wwm = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %dpp)
284 store i32 %wwm, ptr addrspace(1) %out
285 ret void
286 }
288 ; Make sure that if we need to use the register for %inactive for something else (in this case a call),
289 ; we save its inactive lanes for later use in set.inactive.chain.arg.
290 define amdgpu_cs_chain void @set_inactive_chain_arg_call(ptr addrspace(1) %out, i32 %inactive, i32 %active) {
291 ; GISEL11-LABEL: set_inactive_chain_arg_call:
292 ; GISEL11: ; %bb.0:
293 ; GISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
294 ; GISEL11-NEXT: s_mov_b32 s32, 0
295 ; GISEL11-NEXT: v_dual_mov_b32 v41, v8 :: v_dual_mov_b32 v42, v9
296 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
297 ; GISEL11-NEXT: v_mov_b32_e32 v40, v10
298 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
299 ; GISEL11-NEXT: s_getpc_b64 s[0:1]
300 ; GISEL11-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
301 ; GISEL11-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
302 ; GISEL11-NEXT: v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v0, 0
303 ; GISEL11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
304 ; GISEL11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
305 ; GISEL11-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v4, 0
306 ; GISEL11-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_mov_b32 v6, 0
307 ; GISEL11-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v8, 0
308 ; GISEL11-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v10, 0
309 ; GISEL11-NEXT: v_mov_b32_e32 v11, 0
310 ; GISEL11-NEXT: s_waitcnt lgkmcnt(0)
311 ; GISEL11-NEXT: s_swappc_b64 s[30:31], s[0:1]
312 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
313 ; GISEL11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
314 ; GISEL11-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
315 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
316 ; GISEL11-NEXT: v_mov_b32_e32 v0, v12
317 ; GISEL11-NEXT: global_store_b32 v[41:42], v0, off
318 ; GISEL11-NEXT: s_endpgm
319 ;
320 ; DAGISEL11-LABEL: set_inactive_chain_arg_call:
321 ; DAGISEL11: ; %bb.0:
322 ; DAGISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
323 ; DAGISEL11-NEXT: s_mov_b32 s32, 0
324 ; DAGISEL11-NEXT: v_mov_b32_e32 v43, v11
325 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
326 ; DAGISEL11-NEXT: v_mov_b32_e32 v40, v10
327 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
328 ; DAGISEL11-NEXT: s_getpc_b64 s[0:1]
329 ; DAGISEL11-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
330 ; DAGISEL11-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
331 ; DAGISEL11-NEXT: v_dual_mov_b32 v42, v9 :: v_dual_mov_b32 v41, v8
332 ; DAGISEL11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
333 ; DAGISEL11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0
334 ; DAGISEL11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v3, 0
335 ; DAGISEL11-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 0
336 ; DAGISEL11-NEXT: v_dual_mov_b32 v6, 0 :: v_dual_mov_b32 v7, 0
337 ; DAGISEL11-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v9, 0
338 ; DAGISEL11-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v11, 0
339 ; DAGISEL11-NEXT: s_waitcnt lgkmcnt(0)
340 ; DAGISEL11-NEXT: s_swappc_b64 s[30:31], s[0:1]
341 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
342 ; DAGISEL11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
343 ; DAGISEL11-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
344 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
345 ; DAGISEL11-NEXT: v_mov_b32_e32 v0, v12
346 ; DAGISEL11-NEXT: global_store_b32 v[41:42], v0, off
347 ; DAGISEL11-NEXT: s_endpgm
348 ;
349 ; GISEL10-LABEL: set_inactive_chain_arg_call:
350 ; GISEL10: ; %bb.0:
351 ; GISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
352 ; GISEL10-NEXT: s_mov_b32 s32, 0
353 ; GISEL10-NEXT: v_mov_b32_e32 v41, v8
354 ; GISEL10-NEXT: v_mov_b32_e32 v42, v9
