| blob | af713179a888dd52e97da0ba24d5dce5b0c0dd54 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
2 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1030 | FileCheck %s -check-prefixes=GFX10
3 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 | FileCheck %s -check-prefix=GFX11
4 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 | FileCheck %s -check-prefix=GFX12
6 ;; Check that consecutive store operations are grouped greedily into
7 ;; hard clauses of the appropriate length for each target.
8 ;; This test uses <4 x i32> stores in order to prevent the stores from
9 ;; being combined into larger operations due to their adjecency.
10 define amdgpu_kernel void @long_store_chain(ptr addrspace(1) %p) {
11 ; GFX10-LABEL: long_store_chain:
12 ; GFX10: ; %bb.0:
13 ; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x24
14 ; GFX10-NEXT: s_mov_b32 s0, 0
15 ; GFX10-NEXT: v_mov_b32_e32 v4, 0
16 ; GFX10-NEXT: s_mov_b32 s1, s0
17 ; GFX10-NEXT: s_mov_b32 s2, s0
18 ; GFX10-NEXT: s_mov_b32 s3, s0
19 ; GFX10-NEXT: v_mov_b32_e32 v0, s0
20 ; GFX10-NEXT: v_mov_b32_e32 v1, s1
21 ; GFX10-NEXT: v_mov_b32_e32 v2, s2
22 ; GFX10-NEXT: v_mov_b32_e32 v3, s3
23 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
24 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5]
25 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:16
26 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:32
27 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:48
28 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:64
29 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:80
30 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:96
31 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:112
32 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:128
33 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:144
34 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:160
35 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:176
36 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:192
37 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:208
38 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:224
39 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:240
40 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:256
41 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:272
42 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:288
43 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:304
44 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:320
45 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:336
46 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:352
47 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:368
48 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:384
49 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:400
50 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:416
51 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:432
52 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:448
53 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:464
54 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:480
55 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:496
56 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:512
57 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:528
58 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:544
59 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:560
60 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:576
61 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:592
62 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:608
63 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:624
64 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:640
65 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:656
66 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:672
67 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:688
68 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:704
69 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:720
70 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:736
71 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:752
72 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:768
73 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:784
74 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:800
75 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:816
76 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:832
77 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:848
78 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:864
79 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:880
80 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:896
81 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:912
82 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:928
83 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:944
84 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:960
85 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:976
86 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:992
87 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:1008
88 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:1024
89 ; GFX10-NEXT: global_store_dwordx4 v4, v[0:3], s[4:5] offset:1040
90 ; GFX10-NEXT: s_endpgm
91 ;
92 ; GFX11-LABEL: long_store_chain:
93 ; GFX11: ; %bb.0:
94 ; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x24
95 ; GFX11-NEXT: s_mov_b32 s0, 0
96 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
97 ; GFX11-NEXT: s_mov_b32 s1, s0
98 ; GFX11-NEXT: s_mov_b32 s2, s0
99 ; GFX11-NEXT: s_mov_b32 s3, s0
100 ; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
101 ; GFX11-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s1
102 ; GFX11-NEXT: v_mov_b32_e32 v2, s2
103 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
104 ; GFX11-NEXT: s_clause 0x1f
105 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5]
106 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:16
107 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:32
108 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:48
109 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:64
110 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:80
111 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:96
112 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:112
113 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:128
114 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:144
115 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:160
116 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:176
117 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:192
118 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:208
119 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:224
120 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:240
121 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:256
122 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:272
123 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:288
124 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:304
125 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:320
126 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:336
127 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:352
128 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:368
129 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:384
130 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:400
131 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:416
132 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:432
133 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:448
134 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:464
135 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:480
136 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:496
137 ; GFX11-NEXT: s_clause 0x1f
138 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:512
139 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:528
140 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:544
141 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:560
142 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:576
143 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:592
144 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:608
145 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:624
146 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:640
147 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:656
148 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:672
149 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:688
150 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:704
151 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:720
152 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:736
153 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:752
154 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:768
155 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:784
156 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:800
157 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:816
158 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:832
159 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:848
160 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:864
161 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:880
162 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:896
163 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:912
164 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:928
165 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:944
166 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:960
167 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:976
168 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:992
169 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1008
170 ; GFX11-NEXT: s_clause 0x1
171 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1024
172 ; GFX11-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1040
173 ; GFX11-NEXT: s_endpgm
174 ;
175 ; GFX12-LABEL: long_store_chain:
176 ; GFX12: ; %bb.0:
177 ; GFX12-NEXT: s_load_b64 s[4:5], s[4:5], 0x24
178 ; GFX12-NEXT: s_mov_b32 s0, 0
179 ; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
180 ; GFX12-NEXT: s_mov_b32 s1, s0
181 ; GFX12-NEXT: s_mov_b32 s2, s0
182 ; GFX12-NEXT: s_mov_b32 s3, s0
183 ; GFX12-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3
184 ; GFX12-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s1
185 ; GFX12-NEXT: v_mov_b32_e32 v2, s2
186 ; GFX12-NEXT: s_wait_kmcnt 0x0
187 ; GFX12-NEXT: s_clause 0x1f
188 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5]
189 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:16
190 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:32
191 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:48
192 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:64
193 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:80
194 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:96
195 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:112
196 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:128
197 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:144
198 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:160
199 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:176
200 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:192
201 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:208
202 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:224
203 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:240
204 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:256
205 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:272
206 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:288
207 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:304
208 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:320
209 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:336
210 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:352
211 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:368
212 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:384
213 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:400
214 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:416
215 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:432
216 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:448
217 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:464
218 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:480
219 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:496
220 ; GFX12-NEXT: s_clause 0x1f
221 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:512
222 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:528
223 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:544
224 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:560
225 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:576
226 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:592
227 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:608
228 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:624
229 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:640
230 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:656
231 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:672
232 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:688
233 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:704
234 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:720
235 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:736
236 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:752
237 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:768
238 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:784
239 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:800
240 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:816
241 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:832
242 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:848
243 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:864
244 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:880
245 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:896
246 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:912
247 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:928
248 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:944
249 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:960
250 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:976
251 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:992
252 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1008
253 ; GFX12-NEXT: s_clause 0x1
254 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1024
255 ; GFX12-NEXT: global_store_b128 v4, v[0:3], s[4:5] offset:1040
256 ; GFX12-NEXT: s_endpgm
257 store <4 x i32> zeroinitializer, ptr addrspace(1) %p
258 %ptr1 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 1
259 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr1
260 %ptr2 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 2
261 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr2
262 %ptr3 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 3
263 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr3
264 %ptr4 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 4
265 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr4
266 %ptr5 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 5
267 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr5
268 %ptr6 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 6
269 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr6
270 %ptr7 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 7
271 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr7
272 %ptr8 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 8
273 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr8
274 %ptr9 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 9
275 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr9
276 %ptr10 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 10
277 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr10
278 %ptr11 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 11
279 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr11
280 %ptr12 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 12
281 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr12
282 %ptr13 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 13
283 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr13
284 %ptr14 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 14
285 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr14
286 %ptr15 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 15
287 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr15
288 %ptr16 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 16
289 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr16
290 %ptr17 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 17
291 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr17
292 %ptr18 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 18
293 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr18
294 %ptr19 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 19
295 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr19
296 %ptr20 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 20
297 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr20
298 %ptr21 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 21
299 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr21
300 %ptr22 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 22
301 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr22
302 %ptr23 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 23
303 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr23
304 %ptr24 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 24
305 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr24
306 %ptr25 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 25
307 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr25
308 %ptr26 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 26
309 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr26
310 %ptr27 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 27
311 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr27
312 %ptr28 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 28
313 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr28
314 %ptr29 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 29
315 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr29
316 %ptr30 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 30
317 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr30
318 %ptr31 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 31
319 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr31
320 %ptr32 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 32
321 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr32
322 %ptr33 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 33
323 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr33
324 %ptr34 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 34
325 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr34
326 %ptr35 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 35
327 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr35
328 %ptr36 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 36
329 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr36
330 %ptr37 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 37
331 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr37
332 %ptr38 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 38
333 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr38
334 %ptr39 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 39
335 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr39
336 %ptr40 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 40
337 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr40
338 %ptr41 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 41
339 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr41
340 %ptr42 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 42
341 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr42
342 %ptr43 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 43
343 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr43
344 %ptr44 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 44
345 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr44
346 %ptr45 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 45
347 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr45
348 %ptr46 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 46
349 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr46
350 %ptr47 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 47
351 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr47
352 %ptr48 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 48
353 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr48
354 %ptr49 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 49
355 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr49
356 %ptr50 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 50
357 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr50
358 %ptr51 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 51
359 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr51
360 %ptr52 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 52
361 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr52
362 %ptr53 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 53
363 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr53
364 %ptr54 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 54
365 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr54
366 %ptr55 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 55
367 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr55
368 %ptr56 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 56
369 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr56
370 %ptr57 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 57
371 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr57
372 %ptr58 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 58
373 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr58
374 %ptr59 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 59
375 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr59
376 %ptr60 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 60
377 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr60
378 %ptr61 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 61
379 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr61
380 %ptr62 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 62
381 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr62
382 %ptr63 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 63
383 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr63
384 %ptr64 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 64
385 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr64
386 %ptr65 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 65
387 store <4 x i32> zeroinitializer, ptr addrspace(1) %ptr65
388 ret void
389 }
391 ;; Long chain of loads since gfx10 doesn't cluster stores.
