1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=CIVI,HAWAII %s
3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=CIVI,FIJI %s
4; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope --check-prefix=GFX9 %s
5; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -enable-var-scope --check-prefix=GFX10 %s
6
7define void @local_store_i56(i56 addrspace(3)* %ptr, i56 %arg) #0 {
8; CIVI-LABEL: local_store_i56:
9; CIVI:       ; %bb.0:
10; CIVI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
11; CIVI-NEXT:    s_mov_b32 m0, -1
12; CIVI-NEXT:    ds_write_b16 v0, v2 offset:4
13; CIVI-NEXT:    ds_write_b32 v0, v1
14; CIVI-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
15; CIVI-NEXT:    ds_write_b8 v0, v1 offset:6
16; CIVI-NEXT:    s_waitcnt lgkmcnt(0)
17; CIVI-NEXT:    s_setpc_b64 s[30:31]
18;
19; GFX9-LABEL: local_store_i56:
20; GFX9:       ; %bb.0:
21; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
22; GFX9-NEXT:    ds_write_b8_d16_hi v0, v2 offset:6
23; GFX9-NEXT:    ds_write_b16 v0, v2 offset:4
24; GFX9-NEXT:    ds_write_b32 v0, v1
25; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
26; GFX9-NEXT:    s_setpc_b64 s[30:31]
27;
28; GFX10-LABEL: local_store_i56:
29; GFX10:       ; %bb.0:
30; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
31; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
32; GFX10-NEXT:    ds_write_b8_d16_hi v0, v2 offset:6
33; GFX10-NEXT:    ds_write_b16 v0, v2 offset:4
34; GFX10-NEXT:    ds_write_b32 v0, v1
35; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
36; GFX10-NEXT:    s_setpc_b64 s[30:31]
37  store i56 %arg, i56 addrspace(3)* %ptr, align 8
38  ret void
39}
40
41define amdgpu_kernel void @local_store_i55(i55 addrspace(3)* %ptr, i55 %arg) #0 {
42; HAWAII-LABEL: local_store_i55:
43; HAWAII:       ; %bb.0:
44; HAWAII-NEXT:    s_or_b32 s0, s4, 14
45; HAWAII-NEXT:    v_mov_b32_e32 v0, s0
46; HAWAII-NEXT:    v_mov_b32_e32 v1, s5
47; HAWAII-NEXT:    flat_load_ubyte v0, v[0:1]
48; HAWAII-NEXT:    s_load_dword s2, s[4:5], 0x3
49; HAWAII-NEXT:    s_load_dword s0, s[4:5], 0x0
50; HAWAII-NEXT:    s_load_dword s1, s[4:5], 0x2
51; HAWAII-NEXT:    s_mov_b32 m0, -1
52; HAWAII-NEXT:    s_waitcnt lgkmcnt(0)
53; HAWAII-NEXT:    v_mov_b32_e32 v1, s0
54; HAWAII-NEXT:    v_mov_b32_e32 v3, s2
55; HAWAII-NEXT:    v_mov_b32_e32 v2, s1
56; HAWAII-NEXT:    ds_write_b16 v1, v3 offset:4
57; HAWAII-NEXT:    s_waitcnt vmcnt(0)
58; HAWAII-NEXT:    v_and_b32_e32 v0, 0x7f, v0
59; HAWAII-NEXT:    ds_write_b8 v1, v0 offset:6
60; HAWAII-NEXT:    ds_write_b32 v1, v2
61; HAWAII-NEXT:    s_endpgm
62;
63; FIJI-LABEL: local_store_i55:
64; FIJI:       ; %bb.0:
65; FIJI-NEXT:    s_or_b32 s0, s4, 14
66; FIJI-NEXT:    v_mov_b32_e32 v0, s0
67; FIJI-NEXT:    v_mov_b32_e32 v1, s5
68; FIJI-NEXT:    flat_load_ubyte v0, v[0:1]
69; FIJI-NEXT:    s_load_dword s0, s[4:5], 0x0
70; FIJI-NEXT:    s_load_dword s1, s[4:5], 0x8
71; FIJI-NEXT:    s_load_dword s2, s[4:5], 0xc
72; FIJI-NEXT:    s_mov_b32 m0, -1
73; FIJI-NEXT:    s_waitcnt lgkmcnt(0)
74; FIJI-NEXT:    v_mov_b32_e32 v1, s0
75; FIJI-NEXT:    v_mov_b32_e32 v3, s1
76; FIJI-NEXT:    s_and_b32 s3, s2, 0xffff
77; FIJI-NEXT:    v_mov_b32_e32 v2, s2