355 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
356 ; GISEL10-NEXT: v_mov_b32_e32 v40, v10
357 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
358 ; GISEL10-NEXT: s_getpc_b64 s[0:1]
359 ; GISEL10-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
360 ; GISEL10-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
361 ; GISEL10-NEXT: v_mov_b32_e32 v43, v11
362 ; GISEL10-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
363 ; GISEL10-NEXT: v_mov_b32_e32 v0, 0
364 ; GISEL10-NEXT: v_mov_b32_e32 v1, 0
365 ; GISEL10-NEXT: v_mov_b32_e32 v2, 0
366 ; GISEL10-NEXT: v_mov_b32_e32 v3, 0
367 ; GISEL10-NEXT: v_mov_b32_e32 v4, 0
368 ; GISEL10-NEXT: v_mov_b32_e32 v5, 0
369 ; GISEL10-NEXT: v_mov_b32_e32 v6, 0
370 ; GISEL10-NEXT: v_mov_b32_e32 v7, 0
371 ; GISEL10-NEXT: v_mov_b32_e32 v8, 0
372 ; GISEL10-NEXT: v_mov_b32_e32 v9, 0
373 ; GISEL10-NEXT: v_mov_b32_e32 v10, 0
374 ; GISEL10-NEXT: v_mov_b32_e32 v11, 0
375 ; GISEL10-NEXT: s_mov_b64 s[0:1], s[48:49]
376 ; GISEL10-NEXT: s_mov_b64 s[2:3], s[50:51]
377 ; GISEL10-NEXT: s_waitcnt lgkmcnt(0)
378 ; GISEL10-NEXT: s_swappc_b64 s[30:31], s[4:5]
379 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
380 ; GISEL10-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
381 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
382 ; GISEL10-NEXT: v_mov_b32_e32 v0, v12
383 ; GISEL10-NEXT: global_store_dword v[41:42], v0, off
384 ; GISEL10-NEXT: s_endpgm
385 ;
386 ; DAGISEL10-LABEL: set_inactive_chain_arg_call:
387 ; DAGISEL10: ; %bb.0:
388 ; DAGISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
389 ; DAGISEL10-NEXT: s_mov_b32 s32, 0
390 ; DAGISEL10-NEXT: v_mov_b32_e32 v43, v11
391 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
392 ; DAGISEL10-NEXT: v_mov_b32_e32 v40, v10
393 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
394 ; DAGISEL10-NEXT: s_getpc_b64 s[0:1]
395 ; DAGISEL10-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
396 ; DAGISEL10-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
397 ; DAGISEL10-NEXT: v_mov_b32_e32 v42, v9
398 ; DAGISEL10-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
399 ; DAGISEL10-NEXT: v_mov_b32_e32 v41, v8
400 ; DAGISEL10-NEXT: v_mov_b32_e32 v0, 0
401 ; DAGISEL10-NEXT: v_mov_b32_e32 v1, 0
402 ; DAGISEL10-NEXT: v_mov_b32_e32 v2, 0
403 ; DAGISEL10-NEXT: v_mov_b32_e32 v3, 0
404 ; DAGISEL10-NEXT: v_mov_b32_e32 v4, 0
405 ; DAGISEL10-NEXT: v_mov_b32_e32 v5, 0
406 ; DAGISEL10-NEXT: v_mov_b32_e32 v6, 0
407 ; DAGISEL10-NEXT: v_mov_b32_e32 v7, 0
408 ; DAGISEL10-NEXT: v_mov_b32_e32 v8, 0
409 ; DAGISEL10-NEXT: v_mov_b32_e32 v9, 0
410 ; DAGISEL10-NEXT: v_mov_b32_e32 v10, 0
411 ; DAGISEL10-NEXT: v_mov_b32_e32 v11, 0
412 ; DAGISEL10-NEXT: s_mov_b64 s[0:1], s[48:49]
413 ; DAGISEL10-NEXT: s_mov_b64 s[2:3], s[50:51]
414 ; DAGISEL10-NEXT: s_waitcnt lgkmcnt(0)
415 ; DAGISEL10-NEXT: s_swappc_b64 s[30:31], s[4:5]
416 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
417 ; DAGISEL10-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
418 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
419 ; DAGISEL10-NEXT: v_mov_b32_e32 v0, v12
420 ; DAGISEL10-NEXT: global_store_dword v[41:42], v0, off
421 ; DAGISEL10-NEXT: s_endpgm
422 ;
423 ; GISEL11_W64-LABEL: set_inactive_chain_arg_call:
424 ; GISEL11_W64: ; %bb.0:
425 ; GISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
426 ; GISEL11_W64-NEXT: s_mov_b32 s32, 0
427 ; GISEL11_W64-NEXT: v_mov_b32_e32 v41, v8
428 ; GISEL11_W64-NEXT: v_mov_b32_e32 v42, v9
429 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
430 ; GISEL11_W64-NEXT: v_mov_b32_e32 v40, v10
431 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
432 ; GISEL11_W64-NEXT: s_getpc_b64 s[0:1]
433 ; GISEL11_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
434 ; GISEL11_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
435 ; GISEL11_W64-NEXT: v_mov_b32_e32 v43, v11
