| blob | a74dbe1de0d39edad0a323fc622794b444c4b600 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
2 ; RUN: llc -O0 -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9-O0 %s
3 ; RUN: llc -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9-O3 %s
5 ; NOTE: llvm.amdgcn.wwm is deprecated, use llvm.amdgcn.strict.wwm instead.
7 define amdgpu_cs void @no_cfg(ptr addrspace(8) inreg %tmp14) {
8 ; GFX9-O0-LABEL: no_cfg:
9 ; GFX9-O0: ; %bb.0:
10 ; GFX9-O0-NEXT: s_mov_b32 s6, s2
11 ; GFX9-O0-NEXT: s_mov_b32 s4, s0
12 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
13 ; GFX9-O0-NEXT: s_mov_b32 s7, s3
14 ; GFX9-O0-NEXT: s_mov_b32 s8, s7
15 ; GFX9-O0-NEXT: s_mov_b32 s9, s6
16 ; GFX9-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
17 ; GFX9-O0-NEXT: s_mov_b32 s5, s1
18 ; GFX9-O0-NEXT: s_mov_b32 s10, s5
19 ; GFX9-O0-NEXT: s_mov_b32 s0, s4
20 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
21 ; GFX9-O0-NEXT: s_mov_b32 s1, s10
22 ; GFX9-O0-NEXT: s_mov_b32 s2, s9
23 ; GFX9-O0-NEXT: s_mov_b32 s3, s8
24 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
25 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[5:6], off, s[0:3], s4
26 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
27 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v5
28 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v6
29 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
30 ; GFX9-O0-NEXT: s_not_b64 exec, exec
31 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4
32 ; GFX9-O0-NEXT: s_not_b64 exec, exec
33 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
34 ; GFX9-O0-NEXT: s_not_b64 exec, exec
35 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s4
36 ; GFX9-O0-NEXT: s_not_b64 exec, exec
37 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
38 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s4
39 ; GFX9-O0-NEXT: s_nop 1
40 ; GFX9-O0-NEXT: v_mov_b32_dpp v2, v0 row_bcast:31 row_mask:0xc bank_mask:0xf
41 ; GFX9-O0-NEXT: v_add_u32_e64 v0, v0, v2
42 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
43 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
44 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
45 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4
46 ; GFX9-O0-NEXT: s_nop 1
47 ; GFX9-O0-NEXT: v_mov_b32_dpp v0, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
48 ; GFX9-O0-NEXT: v_add_u32_e64 v0, v1, v0
49 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
50 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v0
51 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v3, v4
52 ; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[6:7]
53 ; GFX9-O0-NEXT: s_mov_b32 s5, 1
54 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s5, v3
55 ; GFX9-O0-NEXT: s_mov_b32 s5, 2
56 ; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s5
57 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s4 offset:4
58 ; GFX9-O0-NEXT: s_endpgm
59 ;
60 ; GFX9-O3-LABEL: no_cfg:
61 ; GFX9-O3: ; %bb.0:
62 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[4:5], off, s[0:3], 0
63 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
64 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, 0
65 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
66 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
67 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
68 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v4
69 ; GFX9-O3-NEXT: s_not_b64 exec, exec
70 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
71 ; GFX9-O3-NEXT: s_not_b64 exec, exec
72 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v5
73 ; GFX9-O3-NEXT: s_not_b64 exec, exec
74 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, 0
75 ; GFX9-O3-NEXT: s_not_b64 exec, exec
76 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
77 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
78 ; GFX9-O3-NEXT: v_mov_b32_dpp v0, v3 row_bcast:31 row_mask:0xc bank_mask:0xf
79 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
80 ; GFX9-O3-NEXT: v_add_u32_e32 v0, v3, v0
81 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
82 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
83 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
84 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
85 ; GFX9-O3-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
86 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 1, v4
87 ; GFX9-O3-NEXT: v_and_b32_e32 v4, 2, v4
88 ; GFX9-O3-NEXT: buffer_store_dword v4, off, s[0:3], 0 offset:4
89 ; GFX9-O3-NEXT: s_endpgm
90 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
91 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
92 %tmp102 = extractelement <2 x i32> %tmp101, i32 0
93 %tmp103 = extractelement <2 x i32> %tmp101, i32 1
94 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
95 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
98 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
99 %tmp121 = add i32 %tmp105, %tmp120
100 %tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
102 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
103 %tmp136 = add i32 %tmp107, %tmp135
104 %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
106 %tmp138 = icmp eq i32 %tmp122, %tmp137
107 %tmp139 = sext i1 %tmp138 to i32
108 %tmp140 = shl nsw i32 %tmp139, 1
109 %tmp141 = and i32 %tmp140, 2
110 %tmp145 = bitcast i32 %tmp141 to float
111 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
112 ret void
113 }
115 define amdgpu_cs void @cfg(ptr addrspace(8) inreg %tmp14, i32 %arg) {
116 ; GFX9-O0-LABEL: cfg:
117 ; GFX9-O0: ; %bb.0: ; %entry
118 ; GFX9-O0-NEXT: s_mov_b32 s16, SCRATCH_RSRC_DWORD0
119 ; GFX9-O0-NEXT: s_mov_b32 s17, SCRATCH_RSRC_DWORD1
120 ; GFX9-O0-NEXT: s_mov_b32 s18, -1
121 ; GFX9-O0-NEXT: s_mov_b32 s19, 0xe00000
122 ; GFX9-O0-NEXT: s_add_u32 s16, s16, s4
123 ; GFX9-O0-NEXT: s_addc_u32 s17, s17, 0
124 ; GFX9-O0-NEXT: ; implicit-def: $vgpr3 : SGPR spill to VGPR lane
125 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
126 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
127 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 ; 4-byte Folded Reload
128 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
129 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
130 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 0
131 ; GFX9-O0-NEXT: s_mov_b32 s4, s1
132 ; GFX9-O0-NEXT: v_readlane_b32 s1, v0, 0
133 ; GFX9-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
134 ; GFX9-O0-NEXT: s_mov_b32 s3, s1
135 ; GFX9-O0-NEXT: s_mov_b32 s8, s3
136 ; GFX9-O0-NEXT: s_mov_b32 s9, s2
137 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
138 ; GFX9-O0-NEXT: s_mov_b32 s1, s4
139 ; GFX9-O0-NEXT: s_mov_b32 s10, s1
140 ; GFX9-O0-NEXT: s_mov_b32 s4, s0
141 ; GFX9-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
142 ; GFX9-O0-NEXT: s_mov_b32 s5, s10
143 ; GFX9-O0-NEXT: s_mov_b32 s6, s9
144 ; GFX9-O0-NEXT: s_mov_b32 s7, s8
145 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 1
146 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 2
147 ; GFX9-O0-NEXT: v_writelane_b32 v0, s0, 3
148 ; GFX9-O0-NEXT: v_writelane_b32 v0, s1, 4
149 ; GFX9-O0-NEXT: s_mov_b32 s0, 0
150 ; GFX9-O0-NEXT: s_nop 2
151 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[4:5], off, s[4:7], s0
152 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
153 ; GFX9-O0-NEXT: buffer_store_dword v4, off, s[16:19], 0 offset:12 ; 4-byte Folded Spill
154 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
155 ; GFX9-O0-NEXT: buffer_store_dword v5, off, s[16:19], 0 offset:16 ; 4-byte Folded Spill
156 ; GFX9-O0-NEXT: ; implicit-def: $sgpr2_sgpr3
157 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v4
158 ; GFX9-O0-NEXT: s_not_b64 exec, exec
159 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s0
160 ; GFX9-O0-NEXT: s_not_b64 exec, exec
161 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
162 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s0
163 ; GFX9-O0-NEXT: s_nop 1
164 ; GFX9-O0-NEXT: v_mov_b32_dpp v2, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
165 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v1, v2
166 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
167 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
168 ; GFX9-O0-NEXT: buffer_store_dword v4, off, s[16:19], 0 offset:8 ; 4-byte Folded Spill
169 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v3, s0
170 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, s0
171 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[16:19], 0 offset:4 ; 4-byte Folded Spill
172 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], exec
173 ; GFX9-O0-NEXT: v_writelane_b32 v0, s0, 5
174 ; GFX9-O0-NEXT: v_writelane_b32 v0, s1, 6
175 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
176 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 ; 4-byte Folded Spill
177 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
178 ; GFX9-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
179 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
180 ; GFX9-O0-NEXT: s_cbranch_execz .LBB1_2
181 ; GFX9-O0-NEXT: ; %bb.1: ; %if
182 ; GFX9-O0-NEXT: buffer_load_dword v3, off, s[16:19], 0 offset:12 ; 4-byte Folded Reload
183 ; GFX9-O0-NEXT: buffer_load_dword v4, off, s[16:19], 0 offset:16 ; 4-byte Folded Reload
184 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
185 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
186 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[0:1], -1
187 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, 0
188 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
189 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
190 ; GFX9-O0-NEXT: s_not_b64 exec, exec
191 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v1
192 ; GFX9-O0-NEXT: s_not_b64 exec, exec
193 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[0:1], -1
194 ; GFX9-O0-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
195 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v2, v1
196 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
197 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1
198 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:4 ; 4-byte Folded Spill
199 ; GFX9-O0-NEXT: .LBB1_2: ; %merge
200 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
201 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 ; 4-byte Folded Reload
202 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
203 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
204 ; GFX9-O0-NEXT: v_readlane_b32 s4, v0, 5
205 ; GFX9-O0-NEXT: v_readlane_b32 s5, v0, 6
206 ; GFX9-O0-NEXT: s_or_b64 exec, exec, s[4:5]
207 ; GFX9-O0-NEXT: v_readlane_b32 s2, v0, 1
208 ; GFX9-O0-NEXT: v_readlane_b32 s3, v0, 2
209 ; GFX9-O0-NEXT: v_readlane_b32 s0, v0, 3
210 ; GFX9-O0-NEXT: v_readlane_b32 s1, v0, 4
211 ; GFX9-O0-NEXT: buffer_load_dword v3, off, s[16:19], 0 offset:8 ; 4-byte Folded Reload
212 ; GFX9-O0-NEXT: buffer_load_dword v4, off, s[16:19], 0 offset:4 ; 4-byte Folded Reload
213 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
214 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v3, v4
215 ; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[4:5]
216 ; GFX9-O0-NEXT: s_mov_b32 s4, 1
217 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s4, v3
218 ; GFX9-O0-NEXT: s_mov_b32 s4, 2
219 ; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s4
220 ; GFX9-O0-NEXT: s_mov_b32 s6, s1
221 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
222 ; GFX9-O0-NEXT: s_mov_b32 s4, s3
223 ; GFX9-O0-NEXT: s_mov_b32 s5, s2
224 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
225 ; GFX9-O0-NEXT: s_mov_b32 s1, s6
226 ; GFX9-O0-NEXT: s_mov_b32 s2, s5
227 ; GFX9-O0-NEXT: s_mov_b32 s3, s4
228 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
