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
3
4; Check lowering of some large insertelement that use the stack
5; instead of register indexing.
6
7define amdgpu_kernel void @v_insert_v64i32_varidx(<64 x i32> addrspace(1)* %out.ptr, <64 x i32> addrspace(1)* %ptr, i32 %val, i32 %idx) #0 {
8; GCN-LABEL: v_insert_v64i32_varidx:
9; GCN:       ; %bb.0:
10; GCN-NEXT:    s_add_u32 s0, s0, s7
11; GCN-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x0
12; GCN-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x10
13; GCN-NEXT:    v_mov_b32_e32 v0, 0x100
14; GCN-NEXT:    s_addc_u32 s1, s1, 0
15; GCN-NEXT:    v_add_u32_e32 v1, 4, v0
16; GCN-NEXT:    s_waitcnt lgkmcnt(0)
17; GCN-NEXT:    s_load_dwordx16 s[12:27], s[10:11], 0x0
18; GCN-NEXT:    s_load_dwordx16 s[68:83], s[10:11], 0x40
19; GCN-NEXT:    s_load_dwordx16 s[52:67], s[10:11], 0x80
20; GCN-NEXT:    s_load_dwordx16 s[36:51], s[10:11], 0xc0
21; GCN-NEXT:    s_movk_i32 s4, 0x50
22; GCN-NEXT:    s_waitcnt lgkmcnt(0)
23; GCN-NEXT:    v_mov_b32_e32 v4, s13
24; GCN-NEXT:    v_mov_b32_e32 v5, s14
25; GCN-NEXT:    v_mov_b32_e32 v6, s15
26; GCN-NEXT:    v_mov_b32_e32 v8, s16
27; GCN-NEXT:    v_mov_b32_e32 v10, s17
28; GCN-NEXT:    v_mov_b32_e32 v12, s18
29; GCN-NEXT:    v_mov_b32_e32 v14, s19
30; GCN-NEXT:    s_movk_i32 s5, 0x60
31; GCN-NEXT:    v_add_u32_e32 v2, 8, v0
32; GCN-NEXT:    v_add_u32_e32 v3, 12, v0
33; GCN-NEXT:    v_add_u32_e32 v7, 16, v0
34; GCN-NEXT:    v_add_u32_e32 v9, 20, v0
35; GCN-NEXT:    v_add_u32_e32 v11, 24, v0
36; GCN-NEXT:    v_add_u32_e32 v13, 28, v0
37; GCN-NEXT:    v_add_u32_e32 v15, 32, v0
38; GCN-NEXT:    v_mov_b32_e32 v16, s20
39; GCN-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen
40; GCN-NEXT:    buffer_store_dword v5, v2, s[0:3], 0 offen
41; GCN-NEXT:    buffer_store_dword v6, v3, s[0:3], 0 offen
42; GCN-NEXT:    buffer_store_dword v8, v7, s[0:3], 0 offen
43; GCN-NEXT:    buffer_store_dword v10, v9, s[0:3], 0 offen
44; GCN-NEXT:    buffer_store_dword v12, v11, s[0:3], 0 offen
45; GCN-NEXT:    buffer_store_dword v14, v13, s[0:3], 0 offen
46; GCN-NEXT:    buffer_store_dword v16, v15, s[0:3], 0 offen
47; GCN-NEXT:    v_add_u32_e32 v17, 36, v0
48; GCN-NEXT:    v_mov_b32_e32 v18, s21
49; GCN-NEXT:    v_mov_b32_e32 v26, s25
50; GCN-NEXT:    v_add_u32_e32 v33, 0x44, v0
51; GCN-NEXT:    v_mov_b32_e32 v34, s69
52; GCN-NEXT:    v_mov_b32_e32 v4, s71
53; GCN-NEXT:    v_add_u32_e32 v19, 40, v0
54; GCN-NEXT:    v_mov_b32_e32 v20, s22
55; GCN-NEXT:    v_add_u32_e32 v21, 44, v0
56; GCN-NEXT:    v_mov_b32_e32 v22, s23
57; GCN-NEXT:    v_add_u32_e32 v23, 48, v0
58; GCN-NEXT:    v_mov_b32_e32 v24, s24
59; GCN-NEXT:    v_add_u32_e32 v25, 52, v0