436 ; GISEL11_W64-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
437 ; GISEL11_W64-NEXT: v_mov_b32_e32 v0, 0
438 ; GISEL11_W64-NEXT: v_mov_b32_e32 v1, 0
439 ; GISEL11_W64-NEXT: v_mov_b32_e32 v2, 0
440 ; GISEL11_W64-NEXT: v_mov_b32_e32 v3, 0
441 ; GISEL11_W64-NEXT: v_mov_b32_e32 v4, 0
442 ; GISEL11_W64-NEXT: v_mov_b32_e32 v5, 0
443 ; GISEL11_W64-NEXT: v_mov_b32_e32 v6, 0
444 ; GISEL11_W64-NEXT: v_mov_b32_e32 v7, 0
445 ; GISEL11_W64-NEXT: v_mov_b32_e32 v8, 0
446 ; GISEL11_W64-NEXT: v_mov_b32_e32 v9, 0
447 ; GISEL11_W64-NEXT: v_mov_b32_e32 v10, 0
448 ; GISEL11_W64-NEXT: v_mov_b32_e32 v11, 0
449 ; GISEL11_W64-NEXT: s_waitcnt lgkmcnt(0)
450 ; GISEL11_W64-NEXT: s_swappc_b64 s[30:31], s[0:1]
451 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
452 ; GISEL11_W64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
453 ; GISEL11_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
454 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
455 ; GISEL11_W64-NEXT: v_mov_b32_e32 v0, v12
456 ; GISEL11_W64-NEXT: global_store_b32 v[41:42], v0, off
457 ; GISEL11_W64-NEXT: s_endpgm
458 ;
459 ; DAGISEL11_W64-LABEL: set_inactive_chain_arg_call:
460 ; DAGISEL11_W64: ; %bb.0:
461 ; DAGISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
462 ; DAGISEL11_W64-NEXT: s_mov_b32 s32, 0
463 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v43, v11
464 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
465 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v40, v10
466 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
467 ; DAGISEL11_W64-NEXT: s_getpc_b64 s[0:1]
468 ; DAGISEL11_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
469 ; DAGISEL11_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
470 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v42, v9
471 ; DAGISEL11_W64-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
472 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v41, v8
473 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v0, 0
474 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v1, 0
475 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v2, 0
476 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v3, 0
477 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v4, 0
478 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v5, 0
479 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v6, 0
480 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v7, 0
481 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v8, 0
482 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v9, 0
483 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v10, 0
484 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v11, 0
485 ; DAGISEL11_W64-NEXT: s_waitcnt lgkmcnt(0)
486 ; DAGISEL11_W64-NEXT: s_swappc_b64 s[30:31], s[0:1]
487 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
488 ; DAGISEL11_W64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
489 ; DAGISEL11_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
490 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
491 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v0, v12
492 ; DAGISEL11_W64-NEXT: global_store_b32 v[41:42], v0, off
493 ; DAGISEL11_W64-NEXT: s_endpgm
494 ;
495 ; GISEL10_W64-LABEL: set_inactive_chain_arg_call:
496 ; GISEL10_W64: ; %bb.0:
497 ; GISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
498 ; GISEL10_W64-NEXT: s_mov_b32 s32, 0
499 ; GISEL10_W64-NEXT: v_mov_b32_e32 v41, v8
500 ; GISEL10_W64-NEXT: v_mov_b32_e32 v42, v9
501 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
502 ; GISEL10_W64-NEXT: v_mov_b32_e32 v40, v10
503 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
504 ; GISEL10_W64-NEXT: s_getpc_b64 s[0:1]
505 ; GISEL10_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
506 ; GISEL10_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
507 ; GISEL10_W64-NEXT: v_mov_b32_e32 v43, v11
508 ; GISEL10_W64-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