229 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s4 offset:4
230 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
231 ; GFX9-O0-NEXT: s_endpgm
232 ;
233 ; GFX9-O3-LABEL: cfg:
234 ; GFX9-O3: ; %bb.0: ; %entry
235 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[3:4], off, s[0:3], 0
236 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, 0
237 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
238 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
239 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
240 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
241 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v3
242 ; GFX9-O3-NEXT: s_not_b64 exec, exec
243 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
244 ; GFX9-O3-NEXT: s_not_b64 exec, exec
245 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
246 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
247 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
248 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
249 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v1
250 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
251 ; GFX9-O3-NEXT: s_and_saveexec_b64 s[4:5], vcc
252 ; GFX9-O3-NEXT: s_cbranch_execz .LBB1_2
253 ; GFX9-O3-NEXT: ; %bb.1: ; %if
254 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[6:7], -1
255 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
256 ; GFX9-O3-NEXT: s_mov_b64 exec, s[6:7]
257 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v4
258 ; GFX9-O3-NEXT: s_not_b64 exec, exec
259 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
260 ; GFX9-O3-NEXT: s_not_b64 exec, exec
261 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[6:7], -1
262 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
263 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
264 ; GFX9-O3-NEXT: s_mov_b64 exec, s[6:7]
265 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v1
266 ; GFX9-O3-NEXT: .LBB1_2: ; %merge
267 ; GFX9-O3-NEXT: s_or_b64 exec, exec, s[4:5]
268 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
269 ; GFX9-O3-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
270 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 1, v0
271 ; GFX9-O3-NEXT: v_and_b32_e32 v0, 2, v0
272 ; GFX9-O3-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4
273 ; GFX9-O3-NEXT: s_endpgm
274 entry:
275 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
276 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
277 %tmp102 = extractelement <2 x i32> %tmp101, i32 0
278 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
280 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
281 %tmp121 = add i32 %tmp105, %tmp120
282 %tmp122 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp121)
284 %cond = icmp eq i32 %arg, 0
285 br i1 %cond, label %if, label %merge
286 if:
287 %tmp103 = extractelement <2 x i32> %tmp101, i32 1
288 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
290 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
291 %tmp136 = add i32 %tmp107, %tmp135
292 %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
293 br label %merge
295 merge:
296 %merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]
297 %tmp138 = icmp eq i32 %tmp122, %merge_value
298 %tmp139 = sext i1 %tmp138 to i32
299 %tmp140 = shl nsw i32 %tmp139, 1
300 %tmp141 = and i32 %tmp140, 2
301 %tmp145 = bitcast i32 %tmp141 to float
302 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
303 ret void
304 }
306 define hidden i32 @called(i32 %a) noinline {
307 ; GFX9-O0-LABEL: called:
308 ; GFX9-O0: ; %bb.0:
309 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
310 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v0, v0
311 ; GFX9-O0-NEXT: v_mul_lo_u32 v0, v1, v0
312 ; GFX9-O0-NEXT: v_sub_u32_e64 v0, v0, v1
313 ; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
314 ;
315 ; GFX9-O3-LABEL: called:
316 ; GFX9-O3: ; %bb.0:
317 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
318 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v0, v0
319 ; GFX9-O3-NEXT: v_mul_lo_u32 v0, v1, v0
320 ; GFX9-O3-NEXT: v_sub_u32_e32 v0, v0, v1
321 ; GFX9-O3-NEXT: s_setpc_b64 s[30:31]
322 %add = add i32 %a, %a
323 %mul = mul i32 %add, %a
324 %sub = sub i32 %mul, %add
325 ret i32 %sub
326 }
328 define amdgpu_kernel void @call(ptr addrspace(8) inreg %tmp14, i32 inreg %arg) {
329 ; GFX9-O0-LABEL: call:
330 ; GFX9-O0: ; %bb.0:
331 ; GFX9-O0-NEXT: s_mov_b32 s32, 0x400
332 ; GFX9-O0-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
333 ; GFX9-O0-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
334 ; GFX9-O0-NEXT: s_mov_b32 s26, -1
335 ; GFX9-O0-NEXT: s_mov_b32 s27, 0xe00000
336 ; GFX9-O0-NEXT: s_add_u32 s24, s24, s9
337 ; GFX9-O0-NEXT: s_addc_u32 s25, s25, 0
338 ; GFX9-O0-NEXT: ; implicit-def: $vgpr7 : SGPR spill to VGPR lane
339 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
340 ; GFX9-O0-NEXT: v_writelane_b32 v7, s10, 0
341 ; GFX9-O0-NEXT: v_writelane_b32 v7, s11, 1
342 ; GFX9-O0-NEXT: s_mov_b32 s14, s8
343 ; GFX9-O0-NEXT: s_mov_b32 s13, s7
344 ; GFX9-O0-NEXT: s_mov_b32 s12, s6
345 ; GFX9-O0-NEXT: s_mov_b64 s[10:11], s[4:5]
346 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[2:3]
347 ; GFX9-O0-NEXT: v_readlane_b32 s2, v7, 0
348 ; GFX9-O0-NEXT: v_readlane_b32 s3, v7, 1
349 ; GFX9-O0-NEXT: v_writelane_b32 v7, s4, 2
350 ; GFX9-O0-NEXT: v_writelane_b32 v7, s5, 3
351 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
352 ; GFX9-O0-NEXT: buffer_store_dword v7, off, s[24:27], 0 ; 4-byte Folded Spill
353 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
354 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[0:1]
355 ; GFX9-O0-NEXT: v_readlane_b32 s0, v7, 2
356 ; GFX9-O0-NEXT: v_readlane_b32 s1, v7, 3
357 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
358 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
359 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
360 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
361 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
362 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
363 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
364 ; GFX9-O0-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x24
365 ; GFX9-O0-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x2c
366 ; GFX9-O0-NEXT: s_load_dword s2, s[0:1], 0x34
367 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
368 ; GFX9-O0-NEXT: s_mov_b32 s3, s7
369 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
370 ; GFX9-O0-NEXT: s_mov_b32 s7, s9
371 ; GFX9-O0-NEXT: s_mov_b32 s16, s8
372 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17_sgpr18_sgpr19
373 ; GFX9-O0-NEXT: s_mov_b32 s17, s7
374 ; GFX9-O0-NEXT: s_mov_b32 s18, s6
375 ; GFX9-O0-NEXT: s_mov_b32 s19, s3
376 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
377 ; GFX9-O0-NEXT: v_writelane_b32 v0, s16, 4
378 ; GFX9-O0-NEXT: v_writelane_b32 v0, s17, 5
379 ; GFX9-O0-NEXT: v_writelane_b32 v0, s18, 6
380 ; GFX9-O0-NEXT: v_writelane_b32 v0, s19, 7
381 ; GFX9-O0-NEXT: s_mov_b32 s3, 0
382 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 8
383 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, s2
384 ; GFX9-O0-NEXT: s_not_b64 exec, exec
385 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, s3
386 ; GFX9-O0-NEXT: s_not_b64 exec, exec
387 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
388 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 9
389 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 10
390 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
391 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[24:27], 0 ; 4-byte Folded Spill
392 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
393 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 56
394 ; GFX9-O0-NEXT: s_mov_b32 s2, s0
395 ; GFX9-O0-NEXT: s_mov_b32 s0, s1
396 ; GFX9-O0-NEXT: s_mov_b32 s3, s6
397 ; GFX9-O0-NEXT: s_mov_b32 s1, s7
398 ; GFX9-O0-NEXT: s_add_u32 s8, s2, s3
399 ; GFX9-O0-NEXT: s_addc_u32 s0, s0, s1
400 ; GFX9-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
401 ; GFX9-O0-NEXT: s_mov_b32 s9, s0
402 ; GFX9-O0-NEXT: s_getpc_b64 s[16:17]
403 ; GFX9-O0-NEXT: s_add_u32 s16, s16, called@rel32@lo+4
404 ; GFX9-O0-NEXT: s_addc_u32 s17, s17, called@rel32@hi+12
405 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[24:25]
406 ; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[26:27]
407 ; GFX9-O0-NEXT: s_mov_b32 s6, 20
408 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s6, v3
409 ; GFX9-O0-NEXT: s_mov_b32 s6, 10
410 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v4, s6, v4
411 ; GFX9-O0-NEXT: v_or3_b32 v3, v5, v4, v3
412 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6_sgpr7
413 ; GFX9-O0-NEXT: ; implicit-def: $sgpr15
414 ; GFX9-O0-NEXT: v_mov_b32_e32 v31, v3
415 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v6
416 ; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[16:17]
417 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
418 ; GFX9-O0-NEXT: buffer_load_dword v1, off, s[24:27], 0 ; 4-byte Folded Reload
419 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
420 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
421 ; GFX9-O0-NEXT: v_readlane_b32 s0, v1, 4
422 ; GFX9-O0-NEXT: v_readlane_b32 s1, v1, 5
423 ; GFX9-O0-NEXT: v_readlane_b32 s2, v1, 6
424 ; GFX9-O0-NEXT: v_readlane_b32 s3, v1, 7
425 ; GFX9-O0-NEXT: v_readlane_b32 s6, v1, 9
426 ; GFX9-O0-NEXT: v_readlane_b32 s7, v1, 10
427 ; GFX9-O0-NEXT: v_readlane_b32 s4, v1, 8
428 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
429 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
430 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
431 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
432 ; GFX9-O0-NEXT: v_add_u32_e64 v3, v3, v6
433 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
434 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
435 ; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s4 offset:4
436 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
437 ; GFX9-O0-NEXT: s_endpgm
438 ;
439 ; GFX9-O3-LABEL: call:
440 ; GFX9-O3: ; %bb.0:
441 ; GFX9-O3-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
442 ; GFX9-O3-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
443 ; GFX9-O3-NEXT: s_mov_b32 s26, -1
444 ; GFX9-O3-NEXT: s_mov_b32 s27, 0xe00000
445 ; GFX9-O3-NEXT: s_add_u32 s24, s24, s9
446 ; GFX9-O3-NEXT: s_mov_b32 s32, 0
447 ; GFX9-O3-NEXT: s_addc_u32 s25, s25, 0
448 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[12:13], -1
449 ; GFX9-O3-NEXT: s_mov_b32 s14, s8
450 ; GFX9-O3-NEXT: s_mov_b64 s[10:11], s[4:5]
451 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v2
452 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
453 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
454 ; GFX9-O3-NEXT: s_mov_b64 exec, s[12:13]
455 ; GFX9-O3-NEXT: s_load_dword s4, s[2:3], 0x34
456 ; GFX9-O3-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
457 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
458 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s4
459 ; GFX9-O3-NEXT: s_not_b64 exec, exec
460 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
461 ; GFX9-O3-NEXT: s_not_b64 exec, exec
462 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[20:21], -1
463 ; GFX9-O3-NEXT: s_add_u32 s8, s2, 56
464 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v3, 20, v3
465 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 10, v4
466 ; GFX9-O3-NEXT: s_addc_u32 s9, s3, 0
467 ; GFX9-O3-NEXT: v_or3_b32 v3, v5, v4, v3
468 ; GFX9-O3-NEXT: s_mov_b64 s[4:5], s[0:1]
469 ; GFX9-O3-NEXT: s_mov_b64 s[0:1], s[24:25]
470 ; GFX9-O3-NEXT: s_mov_b32 s12, s6
471 ; GFX9-O3-NEXT: s_mov_b32 s13, s7
472 ; GFX9-O3-NEXT: v_mov_b32_e32 v31, v3
473 ; GFX9-O3-NEXT: s_mov_b64 s[2:3], s[26:27]