60; GCN-NEXT:    v_add_u32_e32 v27, 56, v0
61; GCN-NEXT:    v_mov_b32_e32 v28, s26
62; GCN-NEXT:    v_add_u32_e32 v29, 60, v0
63; GCN-NEXT:    v_mov_b32_e32 v30, s27
64; GCN-NEXT:    v_add_u32_e32 v31, 64, v0
65; GCN-NEXT:    v_mov_b32_e32 v32, s68
66; GCN-NEXT:    buffer_store_dword v18, v17, s[0:3], 0 offen
67; GCN-NEXT:    buffer_store_dword v20, v19, s[0:3], 0 offen
68; GCN-NEXT:    buffer_store_dword v22, v21, s[0:3], 0 offen
69; GCN-NEXT:    buffer_store_dword v24, v23, s[0:3], 0 offen
70; GCN-NEXT:    buffer_store_dword v26, v25, s[0:3], 0 offen
71; GCN-NEXT:    buffer_store_dword v28, v27, s[0:3], 0 offen
72; GCN-NEXT:    buffer_store_dword v30, v29, s[0:3], 0 offen
73; GCN-NEXT:    buffer_store_dword v32, v31, s[0:3], 0 offen
74; GCN-NEXT:    s_movk_i32 s13, 0x70
75; GCN-NEXT:    v_add_u32_e32 v35, 0x48, v0
76; GCN-NEXT:    v_mov_b32_e32 v36, s70
77; GCN-NEXT:    v_add_u32_e32 v37, 0x4c, v0
78; GCN-NEXT:    v_add_u32_e32 v38, s4, v0
79; GCN-NEXT:    v_mov_b32_e32 v5, s72
80; GCN-NEXT:    v_add_u32_e32 v39, 0x54, v0
81; GCN-NEXT:    v_mov_b32_e32 v6, s73
82; GCN-NEXT:    v_add_u32_e32 v40, 0x58, v0
83; GCN-NEXT:    v_mov_b32_e32 v8, s74
84; GCN-NEXT:    v_add_u32_e32 v41, 0x5c, v0
85; GCN-NEXT:    v_mov_b32_e32 v10, s75
86; GCN-NEXT:    v_add_u32_e32 v42, s5, v0
87; GCN-NEXT:    v_mov_b32_e32 v12, s76
88; GCN-NEXT:    buffer_store_dword v34, v33, s[0:3], 0 offen
89; GCN-NEXT:    buffer_store_dword v36, v35, s[0:3], 0 offen
90; GCN-NEXT:    buffer_store_dword v4, v37, s[0:3], 0 offen
91; GCN-NEXT:    buffer_store_dword v5, v38, s[0:3], 0 offen
92; GCN-NEXT:    buffer_store_dword v6, v39, s[0:3], 0 offen
93; GCN-NEXT:    buffer_store_dword v8, v40, s[0:3], 0 offen
94; GCN-NEXT:    buffer_store_dword v10, v41, s[0:3], 0 offen
95; GCN-NEXT:    buffer_store_dword v12, v42, s[0:3], 0 offen
96; GCN-NEXT:    v_add_u32_e32 v26, 0x64, v0
97; GCN-NEXT:    v_mov_b32_e32 v14, s77
98; GCN-NEXT:    v_mov_b32_e32 v4, s81
99; GCN-NEXT:    s_movk_i32 s14, 0x90
100; GCN-NEXT:    s_movk_i32 s15, 0xa0
101; GCN-NEXT:    v_add_u32_e32 v28, 0x68, v0
102; GCN-NEXT:    v_mov_b32_e32 v16, s78
103; GCN-NEXT:    v_add_u32_e32 v30, 0x6c, v0
104; GCN-NEXT:    v_mov_b32_e32 v18, s79
105; GCN-NEXT:    v_add_u32_e32 v32, s13, v0
106; GCN-NEXT:    v_mov_b32_e32 v20, s80
107; GCN-NEXT:    v_add_u32_e32 v34, 0x74, v0
108; GCN-NEXT:    v_add_u32_e32 v36, 0x78, v0
109; GCN-NEXT:    v_mov_b32_e32 v5, s82
110; GCN-NEXT:    v_add_u32_e32 v43, 0x7c, v0
111; GCN-NEXT:    v_mov_b32_e32 v6, s83
112; GCN-NEXT:    v_add_u32_e32 v44, 0x80, v0
113; GCN-NEXT:    v_mov_b32_e32 v8, s52
114; GCN-NEXT:    buffer_store_dword v14, v26, s[0:3], 0 offen