509 ; GISEL10_W64-NEXT: v_mov_b32_e32 v0, 0
510 ; GISEL10_W64-NEXT: v_mov_b32_e32 v1, 0
511 ; GISEL10_W64-NEXT: v_mov_b32_e32 v2, 0
512 ; GISEL10_W64-NEXT: v_mov_b32_e32 v3, 0
513 ; GISEL10_W64-NEXT: v_mov_b32_e32 v4, 0
514 ; GISEL10_W64-NEXT: v_mov_b32_e32 v5, 0
515 ; GISEL10_W64-NEXT: v_mov_b32_e32 v6, 0
516 ; GISEL10_W64-NEXT: v_mov_b32_e32 v7, 0
517 ; GISEL10_W64-NEXT: v_mov_b32_e32 v8, 0
518 ; GISEL10_W64-NEXT: v_mov_b32_e32 v9, 0
519 ; GISEL10_W64-NEXT: v_mov_b32_e32 v10, 0
520 ; GISEL10_W64-NEXT: v_mov_b32_e32 v11, 0
521 ; GISEL10_W64-NEXT: s_mov_b64 s[0:1], s[48:49]
522 ; GISEL10_W64-NEXT: s_mov_b64 s[2:3], s[50:51]
523 ; GISEL10_W64-NEXT: s_waitcnt lgkmcnt(0)
524 ; GISEL10_W64-NEXT: s_swappc_b64 s[30:31], s[4:5]
525 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
526 ; GISEL10_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
527 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
528 ; GISEL10_W64-NEXT: v_mov_b32_e32 v0, v12
529 ; GISEL10_W64-NEXT: global_store_dword v[41:42], v0, off
530 ; GISEL10_W64-NEXT: s_endpgm
531 ;
532 ; DAGISEL10_W64-LABEL: set_inactive_chain_arg_call:
533 ; DAGISEL10_W64: ; %bb.0:
534 ; DAGISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
535 ; DAGISEL10_W64-NEXT: s_mov_b32 s32, 0
536 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v43, v11
537 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
538 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v40, v10
539 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
540 ; DAGISEL10_W64-NEXT: s_getpc_b64 s[0:1]
541 ; DAGISEL10_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
542 ; DAGISEL10_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
543 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v42, v9
544 ; DAGISEL10_W64-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
545 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v41, v8
546 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v0, 0
547 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v1, 0
548 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v2, 0
549 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v3, 0
550 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v4, 0
551 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v5, 0
552 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v6, 0
553 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v7, 0
554 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v8, 0
555 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v9, 0
556 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v10, 0
557 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v11, 0
558 ; DAGISEL10_W64-NEXT: s_mov_b64 s[0:1], s[48:49]
559 ; DAGISEL10_W64-NEXT: s_mov_b64 s[2:3], s[50:51]
560 ; DAGISEL10_W64-NEXT: s_waitcnt lgkmcnt(0)
561 ; DAGISEL10_W64-NEXT: s_swappc_b64 s[30:31], s[4:5]
562 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
563 ; DAGISEL10_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
564 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
565 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v0, v12
566 ; DAGISEL10_W64-NEXT: global_store_dword v[41:42], v0, off
567 ; DAGISEL10_W64-NEXT: s_endpgm
568 call amdgpu_gfx void @gfx_callee(<12 x i32> zeroinitializer)
569 %tmp = call i32 @llvm.amdgcn.set.inactive.chain.arg.i32(i32 %active, i32 %inactive) #0
570 %wwm = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp)
571 store i32 %wwm, ptr addrspace(1) %out
572 ret void
573 }
575 ; When lowering function arguments, SelectionDAG will put the COPY for the last argument first.
576 ; This used to trigger a bug in si-wqm where the first COPY in the entry block was always skipped
577 ; before entering a strict mode, meaning that we'd only copy the active lanes of the last VGPR
578 ; argument, so we'd end up using arbitrary values for the inactive lanes.
