1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
3; RUN: llc -mtriple=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX8 %s
4; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-NODL %s
5; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9-DL %s
6; RUN: llc -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-DL %s
7; RUN: llc -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-DL %s
8
9; add(mul(S0.x, S1.y),
10;     add (mul (S0.y, S1.y), S3)) -> v_dot2_{I|U}32_{I|U}16(S1, S2, S3)
11
12define amdgpu_kernel void @udot2(<2 x i16> addrspace(1)* %src1,
13; GFX7-LABEL: udot2:
14; GFX7:       ; %bb.0: ; %entry
15; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
16; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
17; GFX7-NEXT:    s_mov_b32 s8, 0xffff
18; GFX7-NEXT:    s_mov_b32 s3, 0xf000
19; GFX7-NEXT:    s_mov_b32 s2, -1
20; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
21; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
22; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
23; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
24; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
25; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
26; GFX7-NEXT:    s_and_b32 s4, s4, s8
27; GFX7-NEXT:    s_and_b32 s5, s5, s8
28; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
29; GFX7-NEXT:    v_mov_b32_e32 v0, s6
30; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
31; GFX7-NEXT:    v_mov_b32_e32 v1, s8
32; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
33; GFX7-NEXT:    v_mov_b32_e32 v1, s4
34; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v1, v0
35; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
36; GFX7-NEXT:    s_endpgm
37;
38; GFX8-LABEL: udot2:
39; GFX8:       ; %bb.0: ; %entry
40; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
41; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
42; GFX8-NEXT:    s_mov_b32 s2, 0xffff
43; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
44; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
45; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
46; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
47; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
48; GFX8-NEXT:    s_and_b32 s6, s3, s2
49; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
50; GFX8-NEXT:    s_and_b32 s2, s4, s2
51; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
52; GFX8-NEXT:    v_mov_b32_e32 v0, s5
53; GFX8-NEXT:    v_mov_b32_e32 v1, s3
54; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
55; GFX8-NEXT:    v_mov_b32_e32 v1, s6
56; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
57; GFX8-NEXT:    v_mov_b32_e32 v0, s0
58; GFX8-NEXT:    v_mov_b32_e32 v1, s1
59; GFX8-NEXT:    flat_store_dword v[0:1], v2
60; GFX8-NEXT:    s_endpgm
61;
62; GFX9-NODL-LABEL: udot2:
63; GFX9-NODL:       ; %bb.0: ; %entry
64; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
65; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
66; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
67; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
68; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
69; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
70; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
71; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
72; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
73; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
74; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
75; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
76; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
77; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
78; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
79; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
80; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
81; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
82; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
83; GFX9-NODL-NEXT:    s_endpgm
84;
85; GFX9-DL-LABEL: udot2:
86; GFX9-DL:       ; %bb.0: ; %entry
87; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
88; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
89; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
90; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
91; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
92; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
93; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
94; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
95; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s2
96; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
97; GFX9-DL-NEXT:    v_dot2_u32_u16 v1, s4, v1, v2
98; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
99; GFX9-DL-NEXT:    s_endpgm
100;
101; GFX10-DL-LABEL: udot2:
102; GFX10-DL:       ; %bb.0: ; %entry
103; GFX10-DL-NEXT:    s_clause 0x1
104; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
105; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
106; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
107; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
108; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
109; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
110; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
111; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
112; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
113; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
114; GFX10-DL-NEXT:    v_dot2_u32_u16 v0, s1, s0, v0
115; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
116; GFX10-DL-NEXT:    s_endpgm
117                                 <2 x i16> addrspace(1)* %src2,
118                                 i32 addrspace(1)* nocapture %dst) {
119entry:
120  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
121  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
122
123  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
124  %conv = zext i16 %s1.elt1 to i32
125  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
126  %conv2 = zext i16 %s2.elt1 to i32
127  %mul1 = mul nuw i32 %conv2, %conv
128
129  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
130  %conv3 = zext i16 %s1.elt2 to i32
131  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
132  %conv4 = zext i16 %s2.elt2 to i32
133  %mul2 = mul nuw i32 %conv4, %conv3
134
135  %s3 = load i32, i32 addrspace(1)* %dst, align 4
136  %add = add i32 %mul2, %s3
137  %add6 = add i32 %add, %mul1
138  store i32 %add6, i32 addrspace(1)* %dst, align 4
139  ret void
140}
141
142; TODO: Support this pattern
143;      add(S3,
144;          add (mul (S0.y, S1.y), mul (S0.y, S1.y))) -> v_dot2_{I|U}32_{I|U}16(S1, S2, S3)
145define amdgpu_kernel void @udot2_MulMul(<2 x i16> addrspace(1)* %src1,
146; GFX7-LABEL: udot2_MulMul:
147; GFX7:       ; %bb.0: ; %entry
148; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
149; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
150; GFX7-NEXT:    s_mov_b32 s8, 0xffff
151; GFX7-NEXT:    s_mov_b32 s3, 0xf000
152; GFX7-NEXT:    s_mov_b32 s2, -1
153; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
154; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
155; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
156; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
157; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
158; GFX7-NEXT:    s_and_b32 s4, s4, s8
159; GFX7-NEXT:    v_mov_b32_e32 v0, s4
160; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
161; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
162; GFX7-NEXT:    s_and_b32 s5, s5, s8
163; GFX7-NEXT:    v_mul_u32_u24_e32 v0, s5, v0
164; GFX7-NEXT:    v_mov_b32_e32 v1, s6
165; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v1, v0
166; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
167; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
168; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
169; GFX7-NEXT:    s_endpgm
170;
171; GFX8-LABEL: udot2_MulMul:
172; GFX8:       ; %bb.0: ; %entry
173; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
174; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
175; GFX8-NEXT:    s_mov_b32 s2, 0xffff
176; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
177; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
178; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
179; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
180; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
181; GFX8-NEXT:    s_and_b32 s6, s3, s2
182; GFX8-NEXT:    s_and_b32 s2, s4, s2
183; GFX8-NEXT:    v_mov_b32_e32 v0, s6
184; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
185; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
186; GFX8-NEXT:    v_mov_b32_e32 v1, s3
187; GFX8-NEXT:    v_mul_u32_u24_e32 v0, s2, v0
188; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
189; GFX8-NEXT:    v_add_u32_e32 v2, vcc, s5, v0
190; GFX8-NEXT:    v_mov_b32_e32 v0, s0
191; GFX8-NEXT:    v_mov_b32_e32 v1, s1
192; GFX8-NEXT:    flat_store_dword v[0:1], v2
193; GFX8-NEXT:    s_endpgm
194;
195; GFX9-NODL-LABEL: udot2_MulMul:
196; GFX9-NODL:       ; %bb.0: ; %entry
197; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
198; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
199; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
200; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
201; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
202; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
203; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
204; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
205; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
206; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
207; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
208; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s6
209; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
210; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
211; GFX9-NODL-NEXT:    v_mul_u32_u24_e32 v1, s2, v1
212; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s3
213; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
214; GFX9-NODL-NEXT:    v_add_u32_e32 v1, s5, v1
215; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
216; GFX9-NODL-NEXT:    s_endpgm
217;
218; GFX9-DL-LABEL: udot2_MulMul:
219; GFX9-DL:       ; %bb.0: ; %entry
220; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
221; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
222; GFX9-DL-NEXT:    s_mov_b32 s2, 0xffff
223; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
224; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
225; GFX9-DL-NEXT:    s_load_dword s3, s[4:5], 0x0
226; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
227; GFX9-DL-NEXT:    s_load_dword s5, s[0:1], 0x0
228; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
229; GFX9-DL-NEXT:    s_and_b32 s6, s3, s2
230; GFX9-DL-NEXT:    s_and_b32 s2, s4, s2
231; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s6
232; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
233; GFX9-DL-NEXT:    s_lshr_b32 s4, s4, 16
234; GFX9-DL-NEXT:    v_mul_u32_u24_e32 v1, s2, v1
235; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
236; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
237; GFX9-DL-NEXT:    v_add_u32_e32 v1, s5, v1
238; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
239; GFX9-DL-NEXT:    s_endpgm
240;
241; GFX10-DL-LABEL: udot2_MulMul:
242; GFX10-DL:       ; %bb.0: ; %entry
243; GFX10-DL-NEXT:    s_clause 0x1
244; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
245; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
246; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
247; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
248; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
249; GFX10-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
250; GFX10-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
251; GFX10-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
252; GFX10-DL-NEXT:    s_mov_b32 s5, 0xffff
253; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
254; GFX10-DL-NEXT:    s_and_b32 s6, s2, s5
255; GFX10-DL-NEXT:    s_and_b32 s5, s3, s5
256; GFX10-DL-NEXT:    s_lshr_b32 s2, s2, 16
257; GFX10-DL-NEXT:    v_mul_u32_u24_e64 v0, s5, s6
258; GFX10-DL-NEXT:    s_lshr_b32 s3, s3, 16
259; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s2, v0
260; GFX10-DL-NEXT:    v_add_nc_u32_e32 v0, s4, v0
261; GFX10-DL-NEXT:    global_store_dword v1, v0, s[0:1]
262; GFX10-DL-NEXT:    s_endpgm
263                                        <2 x i16> addrspace(1)* %src2,
264                                        i32 addrspace(1)* nocapture %dst) {
265entry:
266  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
267  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
268
269  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
270  %conv = zext i16 %s1.elt1 to i32
271  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
272  %conv2 = zext i16 %s2.elt1 to i32
273  %mul1 = mul nuw i32 %conv2, %conv
274
275  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
276  %conv3 = zext i16 %s1.elt2 to i32
277  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
