1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefix=GFX9 %s
3; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii -verify-machineinstrs < %s | FileCheck --check-prefix=GFX7 %s
4; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti -verify-machineinstrs < %s | FileCheck --check-prefix=GFX6 %s
5; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefix=GFX10 %s
6
7define <4 x i32> @load_lds_v4i32(<4 x i32> addrspace(3)* %ptr) {
8; GFX9-LABEL: load_lds_v4i32:
9; GFX9:       ; %bb.0:
10; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
11; GFX9-NEXT:    ds_read_b128 v[0:3], v0
12; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
13; GFX9-NEXT:    s_setpc_b64 s[30:31]
14;
15; GFX7-LABEL: load_lds_v4i32:
16; GFX7:       ; %bb.0:
17; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
18; GFX7-NEXT:    s_mov_b32 m0, -1
19; GFX7-NEXT:    ds_read_b128 v[0:3], v0
20; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
21; GFX7-NEXT:    s_setpc_b64 s[30:31]
22;
23; GFX6-LABEL: load_lds_v4i32:
24; GFX6:       ; %bb.0:
25; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
26; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
27; GFX6-NEXT:    s_mov_b32 m0, -1
28; GFX6-NEXT:    ds_read_b64 v[2:3], v1
29; GFX6-NEXT:    ds_read_b64 v[0:1], v0
30; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
31; GFX6-NEXT:    s_setpc_b64 s[30:31]
32;
33; GFX10-LABEL: load_lds_v4i32:
34; GFX10:       ; %bb.0:
35; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
36; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
37; GFX10-NEXT:    ds_read_b128 v[0:3], v0
38; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
39; GFX10-NEXT:    s_setpc_b64 s[30:31]
40  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr
41  ret <4 x i32> %load
42}
43
44define <4 x i32> @load_lds_v4i32_align1(<4 x i32> addrspace(3)* %ptr) {
45; GFX9-LABEL: load_lds_v4i32_align1:
46; GFX9:       ; %bb.0:
47; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
48; GFX9-NEXT:    ds_read_u8 v1, v0
49; GFX9-NEXT:    ds_read_u8 v2, v0 offset:1
50; GFX9-NEXT:    ds_read_u8 v3, v0 offset:2
51; GFX9-NEXT:    ds_read_u8 v4, v0 offset:3
52; GFX9-NEXT:    ds_read_u8 v5, v0 offset:4
53; GFX9-NEXT:    ds_read_u8 v6, v0 offset:5
54; GFX9-NEXT:    ds_read_u8 v7, v0 offset:6
55; GFX9-NEXT:    ds_read_u8 v8, v0 offset:7
56; GFX9-NEXT:    ds_read_u8 v9, v0 offset:8
57; GFX9-NEXT:    ds_read_u8 v10, v0 offset:9
58; GFX9-NEXT:    ds_read_u8 v11, v0 offset:10
59; GFX9-NEXT:    ds_read_u8 v12, v0 offset:11
60; GFX9-NEXT:    ds_read_u8 v13, v0 offset:12
61; GFX9-NEXT:    ds_read_u8 v14, v0 offset:13
62; GFX9-NEXT:    ds_read_u8 v15, v0 offset:14
63; GFX9-NEXT:    ds_read_u8 v16, v0 offset:15
64; GFX9-NEXT:    s_waitcnt lgkmcnt(14)
65; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 8, v1
66; GFX9-NEXT:    s_waitcnt lgkmcnt(12)
67; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
68; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
69; GFX9-NEXT:    s_waitcnt lgkmcnt(10)
70; GFX9-NEXT:    v_lshl_or_b32 v1, v6, 8, v5
71; GFX9-NEXT:    s_waitcnt lgkmcnt(8)
72; GFX9-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
73; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
74; GFX9-NEXT:    s_waitcnt lgkmcnt(6)
75; GFX9-NEXT:    v_lshl_or_b32 v2, v10, 8, v9
76; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
77; GFX9-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
78; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
79; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
80; GFX9-NEXT:    v_lshl_or_b32 v3, v14, 8, v13
81; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
82; GFX9-NEXT:    v_lshl_or_b32 v4, v16, 8, v15
83; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
84; GFX9-NEXT:    s_setpc_b64 s[30:31]
85;
86; GFX7-LABEL: load_lds_v4i32_align1:
87; GFX7:       ; %bb.0:
88; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
89; GFX7-NEXT:    s_mov_b32 m0, -1
90; GFX7-NEXT:    ds_read_u8 v1, v0 offset:7
91; GFX7-NEXT:    ds_read_u8 v2, v0 offset:6
92; GFX7-NEXT:    ds_read_u8 v3, v0 offset:5
93; GFX7-NEXT:    ds_read_u8 v5, v0 offset:4
94; GFX7-NEXT:    ds_read_u8 v4, v0 offset:3
95; GFX7-NEXT:    ds_read_u8 v6, v0 offset:2
96; GFX7-NEXT:    ds_read_u8 v7, v0 offset:1
97; GFX7-NEXT:    ds_read_u8 v8, v0
98; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
99; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
100; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
101; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
102; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
103; GFX7-NEXT:    v_or_b32_e32 v4, v4, v6
104; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
105; GFX7-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
106; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
107; GFX7-NEXT:    v_or_b32_e32 v1, v1, v2
108; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
109; GFX7-NEXT:    v_or_b32_e32 v7, v7, v8
110; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
111; GFX7-NEXT:    v_or_b32_e32 v3, v3, v5
112; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
113; GFX7-NEXT:    v_or_b32_e32 v4, v4, v7
114; GFX7-NEXT:    v_or_b32_e32 v1, v1, v3
115; GFX7-NEXT:    ds_read_u8 v3, v0 offset:15
116; GFX7-NEXT:    ds_read_u8 v5, v0 offset:14
117; GFX7-NEXT:    ds_read_u8 v6, v0 offset:13
118; GFX7-NEXT:    ds_read_u8 v7, v0 offset:12
119; GFX7-NEXT:    ds_read_u8 v2, v0 offset:11
120; GFX7-NEXT:    ds_read_u8 v8, v0 offset:10
121; GFX7-NEXT:    ds_read_u8 v9, v0 offset:9
122; GFX7-NEXT:    ds_read_u8 v0, v0 offset:8
123; GFX7-NEXT:    s_waitcnt lgkmcnt(7)
124; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
125; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
126; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
127; GFX7-NEXT:    s_waitcnt lgkmcnt(2)
128; GFX7-NEXT:    v_or_b32_e32 v2, v2, v8
129; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
130; GFX7-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
131; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
132; GFX7-NEXT:    v_or_b32_e32 v0, v9, v0
133; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
134; GFX7-NEXT:    v_or_b32_e32 v3, v3, v5
135; GFX7-NEXT:    v_or_b32_e32 v2, v2, v0
136; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
137; GFX7-NEXT:    v_or_b32_e32 v0, v0, v7
138; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
139; GFX7-NEXT:    v_or_b32_e32 v3, v3, v0
140; GFX7-NEXT:    v_mov_b32_e32 v0, v4
141; GFX7-NEXT:    s_setpc_b64 s[30:31]
142;
143; GFX6-LABEL: load_lds_v4i32_align1:
144; GFX6:       ; %bb.0:
145; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
146; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 5, v0
147; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
148; GFX6-NEXT:    v_add_i32_e32 v3, vcc, 7, v0
149; GFX6-NEXT:    v_add_i32_e32 v4, vcc, 6, v0
150; GFX6-NEXT:    v_add_i32_e32 v5, vcc, 9, v0
151; GFX6-NEXT:    v_add_i32_e32 v6, vcc, 8, v0
152; GFX6-NEXT:    v_add_i32_e32 v7, vcc, 11, v0
153; GFX6-NEXT:    s_mov_b32 m0, -1
154; GFX6-NEXT:    ds_read_u8 v2, v2
155; GFX6-NEXT:    ds_read_u8 v3, v3
156; GFX6-NEXT:    ds_read_u8 v4, v4
157; GFX6-NEXT:    ds_read_u8 v5, v5