392 ;; Use i32 loads to save on register pressure
393 define amdgpu_kernel void @long_load_chain(ptr addrspace(1) %p) {
394 ; GFX10-LABEL: long_load_chain:
395 ; GFX10: ; %bb.0:
396 ; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
397 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
398 ; GFX10-NEXT: s_clause 0x3e
399 ; GFX10-NEXT: s_load_dword s2, s[0:1], 0x0
400 ; GFX10-NEXT: s_load_dword s3, s[0:1], 0x10
401 ; GFX10-NEXT: s_load_dword s4, s[0:1], 0x20
402 ; GFX10-NEXT: s_load_dword s5, s[0:1], 0x30
403 ; GFX10-NEXT: s_load_dword s6, s[0:1], 0x40
404 ; GFX10-NEXT: s_load_dword s7, s[0:1], 0x50
405 ; GFX10-NEXT: s_load_dword s8, s[0:1], 0x60
406 ; GFX10-NEXT: s_load_dword s9, s[0:1], 0x70
407 ; GFX10-NEXT: s_load_dword s10, s[0:1], 0x80
408 ; GFX10-NEXT: s_load_dword s11, s[0:1], 0x90
409 ; GFX10-NEXT: s_load_dword s12, s[0:1], 0xa0
410 ; GFX10-NEXT: s_load_dword s13, s[0:1], 0xb0
411 ; GFX10-NEXT: s_load_dword s14, s[0:1], 0xc0
412 ; GFX10-NEXT: s_load_dword s15, s[0:1], 0xd0
413 ; GFX10-NEXT: s_load_dword s16, s[0:1], 0xe0
414 ; GFX10-NEXT: s_load_dword s17, s[0:1], 0xf0
415 ; GFX10-NEXT: s_load_dword s18, s[0:1], 0x100
416 ; GFX10-NEXT: s_load_dword s19, s[0:1], 0x110
417 ; GFX10-NEXT: s_load_dword s20, s[0:1], 0x120
418 ; GFX10-NEXT: s_load_dword s21, s[0:1], 0x130
419 ; GFX10-NEXT: s_load_dword s22, s[0:1], 0x140
420 ; GFX10-NEXT: s_load_dword s23, s[0:1], 0x150
421 ; GFX10-NEXT: s_load_dword s24, s[0:1], 0x160
422 ; GFX10-NEXT: s_load_dword s25, s[0:1], 0x170
423 ; GFX10-NEXT: s_load_dword s26, s[0:1], 0x180
424 ; GFX10-NEXT: s_load_dword s27, s[0:1], 0x190
425 ; GFX10-NEXT: s_load_dword s28, s[0:1], 0x1a0
426 ; GFX10-NEXT: s_load_dword s29, s[0:1], 0x1b0
427 ; GFX10-NEXT: s_load_dword s30, s[0:1], 0x1c0
428 ; GFX10-NEXT: s_load_dword s31, s[0:1], 0x1d0
429 ; GFX10-NEXT: s_load_dword s33, s[0:1], 0x1e0
430 ; GFX10-NEXT: s_load_dword s34, s[0:1], 0x1f0
431 ; GFX10-NEXT: s_load_dword s35, s[0:1], 0x200
432 ; GFX10-NEXT: s_load_dword s36, s[0:1], 0x210
433 ; GFX10-NEXT: s_load_dword s37, s[0:1], 0x220
434 ; GFX10-NEXT: s_load_dword s38, s[0:1], 0x230
435 ; GFX10-NEXT: s_load_dword s39, s[0:1], 0x240
436 ; GFX10-NEXT: s_load_dword s40, s[0:1], 0x250
437 ; GFX10-NEXT: s_load_dword s41, s[0:1], 0x260
438 ; GFX10-NEXT: s_load_dword s42, s[0:1], 0x270
439 ; GFX10-NEXT: s_load_dword s43, s[0:1], 0x280
440 ; GFX10-NEXT: s_load_dword s44, s[0:1], 0x290
441 ; GFX10-NEXT: s_load_dword s45, s[0:1], 0x2a0
442 ; GFX10-NEXT: s_load_dword s46, s[0:1], 0x2b0
443 ; GFX10-NEXT: s_load_dword s47, s[0:1], 0x2c0
444 ; GFX10-NEXT: s_load_dword s48, s[0:1], 0x2d0
445 ; GFX10-NEXT: s_load_dword s49, s[0:1], 0x2e0
446 ; GFX10-NEXT: s_load_dword s50, s[0:1], 0x2f0
447 ; GFX10-NEXT: s_load_dword s51, s[0:1], 0x300
448 ; GFX10-NEXT: s_load_dword s52, s[0:1], 0x310
449 ; GFX10-NEXT: s_load_dword s53, s[0:1], 0x320
450 ; GFX10-NEXT: s_load_dword s54, s[0:1], 0x330
451 ; GFX10-NEXT: s_load_dword s55, s[0:1], 0x340
452 ; GFX10-NEXT: s_load_dword s56, s[0:1], 0x350
453 ; GFX10-NEXT: s_load_dword s57, s[0:1], 0x360
454 ; GFX10-NEXT: s_load_dword s58, s[0:1], 0x370
455 ; GFX10-NEXT: s_load_dword s59, s[0:1], 0x380
456 ; GFX10-NEXT: s_load_dword s60, s[0:1], 0x390
457 ; GFX10-NEXT: s_load_dword s61, s[0:1], 0x3a0
458 ; GFX10-NEXT: s_load_dword s62, s[0:1], 0x3b0
459 ; GFX10-NEXT: s_load_dword s63, s[0:1], 0x3c0
460 ; GFX10-NEXT: s_load_dword s64, s[0:1], 0x3d0
461 ; GFX10-NEXT: s_load_dword s65, s[0:1], 0x3e0
462 ; GFX10-NEXT: s_clause 0x2
463 ; GFX10-NEXT: s_load_dword s66, s[0:1], 0x3f0
464 ; GFX10-NEXT: s_load_dword s67, s[0:1], 0x400
465 ; GFX10-NEXT: s_load_dword s0, s[0:1], 0x410
466 ; GFX10-NEXT: s_waitcnt lgkmcnt(0)
467 ; GFX10-NEXT: ;;#ASMSTART
468 ; GFX10-NEXT: ; use s2
469 ; GFX10-NEXT: ;;#ASMEND
470 ; GFX10-NEXT: ;;#ASMSTART
471 ; GFX10-NEXT: ; use s3
472 ; GFX10-NEXT: ;;#ASMEND
473 ; GFX10-NEXT: ;;#ASMSTART
474 ; GFX10-NEXT: ; use s4
475 ; GFX10-NEXT: ;;#ASMEND
476 ; GFX10-NEXT: ;;#ASMSTART
477 ; GFX10-NEXT: ; use s5
478 ; GFX10-NEXT: ;;#ASMEND
479 ; GFX10-NEXT: ;;#ASMSTART
480 ; GFX10-NEXT: ; use s6
481 ; GFX10-NEXT: ;;#ASMEND
482 ; GFX10-NEXT: ;;#ASMSTART
483 ; GFX10-NEXT: ; use s7
484 ; GFX10-NEXT: ;;#ASMEND
485 ; GFX10-NEXT: ;;#ASMSTART
486 ; GFX10-NEXT: ; use s8
487 ; GFX10-NEXT: ;;#ASMEND
488 ; GFX10-NEXT: ;;#ASMSTART
489 ; GFX10-NEXT: ; use s9
490 ; GFX10-NEXT: ;;#ASMEND
491 ; GFX10-NEXT: ;;#ASMSTART
492 ; GFX10-NEXT: ; use s10
493 ; GFX10-NEXT: ;;#ASMEND
494 ; GFX10-NEXT: ;;#ASMSTART
495 ; GFX10-NEXT: ; use s11
496 ; GFX10-NEXT: ;;#ASMEND
497 ; GFX10-NEXT: ;;#ASMSTART
498 ; GFX10-NEXT: ; use s12
499 ; GFX10-NEXT: ;;#ASMEND
500 ; GFX10-NEXT: ;;#ASMSTART
501 ; GFX10-NEXT: ; use s13
502 ; GFX10-NEXT: ;;#ASMEND
503 ; GFX10-NEXT: ;;#ASMSTART
504 ; GFX10-NEXT: ; use s14
505 ; GFX10-NEXT: ;;#ASMEND
506 ; GFX10-NEXT: ;;#ASMSTART
507 ; GFX10-NEXT: ; use s15
508 ; GFX10-NEXT: ;;#ASMEND
509 ; GFX10-NEXT: ;;#ASMSTART
510 ; GFX10-NEXT: ; use s16
511 ; GFX10-NEXT: ;;#ASMEND
512 ; GFX10-NEXT: ;;#ASMSTART
513 ; GFX10-NEXT: ; use s17
514 ; GFX10-NEXT: ;;#ASMEND
515 ; GFX10-NEXT: ;;#ASMSTART
516 ; GFX10-NEXT: ; use s18
517 ; GFX10-NEXT: ;;#ASMEND
518 ; GFX10-NEXT: ;;#ASMSTART
519 ; GFX10-NEXT: ; use s19
520 ; GFX10-NEXT: ;;#ASMEND
521 ; GFX10-NEXT: ;;#ASMSTART
522 ; GFX10-NEXT: ; use s20
523 ; GFX10-NEXT: ;;#ASMEND
524 ; GFX10-NEXT: ;;#ASMSTART
525 ; GFX10-NEXT: ; use s21
526 ; GFX10-NEXT: ;;#ASMEND
527 ; GFX10-NEXT: ;;#ASMSTART
528 ; GFX10-NEXT: ; use s22
529 ; GFX10-NEXT: ;;#ASMEND
530 ; GFX10-NEXT: ;;#ASMSTART
531 ; GFX10-NEXT: ; use s23
532 ; GFX10-NEXT: ;;#ASMEND
533 ; GFX10-NEXT: ;;#ASMSTART