474 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v6
475 ; GFX9-O3-NEXT: s_getpc_b64 s[22:23]
476 ; GFX9-O3-NEXT: s_add_u32 s22, s22, called@rel32@lo+4
477 ; GFX9-O3-NEXT: s_addc_u32 s23, s23, called@rel32@hi+12
478 ; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[22:23]
479 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v0
480 ; GFX9-O3-NEXT: v_add_u32_e32 v3, v3, v6
481 ; GFX9-O3-NEXT: s_mov_b64 exec, s[20:21]
482 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v3
483 ; GFX9-O3-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:4
484 ; GFX9-O3-NEXT: s_endpgm
487 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)
488 %tmp134 = call i32 @called(i32 %tmp107)
489 %tmp136 = add i32 %tmp134, %tmp107
490 %tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
491 call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
492 ret void
493 }
495 define i64 @called_i64(i64 %a) noinline {
496 ; GFX9-O0-LABEL: called_i64:
497 ; GFX9-O0: ; %bb.0:
498 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
499 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
500 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4
501 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4
502 ; GFX9-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
503 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1
504 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
505 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2
506 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v2
507 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v3
508 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
509 ; GFX9-O0-NEXT: v_add_co_u32_e64 v4, s[4:5], v4, v5
510 ; GFX9-O0-NEXT: v_addc_co_u32_e64 v0, s[4:5], v0, v1, s[4:5]
511 ; GFX9-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
512 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
513 ; GFX9-O0-NEXT: s_mov_b32 s4, 32
514 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2
515 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
516 ; GFX9-O0-NEXT: v_lshrrev_b64 v[0:1], s4, v[0:1]
517 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v0
518 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
519 ; GFX9-O0-NEXT: v_mul_lo_u32 v1, v0, v1
520 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
521 ; GFX9-O0-NEXT: v_lshrrev_b64 v[6:7], s4, v[4:5]
522 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
523 ; GFX9-O0-NEXT: v_mul_lo_u32 v2, v2, v3
524 ; GFX9-O0-NEXT: v_mad_u64_u32 v[6:7], s[6:7], v0, v3, 0
525 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
526 ; GFX9-O0-NEXT: v_add3_u32 v0, v0, v1, v2
527 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
528 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
529 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
530 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s5
531 ; GFX9-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
532 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v2
533 ; GFX9-O0-NEXT: v_lshlrev_b64 v[1:2], s4, v[0:1]
534 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
535 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec
536 ; GFX9-O0-NEXT: s_mov_b32 s5, 0
537 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
538 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, 0
539 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
540 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0
541 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
542 ; GFX9-O0-NEXT: v_or_b32_e64 v0, v0, v3
543 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v1
544 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6
545 ; GFX9-O0-NEXT: v_or_b32_e64 v6, v1, v2
546 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
547 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0
548 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6
549 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v4
550 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
551 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v5
552 ; GFX9-O0-NEXT: v_sub_co_u32_e64 v1, s[6:7], v1, v3
553 ; GFX9-O0-NEXT: v_subb_co_u32_e64 v0, s[6:7], v0, v2, s[6:7]
554 ; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
555 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
556 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1
557 ; GFX9-O0-NEXT: v_lshrrev_b64 v[1:2], s4, v[1:2]
558 ; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec
559 ; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
560 ;
561 ; GFX9-O3-LABEL: called_i64:
562 ; GFX9-O3: ; %bb.0:
563 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
564 ; GFX9-O3-NEXT: v_add_co_u32_e32 v2, vcc, v0, v0
565 ; GFX9-O3-NEXT: v_addc_co_u32_e32 v3, vcc, v1, v1, vcc
566 ; GFX9-O3-NEXT: v_mul_lo_u32 v4, v3, v0
567 ; GFX9-O3-NEXT: v_mul_lo_u32 v5, v2, v1
568 ; GFX9-O3-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, 0
569 ; GFX9-O3-NEXT: v_add3_u32 v1, v1, v5, v4
570 ; GFX9-O3-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v2
571 ; GFX9-O3-NEXT: v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
572 ; GFX9-O3-NEXT: s_setpc_b64 s[30:31]
573 %add = add i64 %a, %a
574 %mul = mul i64 %add, %a
575 %sub = sub i64 %mul, %add
576 ret i64 %sub
577 }
579 define amdgpu_kernel void @call_i64(ptr addrspace(8) inreg %tmp14, i64 inreg %arg) {
580 ; GFX9-O0-LABEL: call_i64:
581 ; GFX9-O0: ; %bb.0:
582 ; GFX9-O0-NEXT: s_mov_b32 s32, 0x400
583 ; GFX9-O0-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
584 ; GFX9-O0-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
585 ; GFX9-O0-NEXT: s_mov_b32 s26, -1
586 ; GFX9-O0-NEXT: s_mov_b32 s27, 0xe00000
587 ; GFX9-O0-NEXT: s_add_u32 s24, s24, s9
588 ; GFX9-O0-NEXT: s_addc_u32 s25, s25, 0
589 ; GFX9-O0-NEXT: ; implicit-def: $vgpr12 : SGPR spill to VGPR lane
590 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
591 ; GFX9-O0-NEXT: v_writelane_b32 v12, s10, 0
592 ; GFX9-O0-NEXT: v_writelane_b32 v12, s11, 1
593 ; GFX9-O0-NEXT: s_mov_b32 s14, s8
594 ; GFX9-O0-NEXT: s_mov_b32 s13, s7
595 ; GFX9-O0-NEXT: s_mov_b32 s12, s6
596 ; GFX9-O0-NEXT: s_mov_b64 s[10:11], s[4:5]
597 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[2:3]
598 ; GFX9-O0-NEXT: v_readlane_b32 s2, v12, 0
599 ; GFX9-O0-NEXT: v_readlane_b32 s3, v12, 1
600 ; GFX9-O0-NEXT: v_writelane_b32 v12, s4, 2
601 ; GFX9-O0-NEXT: v_writelane_b32 v12, s5, 3
602 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
603 ; GFX9-O0-NEXT: buffer_store_dword v12, off, s[24:27], 0 ; 4-byte Folded Spill
604 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
605 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[0:1]
606 ; GFX9-O0-NEXT: v_readlane_b32 s0, v12, 2
607 ; GFX9-O0-NEXT: v_readlane_b32 s1, v12, 3
608 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
609 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
610 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
611 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
612 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
613 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
614 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
615 ; GFX9-O0-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x24
616 ; GFX9-O0-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x2c
617 ; GFX9-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
618 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
619 ; GFX9-O0-NEXT: s_mov_b32 s6, s9
620 ; GFX9-O0-NEXT: s_mov_b32 s7, s8
621 ; GFX9-O0-NEXT: s_mov_b32 s8, s17
622 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 killed $sgpr16_sgpr17
623 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17_sgpr18_sgpr19
624 ; GFX9-O0-NEXT: s_mov_b32 s17, s8
625 ; GFX9-O0-NEXT: s_mov_b32 s18, s7
626 ; GFX9-O0-NEXT: s_mov_b32 s19, s6
627 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
628 ; GFX9-O0-NEXT: v_writelane_b32 v0, s16, 4
629 ; GFX9-O0-NEXT: v_writelane_b32 v0, s17, 5
630 ; GFX9-O0-NEXT: v_writelane_b32 v0, s18, 6
631 ; GFX9-O0-NEXT: v_writelane_b32 v0, s19, 7
632 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 0
633 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, s2
634 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, s3
635 ; GFX9-O0-NEXT: s_not_b64 exec, exec
636 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, s6
637 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, s7
638 ; GFX9-O0-NEXT: s_not_b64 exec, exec
639 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
640 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 8
641 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 9
642 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
643 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[24:27], 0 ; 4-byte Folded Spill
644 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
645 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 60
646 ; GFX9-O0-NEXT: s_mov_b32 s2, s0
647 ; GFX9-O0-NEXT: s_mov_b32 s0, s1
648 ; GFX9-O0-NEXT: s_mov_b32 s3, s6
649 ; GFX9-O0-NEXT: s_mov_b32 s1, s7
650 ; GFX9-O0-NEXT: s_add_u32 s8, s2, s3
651 ; GFX9-O0-NEXT: s_addc_u32 s0, s0, s1
652 ; GFX9-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
653 ; GFX9-O0-NEXT: s_mov_b32 s9, s0
654 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v8
655 ; GFX9-O0-NEXT: s_mov_b32 s0, 32
656 ; GFX9-O0-NEXT: ; implicit-def: $sgpr2_sgpr3
657 ; GFX9-O0-NEXT: v_lshrrev_b64 v[10:11], s0, v[8:9]
658 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v10
659 ; GFX9-O0-NEXT: s_getpc_b64 s[0:1]
660 ; GFX9-O0-NEXT: s_add_u32 s0, s0, called_i64@gotpcrel32@lo+4
661 ; GFX9-O0-NEXT: s_addc_u32 s1, s1, called_i64@gotpcrel32@hi+12
662 ; GFX9-O0-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
663 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[24:25]
664 ; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[26:27]
665 ; GFX9-O0-NEXT: s_mov_b32 s6, 20
666 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s6, v3
667 ; GFX9-O0-NEXT: s_mov_b32 s6, 10
668 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v4, s6, v4
669 ; GFX9-O0-NEXT: v_or3_b32 v3, v5, v4, v3
670 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6_sgpr7
671 ; GFX9-O0-NEXT: ; implicit-def: $sgpr15
672 ; GFX9-O0-NEXT: v_mov_b32_e32 v31, v3
673 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v6
674 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v7
675 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
676 ; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[16:17]
677 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
678 ; GFX9-O0-NEXT: buffer_load_dword v2, off, s[24:27], 0 ; 4-byte Folded Reload
679 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
680 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
681 ; GFX9-O0-NEXT: v_readlane_b32 s0, v2, 4
682 ; GFX9-O0-NEXT: v_readlane_b32 s1, v2, 5
683 ; GFX9-O0-NEXT: v_readlane_b32 s2, v2, 6
684 ; GFX9-O0-NEXT: v_readlane_b32 s3, v2, 7
685 ; GFX9-O0-NEXT: v_readlane_b32 s4, v2, 8
686 ; GFX9-O0-NEXT: v_readlane_b32 s5, v2, 9
687 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
688 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
689 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
690 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
691 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
692 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
693 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
694 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v8
695 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v9
696 ; GFX9-O0-NEXT: v_add_co_u32_e64 v3, s[6:7], v3, v5
697 ; GFX9-O0-NEXT: v_addc_co_u32_e64 v4, s[6:7], v4, v6, s[6:7]
698 ; GFX9-O0-NEXT: s_mov_b64 exec, s[4:5]
699 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
700 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v4
