| blob | faab70cac39cbf6a905ab2d730de27a6bd7f9f23 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
4 ; Check lowering of some large insertelement that use the stack
5 ; instead of register indexing.
7 define amdgpu_kernel void @v_insert_v64i32_varidx(ptr addrspace(1) %out.ptr, ptr addrspace(1) %ptr, i32 %val, i32 %idx) #0 {
8 ; GCN-LABEL: v_insert_v64i32_varidx:
9 ; GCN: ; %bb.0:
10 ; GCN-NEXT: s_load_dwordx4 s[20:23], s[4:5], 0x0
11 ; GCN-NEXT: s_load_dwordx2 s[24:25], s[4:5], 0x10
12 ; GCN-NEXT: s_add_u32 s0, s0, s7
13 ; GCN-NEXT: s_addc_u32 s1, s1, 0
14 ; GCN-NEXT: v_mov_b32_e32 v16, 0x100
15 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
16 ; GCN-NEXT: s_load_dwordx16 s[36:51], s[22:23], 0x0
17 ; GCN-NEXT: s_load_dwordx16 s[52:67], s[22:23], 0x40
18 ; GCN-NEXT: s_load_dwordx16 s[4:19], s[22:23], 0x80
19 ; GCN-NEXT: v_mov_b32_e32 v64, 0
20 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
21 ; GCN-NEXT: v_mov_b32_e32 v0, s36
22 ; GCN-NEXT: v_mov_b32_e32 v1, s37
23 ; GCN-NEXT: v_mov_b32_e32 v2, s38
24 ; GCN-NEXT: v_mov_b32_e32 v3, s39
25 ; GCN-NEXT: v_mov_b32_e32 v4, s40
26 ; GCN-NEXT: v_mov_b32_e32 v5, s41
27 ; GCN-NEXT: v_mov_b32_e32 v6, s42
28 ; GCN-NEXT: v_mov_b32_e32 v7, s43
29 ; GCN-NEXT: v_mov_b32_e32 v8, s44
30 ; GCN-NEXT: v_mov_b32_e32 v9, s45
31 ; GCN-NEXT: v_mov_b32_e32 v10, s46
32 ; GCN-NEXT: v_mov_b32_e32 v11, s47
33 ; GCN-NEXT: v_mov_b32_e32 v12, s48
34 ; GCN-NEXT: v_mov_b32_e32 v13, s49
35 ; GCN-NEXT: v_mov_b32_e32 v14, s50
36 ; GCN-NEXT: v_mov_b32_e32 v15, s51
37 ; GCN-NEXT: s_load_dwordx16 s[36:51], s[22:23], 0xc0
38 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:256
39 ; GCN-NEXT: buffer_store_dword v1, off, s[0:3], 0 offset:260
40 ; GCN-NEXT: buffer_store_dword v2, off, s[0:3], 0 offset:264
41 ; GCN-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:268
42 ; GCN-NEXT: buffer_store_dword v4, off, s[0:3], 0 offset:272
43 ; GCN-NEXT: buffer_store_dword v5, off, s[0:3], 0 offset:276
44 ; GCN-NEXT: buffer_store_dword v6, off, s[0:3], 0 offset:280
45 ; GCN-NEXT: buffer_store_dword v7, off, s[0:3], 0 offset:284
46 ; GCN-NEXT: buffer_store_dword v8, off, s[0:3], 0 offset:288
47 ; GCN-NEXT: buffer_store_dword v9, off, s[0:3], 0 offset:292
48 ; GCN-NEXT: buffer_store_dword v10, off, s[0:3], 0 offset:296
49 ; GCN-NEXT: buffer_store_dword v11, off, s[0:3], 0 offset:300
50 ; GCN-NEXT: buffer_store_dword v12, off, s[0:3], 0 offset:304
51 ; GCN-NEXT: buffer_store_dword v13, off, s[0:3], 0 offset:308
52 ; GCN-NEXT: buffer_store_dword v14, off, s[0:3], 0 offset:312
53 ; GCN-NEXT: buffer_store_dword v15, off, s[0:3], 0 offset:316
54 ; GCN-NEXT: v_mov_b32_e32 v0, s52
55 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:320
56 ; GCN-NEXT: v_mov_b32_e32 v0, s53
57 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:324
58 ; GCN-NEXT: v_mov_b32_e32 v0, s54
59 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:328
60 ; GCN-NEXT: v_mov_b32_e32 v0, s55
61 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:332
62 ; GCN-NEXT: v_mov_b32_e32 v0, s56
63 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:336
64 ; GCN-NEXT: v_mov_b32_e32 v0, s57
65 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:340
66 ; GCN-NEXT: v_mov_b32_e32 v0, s58
67 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:344
68 ; GCN-NEXT: v_mov_b32_e32 v0, s59
69 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:348
70 ; GCN-NEXT: v_mov_b32_e32 v0, s60