158; GFX6-NEXT:    ds_read_u8 v6, v6
159; GFX6-NEXT:    ds_read_u8 v7, v7
160; GFX6-NEXT:    ds_read_u8 v1, v1
161; GFX6-NEXT:    ds_read_u8 v8, v0
162; GFX6-NEXT:    v_add_i32_e32 v9, vcc, 14, v0
163; GFX6-NEXT:    v_add_i32_e32 v10, vcc, 3, v0
164; GFX6-NEXT:    v_add_i32_e32 v11, vcc, 2, v0
165; GFX6-NEXT:    s_waitcnt lgkmcnt(1)
166; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
167; GFX6-NEXT:    v_or_b32_e32 v1, v1, v2
168; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 8, v3
169; GFX6-NEXT:    v_or_b32_e32 v2, v2, v4
170; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
171; GFX6-NEXT:    v_or_b32_e32 v1, v2, v1
172; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
173; GFX6-NEXT:    v_or_b32_e32 v2, v2, v6
174; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 8, v7
175; GFX6-NEXT:    v_add_i32_e32 v4, vcc, 10, v0
176; GFX6-NEXT:    v_add_i32_e32 v5, vcc, 13, v0
177; GFX6-NEXT:    v_add_i32_e32 v6, vcc, 12, v0
178; GFX6-NEXT:    v_add_i32_e32 v7, vcc, 15, v0
179; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 1, v0
180; GFX6-NEXT:    ds_read_u8 v4, v4
181; GFX6-NEXT:    ds_read_u8 v5, v5
182; GFX6-NEXT:    ds_read_u8 v6, v6
183; GFX6-NEXT:    ds_read_u8 v7, v7
184; GFX6-NEXT:    ds_read_u8 v9, v9
185; GFX6-NEXT:    ds_read_u8 v10, v10
186; GFX6-NEXT:    ds_read_u8 v11, v11
187; GFX6-NEXT:    ds_read_u8 v0, v0
188; GFX6-NEXT:    s_waitcnt lgkmcnt(7)
189; GFX6-NEXT:    v_or_b32_e32 v3, v3, v4
190; GFX6-NEXT:    s_waitcnt lgkmcnt(4)
191; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 8, v7
192; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
193; GFX6-NEXT:    s_waitcnt lgkmcnt(3)
194; GFX6-NEXT:    v_or_b32_e32 v4, v4, v9
195; GFX6-NEXT:    v_or_b32_e32 v2, v3, v2
196; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
197; GFX6-NEXT:    v_or_b32_e32 v3, v3, v6
198; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
199; GFX6-NEXT:    v_or_b32_e32 v3, v4, v3
200; GFX6-NEXT:    s_waitcnt lgkmcnt(2)
201; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 8, v10
202; GFX6-NEXT:    s_waitcnt lgkmcnt(1)
203; GFX6-NEXT:    v_or_b32_e32 v4, v4, v11
204; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
205; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
206; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
207; GFX6-NEXT:    v_or_b32_e32 v0, v0, v8
208; GFX6-NEXT:    v_or_b32_e32 v0, v4, v0
209; GFX6-NEXT:    s_setpc_b64 s[30:31]
210;
211; GFX10-LABEL: load_lds_v4i32_align1:
212; GFX10:       ; %bb.0:
213; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
214; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
215; GFX10-NEXT:    ds_read_u8 v1, v0
216; GFX10-NEXT:    ds_read_u8 v2, v0 offset:1
217; GFX10-NEXT:    ds_read_u8 v3, v0 offset:2
218; GFX10-NEXT:    ds_read_u8 v4, v0 offset:3
219; GFX10-NEXT:    ds_read_u8 v5, v0 offset:4
220; GFX10-NEXT:    ds_read_u8 v6, v0 offset:5
221; GFX10-NEXT:    ds_read_u8 v7, v0 offset:6
222; GFX10-NEXT:    ds_read_u8 v8, v0 offset:7
223; GFX10-NEXT:    ds_read_u8 v9, v0 offset:8
224; GFX10-NEXT:    ds_read_u8 v10, v0 offset:9
225; GFX10-NEXT:    ds_read_u8 v11, v0 offset:10
226; GFX10-NEXT:    ds_read_u8 v12, v0 offset:11
227; GFX10-NEXT:    ds_read_u8 v13, v0 offset:12
228; GFX10-NEXT:    ds_read_u8 v14, v0 offset:13
229; GFX10-NEXT:    ds_read_u8 v15, v0 offset:14
230; GFX10-NEXT:    ds_read_u8 v0, v0 offset:15
231; GFX10-NEXT:    s_waitcnt lgkmcnt(14)
232; GFX10-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
233; GFX10-NEXT:    s_waitcnt lgkmcnt(12)
234; GFX10-NEXT:    v_lshl_or_b32 v2, v4, 8, v3