701 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
702 ; GFX9-O0-NEXT: buffer_store_dwordx2 v[1:2], off, s[0:3], s4 offset:4
703 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
704 ; GFX9-O0-NEXT: s_endpgm
705 ;
706 ; GFX9-O3-LABEL: call_i64:
707 ; GFX9-O3: ; %bb.0:
708 ; GFX9-O3-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
709 ; GFX9-O3-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
710 ; GFX9-O3-NEXT: s_mov_b32 s26, -1
711 ; GFX9-O3-NEXT: s_mov_b32 s27, 0xe00000
712 ; GFX9-O3-NEXT: s_add_u32 s24, s24, s9
713 ; GFX9-O3-NEXT: s_mov_b32 s32, 0
714 ; GFX9-O3-NEXT: s_addc_u32 s25, s25, 0
715 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[12:13], -1
716 ; GFX9-O3-NEXT: s_mov_b32 s14, s8
717 ; GFX9-O3-NEXT: s_mov_b64 s[10:11], s[4:5]
718 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v2
719 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
720 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
721 ; GFX9-O3-NEXT: s_mov_b64 exec, s[12:13]
722 ; GFX9-O3-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x34
723 ; GFX9-O3-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
724 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
725 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s4
726 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, s5
727 ; GFX9-O3-NEXT: s_not_b64 exec, exec
728 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
729 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, 0
730 ; GFX9-O3-NEXT: s_not_b64 exec, exec
731 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[20:21], -1
732 ; GFX9-O3-NEXT: s_add_u32 s8, s2, 60
733 ; GFX9-O3-NEXT: s_addc_u32 s9, s3, 0
734 ; GFX9-O3-NEXT: s_getpc_b64 s[2:3]
735 ; GFX9-O3-NEXT: s_add_u32 s2, s2, called_i64@gotpcrel32@lo+4
736 ; GFX9-O3-NEXT: s_addc_u32 s3, s3, called_i64@gotpcrel32@hi+12
737 ; GFX9-O3-NEXT: s_load_dwordx2 s[22:23], s[2:3], 0x0
738 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v3, 20, v3
739 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 10, v4
740 ; GFX9-O3-NEXT: v_or3_b32 v3, v5, v4, v3
741 ; GFX9-O3-NEXT: s_mov_b64 s[4:5], s[0:1]
742 ; GFX9-O3-NEXT: s_mov_b64 s[0:1], s[24:25]
743 ; GFX9-O3-NEXT: s_mov_b32 s12, s6
744 ; GFX9-O3-NEXT: s_mov_b32 s13, s7
745 ; GFX9-O3-NEXT: v_mov_b32_e32 v31, v3
746 ; GFX9-O3-NEXT: s_mov_b64 s[2:3], s[26:27]
747 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v6
748 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v7
749 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
750 ; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[22:23]
751 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v0
752 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
753 ; GFX9-O3-NEXT: v_add_co_u32_e32 v3, vcc, v3, v6
754 ; GFX9-O3-NEXT: v_addc_co_u32_e32 v4, vcc, v4, v7, vcc
755 ; GFX9-O3-NEXT: s_mov_b64 exec, s[20:21]
756 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v3
757 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v4
758 ; GFX9-O3-NEXT: buffer_store_dwordx2 v[0:1], off, s[16:19], 0 offset:4
759 ; GFX9-O3-NEXT: s_endpgm
763 %tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)
764 %tmp134 = call i64 @called_i64(i64 %tmp107)
765 %tmp136 = add i64 %tmp134, %tmp107
766 %tmp137 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp136)
767 %tmp138 = bitcast i64 %tmp137 to <2 x i32>
768 call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %tmp138, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
769 ret void
770 }
772 define amdgpu_cs void @_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {
773 ; GFX9-O0-LABEL: _amdgpu_cs_main:
774 ; GFX9-O0: ; %bb.0:
775 ; GFX9-O0-NEXT: s_mov_b32 s4, s3
776 ; GFX9-O0-NEXT: s_mov_b32 s5, s2
777 ; GFX9-O0-NEXT: s_mov_b32 s6, s1
778 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
779 ; GFX9-O0-NEXT: s_mov_b32 s1, s6
780 ; GFX9-O0-NEXT: s_mov_b32 s2, s5
781 ; GFX9-O0-NEXT: s_mov_b32 s3, s4
782 ; GFX9-O0-NEXT: ; kill: def $sgpr4_sgpr5_sgpr6_sgpr7 killed $sgpr0_sgpr1_sgpr2_sgpr3
783 ; GFX9-O0-NEXT: s_mov_b32 s4, 5
784 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v0, s4, v0
785 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
786 ; GFX9-O0-NEXT: buffer_load_dwordx4 v[10:13], v0, s[0:3], s4 offen
787 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[3:4], v0, s[0:3], s4 offen offset:16
788 ; GFX9-O0-NEXT: s_waitcnt vmcnt(1)
789 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v11
790 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v10
791 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
792 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
793 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
794 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v7
795 ; GFX9-O0-NEXT: s_mov_b32 s5, 0x7fffffff
796 ; GFX9-O0-NEXT: s_mov_b32 s6, -1
797 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
798 ; GFX9-O0-NEXT: s_mov_b32 s7, s5
799 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v5
800 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
801 ; GFX9-O0-NEXT: s_not_b64 exec, exec
802 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
803 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
804 ; GFX9-O0-NEXT: s_not_b64 exec, exec
805 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, v1
806 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, v2
807 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v13
808 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v12
809 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
810 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
811 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
812 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v7
813 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v5
814 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
815 ; GFX9-O0-NEXT: s_not_b64 exec, exec
816 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
817 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
818 ; GFX9-O0-NEXT: s_not_b64 exec, exec
819 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v1
820 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v2
821 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
822 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
823 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v4
824 ; GFX9-O0-NEXT: s_not_b64 exec, exec
825 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
826 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
827 ; GFX9-O0-NEXT: s_not_b64 exec, exec
828 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1
829 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2
830 ; GFX9-O0-NEXT: v_mov_b32_e32 v11, v9
831 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v8
832 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, v7
833 ; GFX9-O0-NEXT: v_mov_b32_e32 v10, v6
834 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
835 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
836 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
837 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
838 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6_vgpr7_vgpr8 killed $exec
839 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v11
840 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v10
841 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, v9
842 ; GFX9-O0-NEXT: buffer_store_dwordx4 v[5:8], v0, s[0:3], s4 offen
843 ; GFX9-O0-NEXT: buffer_store_dwordx2 v[3:4], v0, s[0:3], s4 offen offset:16
844 ; GFX9-O0-NEXT: s_endpgm
845 ;
846 ; GFX9-O3-LABEL: _amdgpu_cs_main:
847 ; GFX9-O3: ; %bb.0:
848 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 5, v0
849 ; GFX9-O3-NEXT: buffer_load_dwordx4 v[7:10], v0, s[0:3], 0 offen
850 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[11:12], v0, s[0:3], 0 offen offset:16
851 ; GFX9-O3-NEXT: s_mov_b32 s4, -1
852 ; GFX9-O3-NEXT: s_brev_b32 s5, -2
853 ; GFX9-O3-NEXT: s_waitcnt vmcnt(1)
854 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v7
855 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v8
856 ; GFX9-O3-NEXT: s_not_b64 exec, exec
857 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, s4
858 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, s5
859 ; GFX9-O3-NEXT: s_not_b64 exec, exec
860 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v9
861 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v10
862 ; GFX9-O3-NEXT: s_not_b64 exec, exec
863 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, s4
864 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, s5
865 ; GFX9-O3-NEXT: s_not_b64 exec, exec
866 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
867 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v11
868 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, v12
869 ; GFX9-O3-NEXT: s_not_b64 exec, exec
870 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, s4
871 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s5
872 ; GFX9-O3-NEXT: s_not_b64 exec, exec
873 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, v1
874 ; GFX9-O3-NEXT: v_mov_b32_e32 v9, v3
875 ; GFX9-O3-NEXT: v_mov_b32_e32 v8, v2
876 ; GFX9-O3-NEXT: v_mov_b32_e32 v10, v4
877 ; GFX9-O3-NEXT: v_mov_b32_e32 v11, v5
878 ; GFX9-O3-NEXT: v_mov_b32_e32 v12, v6
879 ; GFX9-O3-NEXT: buffer_store_dwordx4 v[7:10], v0, s[0:3], 0 offen
880 ; GFX9-O3-NEXT: buffer_store_dwordx2 v[11:12], v0, s[0:3], 0 offen offset:16
881 ; GFX9-O3-NEXT: s_endpgm
882 %tmp17 = shl i32 %index, 5
883 %tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)
884 %.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>
885 %tmp19 = or i32 %tmp17, 16
886 %tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)
887 %.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0
888 %tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)
889 %tmp97 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp22)
890 %.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1
891 %tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)
892 %tmp174 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp99)
893 %.i25 = bitcast <2 x i32> %tmp20 to i64
894 %tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)
895 %tmp251 = tail call i64 @llvm.amdgcn.wwm.i64(i64 %tmp176)
896 %.cast = bitcast i64 %tmp97 to <2 x float>
897 %.cast6 = bitcast i64 %tmp174 to <2 x float>
898 %.cast7 = bitcast i64 %tmp251 to <2 x float>
899 %tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
900 %desc.int = bitcast <4 x i32> %desc to i128
901 %desc.ptr = inttoptr i128 %desc.int to ptr addrspace(8)
902 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %tmp254, ptr addrspace(8) %desc.ptr, i32 %tmp17, i32 0, i32 0)
903 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float> %.cast7, ptr addrspace(8) %desc.ptr, i32 %tmp19, i32 0, i32 0)
904 ret void
905 }
908 define amdgpu_cs void @strict_wwm_no_cfg(ptr addrspace(8) inreg %tmp14) {
909 ; GFX9-O0-LABEL: strict_wwm_no_cfg:
910 ; GFX9-O0: ; %bb.0:
911 ; GFX9-O0-NEXT: s_mov_b32 s6, s2
912 ; GFX9-O0-NEXT: s_mov_b32 s4, s0
913 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
914 ; GFX9-O0-NEXT: s_mov_b32 s7, s3
915 ; GFX9-O0-NEXT: s_mov_b32 s8, s7
916 ; GFX9-O0-NEXT: s_mov_b32 s9, s6
917 ; GFX9-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
918 ; GFX9-O0-NEXT: s_mov_b32 s5, s1
919 ; GFX9-O0-NEXT: s_mov_b32 s10, s5
920 ; GFX9-O0-NEXT: s_mov_b32 s0, s4
921 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
922 ; GFX9-O0-NEXT: s_mov_b32 s1, s10
923 ; GFX9-O0-NEXT: s_mov_b32 s2, s9
924 ; GFX9-O0-NEXT: s_mov_b32 s3, s8
925 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
926 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[5:6], off, s[0:3], s4
927 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
928 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v5