115; GCN-NEXT:    buffer_store_dword v16, v28, s[0:3], 0 offen
116; GCN-NEXT:    buffer_store_dword v18, v30, s[0:3], 0 offen
117; GCN-NEXT:    buffer_store_dword v20, v32, s[0:3], 0 offen
118; GCN-NEXT:    buffer_store_dword v4, v34, s[0:3], 0 offen
119; GCN-NEXT:    buffer_store_dword v5, v36, s[0:3], 0 offen
120; GCN-NEXT:    buffer_store_dword v6, v43, s[0:3], 0 offen
121; GCN-NEXT:    buffer_store_dword v8, v44, s[0:3], 0 offen
122; GCN-NEXT:    v_add_u32_e32 v45, 0x84, v0
123; GCN-NEXT:    v_mov_b32_e32 v4, s53
124; GCN-NEXT:    s_movk_i32 s16, 0xb0
125; GCN-NEXT:    v_add_u32_e32 v46, 0x88, v0
126; GCN-NEXT:    v_mov_b32_e32 v5, s54
127; GCN-NEXT:    v_add_u32_e32 v47, 0x8c, v0
128; GCN-NEXT:    v_mov_b32_e32 v6, s55
129; GCN-NEXT:    v_add_u32_e32 v48, s14, v0
130; GCN-NEXT:    v_mov_b32_e32 v8, s56
131; GCN-NEXT:    v_add_u32_e32 v49, 0x94, v0
132; GCN-NEXT:    v_mov_b32_e32 v10, s57
133; GCN-NEXT:    v_add_u32_e32 v50, 0x98, v0
134; GCN-NEXT:    v_mov_b32_e32 v12, s58
135; GCN-NEXT:    v_add_u32_e32 v51, 0x9c, v0
136; GCN-NEXT:    v_mov_b32_e32 v14, s59
137; GCN-NEXT:    v_add_u32_e32 v52, s15, v0
138; GCN-NEXT:    v_mov_b32_e32 v16, s60
139; GCN-NEXT:    buffer_store_dword v4, v45, s[0:3], 0 offen
140; GCN-NEXT:    buffer_store_dword v5, v46, s[0:3], 0 offen
141; GCN-NEXT:    buffer_store_dword v6, v47, s[0:3], 0 offen
142; GCN-NEXT:    buffer_store_dword v8, v48, s[0:3], 0 offen
143; GCN-NEXT:    buffer_store_dword v10, v49, s[0:3], 0 offen
144; GCN-NEXT:    buffer_store_dword v12, v50, s[0:3], 0 offen
145; GCN-NEXT:    buffer_store_dword v14, v51, s[0:3], 0 offen
146; GCN-NEXT:    buffer_store_dword v16, v52, s[0:3], 0 offen
147; GCN-NEXT:    v_add_u32_e32 v53, 0xa4, v0
148; GCN-NEXT:    v_mov_b32_e32 v4, s61
149; GCN-NEXT:    s_movk_i32 s17, 0xd0
150; GCN-NEXT:    s_movk_i32 s18, 0xe0
151; GCN-NEXT:    v_add_u32_e32 v54, 0xa8, v0
152; GCN-NEXT:    v_mov_b32_e32 v5, s62
153; GCN-NEXT:    v_add_u32_e32 v55, 0xac, v0
154; GCN-NEXT:    v_mov_b32_e32 v6, s63
155; GCN-NEXT:    v_add_u32_e32 v56, s16, v0
156; GCN-NEXT:    v_mov_b32_e32 v8, s64
157; GCN-NEXT:    v_add_u32_e32 v57, 0xb4, v0
158; GCN-NEXT:    v_mov_b32_e32 v10, s65
159; GCN-NEXT:    v_add_u32_e32 v58, 0xb8, v0
160; GCN-NEXT:    v_mov_b32_e32 v12, s66
161; GCN-NEXT:    v_add_u32_e32 v59, 0xbc, v0
162; GCN-NEXT:    v_mov_b32_e32 v14, s67
163; GCN-NEXT:    v_add_u32_e32 v60, 0xc0, v0
164; GCN-NEXT:    v_mov_b32_e32 v16, s36
165; GCN-NEXT:    buffer_store_dword v4, v53, s[0:3], 0 offen
166; GCN-NEXT:    buffer_store_dword v5, v54, s[0:3], 0 offen
167; GCN-NEXT:    buffer_store_dword v6, v55, s[0:3], 0 offen
