xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement-stack-lower.ll (revision 6548b6354d1d990e1c98736f5e7c3de876bedc8e) 
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(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[8:9], 0x0
11; GCN-NEXT:    s_load_dwordx2 s[24:25], s[8:9], 0x10
12; GCN-NEXT:    s_add_u32 s0, s0, s15
13; GCN-NEXT:    s_addc_u32 s1, s1, 0
14; GCN-NEXT:    v_mov_b32_e32 v64, 0
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:    s_waitcnt lgkmcnt(0)
20; GCN-NEXT:    v_mov_b32_e32 v0, s36
21; GCN-NEXT:    v_mov_b32_e32 v1, s37
22; GCN-NEXT:    v_mov_b32_e32 v2, s38
23; GCN-NEXT:    v_mov_b32_e32 v3, s39
24; GCN-NEXT:    v_mov_b32_e32 v4, s40
25; GCN-NEXT:    v_mov_b32_e32 v5, s41
26; GCN-NEXT:    v_mov_b32_e32 v6, s42
27; GCN-NEXT:    v_mov_b32_e32 v7, s43
28; GCN-NEXT:    v_mov_b32_e32 v8, s44
29; GCN-NEXT:    v_mov_b32_e32 v9, s45
30; GCN-NEXT:    v_mov_b32_e32 v10, s46
31; GCN-NEXT:    v_mov_b32_e32 v11, s47
32; GCN-NEXT:    v_mov_b32_e32 v12, s48
33; GCN-NEXT:    v_mov_b32_e32 v13, s49
34; GCN-NEXT:    v_mov_b32_e32 v14, s50
35; GCN-NEXT:    v_mov_b32_e32 v15, s51
36; GCN-NEXT:    s_load_dwordx16 s[36:51], s[22:23], 0xc0
37; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0
38; GCN-NEXT:    buffer_store_dword v1, off, s[0:3], 0 offset:4
39; GCN-NEXT:    buffer_store_dword v2, off, s[0:3], 0 offset:8
40; GCN-NEXT:    buffer_store_dword v3, off, s[0:3], 0 offset:12
41; GCN-NEXT:    buffer_store_dword v4, off, s[0:3], 0 offset:16
42; GCN-NEXT:    buffer_store_dword v5, off, s[0:3], 0 offset:20
43; GCN-NEXT:    buffer_store_dword v6, off, s[0:3], 0 offset:24
44; GCN-NEXT:    buffer_store_dword v7, off, s[0:3], 0 offset:28
45; GCN-NEXT:    buffer_store_dword v8, off, s[0:3], 0 offset:32
46; GCN-NEXT:    buffer_store_dword v9, off, s[0:3], 0 offset:36
47; GCN-NEXT:    buffer_store_dword v10, off, s[0:3], 0 offset:40
48; GCN-NEXT:    buffer_store_dword v11, off, s[0:3], 0 offset:44
49; GCN-NEXT:    buffer_store_dword v12, off, s[0:3], 0 offset:48
50; GCN-NEXT:    buffer_store_dword v13, off, s[0:3], 0 offset:52
51; GCN-NEXT:    buffer_store_dword v14, off, s[0:3], 0 offset:56
52; GCN-NEXT:    buffer_store_dword v15, off, s[0:3], 0 offset:60
53; GCN-NEXT:    v_mov_b32_e32 v0, s52
54; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:64
55; GCN-NEXT:    v_mov_b32_e32 v0, s53
56; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:68
57; GCN-NEXT:    v_mov_b32_e32 v0, s54
58; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:72
59; GCN-NEXT:    v_mov_b32_e32 v0, s55
60; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:76
61; GCN-NEXT:    v_mov_b32_e32 v0, s56
62; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:80
63; GCN-NEXT:    v_mov_b32_e32 v0, s57
64; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:84
65; GCN-NEXT:    v_mov_b32_e32 v0, s58
66; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:88
67; GCN-NEXT:    v_mov_b32_e32 v0, s59
68; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:92
69; GCN-NEXT:    v_mov_b32_e32 v0, s60
70; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:96
71; GCN-NEXT:    v_mov_b32_e32 v0, s61
72; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:100
73; GCN-NEXT:    v_mov_b32_e32 v0, s62
74; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:104
75; GCN-NEXT:    v_mov_b32_e32 v0, s63
76; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:108
77; GCN-NEXT:    v_mov_b32_e32 v0, s64
78; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:112
79; GCN-NEXT:    v_mov_b32_e32 v0, s65
80; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:116
81; GCN-NEXT:    v_mov_b32_e32 v0, s66
82; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:120
83; GCN-NEXT:    v_mov_b32_e32 v0, s67
84; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:124
85; GCN-NEXT:    v_mov_b32_e32 v0, s4
86; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:128