929 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v6
930 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
931 ; GFX9-O0-NEXT: s_not_b64 exec, exec
932 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4
933 ; GFX9-O0-NEXT: s_not_b64 exec, exec
934 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
935 ; GFX9-O0-NEXT: s_not_b64 exec, exec
936 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s4
937 ; GFX9-O0-NEXT: s_not_b64 exec, exec
938 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
939 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s4
940 ; GFX9-O0-NEXT: s_nop 1
941 ; GFX9-O0-NEXT: v_mov_b32_dpp v2, v0 row_bcast:31 row_mask:0xc bank_mask:0xf
942 ; GFX9-O0-NEXT: v_add_u32_e64 v0, v0, v2
943 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
944 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
945 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
946 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4
947 ; GFX9-O0-NEXT: s_nop 1
948 ; GFX9-O0-NEXT: v_mov_b32_dpp v0, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
949 ; GFX9-O0-NEXT: v_add_u32_e64 v0, v1, v0
950 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
951 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v0
952 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v3, v4
953 ; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[6:7]
954 ; GFX9-O0-NEXT: s_mov_b32 s5, 1
955 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s5, v3
956 ; GFX9-O0-NEXT: s_mov_b32 s5, 2
957 ; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s5
958 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s4 offset:4
959 ; GFX9-O0-NEXT: s_endpgm
960 ;
961 ; GFX9-O3-LABEL: strict_wwm_no_cfg:
962 ; GFX9-O3: ; %bb.0:
963 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[4:5], off, s[0:3], 0
964 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
965 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, 0
966 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
967 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
968 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
969 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v4
970 ; GFX9-O3-NEXT: s_not_b64 exec, exec
971 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
972 ; GFX9-O3-NEXT: s_not_b64 exec, exec
973 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v5
974 ; GFX9-O3-NEXT: s_not_b64 exec, exec
975 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, 0
976 ; GFX9-O3-NEXT: s_not_b64 exec, exec
977 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
978 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
979 ; GFX9-O3-NEXT: v_mov_b32_dpp v0, v3 row_bcast:31 row_mask:0xc bank_mask:0xf
980 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
981 ; GFX9-O3-NEXT: v_add_u32_e32 v0, v3, v0
982 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
983 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
984 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
985 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v4, v5
986 ; GFX9-O3-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc
987 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 1, v4
988 ; GFX9-O3-NEXT: v_and_b32_e32 v4, 2, v4
989 ; GFX9-O3-NEXT: buffer_store_dword v4, off, s[0:3], 0 offset:4
990 ; GFX9-O3-NEXT: s_endpgm
991 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
992 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
993 %tmp102 = extractelement <2 x i32> %tmp101, i32 0
994 %tmp103 = extractelement <2 x i32> %tmp101, i32 1
995 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
996 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
999 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
1000 %tmp121 = add i32 %tmp105, %tmp120
1001 %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)
1003 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
1004 %tmp136 = add i32 %tmp107, %tmp135
1005 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
1007 %tmp138 = icmp eq i32 %tmp122, %tmp137
1008 %tmp139 = sext i1 %tmp138 to i32
1009 %tmp140 = shl nsw i32 %tmp139, 1
1010 %tmp141 = and i32 %tmp140, 2
1011 %tmp145 = bitcast i32 %tmp141 to float
1012 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
1013 ret void
1014 }
1016 define amdgpu_cs void @strict_wwm_cfg(ptr addrspace(8) inreg %tmp14, i32 %arg) {
1017 ; GFX9-O0-LABEL: strict_wwm_cfg:
1018 ; GFX9-O0: ; %bb.0: ; %entry
1019 ; GFX9-O0-NEXT: s_mov_b32 s16, SCRATCH_RSRC_DWORD0
1020 ; GFX9-O0-NEXT: s_mov_b32 s17, SCRATCH_RSRC_DWORD1
1021 ; GFX9-O0-NEXT: s_mov_b32 s18, -1
1022 ; GFX9-O0-NEXT: s_mov_b32 s19, 0xe00000
1023 ; GFX9-O0-NEXT: s_add_u32 s16, s16, s4
1024 ; GFX9-O0-NEXT: s_addc_u32 s17, s17, 0
1025 ; GFX9-O0-NEXT: ; implicit-def: $vgpr3 : SGPR spill to VGPR lane
1026 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
1027 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
1028 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 ; 4-byte Folded Reload
1029 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
1030 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1031 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 0
1032 ; GFX9-O0-NEXT: s_mov_b32 s4, s1
1033 ; GFX9-O0-NEXT: v_readlane_b32 s1, v0, 0
1034 ; GFX9-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
1035 ; GFX9-O0-NEXT: s_mov_b32 s3, s1
1036 ; GFX9-O0-NEXT: s_mov_b32 s8, s3
1037 ; GFX9-O0-NEXT: s_mov_b32 s9, s2
1038 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
1039 ; GFX9-O0-NEXT: s_mov_b32 s1, s4
1040 ; GFX9-O0-NEXT: s_mov_b32 s10, s1
1041 ; GFX9-O0-NEXT: s_mov_b32 s4, s0
1042 ; GFX9-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5_sgpr6_sgpr7
1043 ; GFX9-O0-NEXT: s_mov_b32 s5, s10
1044 ; GFX9-O0-NEXT: s_mov_b32 s6, s9
1045 ; GFX9-O0-NEXT: s_mov_b32 s7, s8
1046 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 1
1047 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 2
1048 ; GFX9-O0-NEXT: v_writelane_b32 v0, s0, 3
1049 ; GFX9-O0-NEXT: v_writelane_b32 v0, s1, 4
1050 ; GFX9-O0-NEXT: s_mov_b32 s0, 0
1051 ; GFX9-O0-NEXT: s_nop 2
1052 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[4:5], off, s[4:7], s0
1053 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1054 ; GFX9-O0-NEXT: buffer_store_dword v4, off, s[16:19], 0 offset:12 ; 4-byte Folded Spill
1055 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1056 ; GFX9-O0-NEXT: buffer_store_dword v5, off, s[16:19], 0 offset:16 ; 4-byte Folded Spill
1057 ; GFX9-O0-NEXT: ; implicit-def: $sgpr2_sgpr3
1058 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v4
1059 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1060 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s0
1061 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1062 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
1063 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s0
1064 ; GFX9-O0-NEXT: s_nop 1
1065 ; GFX9-O0-NEXT: v_mov_b32_dpp v2, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
1066 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v1, v2
1067 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
1068 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
1069 ; GFX9-O0-NEXT: buffer_store_dword v4, off, s[16:19], 0 offset:8 ; 4-byte Folded Spill
1070 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v3, s0
1071 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, s0
1072 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[16:19], 0 offset:4 ; 4-byte Folded Spill
1073 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], exec
1074 ; GFX9-O0-NEXT: v_writelane_b32 v0, s0, 5
1075 ; GFX9-O0-NEXT: v_writelane_b32 v0, s1, 6
1076 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
1077 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 ; 4-byte Folded Spill
1078 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
1079 ; GFX9-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
1080 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
1081 ; GFX9-O0-NEXT: s_cbranch_execz .LBB8_2
1082 ; GFX9-O0-NEXT: ; %bb.1: ; %if
1083 ; GFX9-O0-NEXT: buffer_load_dword v3, off, s[16:19], 0 offset:12 ; 4-byte Folded Reload
1084 ; GFX9-O0-NEXT: buffer_load_dword v4, off, s[16:19], 0 offset:16 ; 4-byte Folded Reload
1085 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1086 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
1087 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[0:1], -1
1088 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, 0
1089 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
1090 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
1091 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1092 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v1
1093 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1094 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[0:1], -1
1095 ; GFX9-O0-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
1096 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v2, v1
1097 ; GFX9-O0-NEXT: s_mov_b64 exec, s[0:1]
1098 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1
1099 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:4 ; 4-byte Folded Spill
1100 ; GFX9-O0-NEXT: .LBB8_2: ; %merge
1101 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[12:13], -1
1102 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[16:19], 0 ; 4-byte Folded Reload
1103 ; GFX9-O0-NEXT: s_mov_b64 exec, s[12:13]
1104 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1105 ; GFX9-O0-NEXT: v_readlane_b32 s4, v0, 5
1106 ; GFX9-O0-NEXT: v_readlane_b32 s5, v0, 6
1107 ; GFX9-O0-NEXT: s_or_b64 exec, exec, s[4:5]
1108 ; GFX9-O0-NEXT: v_readlane_b32 s2, v0, 1
1109 ; GFX9-O0-NEXT: v_readlane_b32 s3, v0, 2
1110 ; GFX9-O0-NEXT: v_readlane_b32 s0, v0, 3
1111 ; GFX9-O0-NEXT: v_readlane_b32 s1, v0, 4
1112 ; GFX9-O0-NEXT: buffer_load_dword v3, off, s[16:19], 0 offset:8 ; 4-byte Folded Reload
1113 ; GFX9-O0-NEXT: buffer_load_dword v4, off, s[16:19], 0 offset:4 ; 4-byte Folded Reload
1114 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1115 ; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[4:5], v3, v4
1116 ; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[4:5]
1117 ; GFX9-O0-NEXT: s_mov_b32 s4, 1
1118 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s4, v3
1119 ; GFX9-O0-NEXT: s_mov_b32 s4, 2
1120 ; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s4
1121 ; GFX9-O0-NEXT: s_mov_b32 s6, s1
1122 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 killed $sgpr0_sgpr1
1123 ; GFX9-O0-NEXT: s_mov_b32 s4, s3
1124 ; GFX9-O0-NEXT: s_mov_b32 s5, s2
1125 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1126 ; GFX9-O0-NEXT: s_mov_b32 s1, s6
1127 ; GFX9-O0-NEXT: s_mov_b32 s2, s5
1128 ; GFX9-O0-NEXT: s_mov_b32 s3, s4
1129 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
1130 ; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s4 offset:4
1131 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
1132 ; GFX9-O0-NEXT: s_endpgm
1133 ;
1134 ; GFX9-O3-LABEL: strict_wwm_cfg:
1135 ; GFX9-O3: ; %bb.0: ; %entry
1136 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[3:4], off, s[0:3], 0
1137 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, 0
1138 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
1139 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
1140 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
1141 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
1142 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v3
1143 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1144 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
1145 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1146 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[4:5], -1