534 ; GFX10-NEXT: ; use s24
535 ; GFX10-NEXT: ;;#ASMEND
536 ; GFX10-NEXT: ;;#ASMSTART
537 ; GFX10-NEXT: ; use s25
538 ; GFX10-NEXT: ;;#ASMEND
539 ; GFX10-NEXT: ;;#ASMSTART
540 ; GFX10-NEXT: ; use s26
541 ; GFX10-NEXT: ;;#ASMEND
542 ; GFX10-NEXT: ;;#ASMSTART
543 ; GFX10-NEXT: ; use s27
544 ; GFX10-NEXT: ;;#ASMEND
545 ; GFX10-NEXT: ;;#ASMSTART
546 ; GFX10-NEXT: ; use s28
547 ; GFX10-NEXT: ;;#ASMEND
548 ; GFX10-NEXT: ;;#ASMSTART
549 ; GFX10-NEXT: ; use s29
550 ; GFX10-NEXT: ;;#ASMEND
551 ; GFX10-NEXT: ;;#ASMSTART
552 ; GFX10-NEXT: ; use s30
553 ; GFX10-NEXT: ;;#ASMEND
554 ; GFX10-NEXT: ;;#ASMSTART
555 ; GFX10-NEXT: ; use s31
556 ; GFX10-NEXT: ;;#ASMEND
557 ; GFX10-NEXT: ;;#ASMSTART
558 ; GFX10-NEXT: ; use s33
559 ; GFX10-NEXT: ;;#ASMEND
560 ; GFX10-NEXT: ;;#ASMSTART
561 ; GFX10-NEXT: ; use s34
562 ; GFX10-NEXT: ;;#ASMEND
563 ; GFX10-NEXT: ;;#ASMSTART
564 ; GFX10-NEXT: ; use s35
565 ; GFX10-NEXT: ;;#ASMEND
566 ; GFX10-NEXT: ;;#ASMSTART
567 ; GFX10-NEXT: ; use s36
568 ; GFX10-NEXT: ;;#ASMEND
569 ; GFX10-NEXT: ;;#ASMSTART
570 ; GFX10-NEXT: ; use s37
571 ; GFX10-NEXT: ;;#ASMEND
572 ; GFX10-NEXT: ;;#ASMSTART
573 ; GFX10-NEXT: ; use s38
574 ; GFX10-NEXT: ;;#ASMEND
575 ; GFX10-NEXT: ;;#ASMSTART
576 ; GFX10-NEXT: ; use s39
577 ; GFX10-NEXT: ;;#ASMEND
578 ; GFX10-NEXT: ;;#ASMSTART
579 ; GFX10-NEXT: ; use s40
580 ; GFX10-NEXT: ;;#ASMEND
581 ; GFX10-NEXT: ;;#ASMSTART
582 ; GFX10-NEXT: ; use s41
583 ; GFX10-NEXT: ;;#ASMEND
584 ; GFX10-NEXT: ;;#ASMSTART
585 ; GFX10-NEXT: ; use s42
586 ; GFX10-NEXT: ;;#ASMEND
587 ; GFX10-NEXT: ;;#ASMSTART
588 ; GFX10-NEXT: ; use s43
589 ; GFX10-NEXT: ;;#ASMEND
590 ; GFX10-NEXT: ;;#ASMSTART
591 ; GFX10-NEXT: ; use s44
592 ; GFX10-NEXT: ;;#ASMEND
593 ; GFX10-NEXT: ;;#ASMSTART
594 ; GFX10-NEXT: ; use s45
595 ; GFX10-NEXT: ;;#ASMEND
596 ; GFX10-NEXT: ;;#ASMSTART
597 ; GFX10-NEXT: ; use s46
598 ; GFX10-NEXT: ;;#ASMEND
599 ; GFX10-NEXT: ;;#ASMSTART
600 ; GFX10-NEXT: ; use s47
601 ; GFX10-NEXT: ;;#ASMEND
602 ; GFX10-NEXT: ;;#ASMSTART
603 ; GFX10-NEXT: ; use s48
604 ; GFX10-NEXT: ;;#ASMEND
605 ; GFX10-NEXT: ;;#ASMSTART
606 ; GFX10-NEXT: ; use s49
607 ; GFX10-NEXT: ;;#ASMEND
608 ; GFX10-NEXT: ;;#ASMSTART
609 ; GFX10-NEXT: ; use s50
610 ; GFX10-NEXT: ;;#ASMEND
611 ; GFX10-NEXT: ;;#ASMSTART
612 ; GFX10-NEXT: ; use s51
613 ; GFX10-NEXT: ;;#ASMEND
614 ; GFX10-NEXT: ;;#ASMSTART
615 ; GFX10-NEXT: ; use s52
616 ; GFX10-NEXT: ;;#ASMEND
617 ; GFX10-NEXT: ;;#ASMSTART
618 ; GFX10-NEXT: ; use s53
619 ; GFX10-NEXT: ;;#ASMEND
620 ; GFX10-NEXT: ;;#ASMSTART
621 ; GFX10-NEXT: ; use s54
622 ; GFX10-NEXT: ;;#ASMEND
623 ; GFX10-NEXT: ;;#ASMSTART
624 ; GFX10-NEXT: ; use s55
625 ; GFX10-NEXT: ;;#ASMEND
626 ; GFX10-NEXT: ;;#ASMSTART
627 ; GFX10-NEXT: ; use s56
628 ; GFX10-NEXT: ;;#ASMEND
629 ; GFX10-NEXT: ;;#ASMSTART
630 ; GFX10-NEXT: ; use s57
631 ; GFX10-NEXT: ;;#ASMEND
632 ; GFX10-NEXT: ;;#ASMSTART
633 ; GFX10-NEXT: ; use s58
634 ; GFX10-NEXT: ;;#ASMEND
635 ; GFX10-NEXT: ;;#ASMSTART
636 ; GFX10-NEXT: ; use s59
637 ; GFX10-NEXT: ;;#ASMEND
638 ; GFX10-NEXT: ;;#ASMSTART
639 ; GFX10-NEXT: ; use s60
640 ; GFX10-NEXT: ;;#ASMEND
641 ; GFX10-NEXT: ;;#ASMSTART
642 ; GFX10-NEXT: ; use s61
643 ; GFX10-NEXT: ;;#ASMEND
644 ; GFX10-NEXT: ;;#ASMSTART
645 ; GFX10-NEXT: ; use s62
646 ; GFX10-NEXT: ;;#ASMEND
647 ; GFX10-NEXT: ;;#ASMSTART
648 ; GFX10-NEXT: ; use s63
649 ; GFX10-NEXT: ;;#ASMEND
650 ; GFX10-NEXT: ;;#ASMSTART
651 ; GFX10-NEXT: ; use s64
652 ; GFX10-NEXT: ;;#ASMEND
653 ; GFX10-NEXT: ;;#ASMSTART
654 ; GFX10-NEXT: ; use s65
655 ; GFX10-NEXT: ;;#ASMEND
656 ; GFX10-NEXT: ;;#ASMSTART
657 ; GFX10-NEXT: ; use s66
658 ; GFX10-NEXT: ;;#ASMEND
659 ; GFX10-NEXT: ;;#ASMSTART
660 ; GFX10-NEXT: ; use s67
661 ; GFX10-NEXT: ;;#ASMEND
662 ; GFX10-NEXT: ;;#ASMSTART
663 ; GFX10-NEXT: ; use s0
664 ; GFX10-NEXT: ;;#ASMEND
665 ; GFX10-NEXT: s_endpgm
666 ;
667 ; GFX11-LABEL: long_load_chain:
668 ; GFX11: ; %bb.0:
669 ; GFX11-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
670 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
671 ; GFX11-NEXT: s_clause 0x1f
672 ; GFX11-NEXT: s_load_b32 s2, s[0:1], 0x0
673 ; GFX11-NEXT: s_load_b32 s3, s[0:1], 0x10
674 ; GFX11-NEXT: s_load_b32 s4, s[0:1], 0x20
675 ; GFX11-NEXT: s_load_b32 s5, s[0:1], 0x30
676 ; GFX11-NEXT: s_load_b32 s6, s[0:1], 0x40
677 ; GFX11-NEXT: s_load_b32 s7, s[0:1], 0x50
678 ; GFX11-NEXT: s_load_b32 s8, s[0:1], 0x60
679 ; GFX11-NEXT: s_load_b32 s9, s[0:1], 0x70
680 ; GFX11-NEXT: s_load_b32 s10, s[0:1], 0x80
681 ; GFX11-NEXT: s_load_b32 s11, s[0:1], 0x90
682 ; GFX11-NEXT: s_load_b32 s12, s[0:1], 0xa0
683 ; GFX11-NEXT: s_load_b32 s13, s[0:1], 0xb0
684 ; GFX11-NEXT: s_load_b32 s14, s[0:1], 0xc0
685 ; GFX11-NEXT: s_load_b32 s15, s[0:1], 0xd0
686 ; GFX11-NEXT: s_load_b32 s16, s[0:1], 0xe0
687 ; GFX11-NEXT: s_load_b32 s17, s[0:1], 0xf0
688 ; GFX11-NEXT: s_load_b32 s18, s[0:1], 0x100
689 ; GFX11-NEXT: s_load_b32 s19, s[0:1], 0x110
690 ; GFX11-NEXT: s_load_b32 s20, s[0:1], 0x120
691 ; GFX11-NEXT: s_load_b32 s21, s[0:1], 0x130
692 ; GFX11-NEXT: s_load_b32 s22, s[0:1], 0x140
693 ; GFX11-NEXT: s_load_b32 s23, s[0:1], 0x150
694 ; GFX11-NEXT: s_load_b32 s24, s[0:1], 0x160
695 ; GFX11-NEXT: s_load_b32 s25, s[0:1], 0x170
696 ; GFX11-NEXT: s_load_b32 s26, s[0:1], 0x180
697 ; GFX11-NEXT: s_load_b32 s27, s[0:1], 0x190
698 ; GFX11-NEXT: s_load_b32 s28, s[0:1], 0x1a0
699 ; GFX11-NEXT: s_load_b32 s29, s[0:1], 0x1b0
700 ; GFX11-NEXT: s_load_b32 s30, s[0:1], 0x1c0
701 ; GFX11-NEXT: s_load_b32 s31, s[0:1], 0x1d0
702 ; GFX11-NEXT: s_load_b32 s33, s[0:1], 0x1e0
703 ; GFX11-NEXT: s_load_b32 s34, s[0:1], 0x1f0
704 ; GFX11-NEXT: s_clause 0x1f
705 ; GFX11-NEXT: s_load_b32 s35, s[0:1], 0x200