278  %conv4 = zext i16 %s2.elt2 to i32
279  %mul2 = mul nuw i32 %conv4, %conv3
280  %s3 = load i32, i32 addrspace(1)* %dst, align 4
281  %add = add i32 %mul2, %mul1
282  %add6 = add i32 %add, %s3
283  store i32 %add6, i32 addrspace(1)* %dst, align 4
284  ret void
285}
286
287define amdgpu_kernel void @idot2(<2 x i16> addrspace(1)* %src1,
288; GFX7-LABEL: idot2:
289; GFX7:       ; %bb.0: ; %entry
290; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
291; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
292; GFX7-NEXT:    s_mov_b32 s3, 0xf000
293; GFX7-NEXT:    s_mov_b32 s2, -1
294; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
295; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
296; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
297; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
298; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
299; GFX7-NEXT:    s_sext_i32_i16 s7, s4
300; GFX7-NEXT:    s_ashr_i32 s4, s4, 16
301; GFX7-NEXT:    s_sext_i32_i16 s8, s5
302; GFX7-NEXT:    s_ashr_i32 s5, s5, 16
303; GFX7-NEXT:    v_mov_b32_e32 v0, s4
304; GFX7-NEXT:    v_mov_b32_e32 v1, s6
305; GFX7-NEXT:    v_mad_i32_i24 v0, s5, v0, v1
306; GFX7-NEXT:    v_mov_b32_e32 v1, s7
307; GFX7-NEXT:    v_mad_i32_i24 v0, s8, v1, v0
308; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
309; GFX7-NEXT:    s_endpgm
310;
311; GFX8-LABEL: idot2:
312; GFX8:       ; %bb.0: ; %entry
313; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
314; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
315; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
316; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
317; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
318; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
319; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
320; GFX8-NEXT:    s_sext_i32_i16 s5, s2
321; GFX8-NEXT:    s_ashr_i32 s2, s2, 16
322; GFX8-NEXT:    s_sext_i32_i16 s6, s3
323; GFX8-NEXT:    s_ashr_i32 s3, s3, 16
324; GFX8-NEXT:    v_mov_b32_e32 v0, s4
325; GFX8-NEXT:    v_mov_b32_e32 v1, s2
326; GFX8-NEXT:    v_mov_b32_e32 v2, s5
327; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v1, v0
328; GFX8-NEXT:    v_mad_i32_i24 v2, s6, v2, v0
329; GFX8-NEXT:    v_mov_b32_e32 v0, s0
330; GFX8-NEXT:    v_mov_b32_e32 v1, s1
331; GFX8-NEXT:    flat_store_dword v[0:1], v2
332; GFX8-NEXT:    s_endpgm
333;
334; GFX9-NODL-LABEL: idot2:
335; GFX9-NODL:       ; %bb.0: ; %entry
336; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
337; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
338; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
339; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
340; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
341; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
342; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
343; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
344; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
345; GFX9-NODL-NEXT:    s_ashr_i32 s2, s2, 16
346; GFX9-NODL-NEXT:    s_sext_i32_i16 s6, s3
347; GFX9-NODL-NEXT:    s_ashr_i32 s3, s3, 16
348; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
349; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
350; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
351; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
352; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
353; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
354; GFX9-NODL-NEXT:    s_endpgm
355;
356; GFX9-DL-LABEL: idot2:
357; GFX9-DL:       ; %bb.0: ; %entry
358; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
359; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
360; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
361; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
362; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
363; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
364; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
365; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
366; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s2
367; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
368; GFX9-DL-NEXT:    v_dot2_i32_i16 v1, s4, v1, v2
369; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
370; GFX9-DL-NEXT:    s_endpgm
371;
372; GFX10-DL-LABEL: idot2:
373; GFX10-DL:       ; %bb.0: ; %entry
374; GFX10-DL-NEXT:    s_clause 0x1
375; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
376; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
377; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
378; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
379; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
380; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
381; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
382; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
383; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
384; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
385; GFX10-DL-NEXT:    v_dot2_i32_i16 v0, s1, s0, v0
386; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
387; GFX10-DL-NEXT:    s_endpgm
388                                 <2 x i16> addrspace(1)* %src2,
389                                 i32 addrspace(1)* nocapture %dst) {
390entry:
391  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
392  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
393
394  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
395  %conv = sext i16 %s1.elt1 to i32
396  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
397  %conv2 = sext i16 %s2.elt1 to i32
398  %mul1 = mul nuw i32 %conv2, %conv
399
400  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
401  %conv3 = sext i16 %s1.elt2 to i32
402  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
403  %conv4 = sext i16 %s2.elt2 to i32
404  %mul2 = mul nuw i32 %conv4, %conv3
405
406  %s3 = load i32, i32 addrspace(1)* %dst, align 4
407  %add = add i32 %mul2, %s3
408  %add6 = add i32 %add, %mul1
409  store i32 %add6, i32 addrspace(1)* %dst, align 4
410  ret void
411}
412
413define amdgpu_kernel void @idot2_MixedTypedMul(<2 x i16> addrspace(1)* %src1,
414; GFX7-LABEL: idot2_MixedTypedMul:
415; GFX7:       ; %bb.0: ; %entry
416; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
417; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
418; GFX7-NEXT:    s_mov_b32 s3, 0xf000
419; GFX7-NEXT:    s_mov_b32 s2, -1
420; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
421; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
422; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
423; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
424; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
425; GFX7-NEXT:    s_lshr_b32 s7, s4, 16
426; GFX7-NEXT:    s_lshr_b32 s8, s5, 16
427; GFX7-NEXT:    s_sext_i32_i16 s4, s4
428; GFX7-NEXT:    v_mov_b32_e32 v0, s7
429; GFX7-NEXT:    v_mov_b32_e32 v1, s6
430; GFX7-NEXT:    v_mad_u32_u24 v0, s8, v0, v1
431; GFX7-NEXT:    s_sext_i32_i16 s5, s5
432; GFX7-NEXT:    v_mov_b32_e32 v1, s4
433; GFX7-NEXT:    v_mad_i32_i24 v0, s5, v1, v0
434; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
435; GFX7-NEXT:    s_endpgm
436;
437; GFX8-LABEL: idot2_MixedTypedMul:
438; GFX8:       ; %bb.0: ; %entry
439; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
440; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
441; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
442; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
443; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
444; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
445; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
446; GFX8-NEXT:    s_sext_i32_i16 s5, s2
447; GFX8-NEXT:    s_lshr_b32 s2, s2, 16
448; GFX8-NEXT:    s_sext_i32_i16 s6, s3
449; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
450; GFX8-NEXT:    v_mov_b32_e32 v0, s4
451; GFX8-NEXT:    v_mov_b32_e32 v1, s2
452; GFX8-NEXT:    v_mov_b32_e32 v2, s5
453; GFX8-NEXT:    v_mad_u32_u24 v0, s3, v1, v0
454; GFX8-NEXT:    v_mad_i32_i24 v2, s6, v2, v0
455; GFX8-NEXT:    v_mov_b32_e32 v0, s0
456; GFX8-NEXT:    v_mov_b32_e32 v1, s1
457; GFX8-NEXT:    flat_store_dword v[0:1], v2
458; GFX8-NEXT:    s_endpgm
459;
460; GFX9-NODL-LABEL: idot2_MixedTypedMul:
461; GFX9-NODL:       ; %bb.0: ; %entry
462; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
463; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
464; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
465; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
466; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
467; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
468; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
469; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
470; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
471; GFX9-NODL-NEXT:    s_lshr_b32 s2, s2, 16
472; GFX9-NODL-NEXT:    s_sext_i32_i16 s6, s3
473; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
474; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
475; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
476; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s3, v2, v1
477; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
478; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
479; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
480; GFX9-NODL-NEXT:    s_endpgm
481;
482; GFX9-DL-LABEL: idot2_MixedTypedMul:
483; GFX9-DL:       ; %bb.0: ; %entry
484; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
485; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
486; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
487; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
488; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
489; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
490; GFX9-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
491; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
492; GFX9-DL-NEXT:    s_sext_i32_i16 s5, s2
493; GFX9-DL-NEXT:    s_lshr_b32 s2, s2, 16
494; GFX9-DL-NEXT:    s_sext_i32_i16 s6, s3
495; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
496; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s4
497; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
498; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s3, v2, v1
499; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
500; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
501; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
502; GFX9-DL-NEXT:    s_endpgm
503;
504; GFX10-DL-LABEL: idot2_MixedTypedMul:
505; GFX10-DL:       ; %bb.0: ; %entry
506; GFX10-DL-NEXT:    s_clause 0x1
507; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
508; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
509; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
510; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
511; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
512; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
513; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
514; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
515; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
516; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
517; GFX10-DL-NEXT:    s_lshr_b32 s2, s0, 16
518; GFX10-DL-NEXT:    s_lshr_b32 s3, s1, 16
519; GFX10-DL-NEXT:    s_sext_i32_i16 s0, s0
520; GFX10-DL-NEXT:    s_sext_i32_i16 s1, s1
521; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s2, v0
522; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s1, s0, v0
523; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
524; GFX10-DL-NEXT:    s_endpgm
525                                               <2 x i16> addrspace(1)* %src2,
526                                               i32 addrspace(1)* nocapture %dst) {
527entry:
528  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
529  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
530
531  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
532  %conv = sext i16 %s1.elt1 to i32
533  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
534  %conv2 = sext i16 %s2.elt1 to i32
535  %mul1 = mul nuw i32 %conv2, %conv
536
537  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
538  %conv3 = zext i16 %s1.elt2 to i32
539  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
540  %conv4 = zext i16 %s2.elt2 to i32
541  %mul2 = mul nuw i32 %conv4, %conv3
542
543  %s3 = load i32, i32 addrspace(1)* %dst, align 4
544  %add = add i32 %mul2, %s3
545  %add6 = add i32 %add, %mul1
546  store i32 %add6, i32 addrspace(1)* %dst, align 4
547  ret void
548}
549
550define amdgpu_kernel void @udot2_alt_AddOperands(<2 x i16> addrspace(1)* %src1,
551; GFX7-LABEL: udot2_alt_AddOperands:
552; GFX7:       ; %bb.0: ; %entry
553; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
554; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
555; GFX7-NEXT:    s_mov_b32 s8, 0xffff
556; GFX7-NEXT:    s_mov_b32 s3, 0xf000
557; GFX7-NEXT:    s_mov_b32 s2, -1
558; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
559; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
560; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
561; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
562; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
563; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
564; GFX7-NEXT:    s_and_b32 s4, s4, s8
565; GFX7-NEXT:    s_and_b32 s5, s5, s8
566; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
567; GFX7-NEXT:    v_mov_b32_e32 v0, s6
568; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
569; GFX7-NEXT:    v_mov_b32_e32 v1, s8
570; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
571; GFX7-NEXT:    v_mov_b32_e32 v1, s4
572; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v1, v0
573; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
574; GFX7-NEXT:    s_endpgm
575;
576; GFX8-LABEL: udot2_alt_AddOperands:
577; GFX8:       ; %bb.0: ; %entry
578; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
579; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
580; GFX8-NEXT:    s_mov_b32 s2, 0xffff
581; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
582; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
583; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
584; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
585; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
586; GFX8-NEXT:    s_and_b32 s6, s3, s2
587; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
588; GFX8-NEXT:    s_and_b32 s2, s4, s2
589; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
590; GFX8-NEXT:    v_mov_b32_e32 v0, s5
591; GFX8-NEXT:    v_mov_b32_e32 v1, s3
592; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
593; GFX8-NEXT:    v_mov_b32_e32 v1, s6
594; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
595; GFX8-NEXT:    v_mov_b32_e32 v0, s0
596; GFX8-NEXT:    v_mov_b32_e32 v1, s1
597; GFX8-NEXT:    flat_store_dword v[0:1], v2
598; GFX8-NEXT:    s_endpgm
599;
600; GFX9-NODL-LABEL: udot2_alt_AddOperands:
601; GFX9-NODL:       ; %bb.0: ; %entry
602; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
603; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
604; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
605; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
606; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
607; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
608; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
609; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
610; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
611; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
612; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
613; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
614; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
615; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
616; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
617; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
618; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
619; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
620; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
621; GFX9-NODL-NEXT:    s_endpgm
622;
623; GFX9-DL-LABEL: udot2_alt_AddOperands:
624; GFX9-DL:       ; %bb.0: ; %entry
625; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
626; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
627; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
628; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
629; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
630; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
631; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
632; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
633; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s2
634; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
635; GFX9-DL-NEXT:    v_dot2_u32_u16 v1, s4, v1, v2
636; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
637; GFX9-DL-NEXT:    s_endpgm
638;
639; GFX10-DL-LABEL: udot2_alt_AddOperands:
640; GFX10-DL:       ; %bb.0: ; %entry
641; GFX10-DL-NEXT:    s_clause 0x1
642; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
643; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
644; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
645; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
646; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
647; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
648; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
649; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
650; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
651; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
652; GFX10-DL-NEXT:    v_dot2_u32_u16 v0, s1, s0, v0
653; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
654; GFX10-DL-NEXT:    s_endpgm
655                                                 <2 x i16> addrspace(1)* %src2,
656                                                 i32 addrspace(1)* nocapture %dst) {
657entry:
658  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
659  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
660
661  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
662  %conv = zext i16 %s1.elt1 to i32
663  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
664  %conv2 = zext i16 %s2.elt1 to i32
665  %mul1 = mul nuw i32 %conv2, %conv
666
667  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
668  %conv3 = zext i16 %s1.elt2 to i32
669  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
670  %conv4 = zext i16 %s2.elt2 to i32
671  %mul2 = mul nuw i32 %conv4, %conv3
672
673  %s3 = load i32, i32 addrspace(1)* %dst, align 4
674  %add = add i32 %s3, %mul2
675  %add6 = add i32 %mul1, %add
676  store i32 %add6, i32 addrspace(1)* %dst, align 4
677  ret void
678}
679
680define amdgpu_kernel void @idot2_MixedExt(<2 x i16> addrspace(1)* %src1,
681; GFX7-LABEL: idot2_MixedExt:
682; GFX7:       ; %bb.0: ; %entry
683; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
684; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
685; GFX7-NEXT:    s_mov_b32 s3, 0xf000
686; GFX7-NEXT:    s_mov_b32 s2, -1
687; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
688; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
689; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
690; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
691; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
692; GFX7-NEXT:    s_sext_i32_i16 s7, s4
693; GFX7-NEXT:    s_ashr_i32 s4, s4, 16
694; GFX7-NEXT:    s_and_b32 s8, s5, 0xffff
695; GFX7-NEXT:    s_ashr_i32 s5, s5, 16
696; GFX7-NEXT:    v_mov_b32_e32 v0, s4
697; GFX7-NEXT:    v_mov_b32_e32 v1, s6
698; GFX7-NEXT:    v_mad_i32_i24 v0, s5, v0, v1
699; GFX7-NEXT:    v_mov_b32_e32 v1, s7
700; GFX7-NEXT:    v_mad_i32_i24 v0, s8, v1, v0
701; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
702; GFX7-NEXT:    s_endpgm
703;
704; GFX8-LABEL: idot2_MixedExt:
705; GFX8:       ; %bb.0: ; %entry
706; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
707; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
708; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
709; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
710; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
711; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
712; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
713; GFX8-NEXT:    s_sext_i32_i16 s5, s2
714; GFX8-NEXT:    s_ashr_i32 s2, s2, 16
715; GFX8-NEXT:    s_and_b32 s6, s3, 0xffff
716; GFX8-NEXT:    s_ashr_i32 s3, s3, 16
717; GFX8-NEXT:    v_mov_b32_e32 v0, s4
718; GFX8-NEXT:    v_mov_b32_e32 v1, s2
719; GFX8-NEXT:    v_mov_b32_e32 v2, s5
720; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v1, v0
721; GFX8-NEXT:    v_mad_i32_i24 v2, s6, v2, v0
722; GFX8-NEXT:    v_mov_b32_e32 v0, s0
723; GFX8-NEXT:    v_mov_b32_e32 v1, s1
724; GFX8-NEXT:    flat_store_dword v[0:1], v2
725; GFX8-NEXT:    s_endpgm
726;
727; GFX9-NODL-LABEL: idot2_MixedExt:
728; GFX9-NODL:       ; %bb.0: ; %entry
729; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
730; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
731; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
732; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
733; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
734; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
735; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
736; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
737; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
738; GFX9-NODL-NEXT:    s_ashr_i32 s2, s2, 16
739; GFX9-NODL-NEXT:    s_and_b32 s6, s3, 0xffff
740; GFX9-NODL-NEXT:    s_ashr_i32 s3, s3, 16
741; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
742; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
743; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
744; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
745; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
746; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
747; GFX9-NODL-NEXT:    s_endpgm
748;
749; GFX9-DL-LABEL: idot2_MixedExt:
750; GFX9-DL:       ; %bb.0: ; %entry
751; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
752; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
753; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
754; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
755; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
756; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
757; GFX9-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
758; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
759; GFX9-DL-NEXT:    s_sext_i32_i16 s5, s2
760; GFX9-DL-NEXT:    s_ashr_i32 s2, s2, 16
761; GFX9-DL-NEXT:    s_and_b32 s6, s3, 0xffff
762; GFX9-DL-NEXT:    s_ashr_i32 s3, s3, 16
763; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s4
764; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
765; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
766; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
767; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
768; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
769; GFX9-DL-NEXT:    s_endpgm
770;
771; GFX10-DL-LABEL: idot2_MixedExt:
772; GFX10-DL:       ; %bb.0: ; %entry
773; GFX10-DL-NEXT:    s_clause 0x1
774; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
775; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
776; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
777; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
778; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
779; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
780; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
781; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
782; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
783; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
784; GFX10-DL-NEXT:    s_ashr_i32 s2, s0, 16
785; GFX10-DL-NEXT:    s_ashr_i32 s3, s1, 16
786; GFX10-DL-NEXT:    s_sext_i32_i16 s0, s0
787; GFX10-DL-NEXT:    s_and_b32 s1, s1, 0xffff
788; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
789; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s1, s0, v0
790; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
791; GFX10-DL-NEXT:    s_endpgm
792                                          <2 x i16> addrspace(1)* %src2,
793                                          i32 addrspace(1)* nocapture %dst) {
794entry:
795  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
796  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
797
798  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
799  %conv = sext i16 %s1.elt1 to i32
800  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
801  %conv2 = zext i16 %s2.elt1 to i32
802  %mul1 = mul nuw i32 %conv2, %conv
803
804  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
805  %conv3 = sext i16 %s1.elt2 to i32
806  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
807  %conv4 = sext i16 %s2.elt2 to i32
808  %mul2 = mul nuw i32 %conv4, %conv3
809
810  %s3 = load i32, i32 addrspace(1)* %dst, align 4
811  %add = add i32 %mul2, %s3
812  %add6 = add i32 %add, %mul1
813  store i32 %add6, i32 addrspace(1)* %dst, align 4
814  ret void
815}
816
817define amdgpu_kernel void @notudot2_SameVec(<2 x i16> addrspace(1)* %src1,
818; GFX7-LABEL: notudot2_SameVec:
819; GFX7:       ; %bb.0: ; %entry
820; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
821; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
822; GFX7-NEXT:    s_mov_b32 s3, 0xf000
823; GFX7-NEXT:    s_mov_b32 s2, -1
824; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
825; GFX7-NEXT:    s_load_dword s6, s[6:7], 0x0
826; GFX7-NEXT:    s_load_dword s7, s[0:1], 0x0
827; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
828; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
829; GFX7-NEXT:    s_lshr_b32 s5, s6, 16
830; GFX7-NEXT:    v_mov_b32_e32 v0, s7
831; GFX7-NEXT:    s_and_b32 s4, s4, 0xffff
832; GFX7-NEXT:    v_mad_u32_u24 v0, s5, s5, v0
833; GFX7-NEXT:    v_mad_u32_u24 v0, s4, s4, v0
834; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
835; GFX7-NEXT:    s_endpgm
836;
837; GFX8-LABEL: notudot2_SameVec:
838; GFX8:       ; %bb.0: ; %entry
839; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
840; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
841; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
842; GFX8-NEXT:    s_load_dword s2, s[6:7], 0x0
843; GFX8-NEXT:    s_load_dword s3, s[0:1], 0x0
844; GFX8-NEXT:    s_load_dword s4, s[4:5], 0x0
845; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
846; GFX8-NEXT:    s_lshr_b32 s2, s2, 16
847; GFX8-NEXT:    v_mov_b32_e32 v0, s3
848; GFX8-NEXT:    s_and_b32 s4, s4, 0xffff
849; GFX8-NEXT:    v_mad_u32_u24 v0, s2, s2, v0
850; GFX8-NEXT:    v_mad_u32_u24 v2, s4, s4, v0
851; GFX8-NEXT:    v_mov_b32_e32 v0, s0
852; GFX8-NEXT:    v_mov_b32_e32 v1, s1
853; GFX8-NEXT:    flat_store_dword v[0:1], v2
854; GFX8-NEXT:    s_endpgm
855;
856; GFX9-NODL-LABEL: notudot2_SameVec:
857; GFX9-NODL:       ; %bb.0: ; %entry
858; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
859; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
860; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
861; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
862; GFX9-NODL-NEXT:    s_load_dword s2, s[6:7], 0x0
863; GFX9-NODL-NEXT:    s_load_dword s3, s[0:1], 0x0
864; GFX9-NODL-NEXT:    s_load_dword s4, s[4:5], 0x0
865; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
866; GFX9-NODL-NEXT:    s_lshr_b32 s2, s2, 16
867; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
868; GFX9-NODL-NEXT:    s_and_b32 s4, s4, 0xffff
869; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, s2, v1
870; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, s4, v1
871; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
872; GFX9-NODL-NEXT:    s_endpgm
873;
874; GFX9-DL-LABEL: notudot2_SameVec:
875; GFX9-DL:       ; %bb.0: ; %entry
876; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
877; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
878; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
879; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
880; GFX9-DL-NEXT:    s_load_dword s2, s[6:7], 0x0
881; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
882; GFX9-DL-NEXT:    s_load_dword s4, s[4:5], 0x0
883; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
884; GFX9-DL-NEXT:    s_lshr_b32 s2, s2, 16
885; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
886; GFX9-DL-NEXT:    s_and_b32 s4, s4, 0xffff
887; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s2, s2, v1
888; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, s4, v1
889; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
890; GFX9-DL-NEXT:    s_endpgm
891;
892; GFX10-DL-LABEL: notudot2_SameVec:
893; GFX10-DL:       ; %bb.0: ; %entry
894; GFX10-DL-NEXT:    s_clause 0x1
895; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
896; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
897; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
898; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
899; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
900; GFX10-DL-NEXT:    s_load_dword s2, s[6:7], 0x0
901; GFX10-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
902; GFX10-DL-NEXT:    s_load_dword s4, s[4:5], 0x0
903; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
904; GFX10-DL-NEXT:    s_lshr_b32 s2, s2, 16
905; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s2, s2, s3
906; GFX10-DL-NEXT:    s_and_b32 s2, s4, 0xffff
907; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s2, s2, v0
908; GFX10-DL-NEXT:    global_store_dword v1, v0, s[0:1]
909; GFX10-DL-NEXT:    s_endpgm
910                                            <2 x i16> addrspace(1)* %src2,
911                                            i32 addrspace(1)* nocapture %dst) {
912entry:
913  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
914  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
915
916  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
917  %conv = zext i16 %s1.elt1 to i32
918  %s2.elt1 = extractelement <2 x i16> %vec1, i64 0
919  %conv2 = zext i16 %s2.elt1 to i32
920  %mul1 = mul i32 %conv2, %conv
921
922  %s1.elt2 = extractelement <2 x i16> %vec2, i64 1
923  %conv3 = zext i16 %s1.elt2 to i32
924  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
925  %conv4 = zext i16 %s2.elt2 to i32
926  %mul2 = mul i32 %conv4, %conv3
927
928  %s3 = load i32, i32 addrspace(1)* %dst, align 4
929  %add = add i32 %mul2, %s3
930  %add6 = add i32 %add, %mul1
931  store i32 %add6, i32 addrspace(1)* %dst, align 4
932  ret void
933}
934
935define amdgpu_kernel void @udot2_v4i16(<4 x i16> addrspace(1)* %src1,
936; GFX7-LABEL: udot2_v4i16:
937; GFX7:       ; %bb.0: ; %entry
938; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
939; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
940; GFX7-NEXT:    s_mov_b32 s8, 0xffff
941; GFX7-NEXT:    s_mov_b32 s3, 0xf000
942; GFX7-NEXT:    s_mov_b32 s2, -1
943; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
944; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
945; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
946; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
947; GFX7-NEXT:    s_and_b32 s6, s4, s8
948; GFX7-NEXT:    s_and_b32 s7, s5, s8
949; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
950; GFX7-NEXT:    s_lshr_b32 s4, s4, 16
951; GFX7-NEXT:    s_lshr_b32 s5, s5, 16
952; GFX7-NEXT:    v_mov_b32_e32 v0, s4
953; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
954; GFX7-NEXT:    v_mov_b32_e32 v1, s8
955; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v0, v1
956; GFX7-NEXT:    v_mov_b32_e32 v1, s6
957; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v1, v0
958; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
959; GFX7-NEXT:    s_endpgm
960;
961; GFX8-LABEL: udot2_v4i16:
962; GFX8:       ; %bb.0: ; %entry
963; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
964; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
965; GFX8-NEXT:    s_mov_b32 s2, 0xffff
966; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
967; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
968; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
969; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
970; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
971; GFX8-NEXT:    s_and_b32 s6, s3, s2
972; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
973; GFX8-NEXT:    s_and_b32 s2, s4, s2
974; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
975; GFX8-NEXT:    v_mov_b32_e32 v0, s5
976; GFX8-NEXT:    v_mov_b32_e32 v1, s3
977; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
978; GFX8-NEXT:    v_mov_b32_e32 v1, s6
979; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
980; GFX8-NEXT:    v_mov_b32_e32 v0, s0
981; GFX8-NEXT:    v_mov_b32_e32 v1, s1
982; GFX8-NEXT:    flat_store_dword v[0:1], v2
983; GFX8-NEXT:    s_endpgm
984;
985; GFX9-NODL-LABEL: udot2_v4i16:
986; GFX9-NODL:       ; %bb.0: ; %entry
987; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
988; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
989; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
990; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
991; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
992; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
993; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
994; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
995; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
996; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
997; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
998; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
999; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
1000; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1001; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1002; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
1003; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1004; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
1005; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1006; GFX9-NODL-NEXT:    s_endpgm
1007;
1008; GFX9-DL-LABEL: udot2_v4i16:
1009; GFX9-DL:       ; %bb.0: ; %entry
1010; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1011; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1012; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1013; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1014; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
1015; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
1016; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
1017; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1018; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s2
1019; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
1020; GFX9-DL-NEXT:    v_dot2_u32_u16 v1, s4, v1, v2
1021; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1022; GFX9-DL-NEXT:    s_endpgm
1023;
1024; GFX10-DL-LABEL: udot2_v4i16:
1025; GFX10-DL:       ; %bb.0: ; %entry
1026; GFX10-DL-NEXT:    s_clause 0x1
1027; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1028; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1029; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
1030; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1031; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1032; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1033; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
1034; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
1035; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1036; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1037; GFX10-DL-NEXT:    v_dot2_u32_u16 v0, s1, s0, v0
1038; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
1039; GFX10-DL-NEXT:    s_endpgm
1040                                       <4 x i16> addrspace(1)* %src2,
1041                                       i32 addrspace(1)* nocapture %dst) {
1042entry:
1043  %vec1 = load <4 x i16>, <4 x i16> addrspace(1)* %src1
1044  %vec2 = load <4 x i16>, <4 x i16> addrspace(1)* %src2
1045
1046  %s1.elt1 = extractelement <4 x i16> %vec1, i64 0
1047  %conv = zext i16 %s1.elt1 to i32
1048  %s2.elt1 = extractelement <4 x i16> %vec2, i64 0
1049  %conv2 = zext i16 %s2.elt1 to i32
1050  %mul1 = mul i32 %conv2, %conv
1051
1052  %s1.elt2 = extractelement <4 x i16> %vec1, i64 1
1053  %conv3 = zext i16 %s1.elt2 to i32
1054  %s2.elt2 = extractelement <4 x i16> %vec2, i64 1
1055  %conv4 = zext i16 %s2.elt2 to i32
1056  %mul2 = mul i32 %conv4, %conv3
1057
1058  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1059  %add = add i32 %mul2, %s3
1060  %add6 = add i32 %add, %mul1
1061  store i32 %add6, i32 addrspace(1)* %dst, align 4
1062  ret void
1063}
1064
1065define amdgpu_kernel void @udot2_v4i16_Hi(<4 x i16> addrspace(1)* %src1,
1066; GFX7-LABEL: udot2_v4i16_Hi:
1067; GFX7:       ; %bb.0: ; %entry
1068; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1069; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1070; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1071; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1072; GFX7-NEXT:    s_mov_b32 s2, -1
1073; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1074; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x1
1075; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x1
1076; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1077; GFX7-NEXT:    s_and_b32 s6, s4, s8
1078; GFX7-NEXT:    s_and_b32 s7, s5, s8
1079; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1080; GFX7-NEXT:    s_lshr_b32 s4, s4, 16
1081; GFX7-NEXT:    s_lshr_b32 s5, s5, 16
1082; GFX7-NEXT:    v_mov_b32_e32 v0, s4
1083; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1084; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1085; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v0, v1
1086; GFX7-NEXT:    v_mov_b32_e32 v1, s6
1087; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v1, v0
1088; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1089; GFX7-NEXT:    s_endpgm
1090;
1091; GFX8-LABEL: udot2_v4i16_Hi:
1092; GFX8:       ; %bb.0: ; %entry
1093; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1094; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1095; GFX8-NEXT:    s_mov_b32 s2, 0xffff
1096; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1097; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x4
1098; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x4
1099; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
1100; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1101; GFX8-NEXT:    s_and_b32 s6, s3, s2
1102; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
1103; GFX8-NEXT:    s_and_b32 s2, s4, s2
1104; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
1105; GFX8-NEXT:    v_mov_b32_e32 v0, s5
1106; GFX8-NEXT:    v_mov_b32_e32 v1, s3
1107; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
1108; GFX8-NEXT:    v_mov_b32_e32 v1, s6
1109; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
1110; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1111; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1112; GFX8-NEXT:    flat_store_dword v[0:1], v2
1113; GFX8-NEXT:    s_endpgm
1114;
1115; GFX9-NODL-LABEL: udot2_v4i16_Hi:
1116; GFX9-NODL:       ; %bb.0: ; %entry
1117; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1118; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1119; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