1147 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
1148 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
1149 ; GFX9-O3-NEXT: s_mov_b64 exec, s[4:5]
1150 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v1
1151 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
1152 ; GFX9-O3-NEXT: s_and_saveexec_b64 s[4:5], vcc
1153 ; GFX9-O3-NEXT: s_cbranch_execz .LBB8_2
1154 ; GFX9-O3-NEXT: ; %bb.1: ; %if
1155 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[6:7], -1
1156 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, 0
1157 ; GFX9-O3-NEXT: s_mov_b64 exec, s[6:7]
1158 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v4
1159 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1160 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, 0
1161 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1162 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[6:7], -1
1163 ; GFX9-O3-NEXT: v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
1164 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v2, v1
1165 ; GFX9-O3-NEXT: s_mov_b64 exec, s[6:7]
1166 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v1
1167 ; GFX9-O3-NEXT: .LBB8_2: ; %merge
1168 ; GFX9-O3-NEXT: s_or_b64 exec, exec, s[4:5]
1169 ; GFX9-O3-NEXT: v_cmp_eq_u32_e32 vcc, v3, v5
1170 ; GFX9-O3-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
1171 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 1, v0
1172 ; GFX9-O3-NEXT: v_and_b32_e32 v0, 2, v0
1173 ; GFX9-O3-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4
1174 ; GFX9-O3-NEXT: s_endpgm
1175 entry:
1176 %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
1177 %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
1178 %tmp102 = extractelement <2 x i32> %tmp101, i32 0
1179 %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
1181 %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
1182 %tmp121 = add i32 %tmp105, %tmp120
1183 %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)
1185 %cond = icmp eq i32 %arg, 0
1186 br i1 %cond, label %if, label %merge
1187 if:
1188 %tmp103 = extractelement <2 x i32> %tmp101, i32 1
1189 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
1191 %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
1192 %tmp136 = add i32 %tmp107, %tmp135
1193 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
1194 br label %merge
1196 merge:
1197 %merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]
1198 %tmp138 = icmp eq i32 %tmp122, %merge_value
1199 %tmp139 = sext i1 %tmp138 to i32
1200 %tmp140 = shl nsw i32 %tmp139, 1
1201 %tmp141 = and i32 %tmp140, 2
1202 %tmp145 = bitcast i32 %tmp141 to float
1203 call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
1204 ret void
1205 }
1207 define hidden i32 @strict_wwm_called(i32 %a) noinline {
1208 ; GFX9-O0-LABEL: strict_wwm_called:
1209 ; GFX9-O0: ; %bb.0:
1210 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1211 ; GFX9-O0-NEXT: v_add_u32_e64 v1, v0, v0
1212 ; GFX9-O0-NEXT: v_mul_lo_u32 v0, v1, v0
1213 ; GFX9-O0-NEXT: v_sub_u32_e64 v0, v0, v1
1214 ; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
1215 ;
1216 ; GFX9-O3-LABEL: strict_wwm_called:
1217 ; GFX9-O3: ; %bb.0:
1218 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1219 ; GFX9-O3-NEXT: v_add_u32_e32 v1, v0, v0
1220 ; GFX9-O3-NEXT: v_mul_lo_u32 v0, v1, v0
1221 ; GFX9-O3-NEXT: v_sub_u32_e32 v0, v0, v1
1222 ; GFX9-O3-NEXT: s_setpc_b64 s[30:31]
1223 %add = add i32 %a, %a
1224 %mul = mul i32 %add, %a
1225 %sub = sub i32 %mul, %add
1226 ret i32 %sub
1227 }
1229 define amdgpu_kernel void @strict_wwm_call(ptr addrspace(8) inreg %tmp14, i32 inreg %arg) {
1230 ; GFX9-O0-LABEL: strict_wwm_call:
1231 ; GFX9-O0: ; %bb.0:
1232 ; GFX9-O0-NEXT: s_mov_b32 s32, 0x400
1233 ; GFX9-O0-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
1234 ; GFX9-O0-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
1235 ; GFX9-O0-NEXT: s_mov_b32 s26, -1
1236 ; GFX9-O0-NEXT: s_mov_b32 s27, 0xe00000
1237 ; GFX9-O0-NEXT: s_add_u32 s24, s24, s9
1238 ; GFX9-O0-NEXT: s_addc_u32 s25, s25, 0
1239 ; GFX9-O0-NEXT: ; implicit-def: $vgpr7 : SGPR spill to VGPR lane
1240 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
1241 ; GFX9-O0-NEXT: v_writelane_b32 v7, s10, 0
1242 ; GFX9-O0-NEXT: v_writelane_b32 v7, s11, 1
1243 ; GFX9-O0-NEXT: s_mov_b32 s14, s8
1244 ; GFX9-O0-NEXT: s_mov_b32 s13, s7
1245 ; GFX9-O0-NEXT: s_mov_b32 s12, s6
1246 ; GFX9-O0-NEXT: s_mov_b64 s[10:11], s[4:5]
1247 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[2:3]
1248 ; GFX9-O0-NEXT: v_readlane_b32 s2, v7, 0
1249 ; GFX9-O0-NEXT: v_readlane_b32 s3, v7, 1
1250 ; GFX9-O0-NEXT: v_writelane_b32 v7, s4, 2
1251 ; GFX9-O0-NEXT: v_writelane_b32 v7, s5, 3
1252 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1253 ; GFX9-O0-NEXT: buffer_store_dword v7, off, s[24:27], 0 ; 4-byte Folded Spill
1254 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1255 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[0:1]
1256 ; GFX9-O0-NEXT: v_readlane_b32 s0, v7, 2
1257 ; GFX9-O0-NEXT: v_readlane_b32 s1, v7, 3
1258 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
1259 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
1260 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
1261 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1262 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
1263 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1264 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
1265 ; GFX9-O0-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x24
1266 ; GFX9-O0-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x2c
1267 ; GFX9-O0-NEXT: s_load_dword s2, s[0:1], 0x34
1268 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
1269 ; GFX9-O0-NEXT: s_mov_b32 s3, s7
1270 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 killed $sgpr6_sgpr7
1271 ; GFX9-O0-NEXT: s_mov_b32 s7, s9
1272 ; GFX9-O0-NEXT: s_mov_b32 s16, s8
1273 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17_sgpr18_sgpr19
1274 ; GFX9-O0-NEXT: s_mov_b32 s17, s7
1275 ; GFX9-O0-NEXT: s_mov_b32 s18, s6
1276 ; GFX9-O0-NEXT: s_mov_b32 s19, s3
1277 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1278 ; GFX9-O0-NEXT: v_writelane_b32 v0, s16, 4
1279 ; GFX9-O0-NEXT: v_writelane_b32 v0, s17, 5
1280 ; GFX9-O0-NEXT: v_writelane_b32 v0, s18, 6
1281 ; GFX9-O0-NEXT: v_writelane_b32 v0, s19, 7
1282 ; GFX9-O0-NEXT: s_mov_b32 s3, 0
1283 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 8
1284 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, s2
1285 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1286 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, s3
1287 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1288 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
1289 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 9
1290 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 10
1291 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1292 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[24:27], 0 ; 4-byte Folded Spill
1293 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1294 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 56
1295 ; GFX9-O0-NEXT: s_mov_b32 s2, s0
1296 ; GFX9-O0-NEXT: s_mov_b32 s0, s1
1297 ; GFX9-O0-NEXT: s_mov_b32 s3, s6
1298 ; GFX9-O0-NEXT: s_mov_b32 s1, s7
1299 ; GFX9-O0-NEXT: s_add_u32 s8, s2, s3
1300 ; GFX9-O0-NEXT: s_addc_u32 s0, s0, s1
1301 ; GFX9-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
1302 ; GFX9-O0-NEXT: s_mov_b32 s9, s0
1303 ; GFX9-O0-NEXT: s_getpc_b64 s[16:17]
1304 ; GFX9-O0-NEXT: s_add_u32 s16, s16, strict_wwm_called@rel32@lo+4
1305 ; GFX9-O0-NEXT: s_addc_u32 s17, s17, strict_wwm_called@rel32@hi+12
1306 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[24:25]
1307 ; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[26:27]
1308 ; GFX9-O0-NEXT: s_mov_b32 s6, 20
1309 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s6, v3
1310 ; GFX9-O0-NEXT: s_mov_b32 s6, 10
1311 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v4, s6, v4
1312 ; GFX9-O0-NEXT: v_or3_b32 v3, v5, v4, v3
1313 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6_sgpr7
1314 ; GFX9-O0-NEXT: ; implicit-def: $sgpr15
1315 ; GFX9-O0-NEXT: v_mov_b32_e32 v31, v3
1316 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v6
1317 ; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[16:17]
1318 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1319 ; GFX9-O0-NEXT: buffer_load_dword v1, off, s[24:27], 0 ; 4-byte Folded Reload
1320 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1321 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1322 ; GFX9-O0-NEXT: v_readlane_b32 s0, v1, 4
1323 ; GFX9-O0-NEXT: v_readlane_b32 s1, v1, 5
1324 ; GFX9-O0-NEXT: v_readlane_b32 s2, v1, 6
1325 ; GFX9-O0-NEXT: v_readlane_b32 s3, v1, 7
1326 ; GFX9-O0-NEXT: v_readlane_b32 s6, v1, 9
1327 ; GFX9-O0-NEXT: v_readlane_b32 s7, v1, 10
1328 ; GFX9-O0-NEXT: v_readlane_b32 s4, v1, 8
1329 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
1330 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1331 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
1332 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1333 ; GFX9-O0-NEXT: v_add_u32_e64 v3, v3, v6
1334 ; GFX9-O0-NEXT: s_mov_b64 exec, s[6:7]
1335 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
1336 ; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s4 offset:4
1337 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
1338 ; GFX9-O0-NEXT: s_endpgm
1339 ;
1340 ; GFX9-O3-LABEL: strict_wwm_call:
1341 ; GFX9-O3: ; %bb.0:
1342 ; GFX9-O3-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
1343 ; GFX9-O3-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
1344 ; GFX9-O3-NEXT: s_mov_b32 s26, -1
1345 ; GFX9-O3-NEXT: s_mov_b32 s27, 0xe00000
1346 ; GFX9-O3-NEXT: s_add_u32 s24, s24, s9
1347 ; GFX9-O3-NEXT: s_mov_b32 s32, 0
1348 ; GFX9-O3-NEXT: s_addc_u32 s25, s25, 0
1349 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[12:13], -1
1350 ; GFX9-O3-NEXT: s_mov_b32 s14, s8
1351 ; GFX9-O3-NEXT: s_mov_b64 s[10:11], s[4:5]
1352 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v2
1353 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
1354 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
1355 ; GFX9-O3-NEXT: s_mov_b64 exec, s[12:13]
1356 ; GFX9-O3-NEXT: s_load_dword s4, s[2:3], 0x34
1357 ; GFX9-O3-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
1358 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
1359 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s4
1360 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1361 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
1362 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1363 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[20:21], -1
1364 ; GFX9-O3-NEXT: s_add_u32 s8, s2, 56
1365 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v3, 20, v3
1366 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 10, v4
1367 ; GFX9-O3-NEXT: s_addc_u32 s9, s3, 0
1368 ; GFX9-O3-NEXT: v_or3_b32 v3, v5, v4, v3
1369 ; GFX9-O3-NEXT: s_mov_b64 s[4:5], s[0:1]
1370 ; GFX9-O3-NEXT: s_mov_b64 s[0:1], s[24:25]
1371 ; GFX9-O3-NEXT: s_mov_b32 s12, s6
1372 ; GFX9-O3-NEXT: s_mov_b32 s13, s7
1373 ; GFX9-O3-NEXT: v_mov_b32_e32 v31, v3
1374 ; GFX9-O3-NEXT: s_mov_b64 s[2:3], s[26:27]
1375 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v6
1376 ; GFX9-O3-NEXT: s_getpc_b64 s[22:23]
1377 ; GFX9-O3-NEXT: s_add_u32 s22, s22, strict_wwm_called@rel32@lo+4
1378 ; GFX9-O3-NEXT: s_addc_u32 s23, s23, strict_wwm_called@rel32@hi+12
1379 ; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[22:23]
1380 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v0
1381 ; GFX9-O3-NEXT: v_add_u32_e32 v3, v3, v6
1382 ; GFX9-O3-NEXT: s_mov_b64 exec, s[20:21]
1383 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v3