87; GCN-NEXT:    v_mov_b32_e32 v0, s5
88; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:132
89; GCN-NEXT:    v_mov_b32_e32 v0, s6
90; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:136
91; GCN-NEXT:    v_mov_b32_e32 v0, s7
92; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:140
93; GCN-NEXT:    v_mov_b32_e32 v0, s8
94; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:144
95; GCN-NEXT:    v_mov_b32_e32 v0, s9
96; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:148
97; GCN-NEXT:    v_mov_b32_e32 v0, s10
98; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:152
99; GCN-NEXT:    v_mov_b32_e32 v0, s11
100; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:156
101; GCN-NEXT:    v_mov_b32_e32 v0, s12
102; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:160
103; GCN-NEXT:    v_mov_b32_e32 v0, s13
104; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:164
105; GCN-NEXT:    v_mov_b32_e32 v0, s14
106; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:168
107; GCN-NEXT:    v_mov_b32_e32 v0, s15
108; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:172
109; GCN-NEXT:    v_mov_b32_e32 v0, s16
110; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:176
111; GCN-NEXT:    v_mov_b32_e32 v0, s17
112; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:180
113; GCN-NEXT:    v_mov_b32_e32 v0, s18
114; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:184
115; GCN-NEXT:    v_mov_b32_e32 v0, s19
116; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:188
117; GCN-NEXT:    s_waitcnt lgkmcnt(0)
118; GCN-NEXT:    v_mov_b32_e32 v0, s36
119; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:192
120; GCN-NEXT:    v_mov_b32_e32 v0, s37
121; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:196
122; GCN-NEXT:    v_mov_b32_e32 v0, s38
123; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:200
124; GCN-NEXT:    v_mov_b32_e32 v0, s39
125; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:204
126; GCN-NEXT:    v_mov_b32_e32 v0, s40
127; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:208
128; GCN-NEXT:    v_mov_b32_e32 v0, s41
129; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:212
130; GCN-NEXT:    v_mov_b32_e32 v0, s42
131; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:216
132; GCN-NEXT:    v_mov_b32_e32 v0, s43
133; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:220
134; GCN-NEXT:    v_mov_b32_e32 v0, s44
135; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:224
136; GCN-NEXT:    v_mov_b32_e32 v0, s45
137; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:228
138; GCN-NEXT:    v_mov_b32_e32 v0, s46
139; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:232
140; GCN-NEXT:    v_mov_b32_e32 v0, s47
141; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:236
142; GCN-NEXT:    v_mov_b32_e32 v0, s48
143; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:240
144; GCN-NEXT:    v_mov_b32_e32 v0, s49
145; GCN-NEXT:    s_and_b32 s4, s25, 63
146; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:244
147; GCN-NEXT:    v_mov_b32_e32 v0, s50
148; GCN-NEXT:    s_lshl_b32 s4, s4, 2
149; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:248
150; GCN-NEXT:    v_mov_b32_e32 v0, s51
151; GCN-NEXT:    s_add_u32 s4, 0, s4
152; GCN-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:252
153; GCN-NEXT:    v_mov_b32_e32 v0, s24
154; GCN-NEXT:    v_mov_b32_e32 v1, s4
155; GCN-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
156; GCN-NEXT:    buffer_load_dword v0, off, s[0:3], 0
157; GCN-NEXT:    s_nop 0
158; GCN-NEXT:    buffer_load_dword v1, off, s[0:3], 0 offset:4
159; GCN-NEXT:    buffer_load_dword v2, off, s[0:3], 0 offset:8
160; GCN-NEXT:    buffer_load_dword v3, off, s[0:3], 0 offset:12
161; GCN-NEXT:    buffer_load_dword v4, off, s[0:3], 0 offset:16
162; GCN-NEXT:    buffer_load_dword v5, off, s[0:3], 0 offset:20
163; GCN-NEXT:    buffer_load_dword v6, off, s[0:3], 0 offset:24