71 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:352
72 ; GCN-NEXT: v_mov_b32_e32 v0, s61
73 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:356
74 ; GCN-NEXT: v_mov_b32_e32 v0, s62
75 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:360
76 ; GCN-NEXT: v_mov_b32_e32 v0, s63
77 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:364
78 ; GCN-NEXT: v_mov_b32_e32 v0, s64
79 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:368
80 ; GCN-NEXT: v_mov_b32_e32 v0, s65
81 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:372
82 ; GCN-NEXT: v_mov_b32_e32 v0, s66
83 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:376
84 ; GCN-NEXT: v_mov_b32_e32 v0, s67
85 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:380
86 ; GCN-NEXT: v_mov_b32_e32 v0, s4
87 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:384
88 ; GCN-NEXT: v_mov_b32_e32 v0, s5
89 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:388
90 ; GCN-NEXT: v_mov_b32_e32 v0, s6
91 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:392
92 ; GCN-NEXT: v_mov_b32_e32 v0, s7
93 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:396
94 ; GCN-NEXT: v_mov_b32_e32 v0, s8
95 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:400
96 ; GCN-NEXT: v_mov_b32_e32 v0, s9
97 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:404
98 ; GCN-NEXT: v_mov_b32_e32 v0, s10
99 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:408
100 ; GCN-NEXT: v_mov_b32_e32 v0, s11
101 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:412
102 ; GCN-NEXT: v_mov_b32_e32 v0, s12
103 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:416
104 ; GCN-NEXT: v_mov_b32_e32 v0, s13
105 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:420
106 ; GCN-NEXT: v_mov_b32_e32 v0, s14
107 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:424
108 ; GCN-NEXT: v_mov_b32_e32 v0, s15
109 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:428
110 ; GCN-NEXT: v_mov_b32_e32 v0, s16
111 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:432
112 ; GCN-NEXT: v_mov_b32_e32 v0, s17
113 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:436
114 ; GCN-NEXT: v_mov_b32_e32 v0, s18
115 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:440
116 ; GCN-NEXT: v_mov_b32_e32 v0, s19
117 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:444
118 ; GCN-NEXT: s_waitcnt lgkmcnt(0)
119 ; GCN-NEXT: v_mov_b32_e32 v0, s36
120 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:448
121 ; GCN-NEXT: v_mov_b32_e32 v0, s37
122 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:452
123 ; GCN-NEXT: v_mov_b32_e32 v0, s38
124 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:456
125 ; GCN-NEXT: v_mov_b32_e32 v0, s39
126 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:460
127 ; GCN-NEXT: v_mov_b32_e32 v0, s40
128 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:464
129 ; GCN-NEXT: v_mov_b32_e32 v0, s41
130 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:468
131 ; GCN-NEXT: v_mov_b32_e32 v0, s42
132 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:472
133 ; GCN-NEXT: v_mov_b32_e32 v0, s43
134 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:476
135 ; GCN-NEXT: v_mov_b32_e32 v0, s44
136 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:480
137 ; GCN-NEXT: v_mov_b32_e32 v0, s45
138 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:484
139 ; GCN-NEXT: v_mov_b32_e32 v0, s46
140 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:488