1384 ; GFX9-O3-NEXT: buffer_store_dword v0, off, s[16:19], 0 offset:4
1385 ; GFX9-O3-NEXT: s_endpgm
1388 %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)
1389 %tmp134 = call i32 @strict_wwm_called(i32 %tmp107)
1390 %tmp136 = add i32 %tmp134, %tmp107
1391 %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
1392 call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
1393 ret void
1394 }
1396 define i64 @strict_wwm_called_i64(i64 %a) noinline {
1397 ; GFX9-O0-LABEL: strict_wwm_called_i64:
1398 ; GFX9-O0: ; %bb.0:
1399 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1400 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
1401 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4
1402 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4
1403 ; GFX9-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
1404 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1
1405 ; GFX9-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
1406 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2
1407 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v2
1408 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v3
1409 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
1410 ; GFX9-O0-NEXT: v_add_co_u32_e64 v4, s[4:5], v4, v5
1411 ; GFX9-O0-NEXT: v_addc_co_u32_e64 v0, s[4:5], v0, v1, s[4:5]
1412 ; GFX9-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
1413 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
1414 ; GFX9-O0-NEXT: s_mov_b32 s4, 32
1415 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2
1416 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
1417 ; GFX9-O0-NEXT: v_lshrrev_b64 v[0:1], s4, v[0:1]
1418 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v0
1419 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
1420 ; GFX9-O0-NEXT: v_mul_lo_u32 v1, v0, v1
1421 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
1422 ; GFX9-O0-NEXT: v_lshrrev_b64 v[6:7], s4, v[4:5]
1423 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
1424 ; GFX9-O0-NEXT: v_mul_lo_u32 v2, v2, v3
1425 ; GFX9-O0-NEXT: v_mad_u64_u32 v[6:7], s[6:7], v0, v3, 0
1426 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
1427 ; GFX9-O0-NEXT: v_add3_u32 v0, v0, v1, v2
1428 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1429 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
1430 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
1431 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s5
1432 ; GFX9-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
1433 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v2
1434 ; GFX9-O0-NEXT: v_lshlrev_b64 v[1:2], s4, v[0:1]
1435 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
1436 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec
1437 ; GFX9-O0-NEXT: s_mov_b32 s5, 0
1438 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1439 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, 0
1440 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
1441 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0
1442 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
1443 ; GFX9-O0-NEXT: v_or_b32_e64 v0, v0, v3
1444 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v1
1445 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6
1446 ; GFX9-O0-NEXT: v_or_b32_e64 v6, v1, v2
1447 ; GFX9-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
1448 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v0
1449 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v6
1450 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v4
1451 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v7
1452 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v5
1453 ; GFX9-O0-NEXT: v_sub_co_u32_e64 v1, s[6:7], v1, v3
1454 ; GFX9-O0-NEXT: v_subb_co_u32_e64 v0, s[6:7], v0, v2, s[6:7]
1455 ; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
1456 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
1457 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1
1458 ; GFX9-O0-NEXT: v_lshrrev_b64 v[1:2], s4, v[1:2]
1459 ; GFX9-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec
1460 ; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
1461 ;
1462 ; GFX9-O3-LABEL: strict_wwm_called_i64:
1463 ; GFX9-O3: ; %bb.0:
1464 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1465 ; GFX9-O3-NEXT: v_add_co_u32_e32 v2, vcc, v0, v0
1466 ; GFX9-O3-NEXT: v_addc_co_u32_e32 v3, vcc, v1, v1, vcc
1467 ; GFX9-O3-NEXT: v_mul_lo_u32 v4, v3, v0
1468 ; GFX9-O3-NEXT: v_mul_lo_u32 v5, v2, v1
1469 ; GFX9-O3-NEXT: v_mad_u64_u32 v[0:1], s[4:5], v2, v0, 0
1470 ; GFX9-O3-NEXT: v_add3_u32 v1, v1, v5, v4
1471 ; GFX9-O3-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v2
1472 ; GFX9-O3-NEXT: v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
1473 ; GFX9-O3-NEXT: s_setpc_b64 s[30:31]
1474 %add = add i64 %a, %a
1475 %mul = mul i64 %add, %a
1476 %sub = sub i64 %mul, %add
1477 ret i64 %sub
1478 }
1480 define amdgpu_kernel void @strict_wwm_call_i64(ptr addrspace(8) inreg %tmp14, i64 inreg %arg) {
1481 ; GFX9-O0-LABEL: strict_wwm_call_i64:
1482 ; GFX9-O0: ; %bb.0:
1483 ; GFX9-O0-NEXT: s_mov_b32 s32, 0x400
1484 ; GFX9-O0-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
1485 ; GFX9-O0-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
1486 ; GFX9-O0-NEXT: s_mov_b32 s26, -1
1487 ; GFX9-O0-NEXT: s_mov_b32 s27, 0xe00000
1488 ; GFX9-O0-NEXT: s_add_u32 s24, s24, s9
1489 ; GFX9-O0-NEXT: s_addc_u32 s25, s25, 0
1490 ; GFX9-O0-NEXT: ; implicit-def: $vgpr12 : SGPR spill to VGPR lane
1491 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[10:11], -1
1492 ; GFX9-O0-NEXT: v_writelane_b32 v12, s10, 0
1493 ; GFX9-O0-NEXT: v_writelane_b32 v12, s11, 1
1494 ; GFX9-O0-NEXT: s_mov_b32 s14, s8
1495 ; GFX9-O0-NEXT: s_mov_b32 s13, s7
1496 ; GFX9-O0-NEXT: s_mov_b32 s12, s6
1497 ; GFX9-O0-NEXT: s_mov_b64 s[10:11], s[4:5]
1498 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[2:3]
1499 ; GFX9-O0-NEXT: v_readlane_b32 s2, v12, 0
1500 ; GFX9-O0-NEXT: v_readlane_b32 s3, v12, 1
1501 ; GFX9-O0-NEXT: v_writelane_b32 v12, s4, 2
1502 ; GFX9-O0-NEXT: v_writelane_b32 v12, s5, 3
1503 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1504 ; GFX9-O0-NEXT: buffer_store_dword v12, off, s[24:27], 0 ; 4-byte Folded Spill
1505 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1506 ; GFX9-O0-NEXT: s_mov_b64 s[4:5], s[0:1]
1507 ; GFX9-O0-NEXT: v_readlane_b32 s0, v12, 2
1508 ; GFX9-O0-NEXT: v_readlane_b32 s1, v12, 3
1509 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v2
1510 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
1511 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v0
1512 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1513 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
1514 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1515 ; GFX9-O0-NEXT: s_mov_b64 exec, s[2:3]
1516 ; GFX9-O0-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x24
1517 ; GFX9-O0-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x2c
1518 ; GFX9-O0-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
1519 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
1520 ; GFX9-O0-NEXT: s_mov_b32 s6, s9
1521 ; GFX9-O0-NEXT: s_mov_b32 s7, s8
1522 ; GFX9-O0-NEXT: s_mov_b32 s8, s17
1523 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 killed $sgpr16_sgpr17
1524 ; GFX9-O0-NEXT: ; kill: def $sgpr16 killed $sgpr16 def $sgpr16_sgpr17_sgpr18_sgpr19
1525 ; GFX9-O0-NEXT: s_mov_b32 s17, s8
1526 ; GFX9-O0-NEXT: s_mov_b32 s18, s7
1527 ; GFX9-O0-NEXT: s_mov_b32 s19, s6
1528 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1529 ; GFX9-O0-NEXT: v_writelane_b32 v0, s16, 4
1530 ; GFX9-O0-NEXT: v_writelane_b32 v0, s17, 5
1531 ; GFX9-O0-NEXT: v_writelane_b32 v0, s18, 6
1532 ; GFX9-O0-NEXT: v_writelane_b32 v0, s19, 7
1533 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 0
1534 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, s2
1535 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, s3
1536 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1537 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, s6
1538 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, s7
1539 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1540 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[2:3], -1
1541 ; GFX9-O0-NEXT: v_writelane_b32 v0, s2, 8
1542 ; GFX9-O0-NEXT: v_writelane_b32 v0, s3, 9
1543 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1544 ; GFX9-O0-NEXT: buffer_store_dword v0, off, s[24:27], 0 ; 4-byte Folded Spill
1545 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1546 ; GFX9-O0-NEXT: s_mov_b64 s[6:7], 60
1547 ; GFX9-O0-NEXT: s_mov_b32 s2, s0
1548 ; GFX9-O0-NEXT: s_mov_b32 s0, s1
1549 ; GFX9-O0-NEXT: s_mov_b32 s3, s6
1550 ; GFX9-O0-NEXT: s_mov_b32 s1, s7
1551 ; GFX9-O0-NEXT: s_add_u32 s8, s2, s3
1552 ; GFX9-O0-NEXT: s_addc_u32 s0, s0, s1
1553 ; GFX9-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9
1554 ; GFX9-O0-NEXT: s_mov_b32 s9, s0
1555 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v8
1556 ; GFX9-O0-NEXT: s_mov_b32 s0, 32
1557 ; GFX9-O0-NEXT: ; implicit-def: $sgpr2_sgpr3
1558 ; GFX9-O0-NEXT: v_lshrrev_b64 v[10:11], s0, v[8:9]
1559 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v10
1560 ; GFX9-O0-NEXT: s_getpc_b64 s[0:1]
1561 ; GFX9-O0-NEXT: s_add_u32 s0, s0, strict_wwm_called_i64@gotpcrel32@lo+4
1562 ; GFX9-O0-NEXT: s_addc_u32 s1, s1, strict_wwm_called_i64@gotpcrel32@hi+12
1563 ; GFX9-O0-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
1564 ; GFX9-O0-NEXT: s_mov_b64 s[0:1], s[24:25]
1565 ; GFX9-O0-NEXT: s_mov_b64 s[2:3], s[26:27]
1566 ; GFX9-O0-NEXT: s_mov_b32 s6, 20
1567 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s6, v3
1568 ; GFX9-O0-NEXT: s_mov_b32 s6, 10
1569 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v4, s6, v4
1570 ; GFX9-O0-NEXT: v_or3_b32 v3, v5, v4, v3
1571 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6_sgpr7
1572 ; GFX9-O0-NEXT: ; implicit-def: $sgpr15
1573 ; GFX9-O0-NEXT: v_mov_b32_e32 v31, v3
1574 ; GFX9-O0-NEXT: v_mov_b32_e32 v0, v6
1575 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v7
1576 ; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
1577 ; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[16:17]
1578 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1579 ; GFX9-O0-NEXT: buffer_load_dword v2, off, s[24:27], 0 ; 4-byte Folded Reload
1580 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1581 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1582 ; GFX9-O0-NEXT: v_readlane_b32 s0, v2, 4
1583 ; GFX9-O0-NEXT: v_readlane_b32 s1, v2, 5
1584 ; GFX9-O0-NEXT: v_readlane_b32 s2, v2, 6
1585 ; GFX9-O0-NEXT: v_readlane_b32 s3, v2, 7
1586 ; GFX9-O0-NEXT: v_readlane_b32 s4, v2, 8
1587 ; GFX9-O0-NEXT: v_readlane_b32 s5, v2, 9
1588 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
1589 ; GFX9-O0-NEXT: s_or_saveexec_b64 s[20:21], -1
1590 ; GFX9-O0-NEXT: buffer_load_dword v0, off, s[24:27], 0 ; 4-byte Folded Reload
1591 ; GFX9-O0-NEXT: s_mov_b64 exec, s[20:21]
1592 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
1593 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
1594 ; GFX9-O0-NEXT: ; implicit-def: $sgpr6
1595 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v8
1596 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v9
1597 ; GFX9-O0-NEXT: v_add_co_u32_e64 v3, s[6:7], v3, v5
1598 ; GFX9-O0-NEXT: v_addc_co_u32_e64 v4, s[6:7], v4, v6, s[6:7]
1599 ; GFX9-O0-NEXT: s_mov_b64 exec, s[4:5]
1600 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
1601 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v4
1602 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
1603 ; GFX9-O0-NEXT: buffer_store_dwordx2 v[1:2], off, s[0:3], s4 offset:4
1604 ; GFX9-O0-NEXT: ; kill: killed $vgpr0
1605 ; GFX9-O0-NEXT: s_endpgm
1606 ;