1120; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1121; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1122; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x4
1123; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x4
1124; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
1125; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1126; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
1127; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
1128; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
1129; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
1130; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1131; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1132; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
1133; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1134; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
1135; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1136; GFX9-NODL-NEXT:    s_endpgm
1137;
1138; GFX9-DL-LABEL: udot2_v4i16_Hi:
1139; GFX9-DL:       ; %bb.0: ; %entry
1140; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1141; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1142; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1143; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1144; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x4
1145; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
1146; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x4
1147; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1148; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s2
1149; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
1150; GFX9-DL-NEXT:    v_dot2_u32_u16 v1, s4, v1, v2
1151; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1152; GFX9-DL-NEXT:    s_endpgm
1153;
1154; GFX10-DL-LABEL: udot2_v4i16_Hi:
1155; GFX10-DL:       ; %bb.0: ; %entry
1156; GFX10-DL-NEXT:    s_clause 0x1
1157; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1158; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1159; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
1160; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1161; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1162; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1163; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x4
1164; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x4
1165; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1166; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1167; GFX10-DL-NEXT:    v_dot2_u32_u16 v0, s1, s0, v0
1168; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
1169; GFX10-DL-NEXT:    s_endpgm
1170                                          <4 x i16> addrspace(1)* %src2,
1171                                          i32 addrspace(1)* nocapture %dst) {
1172entry:
1173  %vec1 = load <4 x i16>, <4 x i16> addrspace(1)* %src1
1174  %vec2 = load <4 x i16>, <4 x i16> addrspace(1)* %src2
1175
1176  %s1.elt1 = extractelement <4 x i16> %vec1, i64 2
1177  %conv = zext i16 %s1.elt1 to i32
1178  %s2.elt1 = extractelement <4 x i16> %vec2, i64 2
1179  %conv2 = zext i16 %s2.elt1 to i32
1180  %mul1 = mul i32 %conv2, %conv
1181
1182  %s1.elt2 = extractelement <4 x i16> %vec1, i64 3
1183  %conv3 = zext i16 %s1.elt2 to i32
1184  %s2.elt2 = extractelement <4 x i16> %vec2, i64 3
1185  %conv4 = zext i16 %s2.elt2 to i32
1186  %mul2 = mul i32 %conv4, %conv3
1187
1188  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1189  %add = add i32 %mul2, %s3
1190  %add6 = add i32 %add, %mul1
1191  store i32 %add6, i32 addrspace(1)* %dst, align 4
1192  ret void
1193}
1194
1195define amdgpu_kernel void @notudot2_v4i16_Even(<4 x i16> addrspace(1)* %src1,
1196; GFX7-LABEL: notudot2_v4i16_Even:
1197; GFX7:       ; %bb.0: ; %entry
1198; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1199; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1200; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1201; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1202; GFX7-NEXT:    s_mov_b32 s2, -1
1203; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1204; GFX7-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
1205; GFX7-NEXT:    s_load_dwordx2 s[6:7], s[6:7], 0x0
1206; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1207; GFX7-NEXT:    s_and_b32 s5, s5, s8
1208; GFX7-NEXT:    s_and_b32 s4, s4, s8
1209; GFX7-NEXT:    s_and_b32 s6, s6, s8
1210; GFX7-NEXT:    s_and_b32 s7, s7, s8
1211; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1212; GFX7-NEXT:    v_mov_b32_e32 v0, s5
1213; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1214; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1215; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
1216; GFX7-NEXT:    v_mov_b32_e32 v1, s4
1217; GFX7-NEXT:    v_mad_u32_u24 v0, s6, v1, v0
1218; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1219; GFX7-NEXT:    s_endpgm
1220;
1221; GFX8-LABEL: notudot2_v4i16_Even:
1222; GFX8:       ; %bb.0: ; %entry
1223; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1224; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1225; GFX8-NEXT:    s_mov_b32 s8, 0xffff
1226; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1227; GFX8-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1228; GFX8-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1229; GFX8-NEXT:    s_load_dword s6, s[0:1], 0x0
1230; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1231; GFX8-NEXT:    s_and_b32 s3, s3, s8
1232; GFX8-NEXT:    s_and_b32 s2, s2, s8
1233; GFX8-NEXT:    s_and_b32 s5, s5, s8
1234; GFX8-NEXT:    v_mov_b32_e32 v0, s6
1235; GFX8-NEXT:    v_mov_b32_e32 v1, s3
1236; GFX8-NEXT:    v_mad_u32_u24 v0, s5, v1, v0
1237; GFX8-NEXT:    s_and_b32 s4, s4, s8
1238; GFX8-NEXT:    v_mov_b32_e32 v1, s2
1239; GFX8-NEXT:    v_mad_u32_u24 v2, s4, v1, v0
1240; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1241; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1242; GFX8-NEXT:    flat_store_dword v[0:1], v2
1243; GFX8-NEXT:    s_endpgm
1244;
1245; GFX9-NODL-LABEL: notudot2_v4i16_Even:
1246; GFX9-NODL:       ; %bb.0: ; %entry
1247; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1248; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1249; GFX9-NODL-NEXT:    s_mov_b32 s8, 0xffff
1250; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1251; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1252; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1253; GFX9-NODL-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1254; GFX9-NODL-NEXT:    s_load_dword s6, s[0:1], 0x0
1255; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1256; GFX9-NODL-NEXT:    s_and_b32 s3, s3, s8
1257; GFX9-NODL-NEXT:    s_and_b32 s2, s2, s8
1258; GFX9-NODL-NEXT:    s_and_b32 s5, s5, s8
1259; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1260; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1261; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s5, v1, v2
1262; GFX9-NODL-NEXT:    s_and_b32 s4, s4, s8
1263; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
1264; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
1265; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1266; GFX9-NODL-NEXT:    s_endpgm
1267;
1268; GFX9-DL-LABEL: notudot2_v4i16_Even:
1269; GFX9-DL:       ; %bb.0: ; %entry
1270; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1271; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1272; GFX9-DL-NEXT:    s_mov_b32 s8, 0xffff
1273; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1274; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1275; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1276; GFX9-DL-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1277; GFX9-DL-NEXT:    s_load_dword s6, s[0:1], 0x0
1278; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1279; GFX9-DL-NEXT:    s_and_b32 s3, s3, s8
1280; GFX9-DL-NEXT:    s_and_b32 s2, s2, s8
1281; GFX9-DL-NEXT:    s_and_b32 s5, s5, s8
1282; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
1283; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s6
1284; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s5, v1, v2
1285; GFX9-DL-NEXT:    s_and_b32 s4, s4, s8
1286; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
1287; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
1288; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1289; GFX9-DL-NEXT:    s_endpgm
1290;
1291; GFX10-DL-LABEL: notudot2_v4i16_Even:
1292; GFX10-DL:       ; %bb.0: ; %entry
1293; GFX10-DL-NEXT:    s_clause 0x1
1294; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1295; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1296; GFX10-DL-NEXT:    s_mov_b32 s7, 0xffff
1297; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
1298; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1299; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1300; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1301; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1302; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[2:3], 0x0
1303; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1304; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1305; GFX10-DL-NEXT:    s_and_b32 s1, s1, s7
1306; GFX10-DL-NEXT:    s_and_b32 s3, s3, s7
1307; GFX10-DL-NEXT:    s_and_b32 s0, s0, s7
1308; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s1, v0
1309; GFX10-DL-NEXT:    s_and_b32 s1, s2, s7
1310; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s1, s0, v0
1311; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
1312; GFX10-DL-NEXT:    s_endpgm
1313                                               <4 x i16> addrspace(1)* %src2,
1314                                               i32 addrspace(1)* nocapture %dst) {
1315entry:
1316  %vec1 = load <4 x i16>, <4 x i16> addrspace(1)* %src1
1317  %vec2 = load <4 x i16>, <4 x i16> addrspace(1)* %src2
1318
1319  %s1.elt1 = extractelement <4 x i16> %vec1, i64 0
1320  %conv = zext i16 %s1.elt1 to i32
1321  %s2.elt1 = extractelement <4 x i16> %vec2, i64 0
1322  %conv2 = zext i16 %s2.elt1 to i32
1323  %mul1 = mul i32 %conv2, %conv
1324
1325  %s1.elt2 = extractelement <4 x i16> %vec1, i64 2
1326  %conv3 = zext i16 %s1.elt2 to i32
1327  %s2.elt2 = extractelement <4 x i16> %vec2, i64 2
1328  %conv4 = zext i16 %s2.elt2 to i32
1329  %mul2 = mul i32 %conv4, %conv3
1330
1331  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1332  %add = add i32 %mul2, %s3
1333  %add6 = add i32 %add, %mul1
1334  store i32 %add6, i32 addrspace(1)* %dst, align 4
1335  ret void
1336}
1337
1338define amdgpu_kernel void @notudot2_v4i16_Middle(<4 x i16> addrspace(1)* %src1,
1339; GFX7-LABEL: notudot2_v4i16_Middle:
1340; GFX7:       ; %bb.0: ; %entry
1341; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1342; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1343; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1344; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1345; GFX7-NEXT:    s_mov_b32 s2, -1
1346; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1347; GFX7-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
1348; GFX7-NEXT:    s_load_dwordx2 s[6:7], s[6:7], 0x0
1349; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1350; GFX7-NEXT:    s_and_b32 s5, s5, s8
1351; GFX7-NEXT:    s_and_b32 s7, s7, s8
1352; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1353; GFX7-NEXT:    s_lshr_b32 s4, s4, 16
1354; GFX7-NEXT:    v_mov_b32_e32 v0, s5
1355; GFX7-NEXT:    s_lshr_b32 s6, s6, 16
1356; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1357; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1358; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
1359; GFX7-NEXT:    v_mov_b32_e32 v1, s4
1360; GFX7-NEXT:    v_mad_u32_u24 v0, s6, v1, v0
1361; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1362; GFX7-NEXT:    s_endpgm
1363;
1364; GFX8-LABEL: notudot2_v4i16_Middle:
1365; GFX8:       ; %bb.0: ; %entry
1366; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1367; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1368; GFX8-NEXT:    s_mov_b32 s8, 0xffff
1369; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1370; GFX8-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1371; GFX8-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1372; GFX8-NEXT:    s_load_dword s6, s[0:1], 0x0
1373; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1374; GFX8-NEXT:    s_and_b32 s3, s3, s8
1375; GFX8-NEXT:    s_lshr_b32 s2, s2, 16
1376; GFX8-NEXT:    s_and_b32 s5, s5, s8
1377; GFX8-NEXT:    v_mov_b32_e32 v0, s6
1378; GFX8-NEXT:    v_mov_b32_e32 v1, s3
1379; GFX8-NEXT:    v_mad_u32_u24 v0, s5, v1, v0
1380; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
1381; GFX8-NEXT:    v_mov_b32_e32 v1, s2
1382; GFX8-NEXT:    v_mad_u32_u24 v2, s4, v1, v0
1383; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1384; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1385; GFX8-NEXT:    flat_store_dword v[0:1], v2
1386; GFX8-NEXT:    s_endpgm
1387;
1388; GFX9-NODL-LABEL: notudot2_v4i16_Middle:
1389; GFX9-NODL:       ; %bb.0: ; %entry
1390; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1391; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1392; GFX9-NODL-NEXT:    s_mov_b32 s8, 0xffff