141 ; GCN-NEXT: v_mov_b32_e32 v0, s47
142 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:492
143 ; GCN-NEXT: v_mov_b32_e32 v0, s48
144 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:496
145 ; GCN-NEXT: v_mov_b32_e32 v0, s49
146 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:500
147 ; GCN-NEXT: v_mov_b32_e32 v0, s50
148 ; GCN-NEXT: s_and_b32 s4, s25, 63
149 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:504
150 ; GCN-NEXT: v_mov_b32_e32 v0, s51
151 ; GCN-NEXT: s_lshl_b32 s4, s4, 2
152 ; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:508
153 ; GCN-NEXT: v_add_u32_e32 v0, s4, v16
154 ; GCN-NEXT: v_mov_b32_e32 v1, s24
155 ; GCN-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
156 ; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:256
157 ; GCN-NEXT: s_nop 0
158 ; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:260
159 ; GCN-NEXT: buffer_load_dword v2, off, s[0:3], 0 offset:264
160 ; GCN-NEXT: buffer_load_dword v3, off, s[0:3], 0 offset:268
161 ; GCN-NEXT: buffer_load_dword v4, off, s[0:3], 0 offset:272
162 ; GCN-NEXT: buffer_load_dword v5, off, s[0:3], 0 offset:276
163 ; GCN-NEXT: buffer_load_dword v6, off, s[0:3], 0 offset:280
164 ; GCN-NEXT: buffer_load_dword v7, off, s[0:3], 0 offset:284
165 ; GCN-NEXT: buffer_load_dword v8, off, s[0:3], 0 offset:288
166 ; GCN-NEXT: buffer_load_dword v9, off, s[0:3], 0 offset:292
167 ; GCN-NEXT: buffer_load_dword v10, off, s[0:3], 0 offset:296
168 ; GCN-NEXT: buffer_load_dword v11, off, s[0:3], 0 offset:300
169 ; GCN-NEXT: buffer_load_dword v12, off, s[0:3], 0 offset:304
170 ; GCN-NEXT: buffer_load_dword v13, off, s[0:3], 0 offset:308
171 ; GCN-NEXT: buffer_load_dword v14, off, s[0:3], 0 offset:312
172 ; GCN-NEXT: buffer_load_dword v15, off, s[0:3], 0 offset:316
173 ; GCN-NEXT: buffer_load_dword v16, off, s[0:3], 0 offset:320
174 ; GCN-NEXT: buffer_load_dword v17, off, s[0:3], 0 offset:324
175 ; GCN-NEXT: buffer_load_dword v18, off, s[0:3], 0 offset:328
176 ; GCN-NEXT: buffer_load_dword v19, off, s[0:3], 0 offset:332
177 ; GCN-NEXT: buffer_load_dword v20, off, s[0:3], 0 offset:336
178 ; GCN-NEXT: buffer_load_dword v21, off, s[0:3], 0 offset:340
179 ; GCN-NEXT: buffer_load_dword v22, off, s[0:3], 0 offset:344
180 ; GCN-NEXT: buffer_load_dword v23, off, s[0:3], 0 offset:348
181 ; GCN-NEXT: buffer_load_dword v24, off, s[0:3], 0 offset:352
182 ; GCN-NEXT: buffer_load_dword v25, off, s[0:3], 0 offset:356
183 ; GCN-NEXT: buffer_load_dword v26, off, s[0:3], 0 offset:360
184 ; GCN-NEXT: buffer_load_dword v27, off, s[0:3], 0 offset:364
185 ; GCN-NEXT: buffer_load_dword v28, off, s[0:3], 0 offset:368
186 ; GCN-NEXT: buffer_load_dword v29, off, s[0:3], 0 offset:372
187 ; GCN-NEXT: buffer_load_dword v30, off, s[0:3], 0 offset:376
188 ; GCN-NEXT: buffer_load_dword v31, off, s[0:3], 0 offset:380
189 ; GCN-NEXT: buffer_load_dword v32, off, s[0:3], 0 offset:384
190 ; GCN-NEXT: buffer_load_dword v33, off, s[0:3], 0 offset:388
191 ; GCN-NEXT: buffer_load_dword v34, off, s[0:3], 0 offset:392
192 ; GCN-NEXT: buffer_load_dword v35, off, s[0:3], 0 offset:396
193 ; GCN-NEXT: buffer_load_dword v36, off, s[0:3], 0 offset:400
194 ; GCN-NEXT: buffer_load_dword v37, off, s[0:3], 0 offset:404
195 ; GCN-NEXT: buffer_load_dword v38, off, s[0:3], 0 offset:408
196 ; GCN-NEXT: buffer_load_dword v39, off, s[0:3], 0 offset:412
197 ; GCN-NEXT: buffer_load_dword v40, off, s[0:3], 0 offset:416
198 ; GCN-NEXT: buffer_load_dword v41, off, s[0:3], 0 offset:420