168; GCN-NEXT:    buffer_store_dword v8, v56, s[0:3], 0 offen
169; GCN-NEXT:    buffer_store_dword v10, v57, s[0:3], 0 offen
170; GCN-NEXT:    buffer_store_dword v12, v58, s[0:3], 0 offen
171; GCN-NEXT:    buffer_store_dword v14, v59, s[0:3], 0 offen
172; GCN-NEXT:    buffer_store_dword v16, v60, s[0:3], 0 offen
173; GCN-NEXT:    v_add_u32_e32 v61, 0xc4, v0
174; GCN-NEXT:    v_mov_b32_e32 v4, s37
175; GCN-NEXT:    s_and_b32 s7, s7, 63
176; GCN-NEXT:    s_movk_i32 s19, 0xf0
177; GCN-NEXT:    v_add_u32_e32 v62, 0xc8, v0
178; GCN-NEXT:    v_mov_b32_e32 v5, s38
179; GCN-NEXT:    v_add_u32_e32 v63, 0xcc, v0
180; GCN-NEXT:    v_mov_b32_e32 v6, s39
181; GCN-NEXT:    v_add_u32_e32 v64, s17, v0
182; GCN-NEXT:    v_mov_b32_e32 v8, s40
183; GCN-NEXT:    v_add_u32_e32 v65, 0xd4, v0
184; GCN-NEXT:    v_mov_b32_e32 v10, s41
185; GCN-NEXT:    v_add_u32_e32 v66, 0xd8, v0
186; GCN-NEXT:    v_mov_b32_e32 v12, s42
187; GCN-NEXT:    v_add_u32_e32 v67, 0xdc, v0
188; GCN-NEXT:    v_mov_b32_e32 v14, s43
189; GCN-NEXT:    v_add_u32_e32 v68, s18, v0
190; GCN-NEXT:    v_mov_b32_e32 v16, s44
191; GCN-NEXT:    buffer_store_dword v4, v61, s[0:3], 0 offen
192; GCN-NEXT:    buffer_store_dword v5, v62, s[0:3], 0 offen
193; GCN-NEXT:    buffer_store_dword v6, v63, s[0:3], 0 offen
194; GCN-NEXT:    buffer_store_dword v8, v64, s[0:3], 0 offen
195; GCN-NEXT:    buffer_store_dword v10, v65, s[0:3], 0 offen
196; GCN-NEXT:    buffer_store_dword v12, v66, s[0:3], 0 offen
197; GCN-NEXT:    buffer_store_dword v14, v67, s[0:3], 0 offen
198; GCN-NEXT:    buffer_store_dword v16, v68, s[0:3], 0 offen
199; GCN-NEXT:    v_add_u32_e32 v69, 0xe4, v0
200; GCN-NEXT:    v_mov_b32_e32 v4, s45
201; GCN-NEXT:    v_add_u32_e32 v70, 0xe8, v0
202; GCN-NEXT:    v_mov_b32_e32 v5, s46
203; GCN-NEXT:    v_add_u32_e32 v71, 0xec, v0
204; GCN-NEXT:    v_mov_b32_e32 v6, s47
205; GCN-NEXT:    v_add_u32_e32 v72, s19, v0
206; GCN-NEXT:    v_mov_b32_e32 v8, s48
207; GCN-NEXT:    v_add_u32_e32 v73, 0xf4, v0
208; GCN-NEXT:    v_mov_b32_e32 v10, s49
209; GCN-NEXT:    v_add_u32_e32 v74, 0xf8, v0
210; GCN-NEXT:    v_mov_b32_e32 v12, s50
211; GCN-NEXT:    buffer_store_dword v4, v69, s[0:3], 0 offen
212; GCN-NEXT:    buffer_store_dword v5, v70, s[0:3], 0 offen
213; GCN-NEXT:    buffer_store_dword v6, v71, s[0:3], 0 offen
214; GCN-NEXT:    buffer_store_dword v8, v72, s[0:3], 0 offen
215; GCN-NEXT:    buffer_store_dword v10, v73, s[0:3], 0 offen
216; GCN-NEXT:    buffer_store_dword v12, v74, s[0:3], 0 offen
217; GCN-NEXT:    v_mov_b32_e32 v4, s12
218; GCN-NEXT:    s_lshl_b32 s7, s7, 2
219; GCN-NEXT:    v_add_u32_e32 v75, 0xfc, v0
220; GCN-NEXT:    v_mov_b32_e32 v14, s51