78; FIJI-NEXT:    ds_write_b16 v1, v2 offset:4
79; FIJI-NEXT:    s_waitcnt vmcnt(0)
80; FIJI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
81; FIJI-NEXT:    v_or_b32_e32 v0, s3, v0
82; FIJI-NEXT:    v_bfe_u32 v0, v0, 16, 7
83; FIJI-NEXT:    ds_write_b8 v1, v0 offset:6
84; FIJI-NEXT:    ds_write_b32 v1, v3
85; FIJI-NEXT:    s_endpgm
86;
87; GFX9-LABEL: local_store_i55:
88; GFX9:       ; %bb.0:
89; GFX9-NEXT:    v_mov_b32_e32 v0, 0
90; GFX9-NEXT:    global_load_ubyte_d16_hi v0, v0, s[4:5] offset:14
91; GFX9-NEXT:    s_load_dword s2, s[4:5], 0xc
92; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x0
93; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x8
94; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
95; GFX9-NEXT:    s_and_b32 s3, s2, 0xffff
96; GFX9-NEXT:    v_mov_b32_e32 v1, s0
97; GFX9-NEXT:    v_mov_b32_e32 v2, s2
98; GFX9-NEXT:    v_mov_b32_e32 v3, s1
99; GFX9-NEXT:    ds_write_b16 v1, v2 offset:4
100; GFX9-NEXT:    s_waitcnt vmcnt(0)
101; GFX9-NEXT:    v_or_b32_e32 v0, s3, v0
102; GFX9-NEXT:    v_and_b32_e32 v0, 0x7fffff, v0
103; GFX9-NEXT:    ds_write_b8_d16_hi v1, v0 offset:6
104; GFX9-NEXT:    ds_write_b32 v1, v3
105; GFX9-NEXT:    s_endpgm
106;
107; GFX10-LABEL: local_store_i55:
108; GFX10:       ; %bb.0:
109; GFX10-NEXT:    v_mov_b32_e32 v0, 0
110; GFX10-NEXT:    s_clause 0x2
111; GFX10-NEXT:    s_load_dword s0, s[4:5], 0xc
112; GFX10-NEXT:    s_load_dword s1, s[4:5], 0x0
113; GFX10-NEXT:    s_load_dword s2, s[4:5], 0x8
114; GFX10-NEXT:    global_load_ubyte_d16_hi v0, v0, s[4:5] offset:14
115; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
116; GFX10-NEXT:    s_and_b32 s3, s0, 0xffff
117; GFX10-NEXT:    v_mov_b32_e32 v1, s1
118; GFX10-NEXT:    v_mov_b32_e32 v2, s0
119; GFX10-NEXT:    v_mov_b32_e32 v3, s2
120; GFX10-NEXT:    s_waitcnt vmcnt(0)
121; GFX10-NEXT:    v_or_b32_e32 v0, s3, v0
122; GFX10-NEXT:    v_and_b32_e32 v0, 0x7fffff, v0
123; GFX10-NEXT:    ds_write_b16 v1, v2 offset:4
124; GFX10-NEXT:    ds_write_b8_d16_hi v1, v0 offset:6
125; GFX10-NEXT:    ds_write_b32 v1, v3
126; GFX10-NEXT:    s_endpgm
127  store i55 %arg, i55 addrspace(3)* %ptr, align 8
128  ret void
129}
130
131define amdgpu_kernel void @local_store_i48(i48 addrspace(3)* %ptr, i48 %arg) #0 {
132; HAWAII-LABEL: local_store_i48:
133; HAWAII:       ; %bb.0:
134; HAWAII-NEXT:    s_load_dword s0, s[4:5], 0x0
135; HAWAII-NEXT:    s_load_dword s1, s[4:5], 0x2
136; HAWAII-NEXT:    s_load_dword s2, s[4:5], 0x3
137; HAWAII-NEXT:    s_mov_b32 m0, -1
138; HAWAII-NEXT:    s_waitcnt lgkmcnt(0)
139; HAWAII-NEXT:    v_mov_b32_e32 v0, s0
140; HAWAII-NEXT:    v_mov_b32_e32 v1, s2
141; HAWAII-NEXT:    ds_write_b16 v0, v1 offset:4
142; HAWAII-NEXT:    v_mov_b32_e32 v1, s1
143; HAWAII-NEXT:    ds_write_b32 v0, v1
144; HAWAII-NEXT:    s_endpgm
145;
146; FIJI-LABEL: local_store_i48:
147; FIJI:       ; %bb.0:
148; FIJI-NEXT:    s_load_dword s0, s[4:5], 0x0
149; FIJI-NEXT:    s_load_dword s1, s[4:5], 0x8
150; FIJI-NEXT:    s_load_dword s2, s[4:5], 0xc
151; FIJI-NEXT:    s_mov_b32 m0, -1
152; FIJI-NEXT:    s_waitcnt lgkmcnt(0)
153; FIJI-NEXT:    v_mov_b32_e32 v0, s0
154; FIJI-NEXT:    v_mov_b32_e32 v1, s2
155; FIJI-NEXT:    ds_write_b16 v0, v1 offset:4