199 ; GCN-NEXT: buffer_load_dword v42, off, s[0:3], 0 offset:424
200 ; GCN-NEXT: buffer_load_dword v43, off, s[0:3], 0 offset:428
201 ; GCN-NEXT: buffer_load_dword v44, off, s[0:3], 0 offset:432
202 ; GCN-NEXT: buffer_load_dword v45, off, s[0:3], 0 offset:436
203 ; GCN-NEXT: buffer_load_dword v46, off, s[0:3], 0 offset:440
204 ; GCN-NEXT: buffer_load_dword v47, off, s[0:3], 0 offset:444
205 ; GCN-NEXT: buffer_load_dword v48, off, s[0:3], 0 offset:448
206 ; GCN-NEXT: buffer_load_dword v49, off, s[0:3], 0 offset:452
207 ; GCN-NEXT: buffer_load_dword v50, off, s[0:3], 0 offset:456
208 ; GCN-NEXT: buffer_load_dword v51, off, s[0:3], 0 offset:460
209 ; GCN-NEXT: buffer_load_dword v52, off, s[0:3], 0 offset:464
210 ; GCN-NEXT: buffer_load_dword v53, off, s[0:3], 0 offset:468
211 ; GCN-NEXT: buffer_load_dword v54, off, s[0:3], 0 offset:472
212 ; GCN-NEXT: buffer_load_dword v55, off, s[0:3], 0 offset:476
213 ; GCN-NEXT: buffer_load_dword v56, off, s[0:3], 0 offset:480
214 ; GCN-NEXT: buffer_load_dword v57, off, s[0:3], 0 offset:484
215 ; GCN-NEXT: buffer_load_dword v58, off, s[0:3], 0 offset:488
216 ; GCN-NEXT: buffer_load_dword v59, off, s[0:3], 0 offset:492
217 ; GCN-NEXT: buffer_load_dword v60, off, s[0:3], 0 offset:496
218 ; GCN-NEXT: buffer_load_dword v61, off, s[0:3], 0 offset:500
219 ; GCN-NEXT: buffer_load_dword v62, off, s[0:3], 0 offset:504
220 ; GCN-NEXT: buffer_load_dword v63, off, s[0:3], 0 offset:508
221 ; GCN-NEXT: s_waitcnt vmcnt(60)
222 ; GCN-NEXT: global_store_dwordx4 v64, v[0:3], s[20:21]
223 ; GCN-NEXT: s_waitcnt vmcnt(57)
224 ; GCN-NEXT: global_store_dwordx4 v64, v[4:7], s[20:21] offset:16
225 ; GCN-NEXT: s_waitcnt vmcnt(54)
226 ; GCN-NEXT: global_store_dwordx4 v64, v[8:11], s[20:21] offset:32
227 ; GCN-NEXT: s_waitcnt vmcnt(51)
228 ; GCN-NEXT: global_store_dwordx4 v64, v[12:15], s[20:21] offset:48
229 ; GCN-NEXT: s_waitcnt vmcnt(48)
230 ; GCN-NEXT: global_store_dwordx4 v64, v[16:19], s[20:21] offset:64
231 ; GCN-NEXT: s_waitcnt vmcnt(45)
232 ; GCN-NEXT: global_store_dwordx4 v64, v[20:23], s[20:21] offset:80
233 ; GCN-NEXT: s_waitcnt vmcnt(42)
234 ; GCN-NEXT: global_store_dwordx4 v64, v[24:27], s[20:21] offset:96
235 ; GCN-NEXT: s_waitcnt vmcnt(39)
236 ; GCN-NEXT: global_store_dwordx4 v64, v[28:31], s[20:21] offset:112
237 ; GCN-NEXT: s_waitcnt vmcnt(36)
238 ; GCN-NEXT: global_store_dwordx4 v64, v[32:35], s[20:21] offset:128
239 ; GCN-NEXT: s_waitcnt vmcnt(33)
240 ; GCN-NEXT: global_store_dwordx4 v64, v[36:39], s[20:21] offset:144
241 ; GCN-NEXT: s_waitcnt vmcnt(30)
242 ; GCN-NEXT: global_store_dwordx4 v64, v[40:43], s[20:21] offset:160
243 ; GCN-NEXT: s_waitcnt vmcnt(27)
244 ; GCN-NEXT: global_store_dwordx4 v64, v[44:47], s[20:21] offset:176
245 ; GCN-NEXT: s_waitcnt vmcnt(24)
246 ; GCN-NEXT: global_store_dwordx4 v64, v[48:51], s[20:21] offset:192
247 ; GCN-NEXT: s_waitcnt vmcnt(21)
248 ; GCN-NEXT: global_store_dwordx4 v64, v[52:55], s[20:21] offset:208
249 ; GCN-NEXT: s_waitcnt vmcnt(18)
250 ; GCN-NEXT: global_store_dwordx4 v64, v[56:59], s[20:21] offset:224
251 ; GCN-NEXT: s_waitcnt vmcnt(15)
252 ; GCN-NEXT: global_store_dwordx4 v64, v[60:63], s[20:21] offset:240
253 ; GCN-NEXT: s_endpgm
254 %vec = load <64 x i32>, ptr addrspace(1) %ptr
255 %insert = insertelement <64 x i32> %vec, i32 %val, i32 %idx
256 store <64 x i32> %insert, ptr addrspace(1) %out.ptr
257 ret void
258 }
260 attributes #0 = { "amdgpu-flat-work-group-size"="1,256" "amdgpu-waves-per-eu"="1,10" }