221; GCN-NEXT:    buffer_store_dword v4, off, s[0:3], 0 offset:256
222; GCN-NEXT:    buffer_store_dword v14, v75, s[0:3], 0 offen
223; GCN-NEXT:    v_mov_b32_e32 v4, s6
224; GCN-NEXT:    v_add_u32_e32 v0, s7, v0
225; GCN-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen
226; GCN-NEXT:    buffer_load_dword v1, v1, s[0:3], 0 offen
227; GCN-NEXT:    buffer_load_dword v2, v2, s[0:3], 0 offen
228; GCN-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen
229; GCN-NEXT:    buffer_load_dword v4, v7, s[0:3], 0 offen
230; GCN-NEXT:    buffer_load_dword v5, v9, s[0:3], 0 offen
231; GCN-NEXT:    buffer_load_dword v6, v11, s[0:3], 0 offen
232; GCN-NEXT:    buffer_load_dword v7, v13, s[0:3], 0 offen
233; GCN-NEXT:    buffer_load_dword v8, v15, s[0:3], 0 offen
234; GCN-NEXT:    buffer_load_dword v9, v17, s[0:3], 0 offen
235; GCN-NEXT:    buffer_load_dword v10, v19, s[0:3], 0 offen
236; GCN-NEXT:    buffer_load_dword v11, v21, s[0:3], 0 offen
237; GCN-NEXT:    buffer_load_dword v12, v23, s[0:3], 0 offen
238; GCN-NEXT:    buffer_load_dword v13, v25, s[0:3], 0 offen
239; GCN-NEXT:    buffer_load_dword v14, v27, s[0:3], 0 offen
240; GCN-NEXT:    buffer_load_dword v15, v29, s[0:3], 0 offen
241; GCN-NEXT:    buffer_load_dword v16, v31, s[0:3], 0 offen
242; GCN-NEXT:    buffer_load_dword v17, v33, s[0:3], 0 offen
243; GCN-NEXT:    buffer_load_dword v18, v35, s[0:3], 0 offen
244; GCN-NEXT:    buffer_load_dword v19, v37, s[0:3], 0 offen
245; GCN-NEXT:    buffer_load_dword v20, v38, s[0:3], 0 offen
246; GCN-NEXT:    buffer_load_dword v21, v39, s[0:3], 0 offen
247; GCN-NEXT:    buffer_load_dword v22, v40, s[0:3], 0 offen
248; GCN-NEXT:    buffer_load_dword v23, v41, s[0:3], 0 offen
249; GCN-NEXT:    buffer_load_dword v24, v42, s[0:3], 0 offen
250; GCN-NEXT:    buffer_load_dword v25, v26, s[0:3], 0 offen
251; GCN-NEXT:    buffer_load_dword v26, v28, s[0:3], 0 offen
252; GCN-NEXT:    buffer_load_dword v27, v30, s[0:3], 0 offen
253; GCN-NEXT:    buffer_load_dword v28, v32, s[0:3], 0 offen
254; GCN-NEXT:    buffer_load_dword v29, v34, s[0:3], 0 offen
255; GCN-NEXT:    buffer_load_dword v30, v36, s[0:3], 0 offen
256; GCN-NEXT:    buffer_load_dword v31, v43, s[0:3], 0 offen
257; GCN-NEXT:    buffer_load_dword v32, v44, s[0:3], 0 offen
258; GCN-NEXT:    buffer_load_dword v33, v45, s[0:3], 0 offen
259; GCN-NEXT:    buffer_load_dword v34, v46, s[0:3], 0 offen
260; GCN-NEXT:    buffer_load_dword v35, v47, s[0:3], 0 offen
261; GCN-NEXT:    buffer_load_dword v36, v48, s[0:3], 0 offen
262; GCN-NEXT:    buffer_load_dword v37, v49, s[0:3], 0 offen
263; GCN-NEXT:    buffer_load_dword v38, v50, s[0:3], 0 offen
264; GCN-NEXT:    buffer_load_dword v39, v51, s[0:3], 0 offen