156; FIJI-NEXT:    v_mov_b32_e32 v1, s1
157; FIJI-NEXT:    ds_write_b32 v0, v1
158; FIJI-NEXT:    s_endpgm
159;
160; GFX9-LABEL: local_store_i48:
161; GFX9:       ; %bb.0:
162; GFX9-NEXT:    s_load_dword s0, s[4:5], 0x0
163; GFX9-NEXT:    s_load_dword s1, s[4:5], 0x8
164; GFX9-NEXT:    s_load_dword s2, s[4:5], 0xc
165; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
166; GFX9-NEXT:    v_mov_b32_e32 v0, s0
167; GFX9-NEXT:    v_mov_b32_e32 v2, s1
168; GFX9-NEXT:    v_mov_b32_e32 v1, s2
169; GFX9-NEXT:    ds_write_b16 v0, v1 offset:4
170; GFX9-NEXT:    ds_write_b32 v0, v2
171; GFX9-NEXT:    s_endpgm
172;
173; GFX10-LABEL: local_store_i48:
174; GFX10:       ; %bb.0:
175; GFX10-NEXT:    s_clause 0x2
176; GFX10-NEXT:    s_load_dword s0, s[4:5], 0x0
177; GFX10-NEXT:    s_load_dword s1, s[4:5], 0xc
178; GFX10-NEXT:    s_load_dword s2, s[4:5], 0x8
179; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
180; GFX10-NEXT:    v_mov_b32_e32 v0, s0
181; GFX10-NEXT:    v_mov_b32_e32 v1, s1
182; GFX10-NEXT:    v_mov_b32_e32 v2, s2
183; GFX10-NEXT:    ds_write_b16 v0, v1 offset:4
184; GFX10-NEXT:    ds_write_b32 v0, v2
185; GFX10-NEXT:    s_endpgm
186  store i48 %arg, i48 addrspace(3)* %ptr, align 8
187  ret void
188}
189
190define amdgpu_kernel void @local_store_i65(i65 addrspace(3)* %ptr, i65 %arg) #0 {
191; HAWAII-LABEL: local_store_i65:
192; HAWAII:       ; %bb.0:
193; HAWAII-NEXT:    s_load_dword s2, s[4:5], 0x0
194; HAWAII-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
195; HAWAII-NEXT:    s_load_dword s3, s[4:5], 0x4
196; HAWAII-NEXT:    s_mov_b32 m0, -1
197; HAWAII-NEXT:    s_waitcnt lgkmcnt(0)
198; HAWAII-NEXT:    v_mov_b32_e32 v2, s2
199; HAWAII-NEXT:    s_and_b32 s3, s3, 1
200; HAWAII-NEXT:    v_mov_b32_e32 v0, s3
201; HAWAII-NEXT:    ds_write_b8 v2, v0 offset:8
202; HAWAII-NEXT:    v_mov_b32_e32 v0, s0
203; HAWAII-NEXT:    v_mov_b32_e32 v1, s1
204; HAWAII-NEXT:    ds_write_b64 v2, v[0:1]
205; HAWAII-NEXT:    s_endpgm
206;
207; FIJI-LABEL: local_store_i65:
208; FIJI:       ; %bb.0:
209; FIJI-NEXT:    s_load_dword s2, s[4:5], 0x0
210; FIJI-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
211; FIJI-NEXT:    s_load_dword s3, s[4:5], 0x10
212; FIJI-NEXT:    s_mov_b32 m0, -1
213; FIJI-NEXT:    s_waitcnt lgkmcnt(0)
214; FIJI-NEXT:    v_mov_b32_e32 v2, s2
215; FIJI-NEXT:    s_and_b32 s3, s3, 1
216; FIJI-NEXT:    v_mov_b32_e32 v0, s3
217; FIJI-NEXT:    ds_write_b8 v2, v0 offset:8
218; FIJI-NEXT:    v_mov_b32_e32 v0, s0
219; FIJI-NEXT:    v_mov_b32_e32 v1, s1
220; FIJI-NEXT:    ds_write_b64 v2, v[0:1]
221; FIJI-NEXT:    s_endpgm
222;
223; GFX9-LABEL: local_store_i65:
224; GFX9:       ; %bb.0:
225; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x0
226; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
227; GFX9-NEXT:    s_load_dword s3, s[4:5], 0x10
228; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
229; GFX9-NEXT:    v_mov_b32_e32 v2, s2
230; GFX9-NEXT:    v_mov_b32_e32 v0, s0
231; GFX9-NEXT:    s_and_b32 s3, s3, 1
232; GFX9-NEXT:    v_mov_b32_e32 v3, s3
233; GFX9-NEXT:    v_mov_b32_e32 v1, s1
234; GFX9-NEXT:    ds_write_b8 v2, v3 offset:8
235; GFX9-NEXT:    ds_write_b64 v2, v[0:1]
236; GFX9-NEXT:    s_endpgm
237;
238; GFX10-LABEL: local_store_i65:
