xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.large.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-prefix=GCN %s
3; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
4; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX11 %s
5
6define amdgpu_kernel void @v_insert_v64i32_37(ptr addrspace(1) %ptr.in, ptr addrspace(1) %ptr.out) #0 {
7; GCN-LABEL: v_insert_v64i32_37:
8; GCN:       ; %bb.0:
9; GCN-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
10; GCN-NEXT:    v_lshlrev_b32_e32 v64, 8, v0
11; GCN-NEXT:    s_waitcnt lgkmcnt(0)
12; GCN-NEXT:    global_load_dwordx4 v[0:3], v64, s[0:1]
13; GCN-NEXT:    global_load_dwordx4 v[4:7], v64, s[0:1] offset:16
14; GCN-NEXT:    global_load_dwordx4 v[8:11], v64, s[0:1] offset:32
15; GCN-NEXT:    global_load_dwordx4 v[12:15], v64, s[0:1] offset:48
16; GCN-NEXT:    global_load_dwordx4 v[16:19], v64, s[0:1] offset:64
17; GCN-NEXT:    global_load_dwordx4 v[20:23], v64, s[0:1] offset:80
18; GCN-NEXT:    global_load_dwordx4 v[24:27], v64, s[0:1] offset:96
19; GCN-NEXT:    global_load_dwordx4 v[28:31], v64, s[0:1] offset:112
20; GCN-NEXT:    global_load_dwordx4 v[32:35], v64, s[0:1] offset:128
21; GCN-NEXT:    global_load_dwordx4 v[36:39], v64, s[0:1] offset:144
22; GCN-NEXT:    global_load_dwordx4 v[40:43], v64, s[0:1] offset:160
23; GCN-NEXT:    global_load_dwordx4 v[44:47], v64, s[0:1] offset:176
24; GCN-NEXT:    global_load_dwordx4 v[48:51], v64, s[0:1] offset:192
25; GCN-NEXT:    global_load_dwordx4 v[52:55], v64, s[0:1] offset:208
26; GCN-NEXT:    global_load_dwordx4 v[56:59], v64, s[0:1] offset:224
27; GCN-NEXT:    global_load_dwordx4 v[60:63], v64, s[0:1] offset:240
28; GCN-NEXT:    s_waitcnt vmcnt(6)
29; GCN-NEXT:    v_mov_b32_e32 v37, 0x3e7
30; GCN-NEXT:    global_store_dwordx4 v64, v[0:3], s[2:3]
31; GCN-NEXT:    global_store_dwordx4 v64, v[4:7], s[2:3] offset:16
32; GCN-NEXT:    global_store_dwordx4 v64, v[8:11], s[2:3] offset:32
33; GCN-NEXT:    global_store_dwordx4 v64, v[12:15], s[2:3] offset:48
34; GCN-NEXT:    global_store_dwordx4 v64, v[16:19], s[2:3] offset:64
35; GCN-NEXT:    global_store_dwordx4 v64, v[20:23], s[2:3] offset:80
36; GCN-NEXT:    global_store_dwordx4 v64, v[24:27], s[2:3] offset:96
37; GCN-NEXT:    global_store_dwordx4 v64, v[28:31], s[2:3] offset:112
38; GCN-NEXT:    global_store_dwordx4 v64, v[32:35], s[2:3] offset:128
39; GCN-NEXT:    global_store_dwordx4 v64, v[36:39], s[2:3] offset:144
40; GCN-NEXT:    s_waitcnt vmcnt(15)
41; GCN-NEXT:    global_store_dwordx4 v64, v[40:43], s[2:3] offset:160
42; GCN-NEXT:    s_waitcnt vmcnt(15)
43; GCN-NEXT:    global_store_dwordx4 v64, v[44:47], s[2:3] offset:176
44; GCN-NEXT:    s_waitcnt vmcnt(15)
45; GCN-NEXT:    global_store_dwordx4 v64, v[48:51], s[2:3] offset:192
46; GCN-NEXT:    s_waitcnt vmcnt(15)
47; GCN-NEXT:    global_store_dwordx4 v64, v[52:55], s[2:3] offset:208
48; GCN-NEXT:    s_waitcnt vmcnt(15)
49; GCN-NEXT:    global_store_dwordx4 v64, v[56:59], s[2:3] offset:224
50; GCN-NEXT:    s_waitcnt vmcnt(15)
51; GCN-NEXT:    global_store_dwordx4 v64, v[60:63], s[2:3] offset:240
52; GCN-NEXT:    s_endpgm
53;
54; GFX10-LABEL: v_insert_v64i32_37:
55; GFX10:       ; %bb.0:
56; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