579 define amdgpu_cs_chain void @set_inactive_chain_arg_last_vgpr(ptr addrspace(1) %out, i32 %active, i32 %inactive) {
580 ; GISEL11-LABEL: set_inactive_chain_arg_last_vgpr:
581 ; GISEL11: ; %bb.0:
582 ; GISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
583 ; GISEL11-NEXT: s_mov_b32 s32, 0
584 ; GISEL11-NEXT: v_dual_mov_b32 v41, v8 :: v_dual_mov_b32 v42, v9
585 ; GISEL11-NEXT: v_mov_b32_e32 v43, v10
586 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
587 ; GISEL11-NEXT: v_mov_b32_e32 v40, v11
588 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
589 ; GISEL11-NEXT: s_getpc_b64 s[0:1]
590 ; GISEL11-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
591 ; GISEL11-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
592 ; GISEL11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0
593 ; GISEL11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
594 ; GISEL11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v3, 0
595 ; GISEL11-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 0
596 ; GISEL11-NEXT: v_dual_mov_b32 v6, 0 :: v_dual_mov_b32 v7, 0
597 ; GISEL11-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v9, 0
598 ; GISEL11-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v11, 0
599 ; GISEL11-NEXT: s_waitcnt lgkmcnt(0)
600 ; GISEL11-NEXT: s_swappc_b64 s[30:31], s[0:1]
601 ; GISEL11-NEXT: s_or_saveexec_b32 s0, -1
602 ; GISEL11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
603 ; GISEL11-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
604 ; GISEL11-NEXT: s_mov_b32 exec_lo, s0
605 ; GISEL11-NEXT: v_mov_b32_e32 v0, v12
606 ; GISEL11-NEXT: global_store_b32 v[41:42], v0, off
607 ; GISEL11-NEXT: s_endpgm
608 ;
609 ; DAGISEL11-LABEL: set_inactive_chain_arg_last_vgpr:
610 ; DAGISEL11: ; %bb.0:
611 ; DAGISEL11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
612 ; DAGISEL11-NEXT: s_mov_b32 s32, 0
613 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
614 ; DAGISEL11-NEXT: v_mov_b32_e32 v40, v11
615 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
616 ; DAGISEL11-NEXT: s_getpc_b64 s[0:1]
617 ; DAGISEL11-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
618 ; DAGISEL11-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
619 ; DAGISEL11-NEXT: v_dual_mov_b32 v43, v10 :: v_dual_mov_b32 v42, v9
620 ; DAGISEL11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
621 ; DAGISEL11-NEXT: v_dual_mov_b32 v41, v8 :: v_dual_mov_b32 v0, 0
622 ; DAGISEL11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
623 ; DAGISEL11-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v4, 0
624 ; DAGISEL11-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_mov_b32 v6, 0
625 ; DAGISEL11-NEXT: v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v8, 0
626 ; DAGISEL11-NEXT: v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v10, 0
627 ; DAGISEL11-NEXT: v_mov_b32_e32 v11, 0
628 ; DAGISEL11-NEXT: s_waitcnt lgkmcnt(0)
629 ; DAGISEL11-NEXT: s_swappc_b64 s[30:31], s[0:1]
630 ; DAGISEL11-NEXT: s_or_saveexec_b32 s0, -1
631 ; DAGISEL11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
632 ; DAGISEL11-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
633 ; DAGISEL11-NEXT: s_mov_b32 exec_lo, s0
634 ; DAGISEL11-NEXT: v_mov_b32_e32 v0, v12
635 ; DAGISEL11-NEXT: global_store_b32 v[41:42], v0, off
636 ; DAGISEL11-NEXT: s_endpgm
637 ;
638 ; GISEL10-LABEL: set_inactive_chain_arg_last_vgpr:
639 ; GISEL10: ; %bb.0:
640 ; GISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
641 ; GISEL10-NEXT: s_mov_b32 s32, 0
642 ; GISEL10-NEXT: v_mov_b32_e32 v41, v8
643 ; GISEL10-NEXT: v_mov_b32_e32 v42, v9
644 ; GISEL10-NEXT: v_mov_b32_e32 v43, v10
645 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
646 ; GISEL10-NEXT: v_mov_b32_e32 v40, v11
647 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
648 ; GISEL10-NEXT: s_getpc_b64 s[0:1]