706 ; GFX11-NEXT: s_load_b32 s36, s[0:1], 0x210
707 ; GFX11-NEXT: s_load_b32 s37, s[0:1], 0x220
708 ; GFX11-NEXT: s_load_b32 s38, s[0:1], 0x230
709 ; GFX11-NEXT: s_load_b32 s39, s[0:1], 0x240
710 ; GFX11-NEXT: s_load_b32 s40, s[0:1], 0x250
711 ; GFX11-NEXT: s_load_b32 s41, s[0:1], 0x260
712 ; GFX11-NEXT: s_load_b32 s42, s[0:1], 0x270
713 ; GFX11-NEXT: s_load_b32 s43, s[0:1], 0x280
714 ; GFX11-NEXT: s_load_b32 s44, s[0:1], 0x290
715 ; GFX11-NEXT: s_load_b32 s45, s[0:1], 0x2a0
716 ; GFX11-NEXT: s_load_b32 s46, s[0:1], 0x2b0
717 ; GFX11-NEXT: s_load_b32 s47, s[0:1], 0x2c0
718 ; GFX11-NEXT: s_load_b32 s48, s[0:1], 0x2d0
719 ; GFX11-NEXT: s_load_b32 s49, s[0:1], 0x2e0
720 ; GFX11-NEXT: s_load_b32 s50, s[0:1], 0x2f0
721 ; GFX11-NEXT: s_load_b32 s51, s[0:1], 0x300
722 ; GFX11-NEXT: s_load_b32 s52, s[0:1], 0x310
723 ; GFX11-NEXT: s_load_b32 s53, s[0:1], 0x320
724 ; GFX11-NEXT: s_load_b32 s54, s[0:1], 0x330
725 ; GFX11-NEXT: s_load_b32 s55, s[0:1], 0x340
726 ; GFX11-NEXT: s_load_b32 s56, s[0:1], 0x350
727 ; GFX11-NEXT: s_load_b32 s57, s[0:1], 0x360
728 ; GFX11-NEXT: s_load_b32 s58, s[0:1], 0x370
729 ; GFX11-NEXT: s_load_b32 s59, s[0:1], 0x380
730 ; GFX11-NEXT: s_load_b32 s60, s[0:1], 0x390
731 ; GFX11-NEXT: s_load_b32 s61, s[0:1], 0x3a0
732 ; GFX11-NEXT: s_load_b32 s62, s[0:1], 0x3b0
733 ; GFX11-NEXT: s_load_b32 s63, s[0:1], 0x3c0
734 ; GFX11-NEXT: s_load_b32 s64, s[0:1], 0x3d0
735 ; GFX11-NEXT: s_load_b32 s65, s[0:1], 0x3e0
736 ; GFX11-NEXT: s_load_b32 s66, s[0:1], 0x3f0
737 ; GFX11-NEXT: s_clause 0x1
738 ; GFX11-NEXT: s_load_b32 s67, s[0:1], 0x400
739 ; GFX11-NEXT: s_load_b32 s0, s[0:1], 0x410
740 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
741 ; GFX11-NEXT: ;;#ASMSTART
742 ; GFX11-NEXT: ; use s2
743 ; GFX11-NEXT: ;;#ASMEND
744 ; GFX11-NEXT: ;;#ASMSTART
745 ; GFX11-NEXT: ; use s3
746 ; GFX11-NEXT: ;;#ASMEND
747 ; GFX11-NEXT: ;;#ASMSTART
748 ; GFX11-NEXT: ; use s4
749 ; GFX11-NEXT: ;;#ASMEND
750 ; GFX11-NEXT: ;;#ASMSTART
751 ; GFX11-NEXT: ; use s5
752 ; GFX11-NEXT: ;;#ASMEND
753 ; GFX11-NEXT: ;;#ASMSTART
754 ; GFX11-NEXT: ; use s6
755 ; GFX11-NEXT: ;;#ASMEND
756 ; GFX11-NEXT: ;;#ASMSTART
757 ; GFX11-NEXT: ; use s7
758 ; GFX11-NEXT: ;;#ASMEND
759 ; GFX11-NEXT: ;;#ASMSTART
760 ; GFX11-NEXT: ; use s8
761 ; GFX11-NEXT: ;;#ASMEND
762 ; GFX11-NEXT: ;;#ASMSTART
763 ; GFX11-NEXT: ; use s9
764 ; GFX11-NEXT: ;;#ASMEND
765 ; GFX11-NEXT: ;;#ASMSTART
766 ; GFX11-NEXT: ; use s10
767 ; GFX11-NEXT: ;;#ASMEND
768 ; GFX11-NEXT: ;;#ASMSTART
769 ; GFX11-NEXT: ; use s11
770 ; GFX11-NEXT: ;;#ASMEND
771 ; GFX11-NEXT: ;;#ASMSTART
772 ; GFX11-NEXT: ; use s12
773 ; GFX11-NEXT: ;;#ASMEND
774 ; GFX11-NEXT: ;;#ASMSTART
775 ; GFX11-NEXT: ; use s13
776 ; GFX11-NEXT: ;;#ASMEND
777 ; GFX11-NEXT: ;;#ASMSTART
778 ; GFX11-NEXT: ; use s14
779 ; GFX11-NEXT: ;;#ASMEND
780 ; GFX11-NEXT: ;;#ASMSTART
781 ; GFX11-NEXT: ; use s15
782 ; GFX11-NEXT: ;;#ASMEND
783 ; GFX11-NEXT: ;;#ASMSTART
784 ; GFX11-NEXT: ; use s16
785 ; GFX11-NEXT: ;;#ASMEND
786 ; GFX11-NEXT: ;;#ASMSTART
787 ; GFX11-NEXT: ; use s17
788 ; GFX11-NEXT: ;;#ASMEND
789 ; GFX11-NEXT: ;;#ASMSTART
790 ; GFX11-NEXT: ; use s18
791 ; GFX11-NEXT: ;;#ASMEND
792 ; GFX11-NEXT: ;;#ASMSTART
793 ; GFX11-NEXT: ; use s19
794 ; GFX11-NEXT: ;;#ASMEND
795 ; GFX11-NEXT: ;;#ASMSTART
796 ; GFX11-NEXT: ; use s20
797 ; GFX11-NEXT: ;;#ASMEND
798 ; GFX11-NEXT: ;;#ASMSTART
799 ; GFX11-NEXT: ; use s21
800 ; GFX11-NEXT: ;;#ASMEND
801 ; GFX11-NEXT: ;;#ASMSTART
802 ; GFX11-NEXT: ; use s22
803 ; GFX11-NEXT: ;;#ASMEND
804 ; GFX11-NEXT: ;;#ASMSTART
805 ; GFX11-NEXT: ; use s23
806 ; GFX11-NEXT: ;;#ASMEND
807 ; GFX11-NEXT: ;;#ASMSTART
808 ; GFX11-NEXT: ; use s24
809 ; GFX11-NEXT: ;;#ASMEND
810 ; GFX11-NEXT: ;;#ASMSTART
811 ; GFX11-NEXT: ; use s25
812 ; GFX11-NEXT: ;;#ASMEND
813 ; GFX11-NEXT: ;;#ASMSTART
814 ; GFX11-NEXT: ; use s26
815 ; GFX11-NEXT: ;;#ASMEND
816 ; GFX11-NEXT: ;;#ASMSTART
817 ; GFX11-NEXT: ; use s27
818 ; GFX11-NEXT: ;;#ASMEND
819 ; GFX11-NEXT: ;;#ASMSTART
820 ; GFX11-NEXT: ; use s28
821 ; GFX11-NEXT: ;;#ASMEND
822 ; GFX11-NEXT: ;;#ASMSTART
823 ; GFX11-NEXT: ; use s29
824 ; GFX11-NEXT: ;;#ASMEND
825 ; GFX11-NEXT: ;;#ASMSTART
826 ; GFX11-NEXT: ; use s30
827 ; GFX11-NEXT: ;;#ASMEND
828 ; GFX11-NEXT: ;;#ASMSTART
829 ; GFX11-NEXT: ; use s31
830 ; GFX11-NEXT: ;;#ASMEND
831 ; GFX11-NEXT: ;;#ASMSTART
832 ; GFX11-NEXT: ; use s33
833 ; GFX11-NEXT: ;;#ASMEND
834 ; GFX11-NEXT: ;;#ASMSTART
835 ; GFX11-NEXT: ; use s34
836 ; GFX11-NEXT: ;;#ASMEND
837 ; GFX11-NEXT: ;;#ASMSTART
838 ; GFX11-NEXT: ; use s35
839 ; GFX11-NEXT: ;;#ASMEND
840 ; GFX11-NEXT: ;;#ASMSTART
841 ; GFX11-NEXT: ; use s36
842 ; GFX11-NEXT: ;;#ASMEND
843 ; GFX11-NEXT: ;;#ASMSTART
844 ; GFX11-NEXT: ; use s37
845 ; GFX11-NEXT: ;;#ASMEND
846 ; GFX11-NEXT: ;;#ASMSTART
847 ; GFX11-NEXT: ; use s38
848 ; GFX11-NEXT: ;;#ASMEND
849 ; GFX11-NEXT: ;;#ASMSTART
850 ; GFX11-NEXT: ; use s39
851 ; GFX11-NEXT: ;;#ASMEND
852 ; GFX11-NEXT: ;;#ASMSTART
853 ; GFX11-NEXT: ; use s40
854 ; GFX11-NEXT: ;;#ASMEND
855 ; GFX11-NEXT: ;;#ASMSTART
856 ; GFX11-NEXT: ; use s41
857 ; GFX11-NEXT: ;;#ASMEND
858 ; GFX11-NEXT: ;;#ASMSTART
859 ; GFX11-NEXT: ; use s42
860 ; GFX11-NEXT: ;;#ASMEND
861 ; GFX11-NEXT: ;;#ASMSTART
862 ; GFX11-NEXT: ; use s43
863 ; GFX11-NEXT: ;;#ASMEND
864 ; GFX11-NEXT: ;;#ASMSTART
865 ; GFX11-NEXT: ; use s44
866 ; GFX11-NEXT: ;;#ASMEND
867 ; GFX11-NEXT: ;;#ASMSTART
868 ; GFX11-NEXT: ; use s45
869 ; GFX11-NEXT: ;;#ASMEND
870 ; GFX11-NEXT: ;;#ASMSTART
871 ; GFX11-NEXT: ; use s46
872 ; GFX11-NEXT: ;;#ASMEND
873 ; GFX11-NEXT: ;;#ASMSTART
874 ; GFX11-NEXT: ; use s47
875 ; GFX11-NEXT: ;;#ASMEND
876 ; GFX11-NEXT: ;;#ASMSTART