235; GFX10-NEXT:    s_waitcnt lgkmcnt(10)
236; GFX10-NEXT:    v_lshl_or_b32 v3, v6, 8, v5
237; GFX10-NEXT:    s_waitcnt lgkmcnt(8)
238; GFX10-NEXT:    v_lshl_or_b32 v4, v8, 8, v7
239; GFX10-NEXT:    s_waitcnt lgkmcnt(6)
240; GFX10-NEXT:    v_lshl_or_b32 v5, v10, 8, v9
241; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
242; GFX10-NEXT:    v_lshl_or_b32 v6, v12, 8, v11
243; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
244; GFX10-NEXT:    v_lshl_or_b32 v7, v14, 8, v13
245; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
246; GFX10-NEXT:    v_lshl_or_b32 v8, v0, 8, v15
247; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
248; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
249; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
250; GFX10-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
251; GFX10-NEXT:    s_setpc_b64 s[30:31]
252  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 1
253  ret <4 x i32> %load
254}
255
256define <4 x i32> @load_lds_v4i32_align2(<4 x i32> addrspace(3)* %ptr) {
257; GFX9-LABEL: load_lds_v4i32_align2:
258; GFX9:       ; %bb.0:
259; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
260; GFX9-NEXT:    ds_read_u16 v1, v0
261; GFX9-NEXT:    ds_read_u16 v2, v0 offset:2
262; GFX9-NEXT:    ds_read_u16 v3, v0 offset:4
263; GFX9-NEXT:    ds_read_u16 v4, v0 offset:6
264; GFX9-NEXT:    ds_read_u16 v5, v0 offset:8
265; GFX9-NEXT:    ds_read_u16 v6, v0 offset:10
266; GFX9-NEXT:    ds_read_u16 v7, v0 offset:12
267; GFX9-NEXT:    ds_read_u16 v8, v0 offset:14
268; GFX9-NEXT:    s_waitcnt lgkmcnt(6)
269; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
270; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
271; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
272; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
273; GFX9-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
274; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
275; GFX9-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
276; GFX9-NEXT:    s_setpc_b64 s[30:31]
277;
278; GFX7-LABEL: load_lds_v4i32_align2:
279; GFX7:       ; %bb.0:
280; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
281; GFX7-NEXT:    s_mov_b32 m0, -1
282; GFX7-NEXT:    ds_read_u16 v3, v0 offset:14
283; GFX7-NEXT:    ds_read_u16 v4, v0 offset:12
284; GFX7-NEXT:    ds_read_u16 v2, v0 offset:10
285; GFX7-NEXT:    ds_read_u16 v5, v0 offset:8
286; GFX7-NEXT:    ds_read_u16 v1, v0 offset:6
287; GFX7-NEXT:    ds_read_u16 v6, v0 offset:4
288; GFX7-NEXT:    ds_read_u16 v7, v0 offset:2
289; GFX7-NEXT:    ds_read_u16 v0, v0
290; GFX7-NEXT:    s_waitcnt lgkmcnt(5)
291; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
292; GFX7-NEXT:    s_waitcnt lgkmcnt(3)
293; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
294; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
295; GFX7-NEXT:    s_waitcnt lgkmcnt(1)
296; GFX7-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
297; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
298; GFX7-NEXT:    v_or_b32_e32 v0, v7, v0
299; GFX7-NEXT:    v_or_b32_e32 v1, v1, v6
300; GFX7-NEXT:    v_or_b32_e32 v2, v2, v5
301; GFX7-NEXT:    v_or_b32_e32 v3, v3, v4
302; GFX7-NEXT:    s_setpc_b64 s[30:31]
303;
304; GFX6-LABEL: load_lds_v4i32_align2:
305; GFX6:       ; %bb.0:
306; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
307; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 6, v0
308; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
309; GFX6-NEXT:    v_add_i32_e32 v3, vcc, 10, v0
310; GFX6-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
311; GFX6-NEXT:    v_add_i32_e32 v5, vcc, 14, v0