57; GFX10-NEXT:    v_lshlrev_b32_e32 v64, 8, v0
58; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
59; GFX10-NEXT:    s_clause 0xf
60; GFX10-NEXT:    global_load_dwordx4 v[0:3], v64, s[0:1]
61; GFX10-NEXT:    global_load_dwordx4 v[4:7], v64, s[0:1] offset:16
62; GFX10-NEXT:    global_load_dwordx4 v[8:11], v64, s[0:1] offset:32
63; GFX10-NEXT:    global_load_dwordx4 v[12:15], v64, s[0:1] offset:48
64; GFX10-NEXT:    global_load_dwordx4 v[16:19], v64, s[0:1] offset:64
65; GFX10-NEXT:    global_load_dwordx4 v[20:23], v64, s[0:1] offset:80
66; GFX10-NEXT:    global_load_dwordx4 v[24:27], v64, s[0:1] offset:96
67; GFX10-NEXT:    global_load_dwordx4 v[28:31], v64, s[0:1] offset:112
68; GFX10-NEXT:    global_load_dwordx4 v[32:35], v64, s[0:1] offset:128
69; GFX10-NEXT:    global_load_dwordx4 v[36:39], v64, s[0:1] offset:144
70; GFX10-NEXT:    global_load_dwordx4 v[40:43], v64, s[0:1] offset:160
71; GFX10-NEXT:    global_load_dwordx4 v[44:47], v64, s[0:1] offset:176
72; GFX10-NEXT:    global_load_dwordx4 v[48:51], v64, s[0:1] offset:192
73; GFX10-NEXT:    global_load_dwordx4 v[52:55], v64, s[0:1] offset:208
74; GFX10-NEXT:    global_load_dwordx4 v[56:59], v64, s[0:1] offset:224
75; GFX10-NEXT:    global_load_dwordx4 v[60:63], v64, s[0:1] offset:240
76; GFX10-NEXT:    s_waitcnt vmcnt(6)
77; GFX10-NEXT:    v_mov_b32_e32 v37, 0x3e7
78; GFX10-NEXT:    global_store_dwordx4 v64, v[0:3], s[2:3]
79; GFX10-NEXT:    global_store_dwordx4 v64, v[4:7], s[2:3] offset:16
80; GFX10-NEXT:    global_store_dwordx4 v64, v[8:11], s[2:3] offset:32
81; GFX10-NEXT:    global_store_dwordx4 v64, v[12:15], s[2:3] offset:48
82; GFX10-NEXT:    global_store_dwordx4 v64, v[16:19], s[2:3] offset:64
83; GFX10-NEXT:    global_store_dwordx4 v64, v[20:23], s[2:3] offset:80
84; GFX10-NEXT:    global_store_dwordx4 v64, v[24:27], s[2:3] offset:96
85; GFX10-NEXT:    global_store_dwordx4 v64, v[28:31], s[2:3] offset:112
86; GFX10-NEXT:    global_store_dwordx4 v64, v[32:35], s[2:3] offset:128
87; GFX10-NEXT:    global_store_dwordx4 v64, v[36:39], s[2:3] offset:144
88; GFX10-NEXT:    s_waitcnt vmcnt(5)
89; GFX10-NEXT:    global_store_dwordx4 v64, v[40:43], s[2:3] offset:160
90; GFX10-NEXT:    s_waitcnt vmcnt(4)
91; GFX10-NEXT:    global_store_dwordx4 v64, v[44:47], s[2:3] offset:176
92; GFX10-NEXT:    s_waitcnt vmcnt(3)
93; GFX10-NEXT:    global_store_dwordx4 v64, v[48:51], s[2:3] offset:192
94; GFX10-NEXT:    s_waitcnt vmcnt(2)
95; GFX10-NEXT:    global_store_dwordx4 v64, v[52:55], s[2:3] offset:208
96; GFX10-NEXT:    s_waitcnt vmcnt(1)
97; GFX10-NEXT:    global_store_dwordx4 v64, v[56:59], s[2:3] offset:224
98; GFX10-NEXT:    s_waitcnt vmcnt(0)
99; GFX10-NEXT:    global_store_dwordx4 v64, v[60:63], s[2:3] offset:240
100; GFX10-NEXT:    s_endpgm
101;
102; GFX11-LABEL: v_insert_v64i32_37:
103; GFX11:       ; %bb.0:
104; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
105; GFX11-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
106; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
107; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v0
108; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
109; GFX11-NEXT:    s_clause 0xf