265; GCN-NEXT:    buffer_load_dword v40, v52, s[0:3], 0 offen
266; GCN-NEXT:    buffer_load_dword v41, v53, s[0:3], 0 offen
267; GCN-NEXT:    buffer_load_dword v42, v54, s[0:3], 0 offen
268; GCN-NEXT:    buffer_load_dword v43, v55, s[0:3], 0 offen
269; GCN-NEXT:    buffer_load_dword v44, v56, s[0:3], 0 offen
270; GCN-NEXT:    buffer_load_dword v45, v57, s[0:3], 0 offen
271; GCN-NEXT:    buffer_load_dword v46, v58, s[0:3], 0 offen
272; GCN-NEXT:    buffer_load_dword v47, v59, s[0:3], 0 offen
273; GCN-NEXT:    buffer_load_dword v48, v60, s[0:3], 0 offen
274; GCN-NEXT:    buffer_load_dword v49, v61, s[0:3], 0 offen
275; GCN-NEXT:    buffer_load_dword v50, v62, s[0:3], 0 offen
276; GCN-NEXT:    buffer_load_dword v51, v63, s[0:3], 0 offen
277; GCN-NEXT:    buffer_load_dword v52, v64, s[0:3], 0 offen
278; GCN-NEXT:    buffer_load_dword v53, v65, s[0:3], 0 offen
279; GCN-NEXT:    buffer_load_dword v54, v66, s[0:3], 0 offen
280; GCN-NEXT:    buffer_load_dword v55, v67, s[0:3], 0 offen
281; GCN-NEXT:    buffer_load_dword v56, v68, s[0:3], 0 offen
282; GCN-NEXT:    buffer_load_dword v57, v69, s[0:3], 0 offen
283; GCN-NEXT:    buffer_load_dword v58, v70, s[0:3], 0 offen
284; GCN-NEXT:    buffer_load_dword v59, v71, s[0:3], 0 offen
285; GCN-NEXT:    buffer_load_dword v60, v72, s[0:3], 0 offen
286; GCN-NEXT:    buffer_load_dword v61, v73, s[0:3], 0 offen
287; GCN-NEXT:    buffer_load_dword v62, v74, s[0:3], 0 offen
288; GCN-NEXT:    buffer_load_dword v63, v75, s[0:3], 0 offen
289; GCN-NEXT:    buffer_load_dword v0, off, s[0:3], 0 offset:256
290; GCN-NEXT:    s_add_u32 s6, s8, 16
291; GCN-NEXT:    s_addc_u32 s7, s9, 0
292; GCN-NEXT:    v_mov_b32_e32 v67, s7
293; GCN-NEXT:    v_mov_b32_e32 v66, s6
294; GCN-NEXT:    s_add_u32 s6, s8, 32
295; GCN-NEXT:    s_addc_u32 s7, s9, 0
296; GCN-NEXT:    v_mov_b32_e32 v65, s9
297; GCN-NEXT:    s_add_u32 s10, s8, 48
298; GCN-NEXT:    v_mov_b32_e32 v64, s8
299; GCN-NEXT:    s_addc_u32 s11, s9, 0
300; GCN-NEXT:    s_waitcnt vmcnt(0)
301; GCN-NEXT:    global_store_dwordx4 v[64:65], v[0:3], off
302; GCN-NEXT:    global_store_dwordx4 v[66:67], v[4:7], off
303; GCN-NEXT:    v_mov_b32_e32 v0, s6
304; GCN-NEXT:    v_mov_b32_e32 v1, s7
305; GCN-NEXT:    s_add_u32 s6, s8, 64
306; GCN-NEXT:    v_mov_b32_e32 v2, s10
307; GCN-NEXT:    s_addc_u32 s7, s9, 0
308; GCN-NEXT:    v_mov_b32_e32 v3, s11
309; GCN-NEXT:    s_add_u32 s10, s8, s4
310; GCN-NEXT:    s_addc_u32 s11, s9, 0
311; GCN-NEXT:    s_add_u32 s4, s8, s5
312; GCN-NEXT:    global_store_dwordx4 v[0:1], v[8:11], off
313; GCN-NEXT:    global_store_dwordx4 v[2:3], v[12:15], off
314; GCN-NEXT:    v_mov_b32_e32 v0, s6
315; GCN-NEXT:    s_addc_u32 s5, s9, 0