649 ; GISEL10-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
650 ; GISEL10-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
651 ; GISEL10-NEXT: v_mov_b32_e32 v0, 0
652 ; GISEL10-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
653 ; GISEL10-NEXT: v_mov_b32_e32 v1, 0
654 ; GISEL10-NEXT: v_mov_b32_e32 v2, 0
655 ; GISEL10-NEXT: v_mov_b32_e32 v3, 0
656 ; GISEL10-NEXT: v_mov_b32_e32 v4, 0
657 ; GISEL10-NEXT: v_mov_b32_e32 v5, 0
658 ; GISEL10-NEXT: v_mov_b32_e32 v6, 0
659 ; GISEL10-NEXT: v_mov_b32_e32 v7, 0
660 ; GISEL10-NEXT: v_mov_b32_e32 v8, 0
661 ; GISEL10-NEXT: v_mov_b32_e32 v9, 0
662 ; GISEL10-NEXT: v_mov_b32_e32 v10, 0
663 ; GISEL10-NEXT: v_mov_b32_e32 v11, 0
664 ; GISEL10-NEXT: s_mov_b64 s[0:1], s[48:49]
665 ; GISEL10-NEXT: s_mov_b64 s[2:3], s[50:51]
666 ; GISEL10-NEXT: s_waitcnt lgkmcnt(0)
667 ; GISEL10-NEXT: s_swappc_b64 s[30:31], s[4:5]
668 ; GISEL10-NEXT: s_or_saveexec_b32 s0, -1
669 ; GISEL10-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
670 ; GISEL10-NEXT: s_mov_b32 exec_lo, s0
671 ; GISEL10-NEXT: v_mov_b32_e32 v0, v12
672 ; GISEL10-NEXT: global_store_dword v[41:42], v0, off
673 ; GISEL10-NEXT: s_endpgm
674 ;
675 ; DAGISEL10-LABEL: set_inactive_chain_arg_last_vgpr:
676 ; DAGISEL10: ; %bb.0:
677 ; DAGISEL10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
678 ; DAGISEL10-NEXT: s_mov_b32 s32, 0
679 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
680 ; DAGISEL10-NEXT: v_mov_b32_e32 v40, v11
681 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
682 ; DAGISEL10-NEXT: s_getpc_b64 s[0:1]
683 ; DAGISEL10-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
684 ; DAGISEL10-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
685 ; DAGISEL10-NEXT: v_mov_b32_e32 v43, v10
686 ; DAGISEL10-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
687 ; DAGISEL10-NEXT: v_mov_b32_e32 v42, v9
688 ; DAGISEL10-NEXT: v_mov_b32_e32 v41, v8
689 ; DAGISEL10-NEXT: v_mov_b32_e32 v0, 0
690 ; DAGISEL10-NEXT: v_mov_b32_e32 v1, 0
691 ; DAGISEL10-NEXT: v_mov_b32_e32 v2, 0
692 ; DAGISEL10-NEXT: v_mov_b32_e32 v3, 0
693 ; DAGISEL10-NEXT: v_mov_b32_e32 v4, 0
694 ; DAGISEL10-NEXT: v_mov_b32_e32 v5, 0
695 ; DAGISEL10-NEXT: v_mov_b32_e32 v6, 0
696 ; DAGISEL10-NEXT: v_mov_b32_e32 v7, 0
697 ; DAGISEL10-NEXT: v_mov_b32_e32 v8, 0
698 ; DAGISEL10-NEXT: v_mov_b32_e32 v9, 0
699 ; DAGISEL10-NEXT: v_mov_b32_e32 v10, 0
700 ; DAGISEL10-NEXT: v_mov_b32_e32 v11, 0
701 ; DAGISEL10-NEXT: s_mov_b64 s[0:1], s[48:49]
702 ; DAGISEL10-NEXT: s_mov_b64 s[2:3], s[50:51]
703 ; DAGISEL10-NEXT: s_waitcnt lgkmcnt(0)
704 ; DAGISEL10-NEXT: s_swappc_b64 s[30:31], s[4:5]
705 ; DAGISEL10-NEXT: s_or_saveexec_b32 s0, -1
706 ; DAGISEL10-NEXT: v_cndmask_b32_e64 v12, v40, v43, s0
707 ; DAGISEL10-NEXT: s_mov_b32 exec_lo, s0
708 ; DAGISEL10-NEXT: v_mov_b32_e32 v0, v12
709 ; DAGISEL10-NEXT: global_store_dword v[41:42], v0, off
710 ; DAGISEL10-NEXT: s_endpgm
711 ;
712 ; GISEL11_W64-LABEL: set_inactive_chain_arg_last_vgpr:
713 ; GISEL11_W64: ; %bb.0:
714 ; GISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
715 ; GISEL11_W64-NEXT: s_mov_b32 s32, 0
716 ; GISEL11_W64-NEXT: v_mov_b32_e32 v41, v8
717 ; GISEL11_W64-NEXT: v_mov_b32_e32 v42, v9
718 ; GISEL11_W64-NEXT: v_mov_b32_e32 v43, v10
719 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
720 ; GISEL11_W64-NEXT: v_mov_b32_e32 v40, v11
721 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
722 ; GISEL11_W64-NEXT: s_getpc_b64 s[0:1]
723 ; GISEL11_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
724 ; GISEL11_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
725 ; GISEL11_W64-NEXT: v_mov_b32_e32 v0, 0
726 ; GISEL11_W64-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
727 ; GISEL11_W64-NEXT: v_mov_b32_e32 v1, 0
728 ; GISEL11_W64-NEXT: v_mov_b32_e32 v2, 0
729 ; GISEL11_W64-NEXT: v_mov_b32_e32 v3, 0