110; GFX11-NEXT:    global_load_b128 v[0:3], v64, s[0:1]
111; GFX11-NEXT:    global_load_b128 v[4:7], v64, s[0:1] offset:16
112; GFX11-NEXT:    global_load_b128 v[8:11], v64, s[0:1] offset:32
113; GFX11-NEXT:    global_load_b128 v[12:15], v64, s[0:1] offset:48
114; GFX11-NEXT:    global_load_b128 v[16:19], v64, s[0:1] offset:64
115; GFX11-NEXT:    global_load_b128 v[20:23], v64, s[0:1] offset:80
116; GFX11-NEXT:    global_load_b128 v[24:27], v64, s[0:1] offset:96
117; GFX11-NEXT:    global_load_b128 v[28:31], v64, s[0:1] offset:112
118; GFX11-NEXT:    global_load_b128 v[32:35], v64, s[0:1] offset:128
119; GFX11-NEXT:    global_load_b128 v[36:39], v64, s[0:1] offset:144
120; GFX11-NEXT:    global_load_b128 v[40:43], v64, s[0:1] offset:160
121; GFX11-NEXT:    global_load_b128 v[44:47], v64, s[0:1] offset:176
122; GFX11-NEXT:    global_load_b128 v[48:51], v64, s[0:1] offset:192
123; GFX11-NEXT:    global_load_b128 v[52:55], v64, s[0:1] offset:208
124; GFX11-NEXT:    global_load_b128 v[56:59], v64, s[0:1] offset:224
125; GFX11-NEXT:    global_load_b128 v[60:63], v64, s[0:1] offset:240
126; GFX11-NEXT:    s_waitcnt vmcnt(6)
127; GFX11-NEXT:    v_mov_b32_e32 v37, 0x3e7
128; GFX11-NEXT:    s_clause 0x9
129; GFX11-NEXT:    global_store_b128 v64, v[0:3], s[2:3]
130; GFX11-NEXT:    global_store_b128 v64, v[4:7], s[2:3] offset:16
131; GFX11-NEXT:    global_store_b128 v64, v[8:11], s[2:3] offset:32
132; GFX11-NEXT:    global_store_b128 v64, v[12:15], s[2:3] offset:48
133; GFX11-NEXT:    global_store_b128 v64, v[16:19], s[2:3] offset:64
134; GFX11-NEXT:    global_store_b128 v64, v[20:23], s[2:3] offset:80
135; GFX11-NEXT:    global_store_b128 v64, v[24:27], s[2:3] offset:96
136; GFX11-NEXT:    global_store_b128 v64, v[28:31], s[2:3] offset:112
137; GFX11-NEXT:    global_store_b128 v64, v[32:35], s[2:3] offset:128
138; GFX11-NEXT:    global_store_b128 v64, v[36:39], s[2:3] offset:144
139; GFX11-NEXT:    s_waitcnt vmcnt(5)
140; GFX11-NEXT:    global_store_b128 v64, v[40:43], s[2:3] offset:160
141; GFX11-NEXT:    s_waitcnt vmcnt(4)
142; GFX11-NEXT:    global_store_b128 v64, v[44:47], s[2:3] offset:176
143; GFX11-NEXT:    s_waitcnt vmcnt(3)
144; GFX11-NEXT:    global_store_b128 v64, v[48:51], s[2:3] offset:192
145; GFX11-NEXT:    s_waitcnt vmcnt(2)
146; GFX11-NEXT:    global_store_b128 v64, v[52:55], s[2:3] offset:208
147; GFX11-NEXT:    s_waitcnt vmcnt(1)
148; GFX11-NEXT:    global_store_b128 v64, v[56:59], s[2:3] offset:224
149; GFX11-NEXT:    s_waitcnt vmcnt(0)
150; GFX11-NEXT:    global_store_b128 v64, v[60:63], s[2:3] offset:240
151; GFX11-NEXT:    s_endpgm
152  %id = call i32 @llvm.amdgcn.workitem.id.x()
153  %gep.in = getelementptr <64 x i32>, ptr addrspace(1) %ptr.in, i32 %id
154  %vec = load <64 x i32>, ptr addrspace(1) %gep.in
155  %insert = insertelement <64 x i32> %vec, i32 999, i32 37
156  %gep.out = getelementptr <64 x i32>, ptr addrspace(1) %ptr.out, i32 %id
157  store <64 x i32> %insert, ptr addrspace(1) %gep.out
158  ret void
159}
160
161declare i32 @llvm.amdgcn.workitem.id.x() #1
162
163attributes #0 = { "amdgpu-flat-work-group-size"="1,256" "amdgpu-waves-per-eu"="1,10" }
164attributes #1 = { nounwind readnone speculatable willreturn }
165