877 ; GFX11-NEXT: ; use s48
878 ; GFX11-NEXT: ;;#ASMEND
879 ; GFX11-NEXT: ;;#ASMSTART
880 ; GFX11-NEXT: ; use s49
881 ; GFX11-NEXT: ;;#ASMEND
882 ; GFX11-NEXT: ;;#ASMSTART
883 ; GFX11-NEXT: ; use s50
884 ; GFX11-NEXT: ;;#ASMEND
885 ; GFX11-NEXT: ;;#ASMSTART
886 ; GFX11-NEXT: ; use s51
887 ; GFX11-NEXT: ;;#ASMEND
888 ; GFX11-NEXT: ;;#ASMSTART
889 ; GFX11-NEXT: ; use s52
890 ; GFX11-NEXT: ;;#ASMEND
891 ; GFX11-NEXT: ;;#ASMSTART
892 ; GFX11-NEXT: ; use s53
893 ; GFX11-NEXT: ;;#ASMEND
894 ; GFX11-NEXT: ;;#ASMSTART
895 ; GFX11-NEXT: ; use s54
896 ; GFX11-NEXT: ;;#ASMEND
897 ; GFX11-NEXT: ;;#ASMSTART
898 ; GFX11-NEXT: ; use s55
899 ; GFX11-NEXT: ;;#ASMEND
900 ; GFX11-NEXT: ;;#ASMSTART
901 ; GFX11-NEXT: ; use s56
902 ; GFX11-NEXT: ;;#ASMEND
903 ; GFX11-NEXT: ;;#ASMSTART
904 ; GFX11-NEXT: ; use s57
905 ; GFX11-NEXT: ;;#ASMEND
906 ; GFX11-NEXT: ;;#ASMSTART
907 ; GFX11-NEXT: ; use s58
908 ; GFX11-NEXT: ;;#ASMEND
909 ; GFX11-NEXT: ;;#ASMSTART
910 ; GFX11-NEXT: ; use s59
911 ; GFX11-NEXT: ;;#ASMEND
912 ; GFX11-NEXT: ;;#ASMSTART
913 ; GFX11-NEXT: ; use s60
914 ; GFX11-NEXT: ;;#ASMEND
915 ; GFX11-NEXT: ;;#ASMSTART
916 ; GFX11-NEXT: ; use s61
917 ; GFX11-NEXT: ;;#ASMEND
918 ; GFX11-NEXT: ;;#ASMSTART
919 ; GFX11-NEXT: ; use s62
920 ; GFX11-NEXT: ;;#ASMEND
921 ; GFX11-NEXT: ;;#ASMSTART
922 ; GFX11-NEXT: ; use s63
923 ; GFX11-NEXT: ;;#ASMEND
924 ; GFX11-NEXT: ;;#ASMSTART
925 ; GFX11-NEXT: ; use s64
926 ; GFX11-NEXT: ;;#ASMEND
927 ; GFX11-NEXT: ;;#ASMSTART
928 ; GFX11-NEXT: ; use s65
929 ; GFX11-NEXT: ;;#ASMEND
930 ; GFX11-NEXT: ;;#ASMSTART
931 ; GFX11-NEXT: ; use s66
932 ; GFX11-NEXT: ;;#ASMEND
933 ; GFX11-NEXT: ;;#ASMSTART
934 ; GFX11-NEXT: ; use s67
935 ; GFX11-NEXT: ;;#ASMEND
936 ; GFX11-NEXT: ;;#ASMSTART
937 ; GFX11-NEXT: ; use s0
938 ; GFX11-NEXT: ;;#ASMEND
939 ; GFX11-NEXT: s_endpgm
940 ;
941 ; GFX12-LABEL: long_load_chain:
942 ; GFX12: ; %bb.0:
943 ; GFX12-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
944 ; GFX12-NEXT: s_wait_kmcnt 0x0
945 ; GFX12-NEXT: s_clause 0x1f
946 ; GFX12-NEXT: s_load_b32 s2, s[0:1], 0x0
947 ; GFX12-NEXT: s_load_b32 s3, s[0:1], 0x10
948 ; GFX12-NEXT: s_load_b32 s4, s[0:1], 0x20
949 ; GFX12-NEXT: s_load_b32 s5, s[0:1], 0x30
950 ; GFX12-NEXT: s_load_b32 s6, s[0:1], 0x40
951 ; GFX12-NEXT: s_load_b32 s7, s[0:1], 0x50
952 ; GFX12-NEXT: s_load_b32 s8, s[0:1], 0x60
953 ; GFX12-NEXT: s_load_b32 s9, s[0:1], 0x70
954 ; GFX12-NEXT: s_load_b32 s10, s[0:1], 0x80
955 ; GFX12-NEXT: s_load_b32 s11, s[0:1], 0x90
956 ; GFX12-NEXT: s_load_b32 s12, s[0:1], 0xa0
957 ; GFX12-NEXT: s_load_b32 s13, s[0:1], 0xb0
958 ; GFX12-NEXT: s_load_b32 s14, s[0:1], 0xc0
959 ; GFX12-NEXT: s_load_b32 s15, s[0:1], 0xd0
960 ; GFX12-NEXT: s_load_b32 s16, s[0:1], 0xe0
961 ; GFX12-NEXT: s_load_b32 s17, s[0:1], 0xf0
962 ; GFX12-NEXT: s_load_b32 s18, s[0:1], 0x100
963 ; GFX12-NEXT: s_load_b32 s19, s[0:1], 0x110
964 ; GFX12-NEXT: s_load_b32 s20, s[0:1], 0x120
965 ; GFX12-NEXT: s_load_b32 s21, s[0:1], 0x130
966 ; GFX12-NEXT: s_load_b32 s22, s[0:1], 0x140
967 ; GFX12-NEXT: s_load_b32 s23, s[0:1], 0x150
968 ; GFX12-NEXT: s_load_b32 s24, s[0:1], 0x160
969 ; GFX12-NEXT: s_load_b32 s25, s[0:1], 0x170
970 ; GFX12-NEXT: s_load_b32 s26, s[0:1], 0x180
971 ; GFX12-NEXT: s_load_b32 s27, s[0:1], 0x190
972 ; GFX12-NEXT: s_load_b32 s28, s[0:1], 0x1a0
973 ; GFX12-NEXT: s_load_b32 s29, s[0:1], 0x1b0
974 ; GFX12-NEXT: s_load_b32 s30, s[0:1], 0x1c0
975 ; GFX12-NEXT: s_load_b32 s31, s[0:1], 0x1d0
976 ; GFX12-NEXT: s_load_b32 s33, s[0:1], 0x1e0
977 ; GFX12-NEXT: s_load_b32 s34, s[0:1], 0x1f0
978 ; GFX12-NEXT: s_clause 0x1f
979 ; GFX12-NEXT: s_load_b32 s35, s[0:1], 0x200
980 ; GFX12-NEXT: s_load_b32 s36, s[0:1], 0x210
981 ; GFX12-NEXT: s_load_b32 s37, s[0:1], 0x220
982 ; GFX12-NEXT: s_load_b32 s38, s[0:1], 0x230
983 ; GFX12-NEXT: s_load_b32 s39, s[0:1], 0x240
984 ; GFX12-NEXT: s_load_b32 s40, s[0:1], 0x250
985 ; GFX12-NEXT: s_load_b32 s41, s[0:1], 0x260
986 ; GFX12-NEXT: s_load_b32 s42, s[0:1], 0x270
987 ; GFX12-NEXT: s_load_b32 s43, s[0:1], 0x280
988 ; GFX12-NEXT: s_load_b32 s44, s[0:1], 0x290
989 ; GFX12-NEXT: s_load_b32 s45, s[0:1], 0x2a0
990 ; GFX12-NEXT: s_load_b32 s46, s[0:1], 0x2b0
991 ; GFX12-NEXT: s_load_b32 s47, s[0:1], 0x2c0
992 ; GFX12-NEXT: s_load_b32 s48, s[0:1], 0x2d0
993 ; GFX12-NEXT: s_load_b32 s49, s[0:1], 0x2e0
994 ; GFX12-NEXT: s_load_b32 s50, s[0:1], 0x2f0
995 ; GFX12-NEXT: s_load_b32 s51, s[0:1], 0x300
996 ; GFX12-NEXT: s_load_b32 s52, s[0:1], 0x310
997 ; GFX12-NEXT: s_load_b32 s53, s[0:1], 0x320
998 ; GFX12-NEXT: s_load_b32 s54, s[0:1], 0x330
999 ; GFX12-NEXT: s_load_b32 s55, s[0:1], 0x340
1000 ; GFX12-NEXT: s_load_b32 s56, s[0:1], 0x350
1001 ; GFX12-NEXT: s_load_b32 s57, s[0:1], 0x360
1002 ; GFX12-NEXT: s_load_b32 s58, s[0:1], 0x370
1003 ; GFX12-NEXT: s_load_b32 s59, s[0:1], 0x380
1004 ; GFX12-NEXT: s_load_b32 s60, s[0:1], 0x390
1005 ; GFX12-NEXT: s_load_b32 s61, s[0:1], 0x3a0
1006 ; GFX12-NEXT: s_load_b32 s62, s[0:1], 0x3b0
1007 ; GFX12-NEXT: s_load_b32 s63, s[0:1], 0x3c0
1008 ; GFX12-NEXT: s_load_b32 s64, s[0:1], 0x3d0
1009 ; GFX12-NEXT: s_load_b32 s65, s[0:1], 0x3e0
1010 ; GFX12-NEXT: s_load_b32 s66, s[0:1], 0x3f0
1011 ; GFX12-NEXT: s_clause 0x1
1012 ; GFX12-NEXT: s_load_b32 s67, s[0:1], 0x400
1013 ; GFX12-NEXT: s_load_b32 s0, s[0:1], 0x410
1014 ; GFX12-NEXT: s_wait_kmcnt 0x0
1015 ; GFX12-NEXT: ;;#ASMSTART
1016 ; GFX12-NEXT: ; use s2
1017 ; GFX12-NEXT: ;;#ASMEND
1018 ; GFX12-NEXT: ;;#ASMSTART
1019 ; GFX12-NEXT: ; use s3
1020 ; GFX12-NEXT: ;;#ASMEND
1021 ; GFX12-NEXT: ;;#ASMSTART
1022 ; GFX12-NEXT: ; use s4
1023 ; GFX12-NEXT: ;;#ASMEND
1024 ; GFX12-NEXT: ;;#ASMSTART
1025 ; GFX12-NEXT: ; use s5