1393; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1394; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1395; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1396; GFX9-NODL-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1397; GFX9-NODL-NEXT:    s_load_dword s6, s[0:1], 0x0
1398; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1399; GFX9-NODL-NEXT:    s_and_b32 s3, s3, s8
1400; GFX9-NODL-NEXT:    s_lshr_b32 s2, s2, 16
1401; GFX9-NODL-NEXT:    s_and_b32 s5, s5, s8
1402; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1403; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1404; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s5, v1, v2
1405; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
1406; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
1407; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
1408; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1409; GFX9-NODL-NEXT:    s_endpgm
1410;
1411; GFX9-DL-LABEL: notudot2_v4i16_Middle:
1412; GFX9-DL:       ; %bb.0: ; %entry
1413; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1414; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1415; GFX9-DL-NEXT:    s_mov_b32 s8, 0xffff
1416; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1417; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1418; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
1419; GFX9-DL-NEXT:    s_load_dwordx2 s[4:5], s[6:7], 0x0
1420; GFX9-DL-NEXT:    s_load_dword s6, s[0:1], 0x0
1421; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1422; GFX9-DL-NEXT:    s_and_b32 s3, s3, s8
1423; GFX9-DL-NEXT:    s_lshr_b32 s2, s2, 16
1424; GFX9-DL-NEXT:    s_and_b32 s5, s5, s8
1425; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
1426; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s6
1427; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s5, v1, v2
1428; GFX9-DL-NEXT:    s_lshr_b32 s4, s4, 16
1429; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
1430; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, v2, v1
1431; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1432; GFX9-DL-NEXT:    s_endpgm
1433;
1434; GFX10-DL-LABEL: notudot2_v4i16_Middle:
1435; GFX10-DL:       ; %bb.0: ; %entry
1436; GFX10-DL-NEXT:    s_clause 0x1
1437; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1438; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1439; GFX10-DL-NEXT:    s_mov_b32 s7, 0xffff
1440; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
1441; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1442; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1443; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1444; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1445; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[2:3], 0x0
1446; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1447; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1448; GFX10-DL-NEXT:    s_and_b32 s1, s1, s7
1449; GFX10-DL-NEXT:    s_and_b32 s3, s3, s7
1450; GFX10-DL-NEXT:    s_lshr_b32 s0, s0, 16
1451; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s1, v0
1452; GFX10-DL-NEXT:    s_lshr_b32 s1, s2, 16
1453; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s1, s0, v0
1454; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
1455; GFX10-DL-NEXT:    s_endpgm
1456                                                 <4 x i16> addrspace(1)* %src2,
1457                                                 i32 addrspace(1)* nocapture %dst) {
1458entry:
1459  %vec1 = load <4 x i16>, <4 x i16> addrspace(1)* %src1
1460  %vec2 = load <4 x i16>, <4 x i16> addrspace(1)* %src2
1461
1462  %s1.elt1 = extractelement <4 x i16> %vec1, i64 1
1463  %conv = zext i16 %s1.elt1 to i32
1464  %s2.elt1 = extractelement <4 x i16> %vec2, i64 1
1465  %conv2 = zext i16 %s2.elt1 to i32
1466  %mul1 = mul i32 %conv2, %conv
1467
1468  %s1.elt2 = extractelement <4 x i16> %vec1, i64 2
1469  %conv3 = zext i16 %s1.elt2 to i32
1470  %s2.elt2 = extractelement <4 x i16> %vec2, i64 2
1471  %conv4 = zext i16 %s2.elt2 to i32
1472  %mul2 = mul i32 %conv4, %conv3
1473
1474  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1475  %add = add i32 %mul2, %s3
1476  %add6 = add i32 %add, %mul1
1477  store i32 %add6, i32 addrspace(1)* %dst, align 4
1478  ret void
1479}
1480
1481define amdgpu_kernel void @notudot2_DiffIndex(<2 x i16> addrspace(1)* %src1,
1482; GFX7-LABEL: notudot2_DiffIndex:
1483; GFX7:       ; %bb.0: ; %entry
1484; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1485; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1486; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1487; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1488; GFX7-NEXT:    s_mov_b32 s2, -1
1489; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1490; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
1491; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
1492; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1493; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
1494; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
1495; GFX7-NEXT:    s_and_b32 s4, s4, s8
1496; GFX7-NEXT:    s_and_b32 s5, s5, s8
1497; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1498; GFX7-NEXT:    v_mov_b32_e32 v0, s6
1499; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1500; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1501; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v0, v1
1502; GFX7-NEXT:    v_mov_b32_e32 v1, s4
1503; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v1, v0
1504; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1505; GFX7-NEXT:    s_endpgm
1506;
1507; GFX8-LABEL: notudot2_DiffIndex:
1508; GFX8:       ; %bb.0: ; %entry
1509; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1510; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1511; GFX8-NEXT:    s_mov_b32 s2, 0xffff
1512; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1513; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
1514; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
1515; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
1516; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1517; GFX8-NEXT:    s_and_b32 s6, s3, s2
1518; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
1519; GFX8-NEXT:    s_and_b32 s2, s4, s2
1520; GFX8-NEXT:    v_mov_b32_e32 v0, s5
1521; GFX8-NEXT:    v_mov_b32_e32 v1, s3
1522; GFX8-NEXT:    v_mad_u32_u24 v0, s2, v1, v0
1523; GFX8-NEXT:    s_lshr_b32 s7, s4, 16
1524; GFX8-NEXT:    v_mov_b32_e32 v1, s6
1525; GFX8-NEXT:    v_mad_u32_u24 v2, s7, v1, v0
1526; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1527; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1528; GFX8-NEXT:    flat_store_dword v[0:1], v2
1529; GFX8-NEXT:    s_endpgm
1530;
1531; GFX9-NODL-LABEL: notudot2_DiffIndex:
1532; GFX9-NODL:       ; %bb.0: ; %entry
1533; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1534; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1535; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
1536; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1537; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1538; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
1539; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
1540; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
1541; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1542; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
1543; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
1544; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
1545; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1546; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1547; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v1, v2
1548; GFX9-NODL-NEXT:    s_lshr_b32 s7, s4, 16
1549; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1550; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s7, v2, v1
1551; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1552; GFX9-NODL-NEXT:    s_endpgm
1553;
1554; GFX9-DL-LABEL: notudot2_DiffIndex:
1555; GFX9-DL:       ; %bb.0: ; %entry
1556; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1557; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1558; GFX9-DL-NEXT:    s_mov_b32 s2, 0xffff
1559; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1560; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1561; GFX9-DL-NEXT:    s_load_dword s3, s[4:5], 0x0
1562; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
1563; GFX9-DL-NEXT:    s_load_dword s5, s[0:1], 0x0
1564; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1565; GFX9-DL-NEXT:    s_and_b32 s6, s3, s2
1566; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
1567; GFX9-DL-NEXT:    s_and_b32 s2, s4, s2
1568; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
1569; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
1570; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s2, v1, v2
1571; GFX9-DL-NEXT:    s_lshr_b32 s7, s4, 16
1572; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s6
1573; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s7, v2, v1
1574; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1575; GFX9-DL-NEXT:    s_endpgm
1576;
1577; GFX10-DL-LABEL: notudot2_DiffIndex:
1578; GFX10-DL:       ; %bb.0: ; %entry
1579; GFX10-DL-NEXT:    s_clause 0x1
1580; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1581; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1582; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
1583; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1584; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1585; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1586; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
1587; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
1588; GFX10-DL-NEXT:    s_mov_b32 s2, 0xffff
1589; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1590; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1591; GFX10-DL-NEXT:    s_lshr_b32 s3, s0, 16
1592; GFX10-DL-NEXT:    s_and_b32 s6, s1, s2
1593; GFX10-DL-NEXT:    s_and_b32 s0, s0, s2
1594; GFX10-DL-NEXT:    s_lshr_b32 s1, s1, 16
1595; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s6, s3, v0
1596; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s1, s0, v0
1597; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
1598; GFX10-DL-NEXT:    s_endpgm
1599                                              <2 x i16> addrspace(1)* %src2,
1600                                              i32 addrspace(1)* nocapture %dst) {
1601entry:
1602  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
1603  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
1604
1605  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
1606  %conv = zext i16 %s1.elt1 to i32
1607  %s2.elt1 = extractelement <2 x i16> %vec2, i64 1
1608  %conv2 = zext i16 %s2.elt1 to i32
1609  %mul1 = mul i32 %conv2, %conv
1610
1611  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
1612  %conv3 = zext i16 %s1.elt2 to i32
1613  %s2.elt2 = extractelement <2 x i16> %vec2, i64 0
1614  %conv4 = zext i16 %s2.elt2 to i32
1615  %mul2 = mul i32 %conv4, %conv3
1616
1617  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1618  %add = add i32 %mul2, %s3
1619  %add6 = add i32 %add, %mul1
1620  store i32 %add6, i32 addrspace(1)* %dst, align 4
1621  ret void
1622}
1623
1624define amdgpu_kernel void @udot2_MultipleUses_add1(<2 x i16> addrspace(1)* %src1,
1625; GFX7-LABEL: udot2_MultipleUses_add1:
1626; GFX7:       ; %bb.0: ; %entry
1627; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1628; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1629; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1630; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1631; GFX7-NEXT:    s_mov_b32 s2, -1
1632; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1633; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
1634; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
1635; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1636; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
1637; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
1638; GFX7-NEXT:    s_and_b32 s4, s4, s8
1639; GFX7-NEXT:    s_and_b32 s5, s5, s8
1640; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1641; GFX7-NEXT:    v_mov_b32_e32 v0, s6
1642; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1643; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1644; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
1645; GFX7-NEXT:    v_mov_b32_e32 v1, s4
1646; GFX7-NEXT:    v_mad_u32_u24 v1, s5, v1, v0
1647; GFX7-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
1648; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1649; GFX7-NEXT:    s_endpgm
1650;
1651; GFX8-LABEL: udot2_MultipleUses_add1:
1652; GFX8:       ; %bb.0: ; %entry
1653; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1654; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1655; GFX8-NEXT:    s_mov_b32 s2, 0xffff
1656; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1657; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
1658; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