239; GFX10:       ; %bb.0:
240; GFX10-NEXT:    s_clause 0x2
241; GFX10-NEXT:    s_load_dword s2, s[4:5], 0x10
242; GFX10-NEXT:    s_load_dword s3, s[4:5], 0x0
243; GFX10-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
244; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
245; GFX10-NEXT:    s_and_b32 s2, s2, 1
246; GFX10-NEXT:    v_mov_b32_e32 v2, s3
247; GFX10-NEXT:    v_mov_b32_e32 v0, s0
248; GFX10-NEXT:    v_mov_b32_e32 v3, s2
249; GFX10-NEXT:    v_mov_b32_e32 v1, s1
250; GFX10-NEXT:    ds_write_b8 v2, v3 offset:8
251; GFX10-NEXT:    ds_write_b64 v2, v[0:1]
252; GFX10-NEXT:    s_endpgm
253  store i65 %arg, i65 addrspace(3)* %ptr, align 8
254  ret void
255}
256
257define void @local_store_i13(i13 addrspace(3)* %ptr, i13 %arg) #0 {
258; CIVI-LABEL: local_store_i13:
259; CIVI:       ; %bb.0:
260; CIVI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
261; CIVI-NEXT:    v_and_b32_e32 v1, 0x1fff, v1
262; CIVI-NEXT:    s_mov_b32 m0, -1
263; CIVI-NEXT:    ds_write_b16 v0, v1
264; CIVI-NEXT:    s_waitcnt lgkmcnt(0)
265; CIVI-NEXT:    s_setpc_b64 s[30:31]
266;
267; GFX9-LABEL: local_store_i13:
268; GFX9:       ; %bb.0:
269; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
270; GFX9-NEXT:    v_and_b32_e32 v1, 0x1fff, v1
271; GFX9-NEXT:    ds_write_b16 v0, v1
272; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
273; GFX9-NEXT:    s_setpc_b64 s[30:31]
274;
275; GFX10-LABEL: local_store_i13:
276; GFX10:       ; %bb.0:
277; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
278; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
279; GFX10-NEXT:    v_and_b32_e32 v1, 0x1fff, v1
280; GFX10-NEXT:    ds_write_b16 v0, v1
281; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
282; GFX10-NEXT:    s_setpc_b64 s[30:31]
283  store i13 %arg, i13 addrspace(3)* %ptr, align 8
284  ret void
285}
286
287define void @local_store_i17(i17 addrspace(3)* %ptr, i17 %arg) #0 {
288; CIVI-LABEL: local_store_i17:
289; CIVI:       ; %bb.0:
290; CIVI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
291; CIVI-NEXT:    s_mov_b32 m0, -1
292; CIVI-NEXT:    ds_write_b16 v0, v1
293; CIVI-NEXT:    v_bfe_u32 v1, v1, 16, 1
294; CIVI-NEXT:    ds_write_b8 v0, v1 offset:2
295; CIVI-NEXT:    s_waitcnt lgkmcnt(0)
296; CIVI-NEXT:    s_setpc_b64 s[30:31]
297;
298; GFX9-LABEL: local_store_i17:
299; GFX9:       ; %bb.0:
300; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
301; GFX9-NEXT:    ds_write_b16 v0, v1
302; GFX9-NEXT:    v_and_b32_e32 v1, 0x1ffff, v1
303; GFX9-NEXT:    ds_write_b8_d16_hi v0, v1 offset:2
304; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
305; GFX9-NEXT:    s_setpc_b64 s[30:31]
306;
307; GFX10-LABEL: local_store_i17:
308; GFX10:       ; %bb.0:
309; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
310; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
311; GFX10-NEXT:    v_and_b32_e32 v2, 0x1ffff, v1
312; GFX10-NEXT:    ds_write_b16 v0, v1
313; GFX10-NEXT:    ds_write_b8_d16_hi v0, v2 offset:2
314; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
315; GFX10-NEXT:    s_setpc_b64 s[30:31]
316  store i17 %arg, i17 addrspace(3)* %ptr, align 8
317  ret void
318}
319
320attributes #0 = { nounwind }
321