1026 ; GFX12-NEXT: ;;#ASMEND
1027 ; GFX12-NEXT: ;;#ASMSTART
1028 ; GFX12-NEXT: ; use s6
1029 ; GFX12-NEXT: ;;#ASMEND
1030 ; GFX12-NEXT: ;;#ASMSTART
1031 ; GFX12-NEXT: ; use s7
1032 ; GFX12-NEXT: ;;#ASMEND
1033 ; GFX12-NEXT: ;;#ASMSTART
1034 ; GFX12-NEXT: ; use s8
1035 ; GFX12-NEXT: ;;#ASMEND
1036 ; GFX12-NEXT: ;;#ASMSTART
1037 ; GFX12-NEXT: ; use s9
1038 ; GFX12-NEXT: ;;#ASMEND
1039 ; GFX12-NEXT: ;;#ASMSTART
1040 ; GFX12-NEXT: ; use s10
1041 ; GFX12-NEXT: ;;#ASMEND
1042 ; GFX12-NEXT: ;;#ASMSTART
1043 ; GFX12-NEXT: ; use s11
1044 ; GFX12-NEXT: ;;#ASMEND
1045 ; GFX12-NEXT: ;;#ASMSTART
1046 ; GFX12-NEXT: ; use s12
1047 ; GFX12-NEXT: ;;#ASMEND
1048 ; GFX12-NEXT: ;;#ASMSTART
1049 ; GFX12-NEXT: ; use s13
1050 ; GFX12-NEXT: ;;#ASMEND
1051 ; GFX12-NEXT: ;;#ASMSTART
1052 ; GFX12-NEXT: ; use s14
1053 ; GFX12-NEXT: ;;#ASMEND
1054 ; GFX12-NEXT: ;;#ASMSTART
1055 ; GFX12-NEXT: ; use s15
1056 ; GFX12-NEXT: ;;#ASMEND
1057 ; GFX12-NEXT: ;;#ASMSTART
1058 ; GFX12-NEXT: ; use s16
1059 ; GFX12-NEXT: ;;#ASMEND
1060 ; GFX12-NEXT: ;;#ASMSTART
1061 ; GFX12-NEXT: ; use s17
1062 ; GFX12-NEXT: ;;#ASMEND
1063 ; GFX12-NEXT: ;;#ASMSTART
1064 ; GFX12-NEXT: ; use s18
1065 ; GFX12-NEXT: ;;#ASMEND
1066 ; GFX12-NEXT: ;;#ASMSTART
1067 ; GFX12-NEXT: ; use s19
1068 ; GFX12-NEXT: ;;#ASMEND
1069 ; GFX12-NEXT: ;;#ASMSTART
1070 ; GFX12-NEXT: ; use s20
1071 ; GFX12-NEXT: ;;#ASMEND
1072 ; GFX12-NEXT: ;;#ASMSTART
1073 ; GFX12-NEXT: ; use s21
1074 ; GFX12-NEXT: ;;#ASMEND
1075 ; GFX12-NEXT: ;;#ASMSTART
1076 ; GFX12-NEXT: ; use s22
1077 ; GFX12-NEXT: ;;#ASMEND
1078 ; GFX12-NEXT: ;;#ASMSTART
1079 ; GFX12-NEXT: ; use s23
1080 ; GFX12-NEXT: ;;#ASMEND
1081 ; GFX12-NEXT: ;;#ASMSTART
1082 ; GFX12-NEXT: ; use s24
1083 ; GFX12-NEXT: ;;#ASMEND
1084 ; GFX12-NEXT: ;;#ASMSTART
1085 ; GFX12-NEXT: ; use s25
1086 ; GFX12-NEXT: ;;#ASMEND
1087 ; GFX12-NEXT: ;;#ASMSTART
1088 ; GFX12-NEXT: ; use s26
1089 ; GFX12-NEXT: ;;#ASMEND
1090 ; GFX12-NEXT: ;;#ASMSTART
1091 ; GFX12-NEXT: ; use s27
1092 ; GFX12-NEXT: ;;#ASMEND
1093 ; GFX12-NEXT: ;;#ASMSTART
1094 ; GFX12-NEXT: ; use s28
1095 ; GFX12-NEXT: ;;#ASMEND
1096 ; GFX12-NEXT: ;;#ASMSTART
1097 ; GFX12-NEXT: ; use s29
1098 ; GFX12-NEXT: ;;#ASMEND
1099 ; GFX12-NEXT: ;;#ASMSTART
1100 ; GFX12-NEXT: ; use s30
1101 ; GFX12-NEXT: ;;#ASMEND
1102 ; GFX12-NEXT: ;;#ASMSTART
1103 ; GFX12-NEXT: ; use s31
1104 ; GFX12-NEXT: ;;#ASMEND
1105 ; GFX12-NEXT: ;;#ASMSTART
1106 ; GFX12-NEXT: ; use s33
1107 ; GFX12-NEXT: ;;#ASMEND
1108 ; GFX12-NEXT: ;;#ASMSTART
1109 ; GFX12-NEXT: ; use s34
1110 ; GFX12-NEXT: ;;#ASMEND
1111 ; GFX12-NEXT: ;;#ASMSTART
1112 ; GFX12-NEXT: ; use s35
1113 ; GFX12-NEXT: ;;#ASMEND
1114 ; GFX12-NEXT: ;;#ASMSTART
1115 ; GFX12-NEXT: ; use s36
1116 ; GFX12-NEXT: ;;#ASMEND
1117 ; GFX12-NEXT: ;;#ASMSTART
1118 ; GFX12-NEXT: ; use s37
1119 ; GFX12-NEXT: ;;#ASMEND
1120 ; GFX12-NEXT: ;;#ASMSTART
1121 ; GFX12-NEXT: ; use s38
1122 ; GFX12-NEXT: ;;#ASMEND
1123 ; GFX12-NEXT: ;;#ASMSTART
1124 ; GFX12-NEXT: ; use s39
1125 ; GFX12-NEXT: ;;#ASMEND
1126 ; GFX12-NEXT: ;;#ASMSTART
1127 ; GFX12-NEXT: ; use s40
1128 ; GFX12-NEXT: ;;#ASMEND
1129 ; GFX12-NEXT: ;;#ASMSTART
1130 ; GFX12-NEXT: ; use s41
1131 ; GFX12-NEXT: ;;#ASMEND
1132 ; GFX12-NEXT: ;;#ASMSTART
1133 ; GFX12-NEXT: ; use s42
1134 ; GFX12-NEXT: ;;#ASMEND
1135 ; GFX12-NEXT: ;;#ASMSTART
1136 ; GFX12-NEXT: ; use s43
1137 ; GFX12-NEXT: ;;#ASMEND
1138 ; GFX12-NEXT: ;;#ASMSTART
1139 ; GFX12-NEXT: ; use s44
1140 ; GFX12-NEXT: ;;#ASMEND
1141 ; GFX12-NEXT: ;;#ASMSTART
1142 ; GFX12-NEXT: ; use s45
1143 ; GFX12-NEXT: ;;#ASMEND
1144 ; GFX12-NEXT: ;;#ASMSTART
1145 ; GFX12-NEXT: ; use s46
1146 ; GFX12-NEXT: ;;#ASMEND
1147 ; GFX12-NEXT: ;;#ASMSTART
1148 ; GFX12-NEXT: ; use s47
1149 ; GFX12-NEXT: ;;#ASMEND
1150 ; GFX12-NEXT: ;;#ASMSTART
1151 ; GFX12-NEXT: ; use s48
1152 ; GFX12-NEXT: ;;#ASMEND
1153 ; GFX12-NEXT: ;;#ASMSTART
1154 ; GFX12-NEXT: ; use s49
1155 ; GFX12-NEXT: ;;#ASMEND
1156 ; GFX12-NEXT: ;;#ASMSTART
1157 ; GFX12-NEXT: ; use s50
1158 ; GFX12-NEXT: ;;#ASMEND
1159 ; GFX12-NEXT: ;;#ASMSTART
1160 ; GFX12-NEXT: ; use s51
1161 ; GFX12-NEXT: ;;#ASMEND
1162 ; GFX12-NEXT: ;;#ASMSTART
1163 ; GFX12-NEXT: ; use s52
1164 ; GFX12-NEXT: ;;#ASMEND
1165 ; GFX12-NEXT: ;;#ASMSTART
1166 ; GFX12-NEXT: ; use s53
1167 ; GFX12-NEXT: ;;#ASMEND
1168 ; GFX12-NEXT: ;;#ASMSTART
1169 ; GFX12-NEXT: ; use s54
1170 ; GFX12-NEXT: ;;#ASMEND
1171 ; GFX12-NEXT: ;;#ASMSTART
1172 ; GFX12-NEXT: ; use s55
1173 ; GFX12-NEXT: ;;#ASMEND
1174 ; GFX12-NEXT: ;;#ASMSTART
1175 ; GFX12-NEXT: ; use s56
1176 ; GFX12-NEXT: ;;#ASMEND
1177 ; GFX12-NEXT: ;;#ASMSTART
1178 ; GFX12-NEXT: ; use s57
1179 ; GFX12-NEXT: ;;#ASMEND
1180 ; GFX12-NEXT: ;;#ASMSTART
1181 ; GFX12-NEXT: ; use s58
1182 ; GFX12-NEXT: ;;#ASMEND
1183 ; GFX12-NEXT: ;;#ASMSTART
1184 ; GFX12-NEXT: ; use s59
1185 ; GFX12-NEXT: ;;#ASMEND
1186 ; GFX12-NEXT: ;;#ASMSTART
1187 ; GFX12-NEXT: ; use s60
1188 ; GFX12-NEXT: ;;#ASMEND
1189 ; GFX12-NEXT: ;;#ASMSTART
1190 ; GFX12-NEXT: ; use s61
1191 ; GFX12-NEXT: ;;#ASMEND
1192 ; GFX12-NEXT: ;;#ASMSTART
1193 ; GFX12-NEXT: ; use s62
1194 ; GFX12-NEXT: ;;#ASMEND
1195 ; GFX12-NEXT: ;;#ASMSTART
1196 ; GFX12-NEXT: ; use s63
1197 ; GFX12-NEXT: ;;#ASMEND
1198 ; GFX12-NEXT: ;;#ASMSTART
1199 ; GFX12-NEXT: ; use s64
1200 ; GFX12-NEXT: ;;#ASMEND
1201 ; GFX12-NEXT: ;;#ASMSTART
1202 ; GFX12-NEXT: ; use s65
1203 ; GFX12-NEXT: ;;#ASMEND
1204 ; GFX12-NEXT: ;;#ASMSTART
1205 ; GFX12-NEXT: ; use s66
1206 ; GFX12-NEXT: ;;#ASMEND
1207 ; GFX12-NEXT: ;;#ASMSTART
1208 ; GFX12-NEXT: ; use s67
1209 ; GFX12-NEXT: ;;#ASMEND
1210 ; GFX12-NEXT: ;;#ASMSTART
1211 ; GFX12-NEXT: ; use s0
1212 ; GFX12-NEXT: ;;#ASMEND
1213 ; GFX12-NEXT: s_endpgm
1214 %v0 = load i32, ptr addrspace(1) %p