730 ; GISEL11_W64-NEXT: v_mov_b32_e32 v4, 0
731 ; GISEL11_W64-NEXT: v_mov_b32_e32 v5, 0
732 ; GISEL11_W64-NEXT: v_mov_b32_e32 v6, 0
733 ; GISEL11_W64-NEXT: v_mov_b32_e32 v7, 0
734 ; GISEL11_W64-NEXT: v_mov_b32_e32 v8, 0
735 ; GISEL11_W64-NEXT: v_mov_b32_e32 v9, 0
736 ; GISEL11_W64-NEXT: v_mov_b32_e32 v10, 0
737 ; GISEL11_W64-NEXT: v_mov_b32_e32 v11, 0
738 ; GISEL11_W64-NEXT: s_waitcnt lgkmcnt(0)
739 ; GISEL11_W64-NEXT: s_swappc_b64 s[30:31], s[0:1]
740 ; GISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
741 ; GISEL11_W64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
742 ; GISEL11_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
743 ; GISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
744 ; GISEL11_W64-NEXT: v_mov_b32_e32 v0, v12
745 ; GISEL11_W64-NEXT: global_store_b32 v[41:42], v0, off
746 ; GISEL11_W64-NEXT: s_endpgm
747 ;
748 ; DAGISEL11_W64-LABEL: set_inactive_chain_arg_last_vgpr:
749 ; DAGISEL11_W64: ; %bb.0:
750 ; DAGISEL11_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
751 ; DAGISEL11_W64-NEXT: s_mov_b32 s32, 0
752 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
753 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v40, v11
754 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
755 ; DAGISEL11_W64-NEXT: s_getpc_b64 s[0:1]
756 ; DAGISEL11_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
757 ; DAGISEL11_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
758 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v43, v10
759 ; DAGISEL11_W64-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
760 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v42, v9
761 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v41, v8
762 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v0, 0
763 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v1, 0
764 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v2, 0
765 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v3, 0
766 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v4, 0
767 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v5, 0
768 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v6, 0
769 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v7, 0
770 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v8, 0
771 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v9, 0
772 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v10, 0
773 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v11, 0
774 ; DAGISEL11_W64-NEXT: s_waitcnt lgkmcnt(0)
775 ; DAGISEL11_W64-NEXT: s_swappc_b64 s[30:31], s[0:1]
776 ; DAGISEL11_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
777 ; DAGISEL11_W64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
778 ; DAGISEL11_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
779 ; DAGISEL11_W64-NEXT: s_mov_b64 exec, s[0:1]
780 ; DAGISEL11_W64-NEXT: v_mov_b32_e32 v0, v12
781 ; DAGISEL11_W64-NEXT: global_store_b32 v[41:42], v0, off
782 ; DAGISEL11_W64-NEXT: s_endpgm
783 ;
784 ; GISEL10_W64-LABEL: set_inactive_chain_arg_last_vgpr:
785 ; GISEL10_W64: ; %bb.0:
786 ; GISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
787 ; GISEL10_W64-NEXT: s_mov_b32 s32, 0
788 ; GISEL10_W64-NEXT: v_mov_b32_e32 v41, v8
789 ; GISEL10_W64-NEXT: v_mov_b32_e32 v42, v9
790 ; GISEL10_W64-NEXT: v_mov_b32_e32 v43, v10
791 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
792 ; GISEL10_W64-NEXT: v_mov_b32_e32 v40, v11
793 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
794 ; GISEL10_W64-NEXT: s_getpc_b64 s[0:1]
795 ; GISEL10_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
796 ; GISEL10_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
797 ; GISEL10_W64-NEXT: v_mov_b32_e32 v0, 0
798 ; GISEL10_W64-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
799 ; GISEL10_W64-NEXT: v_mov_b32_e32 v1, 0
800 ; GISEL10_W64-NEXT: v_mov_b32_e32 v2, 0