316; GCN-NEXT:    v_mov_b32_e32 v1, s7
317; GCN-NEXT:    s_add_u32 s6, s8, s13
318; GCN-NEXT:    v_mov_b32_e32 v2, s10
319; GCN-NEXT:    v_mov_b32_e32 v3, s11
320; GCN-NEXT:    global_store_dwordx4 v[0:1], v[16:19], off
321; GCN-NEXT:    global_store_dwordx4 v[2:3], v[20:23], off
322; GCN-NEXT:    s_addc_u32 s7, s9, 0
323; GCN-NEXT:    v_mov_b32_e32 v0, s4
324; GCN-NEXT:    v_mov_b32_e32 v1, s5
325; GCN-NEXT:    s_add_u32 s4, s8, 0x80
326; GCN-NEXT:    v_mov_b32_e32 v2, s6
327; GCN-NEXT:    s_addc_u32 s5, s9, 0
328; GCN-NEXT:    v_mov_b32_e32 v3, s7
329; GCN-NEXT:    s_add_u32 s6, s8, s14
330; GCN-NEXT:    global_store_dwordx4 v[0:1], v[24:27], off
331; GCN-NEXT:    global_store_dwordx4 v[2:3], v[28:31], off
332; GCN-NEXT:    s_addc_u32 s7, s9, 0
333; GCN-NEXT:    v_mov_b32_e32 v0, s4
334; GCN-NEXT:    v_mov_b32_e32 v1, s5
335; GCN-NEXT:    s_add_u32 s4, s8, s15
336; GCN-NEXT:    v_mov_b32_e32 v2, s6
337; GCN-NEXT:    s_addc_u32 s5, s9, 0
338; GCN-NEXT:    v_mov_b32_e32 v3, s7
339; GCN-NEXT:    s_add_u32 s6, s8, s16
340; GCN-NEXT:    global_store_dwordx4 v[0:1], v[32:35], off
341; GCN-NEXT:    global_store_dwordx4 v[2:3], v[36:39], off
342; GCN-NEXT:    s_addc_u32 s7, s9, 0
343; GCN-NEXT:    v_mov_b32_e32 v0, s4
344; GCN-NEXT:    v_mov_b32_e32 v1, s5
345; GCN-NEXT:    s_add_u32 s4, s8, 0xc0
346; GCN-NEXT:    v_mov_b32_e32 v2, s6
347; GCN-NEXT:    v_mov_b32_e32 v3, s7
348; GCN-NEXT:    global_store_dwordx4 v[0:1], v[40:43], off
349; GCN-NEXT:    global_store_dwordx4 v[2:3], v[44:47], off
350; GCN-NEXT:    s_addc_u32 s5, s9, 0
351; GCN-NEXT:    v_mov_b32_e32 v0, s4
352; GCN-NEXT:    v_mov_b32_e32 v1, s5
353; GCN-NEXT:    s_add_u32 s4, s8, s17
354; GCN-NEXT:    s_addc_u32 s5, s9, 0
355; GCN-NEXT:    v_mov_b32_e32 v2, s4
356; GCN-NEXT:    v_mov_b32_e32 v3, s5
357; GCN-NEXT:    s_add_u32 s4, s8, s18
358; GCN-NEXT:    global_store_dwordx4 v[0:1], v[48:51], off
359; GCN-NEXT:    global_store_dwordx4 v[2:3], v[52:55], off
360; GCN-NEXT:    s_addc_u32 s5, s9, 0
361; GCN-NEXT:    v_mov_b32_e32 v0, s4
362; GCN-NEXT:    v_mov_b32_e32 v1, s5
363; GCN-NEXT:    s_add_u32 s4, s8, s19
364; GCN-NEXT:    s_addc_u32 s5, s9, 0
365; GCN-NEXT:    v_mov_b32_e32 v2, s4
366; GCN-NEXT:    v_mov_b32_e32 v3, s5
367; GCN-NEXT:    global_store_dwordx4 v[0:1], v[56:59], off
368; GCN-NEXT:    global_store_dwordx4 v[2:3], v[60:63], off
369; GCN-NEXT:    s_endpgm
370  %vec = load <64 x i32>, <64 x i32> addrspace(1)* %ptr
371  %insert = insertelement <64 x i32> %vec, i32 %val, i32 %idx
372  store <64 x i32> %insert, <64 x i32> addrspace(1)* %out.ptr
373  ret void
374}
375
376attributes #0 = { "amdgpu-waves-per-eu"="1,10" }
377