1607 ; GFX9-O3-LABEL: strict_wwm_call_i64:
1608 ; GFX9-O3: ; %bb.0:
1609 ; GFX9-O3-NEXT: s_mov_b32 s24, SCRATCH_RSRC_DWORD0
1610 ; GFX9-O3-NEXT: s_mov_b32 s25, SCRATCH_RSRC_DWORD1
1611 ; GFX9-O3-NEXT: s_mov_b32 s26, -1
1612 ; GFX9-O3-NEXT: s_mov_b32 s27, 0xe00000
1613 ; GFX9-O3-NEXT: s_add_u32 s24, s24, s9
1614 ; GFX9-O3-NEXT: s_mov_b32 s32, 0
1615 ; GFX9-O3-NEXT: s_addc_u32 s25, s25, 0
1616 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[12:13], -1
1617 ; GFX9-O3-NEXT: s_mov_b32 s14, s8
1618 ; GFX9-O3-NEXT: s_mov_b64 s[10:11], s[4:5]
1619 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v2
1620 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
1621 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v0
1622 ; GFX9-O3-NEXT: s_mov_b64 exec, s[12:13]
1623 ; GFX9-O3-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x34
1624 ; GFX9-O3-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
1625 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
1626 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s4
1627 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, s5
1628 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1629 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, 0
1630 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, 0
1631 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1632 ; GFX9-O3-NEXT: s_or_saveexec_b64 s[20:21], -1
1633 ; GFX9-O3-NEXT: s_add_u32 s8, s2, 60
1634 ; GFX9-O3-NEXT: s_addc_u32 s9, s3, 0
1635 ; GFX9-O3-NEXT: s_getpc_b64 s[2:3]
1636 ; GFX9-O3-NEXT: s_add_u32 s2, s2, strict_wwm_called_i64@gotpcrel32@lo+4
1637 ; GFX9-O3-NEXT: s_addc_u32 s3, s3, strict_wwm_called_i64@gotpcrel32@hi+12
1638 ; GFX9-O3-NEXT: s_load_dwordx2 s[22:23], s[2:3], 0x0
1639 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v3, 20, v3
1640 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v4, 10, v4
1641 ; GFX9-O3-NEXT: v_or3_b32 v3, v5, v4, v3
1642 ; GFX9-O3-NEXT: s_mov_b64 s[4:5], s[0:1]
1643 ; GFX9-O3-NEXT: s_mov_b64 s[0:1], s[24:25]
1644 ; GFX9-O3-NEXT: s_mov_b32 s12, s6
1645 ; GFX9-O3-NEXT: s_mov_b32 s13, s7
1646 ; GFX9-O3-NEXT: v_mov_b32_e32 v31, v3
1647 ; GFX9-O3-NEXT: s_mov_b64 s[2:3], s[26:27]
1648 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v6
1649 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v7
1650 ; GFX9-O3-NEXT: s_waitcnt lgkmcnt(0)
1651 ; GFX9-O3-NEXT: s_swappc_b64 s[30:31], s[22:23]
1652 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v0
1653 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v1
1654 ; GFX9-O3-NEXT: v_add_co_u32_e32 v3, vcc, v3, v6
1655 ; GFX9-O3-NEXT: v_addc_co_u32_e32 v4, vcc, v4, v7, vcc
1656 ; GFX9-O3-NEXT: s_mov_b64 exec, s[20:21]
1657 ; GFX9-O3-NEXT: v_mov_b32_e32 v0, v3
1658 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v4
1659 ; GFX9-O3-NEXT: buffer_store_dwordx2 v[0:1], off, s[16:19], 0 offset:4
1660 ; GFX9-O3-NEXT: s_endpgm
1664 %tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)
1665 %tmp134 = call i64 @strict_wwm_called_i64(i64 %tmp107)
1666 %tmp136 = add i64 %tmp134, %tmp107
1667 %tmp137 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp136)
1668 %tmp138 = bitcast i64 %tmp137 to <2 x i32>
1669 call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %tmp138, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
1670 ret void
1671 }
1673 define amdgpu_cs void @strict_wwm_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {
1674 ; GFX9-O0-LABEL: strict_wwm_amdgpu_cs_main:
1675 ; GFX9-O0: ; %bb.0:
1676 ; GFX9-O0-NEXT: s_mov_b32 s4, s3
1677 ; GFX9-O0-NEXT: s_mov_b32 s5, s2
1678 ; GFX9-O0-NEXT: s_mov_b32 s6, s1
1679 ; GFX9-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1_sgpr2_sgpr3
1680 ; GFX9-O0-NEXT: s_mov_b32 s1, s6
1681 ; GFX9-O0-NEXT: s_mov_b32 s2, s5
1682 ; GFX9-O0-NEXT: s_mov_b32 s3, s4
1683 ; GFX9-O0-NEXT: ; kill: def $sgpr4_sgpr5_sgpr6_sgpr7 killed $sgpr0_sgpr1_sgpr2_sgpr3
1684 ; GFX9-O0-NEXT: s_mov_b32 s4, 5
1685 ; GFX9-O0-NEXT: v_lshlrev_b32_e64 v0, s4, v0
1686 ; GFX9-O0-NEXT: s_mov_b32 s4, 0
1687 ; GFX9-O0-NEXT: buffer_load_dwordx4 v[10:13], v0, s[0:3], s4 offen
1688 ; GFX9-O0-NEXT: buffer_load_dwordx2 v[3:4], v0, s[0:3], s4 offen offset:16
1689 ; GFX9-O0-NEXT: s_waitcnt vmcnt(1)
1690 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v11
1691 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v10
1692 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1693 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1694 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
1695 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v7
1696 ; GFX9-O0-NEXT: s_mov_b32 s5, 0x7fffffff
1697 ; GFX9-O0-NEXT: s_mov_b32 s6, -1
1698 ; GFX9-O0-NEXT: ; kill: def $sgpr6 killed $sgpr6 def $sgpr6_sgpr7
1699 ; GFX9-O0-NEXT: s_mov_b32 s7, s5
1700 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v5
1701 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
1702 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1703 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
1704 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
1705 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1706 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, v1
1707 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, v2
1708 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v13
1709 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v12
1710 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1711 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1712 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
1713 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v7
1714 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v5
1715 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v6
1716 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1717 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
1718 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
1719 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1720 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v1
1721 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v2
1722 ; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
1723 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
1724 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, v4
1725 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1726 ; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
1727 ; GFX9-O0-NEXT: v_mov_b32_e32 v2, s7
1728 ; GFX9-O0-NEXT: s_not_b64 exec, exec
1729 ; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1
1730 ; GFX9-O0-NEXT: v_mov_b32_e32 v4, v2
1731 ; GFX9-O0-NEXT: v_mov_b32_e32 v11, v9
1732 ; GFX9-O0-NEXT: v_mov_b32_e32 v5, v8
1733 ; GFX9-O0-NEXT: v_mov_b32_e32 v9, v7
1734 ; GFX9-O0-NEXT: v_mov_b32_e32 v10, v6
1735 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1736 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1737 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1738 ; GFX9-O0-NEXT: ; implicit-def: $sgpr5
1739 ; GFX9-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6_vgpr7_vgpr8 killed $exec
1740 ; GFX9-O0-NEXT: v_mov_b32_e32 v6, v11
1741 ; GFX9-O0-NEXT: v_mov_b32_e32 v7, v10
1742 ; GFX9-O0-NEXT: v_mov_b32_e32 v8, v9
1743 ; GFX9-O0-NEXT: buffer_store_dwordx4 v[5:8], v0, s[0:3], s4 offen
1744 ; GFX9-O0-NEXT: buffer_store_dwordx2 v[3:4], v0, s[0:3], s4 offen offset:16
1745 ; GFX9-O0-NEXT: s_endpgm
1746 ;
1747 ; GFX9-O3-LABEL: strict_wwm_amdgpu_cs_main:
1748 ; GFX9-O3: ; %bb.0:
1749 ; GFX9-O3-NEXT: v_lshlrev_b32_e32 v0, 5, v0
1750 ; GFX9-O3-NEXT: buffer_load_dwordx4 v[7:10], v0, s[0:3], 0 offen
1751 ; GFX9-O3-NEXT: buffer_load_dwordx2 v[11:12], v0, s[0:3], 0 offen offset:16
1752 ; GFX9-O3-NEXT: s_mov_b32 s4, -1
1753 ; GFX9-O3-NEXT: s_brev_b32 s5, -2
1754 ; GFX9-O3-NEXT: s_waitcnt vmcnt(1)
1755 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, v7
1756 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, v8
1757 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1758 ; GFX9-O3-NEXT: v_mov_b32_e32 v1, s4
1759 ; GFX9-O3-NEXT: v_mov_b32_e32 v2, s5
1760 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1761 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, v9
1762 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, v10
1763 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1764 ; GFX9-O3-NEXT: v_mov_b32_e32 v3, s4
1765 ; GFX9-O3-NEXT: v_mov_b32_e32 v4, s5
1766 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1767 ; GFX9-O3-NEXT: s_waitcnt vmcnt(0)
1768 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, v11
1769 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, v12
1770 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1771 ; GFX9-O3-NEXT: v_mov_b32_e32 v5, s4
1772 ; GFX9-O3-NEXT: v_mov_b32_e32 v6, s5
1773 ; GFX9-O3-NEXT: s_not_b64 exec, exec
1774 ; GFX9-O3-NEXT: v_mov_b32_e32 v7, v1
1775 ; GFX9-O3-NEXT: v_mov_b32_e32 v9, v3
1776 ; GFX9-O3-NEXT: v_mov_b32_e32 v8, v2
1777 ; GFX9-O3-NEXT: v_mov_b32_e32 v10, v4
1778 ; GFX9-O3-NEXT: v_mov_b32_e32 v11, v5
1779 ; GFX9-O3-NEXT: v_mov_b32_e32 v12, v6
1780 ; GFX9-O3-NEXT: buffer_store_dwordx4 v[7:10], v0, s[0:3], 0 offen
1781 ; GFX9-O3-NEXT: buffer_store_dwordx2 v[11:12], v0, s[0:3], 0 offen offset:16
1782 ; GFX9-O3-NEXT: s_endpgm
1783 %tmp17 = shl i32 %index, 5
1784 %tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)
1785 %.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>
1786 %tmp19 = or i32 %tmp17, 16
1787 %tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)
1788 %.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0
1789 %tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)
1790 %tmp97 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp22)
1791 %.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1
1792 %tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)
1793 %tmp174 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp99)
1794 %.i25 = bitcast <2 x i32> %tmp20 to i64
1795 %tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)
1796 %tmp251 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp176)
1797 %.cast = bitcast i64 %tmp97 to <2 x float>
1798 %.cast6 = bitcast i64 %tmp174 to <2 x float>
1799 %.cast7 = bitcast i64 %tmp251 to <2 x float>
1800 %tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1801 %desc.int = bitcast <4 x i32> %desc to i128
1802 %desc.ptr = inttoptr i128 %desc.int to ptr addrspace(8)
1803 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %tmp254, ptr addrspace(8) %desc.ptr, i32 %tmp17, i32 0, i32 0)
1804 tail call void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float> %.cast7, ptr addrspace(8)%desc.ptr, i32 %tmp19, i32 0, i32 0)
1805 ret void
1806 }
1808 declare i32 @llvm.amdgcn.strict.wwm.i32(i32)
1809 declare i64 @llvm.amdgcn.strict.wwm.i64(i64)
1810 declare i32 @llvm.amdgcn.wwm.i32(i32)
1811 declare i64 @llvm.amdgcn.wwm.i64(i64)
1812 declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32)
1813 declare i64 @llvm.amdgcn.set.inactive.i64(i64, i64)
1814 declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1)
1815 declare <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8), i32, i32, i32)
1816 declare void @llvm.amdgcn.raw.ptr.buffer.store.f32(float, ptr addrspace(8), i32, i32, i32)
1817 declare void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32, ptr addrspace(8), i32, i32, i32)
1818 declare void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32>, ptr addrspace(8), i32, i32, i32)
1819 declare void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float>, ptr addrspace(8), i32, i32, i32)
1820 declare void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32)
1821 declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
1822 declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)
1824 !llvm.module.flags = !{!0}
1825 !0 = !{i32 1, !"amdhsa_code_object_version", i32 500}