801 ; GISEL10_W64-NEXT: v_mov_b32_e32 v3, 0
802 ; GISEL10_W64-NEXT: v_mov_b32_e32 v4, 0
803 ; GISEL10_W64-NEXT: v_mov_b32_e32 v5, 0
804 ; GISEL10_W64-NEXT: v_mov_b32_e32 v6, 0
805 ; GISEL10_W64-NEXT: v_mov_b32_e32 v7, 0
806 ; GISEL10_W64-NEXT: v_mov_b32_e32 v8, 0
807 ; GISEL10_W64-NEXT: v_mov_b32_e32 v9, 0
808 ; GISEL10_W64-NEXT: v_mov_b32_e32 v10, 0
809 ; GISEL10_W64-NEXT: v_mov_b32_e32 v11, 0
810 ; GISEL10_W64-NEXT: s_mov_b64 s[0:1], s[48:49]
811 ; GISEL10_W64-NEXT: s_mov_b64 s[2:3], s[50:51]
812 ; GISEL10_W64-NEXT: s_waitcnt lgkmcnt(0)
813 ; GISEL10_W64-NEXT: s_swappc_b64 s[30:31], s[4:5]
814 ; GISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
815 ; GISEL10_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
816 ; GISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
817 ; GISEL10_W64-NEXT: v_mov_b32_e32 v0, v12
818 ; GISEL10_W64-NEXT: global_store_dword v[41:42], v0, off
819 ; GISEL10_W64-NEXT: s_endpgm
820 ;
821 ; DAGISEL10_W64-LABEL: set_inactive_chain_arg_last_vgpr:
822 ; DAGISEL10_W64: ; %bb.0:
823 ; DAGISEL10_W64-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
824 ; DAGISEL10_W64-NEXT: s_mov_b32 s32, 0
825 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
826 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v40, v11
827 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
828 ; DAGISEL10_W64-NEXT: s_getpc_b64 s[0:1]
829 ; DAGISEL10_W64-NEXT: s_add_u32 s0, s0, gfx_callee@gotpcrel32@lo+4
830 ; DAGISEL10_W64-NEXT: s_addc_u32 s1, s1, gfx_callee@gotpcrel32@hi+12
831 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v43, v10
832 ; DAGISEL10_W64-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
833 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v42, v9
834 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v41, v8
835 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v0, 0
836 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v1, 0
837 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v2, 0
838 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v3, 0
839 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v4, 0
840 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v5, 0
841 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v6, 0
842 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v7, 0
843 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v8, 0
844 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v9, 0
845 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v10, 0
846 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v11, 0
847 ; DAGISEL10_W64-NEXT: s_mov_b64 s[0:1], s[48:49]
848 ; DAGISEL10_W64-NEXT: s_mov_b64 s[2:3], s[50:51]
849 ; DAGISEL10_W64-NEXT: s_waitcnt lgkmcnt(0)
850 ; DAGISEL10_W64-NEXT: s_swappc_b64 s[30:31], s[4:5]
851 ; DAGISEL10_W64-NEXT: s_or_saveexec_b64 s[0:1], -1
852 ; DAGISEL10_W64-NEXT: v_cndmask_b32_e64 v12, v40, v43, s[0:1]
853 ; DAGISEL10_W64-NEXT: s_mov_b64 exec, s[0:1]
854 ; DAGISEL10_W64-NEXT: v_mov_b32_e32 v0, v12
855 ; DAGISEL10_W64-NEXT: global_store_dword v[41:42], v0, off
856 ; DAGISEL10_W64-NEXT: s_endpgm
857 call amdgpu_gfx void @gfx_callee(<12 x i32> zeroinitializer)
858 %tmp = call i32 @llvm.amdgcn.set.inactive.chain.arg.i32(i32 %active, i32 %inactive) #0
859 %wwm = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp)
860 store i32 %wwm, ptr addrspace(1) %out
861 ret void
862 }
864 declare i32 @llvm.amdgcn.set.inactive.chain.arg.i32(i32, i32) #0
865 declare i64 @llvm.amdgcn.set.inactive.chain.arg.i64(i64, i64) #0
866 declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32 immarg, i32 immarg, i32 immarg, i1 immarg)
867 declare i32 @llvm.amdgcn.strict.wwm.i32(i32)
868 declare i64 @llvm.amdgcn.strict.wwm.i64(i64)
869 declare amdgpu_gfx void @gfx_callee(<12 x i32>)
871 attributes #0 = { convergent readnone willreturn nocallback nofree}