312; GFX6-NEXT:    v_add_i32_e32 v6, vcc, 12, v0
313; GFX6-NEXT:    v_add_i32_e32 v7, vcc, 2, v0
314; GFX6-NEXT:    s_mov_b32 m0, -1
315; GFX6-NEXT:    ds_read_u16 v2, v2
316; GFX6-NEXT:    ds_read_u16 v3, v3
317; GFX6-NEXT:    ds_read_u16 v4, v4
318; GFX6-NEXT:    ds_read_u16 v5, v5
319; GFX6-NEXT:    ds_read_u16 v6, v6
320; GFX6-NEXT:    ds_read_u16 v7, v7
321; GFX6-NEXT:    ds_read_u16 v1, v1
322; GFX6-NEXT:    ds_read_u16 v0, v0
323; GFX6-NEXT:    s_waitcnt lgkmcnt(1)
324; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
325; GFX6-NEXT:    v_or_b32_e32 v1, v1, v2
326; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
327; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
328; GFX6-NEXT:    v_or_b32_e32 v2, v2, v4
329; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v7
330; GFX6-NEXT:    v_or_b32_e32 v3, v3, v6
331; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
332; GFX6-NEXT:    v_or_b32_e32 v0, v4, v0
333; GFX6-NEXT:    s_setpc_b64 s[30:31]
334;
335; GFX10-LABEL: load_lds_v4i32_align2:
336; GFX10:       ; %bb.0:
337; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
338; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
339; GFX10-NEXT:    ds_read_u16 v1, v0
340; GFX10-NEXT:    ds_read_u16 v2, v0 offset:2
341; GFX10-NEXT:    ds_read_u16 v3, v0 offset:4
342; GFX10-NEXT:    ds_read_u16 v4, v0 offset:6
343; GFX10-NEXT:    ds_read_u16 v5, v0 offset:8
344; GFX10-NEXT:    ds_read_u16 v6, v0 offset:10
345; GFX10-NEXT:    ds_read_u16 v7, v0 offset:12
346; GFX10-NEXT:    ds_read_u16 v8, v0 offset:14
347; GFX10-NEXT:    s_waitcnt lgkmcnt(6)
348; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
349; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
350; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
351; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
352; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
353; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
354; GFX10-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
355; GFX10-NEXT:    s_setpc_b64 s[30:31]
356  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 2
357  ret <4 x i32> %load
358}
359
360define <4 x i32> @load_lds_v4i32_align4(<4 x i32> addrspace(3)* %ptr) {
361; GFX9-LABEL: load_lds_v4i32_align4:
362; GFX9:       ; %bb.0:
363; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
364; GFX9-NEXT:    v_mov_b32_e32 v2, v0
365; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
366; GFX9-NEXT:    ds_read2_b32 v[2:3], v2 offset0:2 offset1:3
367; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
368; GFX9-NEXT:    s_setpc_b64 s[30:31]
369;
370; GFX7-LABEL: load_lds_v4i32_align4:
371; GFX7:       ; %bb.0:
372; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
373; GFX7-NEXT:    v_mov_b32_e32 v2, v0
374; GFX7-NEXT:    s_mov_b32 m0, -1
375; GFX7-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
376; GFX7-NEXT:    ds_read2_b32 v[2:3], v2 offset0:2 offset1:3
377; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
378; GFX7-NEXT:    s_setpc_b64 s[30:31]
379;
380; GFX6-LABEL: load_lds_v4i32_align4:
381; GFX6:       ; %bb.0:
382; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
383; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 4, v0
384; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
385; GFX6-NEXT:    v_add_i32_e32 v3, vcc, 12, v0
386; GFX6-NEXT:    s_mov_b32 m0, -1
387; GFX6-NEXT:    ds_read_b32 v2, v2
388; GFX6-NEXT:    ds_read_b32 v3, v3
389; GFX6-NEXT:    ds_read_b32 v1, v1
390; GFX6-NEXT:    ds_read_b32 v0, v0
391; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