1659; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
1660; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1661; GFX8-NEXT:    s_and_b32 s6, s3, s2
1662; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
1663; GFX8-NEXT:    s_and_b32 s2, s4, s2
1664; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
1665; GFX8-NEXT:    v_mov_b32_e32 v0, s5
1666; GFX8-NEXT:    v_mov_b32_e32 v1, s3
1667; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
1668; GFX8-NEXT:    v_mov_b32_e32 v1, s6
1669; GFX8-NEXT:    v_mad_u32_u24 v1, s2, v1, v0
1670; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v0, v1
1671; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1672; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1673; GFX8-NEXT:    flat_store_dword v[0:1], v2
1674; GFX8-NEXT:    s_endpgm
1675;
1676; GFX9-NODL-LABEL: udot2_MultipleUses_add1:
1677; GFX9-NODL:       ; %bb.0: ; %entry
1678; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1679; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1680; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
1681; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1682; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1683; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
1684; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
1685; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
1686; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1687; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
1688; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
1689; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
1690; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
1691; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
1692; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1693; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
1694; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
1695; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, s2, v2, v1
1696; GFX9-NODL-NEXT:    v_add_u32_e32 v1, v2, v1
1697; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1698; GFX9-NODL-NEXT:    s_endpgm
1699;
1700; GFX9-DL-LABEL: udot2_MultipleUses_add1:
1701; GFX9-DL:       ; %bb.0: ; %entry
1702; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1703; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1704; GFX9-DL-NEXT:    s_mov_b32 s2, 0xffff
1705; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1706; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1707; GFX9-DL-NEXT:    s_load_dword s3, s[4:5], 0x0
1708; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
1709; GFX9-DL-NEXT:    s_load_dword s5, s[0:1], 0x0
1710; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1711; GFX9-DL-NEXT:    s_and_b32 s6, s3, s2
1712; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
1713; GFX9-DL-NEXT:    s_and_b32 s2, s4, s2
1714; GFX9-DL-NEXT:    s_lshr_b32 s4, s4, 16
1715; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
1716; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
1717; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
1718; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s6
1719; GFX9-DL-NEXT:    v_mad_u32_u24 v2, s2, v2, v1
1720; GFX9-DL-NEXT:    v_add_u32_e32 v1, v2, v1
1721; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1722; GFX9-DL-NEXT:    s_endpgm
1723;
1724; GFX10-DL-LABEL: udot2_MultipleUses_add1:
1725; GFX10-DL:       ; %bb.0: ; %entry
1726; GFX10-DL-NEXT:    s_clause 0x1
1727; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1728; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1729; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
1730; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1731; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1732; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1733; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
1734; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
1735; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1736; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1737; GFX10-DL-NEXT:    s_lshr_b32 s2, s0, 16
1738; GFX10-DL-NEXT:    s_lshr_b32 s3, s1, 16
1739; GFX10-DL-NEXT:    s_mov_b32 s6, 0xffff
1740; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s2, v0
1741; GFX10-DL-NEXT:    s_and_b32 s0, s0, s6
1742; GFX10-DL-NEXT:    s_and_b32 s1, s1, s6
1743; GFX10-DL-NEXT:    v_mad_u32_u24 v1, s1, s0, v0
1744; GFX10-DL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
1745; GFX10-DL-NEXT:    global_store_dword v2, v0, s[4:5]
1746; GFX10-DL-NEXT:    s_endpgm
1747                                                   <2 x i16> addrspace(1)* %src2,
1748                                                   i32 addrspace(1)* nocapture %dst) {
1749entry:
1750  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
1751  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
1752
1753  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
1754  %conv = zext i16 %s1.elt1 to i32
1755  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
1756  %conv2 = zext i16 %s2.elt1 to i32
1757  %mul1 = mul i32 %conv2, %conv
1758
1759  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
1760  %conv3 = zext i16 %s1.elt2 to i32
1761  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
1762  %conv4 = zext i16 %s2.elt2 to i32
1763  %mul2 = mul i32 %conv4, %conv3
1764
1765  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1766  %add1 = add i32 %mul2, %s3
1767  %add2 = add i32 %add1, %mul1
1768
1769  %res = add i32 %add2, %add1
1770  store i32 %res, i32 addrspace(1)* %dst, align 4
1771  ret void
1772}
1773
1774define amdgpu_kernel void @idot2_MultipleUses_add1(<2 x i16> addrspace(1)* %src1,
1775; GFX7-LABEL: idot2_MultipleUses_add1:
1776; GFX7:       ; %bb.0: ; %entry
1777; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1778; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1779; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1780; GFX7-NEXT:    s_mov_b32 s2, -1
1781; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1782; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
1783; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
1784; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
1785; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1786; GFX7-NEXT:    s_sext_i32_i16 s7, s4
1787; GFX7-NEXT:    s_ashr_i32 s4, s4, 16
1788; GFX7-NEXT:    s_sext_i32_i16 s8, s5
1789; GFX7-NEXT:    s_ashr_i32 s5, s5, 16
1790; GFX7-NEXT:    v_mov_b32_e32 v0, s4
1791; GFX7-NEXT:    v_mov_b32_e32 v1, s6
1792; GFX7-NEXT:    v_mad_i32_i24 v0, s5, v0, v1
1793; GFX7-NEXT:    v_mov_b32_e32 v1, s7
1794; GFX7-NEXT:    v_mad_i32_i24 v1, s8, v1, v0
1795; GFX7-NEXT:    v_add_i32_e32 v0, vcc, v0, v1
1796; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1797; GFX7-NEXT:    s_endpgm
1798;
1799; GFX8-LABEL: idot2_MultipleUses_add1:
1800; GFX8:       ; %bb.0: ; %entry
1801; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1802; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1803; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1804; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
1805; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
1806; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
1807; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1808; GFX8-NEXT:    s_sext_i32_i16 s5, s2
1809; GFX8-NEXT:    s_ashr_i32 s2, s2, 16
1810; GFX8-NEXT:    s_sext_i32_i16 s6, s3
1811; GFX8-NEXT:    s_ashr_i32 s3, s3, 16
1812; GFX8-NEXT:    v_mov_b32_e32 v0, s4
1813; GFX8-NEXT:    v_mov_b32_e32 v1, s2
1814; GFX8-NEXT:    v_mov_b32_e32 v2, s5
1815; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v1, v0
1816; GFX8-NEXT:    v_mad_i32_i24 v1, s6, v2, v0
1817; GFX8-NEXT:    v_add_u32_e32 v2, vcc, v0, v1
1818; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1819; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1820; GFX8-NEXT:    flat_store_dword v[0:1], v2
1821; GFX8-NEXT:    s_endpgm
1822;
1823; GFX9-NODL-LABEL: idot2_MultipleUses_add1:
1824; GFX9-NODL:       ; %bb.0: ; %entry
1825; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1826; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1827; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1828; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1829; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
1830; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
1831; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
1832; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1833; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
1834; GFX9-NODL-NEXT:    s_ashr_i32 s2, s2, 16
1835; GFX9-NODL-NEXT:    s_sext_i32_i16 s6, s3
1836; GFX9-NODL-NEXT:    s_ashr_i32 s3, s3, 16
1837; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
1838; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
1839; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
1840; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1841; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, s6, v2, v1
1842; GFX9-NODL-NEXT:    v_add_u32_e32 v1, v2, v1
1843; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1844; GFX9-NODL-NEXT:    s_endpgm
1845;
1846; GFX9-DL-LABEL: idot2_MultipleUses_add1:
1847; GFX9-DL:       ; %bb.0: ; %entry
1848; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1849; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1850; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
1851; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1852; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
1853; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
1854; GFX9-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
1855; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
1856; GFX9-DL-NEXT:    s_sext_i32_i16 s5, s2
1857; GFX9-DL-NEXT:    s_ashr_i32 s2, s2, 16
1858; GFX9-DL-NEXT:    s_sext_i32_i16 s6, s3
1859; GFX9-DL-NEXT:    s_ashr_i32 s3, s3, 16
1860; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s4
1861; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
1862; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
1863; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
1864; GFX9-DL-NEXT:    v_mad_i32_i24 v2, s6, v2, v1
1865; GFX9-DL-NEXT:    v_add_u32_e32 v1, v2, v1
1866; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
1867; GFX9-DL-NEXT:    s_endpgm
1868;
1869; GFX10-DL-LABEL: idot2_MultipleUses_add1:
1870; GFX10-DL:       ; %bb.0: ; %entry
1871; GFX10-DL-NEXT:    s_clause 0x1
1872; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
1873; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
1874; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
1875; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
1876; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1877; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
1878; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
1879; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
1880; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
1881; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
1882; GFX10-DL-NEXT:    s_ashr_i32 s2, s0, 16
1883; GFX10-DL-NEXT:    s_ashr_i32 s3, s1, 16
1884; GFX10-DL-NEXT:    s_sext_i32_i16 s0, s0
1885; GFX10-DL-NEXT:    s_sext_i32_i16 s1, s1
1886; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
1887; GFX10-DL-NEXT:    v_mad_i32_i24 v1, s1, s0, v0
1888; GFX10-DL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
1889; GFX10-DL-NEXT:    global_store_dword v2, v0, s[4:5]
1890; GFX10-DL-NEXT:    s_endpgm
1891                                                   <2 x i16> addrspace(1)* %src2,
1892                                                   i32 addrspace(1)* nocapture %dst) {
1893entry:
1894  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
1895  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
1896
1897  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
1898  %conv = sext i16 %s1.elt1 to i32
1899  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
1900  %conv2 = sext i16 %s2.elt1 to i32
1901  %mul1 = mul i32 %conv2, %conv
1902
1903  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
1904  %conv3 = sext i16 %s1.elt2 to i32
1905  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
1906  %conv4 = sext i16 %s2.elt2 to i32
1907  %mul2 = mul i32 %conv4, %conv3
1908
1909  %s3 = load i32, i32 addrspace(1)* %dst, align 4
1910  %add1 = add i32 %mul2, %s3
1911  %add2 = add i32 %add1, %mul1
1912
1913  %res = add i32 %add2, %add1
1914  store i32 %res, i32 addrspace(1)* %dst, align 4
1915  ret void
1916}
1917
1918define amdgpu_kernel void @udot2_MultipleUses_mul1(<2 x i16> addrspace(1)* %src1,
1919; GFX7-LABEL: udot2_MultipleUses_mul1:
1920; GFX7:       ; %bb.0: ; %entry
1921; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
1922; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
1923; GFX7-NEXT:    s_mov_b32 s8, 0xffff
1924; GFX7-NEXT:    s_mov_b32 s3, 0xf000
1925; GFX7-NEXT:    s_mov_b32 s2, -1
1926; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1927; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
1928; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
1929; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1930; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