1215 %ptr1 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 1
1216 %v1 = load i32, ptr addrspace(1) %ptr1
1217 %ptr2 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 2
1218 %v2 = load i32, ptr addrspace(1) %ptr2
1219 %ptr3 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 3
1220 %v3 = load i32, ptr addrspace(1) %ptr3
1221 %ptr4 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 4
1222 %v4 = load i32, ptr addrspace(1) %ptr4
1223 %ptr5 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 5
1224 %v5 = load i32, ptr addrspace(1) %ptr5
1225 %ptr6 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 6
1226 %v6 = load i32, ptr addrspace(1) %ptr6
1227 %ptr7 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 7
1228 %v7 = load i32, ptr addrspace(1) %ptr7
1229 %ptr8 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 8
1230 %v8 = load i32, ptr addrspace(1) %ptr8
1231 %ptr9 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 9
1232 %v9 = load i32, ptr addrspace(1) %ptr9
1233 %ptr10 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 10
1234 %v10 = load i32, ptr addrspace(1) %ptr10
1235 %ptr11 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 11
1236 %v11 = load i32, ptr addrspace(1) %ptr11
1237 %ptr12 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 12
1238 %v12 = load i32, ptr addrspace(1) %ptr12
1239 %ptr13 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 13
1240 %v13 = load i32, ptr addrspace(1) %ptr13
1241 %ptr14 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 14
1242 %v14 = load i32, ptr addrspace(1) %ptr14
1243 %ptr15 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 15
1244 %v15 = load i32, ptr addrspace(1) %ptr15
1245 %ptr16 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 16
1246 %v16 = load i32, ptr addrspace(1) %ptr16
1247 %ptr17 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 17
1248 %v17 = load i32, ptr addrspace(1) %ptr17
1249 %ptr18 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 18
1250 %v18 = load i32, ptr addrspace(1) %ptr18
1251 %ptr19 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 19
1252 %v19 = load i32, ptr addrspace(1) %ptr19
1253 %ptr20 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 20
1254 %v20 = load i32, ptr addrspace(1) %ptr20
1255 %ptr21 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 21
1256 %v21 = load i32, ptr addrspace(1) %ptr21
1257 %ptr22 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 22
1258 %v22 = load i32, ptr addrspace(1) %ptr22
1259 %ptr23 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 23
1260 %v23 = load i32, ptr addrspace(1) %ptr23
1261 %ptr24 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 24
1262 %v24 = load i32, ptr addrspace(1) %ptr24
1263 %ptr25 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 25
1264 %v25 = load i32, ptr addrspace(1) %ptr25
1265 %ptr26 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 26
1266 %v26 = load i32, ptr addrspace(1) %ptr26
1267 %ptr27 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 27
1268 %v27 = load i32, ptr addrspace(1) %ptr27
1269 %ptr28 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 28
1270 %v28 = load i32, ptr addrspace(1) %ptr28
1271 %ptr29 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 29
1272 %v29 = load i32, ptr addrspace(1) %ptr29
1273 %ptr30 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 30
1274 %v30 = load i32, ptr addrspace(1) %ptr30
1275 %ptr31 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 31
1276 %v31 = load i32, ptr addrspace(1) %ptr31
1277 %ptr32 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 32
1278 %v32 = load i32, ptr addrspace(1) %ptr32
1279 %ptr33 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 33
1280 %v33 = load i32, ptr addrspace(1) %ptr33
1281 %ptr34 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 34
1282 %v34 = load i32, ptr addrspace(1) %ptr34
1283 %ptr35 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 35
1284 %v35 = load i32, ptr addrspace(1) %ptr35
1285 %ptr36 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 36
1286 %v36 = load i32, ptr addrspace(1) %ptr36
1287 %ptr37 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 37
1288 %v37 = load i32, ptr addrspace(1) %ptr37
1289 %ptr38 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 38
1290 %v38 = load i32, ptr addrspace(1) %ptr38
1291 %ptr39 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 39
1292 %v39 = load i32, ptr addrspace(1) %ptr39
1293 %ptr40 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 40
1294 %v40 = load i32, ptr addrspace(1) %ptr40
1295 %ptr41 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 41
1296 %v41 = load i32, ptr addrspace(1) %ptr41
1297 %ptr42 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 42
1298 %v42 = load i32, ptr addrspace(1) %ptr42
1299 %ptr43 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 43
1300 %v43 = load i32, ptr addrspace(1) %ptr43
1301 %ptr44 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 44
1302 %v44 = load i32, ptr addrspace(1) %ptr44
1303 %ptr45 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 45
1304 %v45 = load i32, ptr addrspace(1) %ptr45
1305 %ptr46 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 46
1306 %v46 = load i32, ptr addrspace(1) %ptr46
1307 %ptr47 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 47
1308 %v47 = load i32, ptr addrspace(1) %ptr47
1309 %ptr48 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 48
1310 %v48 = load i32, ptr addrspace(1) %ptr48
1311 %ptr49 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 49