392; GFX6-NEXT:    s_setpc_b64 s[30:31]
393;
394; GFX10-LABEL: load_lds_v4i32_align4:
395; GFX10:       ; %bb.0:
396; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
397; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
398; GFX10-NEXT:    v_mov_b32_e32 v2, v0
399; GFX10-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
400; GFX10-NEXT:    ds_read2_b32 v[2:3], v2 offset0:2 offset1:3
401; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
402; GFX10-NEXT:    s_setpc_b64 s[30:31]
403  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 4
404  ret <4 x i32> %load
405}
406
407define <4 x i32> @load_lds_v4i32_align8(<4 x i32> addrspace(3)* %ptr) {
408; GFX9-LABEL: load_lds_v4i32_align8:
409; GFX9:       ; %bb.0:
410; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
411; GFX9-NEXT:    ds_read2_b64 v[0:3], v0 offset1:1
412; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
413; GFX9-NEXT:    s_setpc_b64 s[30:31]
414;
415; GFX7-LABEL: load_lds_v4i32_align8:
416; GFX7:       ; %bb.0:
417; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
418; GFX7-NEXT:    s_mov_b32 m0, -1
419; GFX7-NEXT:    ds_read2_b64 v[0:3], v0 offset1:1
420; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
421; GFX7-NEXT:    s_setpc_b64 s[30:31]
422;
423; GFX6-LABEL: load_lds_v4i32_align8:
424; GFX6:       ; %bb.0:
425; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
426; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
427; GFX6-NEXT:    s_mov_b32 m0, -1
428; GFX6-NEXT:    ds_read_b64 v[2:3], v1
429; GFX6-NEXT:    ds_read_b64 v[0:1], v0
430; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
431; GFX6-NEXT:    s_setpc_b64 s[30:31]
432;
433; GFX10-LABEL: load_lds_v4i32_align8:
434; GFX10:       ; %bb.0:
435; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
436; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
437; GFX10-NEXT:    ds_read2_b64 v[0:3], v0 offset1:1
438; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
439; GFX10-NEXT:    s_setpc_b64 s[30:31]
440  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 8
441  ret <4 x i32> %load
442}
443
444define <4 x i32> @load_lds_v4i32_align16(<4 x i32> addrspace(3)* %ptr) {
445; GFX9-LABEL: load_lds_v4i32_align16:
446; GFX9:       ; %bb.0:
447; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
448; GFX9-NEXT:    ds_read_b128 v[0:3], v0
449; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
450; GFX9-NEXT:    s_setpc_b64 s[30:31]
451;
452; GFX7-LABEL: load_lds_v4i32_align16:
453; GFX7:       ; %bb.0:
454; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
455; GFX7-NEXT:    s_mov_b32 m0, -1
456; GFX7-NEXT:    ds_read_b128 v[0:3], v0
457; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
458; GFX7-NEXT:    s_setpc_b64 s[30:31]
459;
460; GFX6-LABEL: load_lds_v4i32_align16:
461; GFX6:       ; %bb.0:
462; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
463; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
464; GFX6-NEXT:    s_mov_b32 m0, -1
465; GFX6-NEXT:    ds_read_b64 v[2:3], v1
466; GFX6-NEXT:    ds_read_b64 v[0:1], v0
467; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
468; GFX6-NEXT:    s_setpc_b64 s[30:31]
469;
470; GFX10-LABEL: load_lds_v4i32_align16:
471; GFX10:       ; %bb.0:
472; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
473; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
474; GFX10-NEXT:    ds_read_b128 v[0:3], v0
475; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
476; GFX10-NEXT:    s_setpc_b64 s[30:31]
477  %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 16
478  ret <4 x i32> %load
479}
480