164; GCN-NEXT:    buffer_load_dword v7, off, s[0:3], 0 offset:28
165; GCN-NEXT:    buffer_load_dword v8, off, s[0:3], 0 offset:32
166; GCN-NEXT:    buffer_load_dword v9, off, s[0:3], 0 offset:36
167; GCN-NEXT:    buffer_load_dword v10, off, s[0:3], 0 offset:40
168; GCN-NEXT:    buffer_load_dword v11, off, s[0:3], 0 offset:44
169; GCN-NEXT:    buffer_load_dword v12, off, s[0:3], 0 offset:48
170; GCN-NEXT:    buffer_load_dword v13, off, s[0:3], 0 offset:52
171; GCN-NEXT:    buffer_load_dword v14, off, s[0:3], 0 offset:56
172; GCN-NEXT:    buffer_load_dword v15, off, s[0:3], 0 offset:60
173; GCN-NEXT:    buffer_load_dword v16, off, s[0:3], 0 offset:64
174; GCN-NEXT:    buffer_load_dword v17, off, s[0:3], 0 offset:68
175; GCN-NEXT:    buffer_load_dword v18, off, s[0:3], 0 offset:72
176; GCN-NEXT:    buffer_load_dword v19, off, s[0:3], 0 offset:76
177; GCN-NEXT:    buffer_load_dword v20, off, s[0:3], 0 offset:80
178; GCN-NEXT:    buffer_load_dword v21, off, s[0:3], 0 offset:84
179; GCN-NEXT:    buffer_load_dword v22, off, s[0:3], 0 offset:88
180; GCN-NEXT:    buffer_load_dword v23, off, s[0:3], 0 offset:92
181; GCN-NEXT:    buffer_load_dword v24, off, s[0:3], 0 offset:96
182; GCN-NEXT:    buffer_load_dword v25, off, s[0:3], 0 offset:100
183; GCN-NEXT:    buffer_load_dword v26, off, s[0:3], 0 offset:104
184; GCN-NEXT:    buffer_load_dword v27, off, s[0:3], 0 offset:108
185; GCN-NEXT:    buffer_load_dword v28, off, s[0:3], 0 offset:112
186; GCN-NEXT:    buffer_load_dword v29, off, s[0:3], 0 offset:116
187; GCN-NEXT:    buffer_load_dword v30, off, s[0:3], 0 offset:120
188; GCN-NEXT:    buffer_load_dword v31, off, s[0:3], 0 offset:124
189; GCN-NEXT:    buffer_load_dword v32, off, s[0:3], 0 offset:128
190; GCN-NEXT:    buffer_load_dword v33, off, s[0:3], 0 offset:132
191; GCN-NEXT:    buffer_load_dword v34, off, s[0:3], 0 offset:136
192; GCN-NEXT:    buffer_load_dword v35, off, s[0:3], 0 offset:140
193; GCN-NEXT:    buffer_load_dword v36, off, s[0:3], 0 offset:144
194; GCN-NEXT:    buffer_load_dword v37, off, s[0:3], 0 offset:148
195; GCN-NEXT:    buffer_load_dword v38, off, s[0:3], 0 offset:152
196; GCN-NEXT:    buffer_load_dword v39, off, s[0:3], 0 offset:156
197; GCN-NEXT:    buffer_load_dword v40, off, s[0:3], 0 offset:160
198; GCN-NEXT:    buffer_load_dword v41, off, s[0:3], 0 offset:164
199; GCN-NEXT:    buffer_load_dword v42, off, s[0:3], 0 offset:168
200; GCN-NEXT:    buffer_load_dword v43, off, s[0:3], 0 offset:172
201; GCN-NEXT:    buffer_load_dword v44, off, s[0:3], 0 offset:176
202; GCN-NEXT:    buffer_load_dword v45, off, s[0:3], 0 offset:180
203; GCN-NEXT:    buffer_load_dword v46, off, s[0:3], 0 offset:184
204; GCN-NEXT:    buffer_load_dword v47, off, s[0:3], 0 offset:188
205; GCN-NEXT:    buffer_load_dword v48, off, s[0:3], 0 offset:192
206; GCN-NEXT:    buffer_load_dword v49, off, s[0:3], 0 offset:196
207; GCN-NEXT:    buffer_load_dword v50, off, s[0:3], 0 offset:200
208; GCN-NEXT:    buffer_load_dword v51, off, s[0:3], 0 offset:204
209; GCN-NEXT:    buffer_load_dword v52, off, s[0:3], 0 offset:208
210; GCN-NEXT:    buffer_load_dword v53, off, s[0:3], 0 offset:212
211; GCN-NEXT:    buffer_load_dword v54, off, s[0:3], 0 offset:216
212; GCN-NEXT:    buffer_load_dword v55, off, s[0:3], 0 offset:220
213; GCN-NEXT:    buffer_load_dword v56, off, s[0:3], 0 offset:224
214; GCN-NEXT:    buffer_load_dword v57, off, s[0:3], 0 offset:228
215; GCN-NEXT:    buffer_load_dword v58, off, s[0:3], 0 offset:232
216; GCN-NEXT:    buffer_load_dword v59, off, s[0:3], 0 offset:236
217; GCN-NEXT:    buffer_load_dword v60, off, s[0:3], 0 offset:240
218; GCN-NEXT:    buffer_load_dword v61, off, s[0:3], 0 offset:244
219; GCN-NEXT:    buffer_load_dword v62, off, s[0:3], 0 offset:248
220; GCN-NEXT:    buffer_load_dword v63, off, s[0:3], 0 offset:252
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}
259
260attributes #0 = { "amdgpu-flat-work-group-size"="1,256" "amdgpu-waves-per-eu"="1,10" }
261