1312 %v49 = load i32, ptr addrspace(1) %ptr49
1313 %ptr50 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 50
1314 %v50 = load i32, ptr addrspace(1) %ptr50
1315 %ptr51 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 51
1316 %v51 = load i32, ptr addrspace(1) %ptr51
1317 %ptr52 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 52
1318 %v52 = load i32, ptr addrspace(1) %ptr52
1319 %ptr53 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 53
1320 %v53 = load i32, ptr addrspace(1) %ptr53
1321 %ptr54 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 54
1322 %v54 = load i32, ptr addrspace(1) %ptr54
1323 %ptr55 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 55
1324 %v55 = load i32, ptr addrspace(1) %ptr55
1325 %ptr56 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 56
1326 %v56 = load i32, ptr addrspace(1) %ptr56
1327 %ptr57 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 57
1328 %v57 = load i32, ptr addrspace(1) %ptr57
1329 %ptr58 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 58
1330 %v58 = load i32, ptr addrspace(1) %ptr58
1331 %ptr59 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 59
1332 %v59 = load i32, ptr addrspace(1) %ptr59
1333 %ptr60 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 60
1334 %v60 = load i32, ptr addrspace(1) %ptr60
1335 %ptr61 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 61
1336 %v61 = load i32, ptr addrspace(1) %ptr61
1337 %ptr62 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 62
1338 %v62 = load i32, ptr addrspace(1) %ptr62
1339 %ptr63 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 63
1340 %v63 = load i32, ptr addrspace(1) %ptr63
1341 %ptr64 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 64
1342 %v64 = load i32, ptr addrspace(1) %ptr64
1343 %ptr65 = getelementptr <4 x i32>, ptr addrspace(1) %p, i64 65
1344 %v65 = load i32, ptr addrspace(1) %ptr65
1346 call void asm sideeffect "; use $0", "s"(i32 %v0)
1347 call void asm sideeffect "; use $0", "s"(i32 %v1)
1348 call void asm sideeffect "; use $0", "s"(i32 %v2)
1349 call void asm sideeffect "; use $0", "s"(i32 %v3)
1350 call void asm sideeffect "; use $0", "s"(i32 %v4)
1351 call void asm sideeffect "; use $0", "s"(i32 %v5)
1352 call void asm sideeffect "; use $0", "s"(i32 %v6)
1353 call void asm sideeffect "; use $0", "s"(i32 %v7)
1354 call void asm sideeffect "; use $0", "s"(i32 %v8)
1355 call void asm sideeffect "; use $0", "s"(i32 %v9)
1356 call void asm sideeffect "; use $0", "s"(i32 %v10)
1357 call void asm sideeffect "; use $0", "s"(i32 %v11)
1358 call void asm sideeffect "; use $0", "s"(i32 %v12)
1359 call void asm sideeffect "; use $0", "s"(i32 %v13)
1360 call void asm sideeffect "; use $0", "s"(i32 %v14)
1361 call void asm sideeffect "; use $0", "s"(i32 %v15)
1362 call void asm sideeffect "; use $0", "s"(i32 %v16)
1363 call void asm sideeffect "; use $0", "s"(i32 %v17)
1364 call void asm sideeffect "; use $0", "s"(i32 %v18)
1365 call void asm sideeffect "; use $0", "s"(i32 %v19)
1366 call void asm sideeffect "; use $0", "s"(i32 %v20)
1367 call void asm sideeffect "; use $0", "s"(i32 %v21)
1368 call void asm sideeffect "; use $0", "s"(i32 %v22)
1369 call void asm sideeffect "; use $0", "s"(i32 %v23)
1370 call void asm sideeffect "; use $0", "s"(i32 %v24)
1371 call void asm sideeffect "; use $0", "s"(i32 %v25)
1372 call void asm sideeffect "; use $0", "s"(i32 %v26)
1373 call void asm sideeffect "; use $0", "s"(i32 %v27)
1374 call void asm sideeffect "; use $0", "s"(i32 %v28)
1375 call void asm sideeffect "; use $0", "s"(i32 %v29)
1376 call void asm sideeffect "; use $0", "s"(i32 %v30)
1377 call void asm sideeffect "; use $0", "s"(i32 %v31)
1378 call void asm sideeffect "; use $0", "s"(i32 %v32)
1379 call void asm sideeffect "; use $0", "s"(i32 %v33)
1380 call void asm sideeffect "; use $0", "s"(i32 %v34)
1381 call void asm sideeffect "; use $0", "s"(i32 %v35)
1382 call void asm sideeffect "; use $0", "s"(i32 %v36)
1383 call void asm sideeffect "; use $0", "s"(i32 %v37)
1384 call void asm sideeffect "; use $0", "s"(i32 %v38)
1385 call void asm sideeffect "; use $0", "s"(i32 %v39)
1386 call void asm sideeffect "; use $0", "s"(i32 %v40)
1387 call void asm sideeffect "; use $0", "s"(i32 %v41)
1388 call void asm sideeffect "; use $0", "s"(i32 %v42)
1389 call void asm sideeffect "; use $0", "s"(i32 %v43)
1390 call void asm sideeffect "; use $0", "s"(i32 %v44)
1391 call void asm sideeffect "; use $0", "s"(i32 %v45)
1392 call void asm sideeffect "; use $0", "s"(i32 %v46)
1393 call void asm sideeffect "; use $0", "s"(i32 %v47)
1394 call void asm sideeffect "; use $0", "s"(i32 %v48)
1395 call void asm sideeffect "; use $0", "s"(i32 %v49)
1396 call void asm sideeffect "; use $0", "s"(i32 %v50)
1397 call void asm sideeffect "; use $0", "s"(i32 %v51)
1398 call void asm sideeffect "; use $0", "s"(i32 %v52)
1399 call void asm sideeffect "; use $0", "s"(i32 %v53)
1400 call void asm sideeffect "; use $0", "s"(i32 %v54)
1401 call void asm sideeffect "; use $0", "s"(i32 %v55)
1402 call void asm sideeffect "; use $0", "s"(i32 %v56)
1403 call void asm sideeffect "; use $0", "s"(i32 %v57)
1404 call void asm sideeffect "; use $0", "s"(i32 %v58)
1405 call void asm sideeffect "; use $0", "s"(i32 %v59)
1406 call void asm sideeffect "; use $0", "s"(i32 %v60)
1407 call void asm sideeffect "; use $0", "s"(i32 %v61)
1408 call void asm sideeffect "; use $0", "s"(i32 %v62)
1409 call void asm sideeffect "; use $0", "s"(i32 %v63)
1410 call void asm sideeffect "; use $0", "s"(i32 %v64)
1411 call void asm sideeffect "; use $0", "s"(i32 %v65)
1413 ret void
1414 }