1931; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
1932; GFX7-NEXT:    s_and_b32 s4, s4, s8
1933; GFX7-NEXT:    s_and_b32 s5, s5, s8
1934; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
1935; GFX7-NEXT:    v_mov_b32_e32 v0, s4
1936; GFX7-NEXT:    v_mov_b32_e32 v2, s6
1937; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
1938; GFX7-NEXT:    v_mov_b32_e32 v1, s8
1939; GFX7-NEXT:    v_mad_u32_u24 v1, s5, v0, v1
1940; GFX7-NEXT:    v_mad_u32_u24 v1, s7, v2, v1
1941; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v0, v1
1942; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
1943; GFX7-NEXT:    s_endpgm
1944;
1945; GFX8-LABEL: udot2_MultipleUses_mul1:
1946; GFX8:       ; %bb.0: ; %entry
1947; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1948; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1949; GFX8-NEXT:    s_mov_b32 s2, 0xffff
1950; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1951; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
1952; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
1953; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
1954; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
1955; GFX8-NEXT:    s_and_b32 s6, s3, s2
1956; GFX8-NEXT:    s_and_b32 s2, s4, s2
1957; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
1958; GFX8-NEXT:    v_mov_b32_e32 v0, s5
1959; GFX8-NEXT:    v_mov_b32_e32 v1, s6
1960; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
1961; GFX8-NEXT:    v_mad_u32_u24 v0, s2, v1, v0
1962; GFX8-NEXT:    v_mov_b32_e32 v2, s3
1963; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v2, v0
1964; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
1965; GFX8-NEXT:    v_mov_b32_e32 v0, s0
1966; GFX8-NEXT:    v_mov_b32_e32 v1, s1
1967; GFX8-NEXT:    flat_store_dword v[0:1], v2
1968; GFX8-NEXT:    s_endpgm
1969;
1970; GFX9-NODL-LABEL: udot2_MultipleUses_mul1:
1971; GFX9-NODL:       ; %bb.0: ; %entry
1972; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1973; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1974; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
1975; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
1976; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1977; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
1978; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
1979; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
1980; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
1981; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
1982; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
1983; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
1984; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s6
1985; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
1986; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
1987; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, s2, v1, v2
1988; GFX9-NODL-NEXT:    v_mov_b32_e32 v3, s3
1989; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, s4, v3, v2
1990; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v1, v2
1991; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
1992; GFX9-NODL-NEXT:    s_endpgm
1993;
1994; GFX9-DL-LABEL: udot2_MultipleUses_mul1:
1995; GFX9-DL:       ; %bb.0: ; %entry
1996; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
1997; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
1998; GFX9-DL-NEXT:    s_mov_b32 s2, 0xffff
1999; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2000; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2001; GFX9-DL-NEXT:    s_load_dword s3, s[4:5], 0x0
2002; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
2003; GFX9-DL-NEXT:    s_load_dword s5, s[0:1], 0x0
2004; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2005; GFX9-DL-NEXT:    s_and_b32 s6, s3, s2
2006; GFX9-DL-NEXT:    s_and_b32 s2, s4, s2
2007; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
2008; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s6
2009; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
2010; GFX9-DL-NEXT:    s_lshr_b32 s4, s4, 16
2011; GFX9-DL-NEXT:    v_mad_u32_u24 v2, s2, v1, v2
2012; GFX9-DL-NEXT:    v_mov_b32_e32 v3, s3
2013; GFX9-DL-NEXT:    v_mad_u32_u24 v2, s4, v3, v2
2014; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s2, v1, v2
2015; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2016; GFX9-DL-NEXT:    s_endpgm
2017;
2018; GFX10-DL-LABEL: udot2_MultipleUses_mul1:
2019; GFX10-DL:       ; %bb.0: ; %entry
2020; GFX10-DL-NEXT:    s_clause 0x1
2021; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
2022; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
2023; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
2024; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2025; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2026; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
2027; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
2028; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
2029; GFX10-DL-NEXT:    s_mov_b32 s2, 0xffff
2030; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2031; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
2032; GFX10-DL-NEXT:    s_and_b32 s3, s0, s2
2033; GFX10-DL-NEXT:    s_and_b32 s2, s1, s2
2034; GFX10-DL-NEXT:    s_lshr_b32 s0, s0, 16
2035; GFX10-DL-NEXT:    s_lshr_b32 s1, s1, 16
2036; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s2, s3, v0
2037; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s1, s0, v0
2038; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s2, s3, v0
2039; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
2040; GFX10-DL-NEXT:    s_endpgm
2041                                                   <2 x i16> addrspace(1)* %src2,
2042                                                   i32 addrspace(1)* nocapture %dst) {
2043entry:
2044  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
2045  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
2046
2047  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
2048  %conv = zext i16 %s1.elt1 to i32
2049  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
2050  %conv2 = zext i16 %s2.elt1 to i32
2051  %mul1 = mul i32 %conv2, %conv
2052
2053  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
2054  %conv3 = zext i16 %s1.elt2 to i32
2055  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
2056  %conv4 = zext i16 %s2.elt2 to i32
2057  %mul2 = mul i32 %conv4, %conv3
2058
2059  %s3 = load i32, i32 addrspace(1)* %dst, align 4
2060  %add0 = add i32 %mul1, %s3
2061
2062  %add1 = add i32 %mul2, %add0
2063  %add2 = add i32 %add1, %mul1
2064
2065  store i32 %add2, i32 addrspace(1)* %dst, align 4
2066  ret void
2067}
2068
2069define amdgpu_kernel void @idot2_MultipleUses_mul1(<2 x i16> addrspace(1)* %src1,
2070; GFX7-LABEL: idot2_MultipleUses_mul1:
2071; GFX7:       ; %bb.0: ; %entry
2072; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
2073; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
2074; GFX7-NEXT:    s_mov_b32 s3, 0xf000
2075; GFX7-NEXT:    s_mov_b32 s2, -1
2076; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2077; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
2078; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
2079; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
2080; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2081; GFX7-NEXT:    s_sext_i32_i16 s7, s4
2082; GFX7-NEXT:    s_sext_i32_i16 s8, s5
2083; GFX7-NEXT:    s_ashr_i32 s4, s4, 16
2084; GFX7-NEXT:    v_mov_b32_e32 v0, s7
2085; GFX7-NEXT:    v_mov_b32_e32 v1, s6
2086; GFX7-NEXT:    s_ashr_i32 s5, s5, 16
2087; GFX7-NEXT:    v_mad_i32_i24 v1, s8, v0, v1
2088; GFX7-NEXT:    v_mov_b32_e32 v2, s4
2089; GFX7-NEXT:    v_mad_i32_i24 v1, s5, v2, v1
2090; GFX7-NEXT:    v_mad_i32_i24 v0, s8, v0, v1
2091; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
2092; GFX7-NEXT:    s_endpgm
2093;
2094; GFX8-LABEL: idot2_MultipleUses_mul1:
2095; GFX8:       ; %bb.0: ; %entry
2096; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2097; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2098; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2099; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
2100; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
2101; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
2102; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2103; GFX8-NEXT:    s_sext_i32_i16 s5, s2
2104; GFX8-NEXT:    s_sext_i32_i16 s6, s3
2105; GFX8-NEXT:    s_ashr_i32 s2, s2, 16
2106; GFX8-NEXT:    v_mov_b32_e32 v0, s4
2107; GFX8-NEXT:    v_mov_b32_e32 v1, s5
2108; GFX8-NEXT:    s_ashr_i32 s3, s3, 16
2109; GFX8-NEXT:    v_mov_b32_e32 v2, s2
2110; GFX8-NEXT:    v_mad_i32_i24 v0, s6, v1, v0
2111; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v2, v0
2112; GFX8-NEXT:    v_mad_i32_i24 v2, s6, v1, v0
2113; GFX8-NEXT:    v_mov_b32_e32 v0, s0
2114; GFX8-NEXT:    v_mov_b32_e32 v1, s1
2115; GFX8-NEXT:    flat_store_dword v[0:1], v2
2116; GFX8-NEXT:    s_endpgm
2117;
2118; GFX9-NODL-LABEL: idot2_MultipleUses_mul1:
2119; GFX9-NODL:       ; %bb.0: ; %entry
2120; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2121; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2122; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
2123; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2124; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
2125; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
2126; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
2127; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2128; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
2129; GFX9-NODL-NEXT:    s_sext_i32_i16 s6, s3
2130; GFX9-NODL-NEXT:    s_ashr_i32 s2, s2, 16
2131; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
2132; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
2133; GFX9-NODL-NEXT:    s_ashr_i32 s3, s3, 16
2134; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2135; GFX9-NODL-NEXT:    v_mov_b32_e32 v3, s2
2136; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v3, v1
2137; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2138; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
2139; GFX9-NODL-NEXT:    s_endpgm
2140;
2141; GFX9-DL-LABEL: idot2_MultipleUses_mul1:
2142; GFX9-DL:       ; %bb.0: ; %entry
2143; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2144; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2145; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2146; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2147; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
2148; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
2149; GFX9-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
2150; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2151; GFX9-DL-NEXT:    s_sext_i32_i16 s5, s2
2152; GFX9-DL-NEXT:    s_sext_i32_i16 s6, s3
2153; GFX9-DL-NEXT:    s_ashr_i32 s2, s2, 16
2154; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s4
2155; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
2156; GFX9-DL-NEXT:    s_ashr_i32 s3, s3, 16
2157; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2158; GFX9-DL-NEXT:    v_mov_b32_e32 v3, s2
2159; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s3, v3, v1
2160; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2161; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2162; GFX9-DL-NEXT:    s_endpgm
2163;
2164; GFX10-DL-LABEL: idot2_MultipleUses_mul1:
2165; GFX10-DL:       ; %bb.0: ; %entry
2166; GFX10-DL-NEXT:    s_clause 0x1
2167; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
2168; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
2169; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
2170; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2171; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2172; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
2173; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
2174; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
2175; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2176; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
2177; GFX10-DL-NEXT:    s_sext_i32_i16 s2, s0
2178; GFX10-DL-NEXT:    s_sext_i32_i16 s3, s1
2179; GFX10-DL-NEXT:    s_ashr_i32 s0, s0, 16
2180; GFX10-DL-NEXT:    s_ashr_i32 s1, s1, 16
2181; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
2182; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s1, s0, v0
2183; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
2184; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
2185; GFX10-DL-NEXT:    s_endpgm
2186                                                   <2 x i16> addrspace(1)* %src2,
2187                                                   i32 addrspace(1)* nocapture %dst) {
2188entry:
2189  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
2190  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
2191
2192  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
2193  %conv = sext i16 %s1.elt1 to i32
2194  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
2195  %conv2 = sext i16 %s2.elt1 to i32
2196  %mul1 = mul i32 %conv2, %conv
2197
2198  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
2199  %conv3 = sext i16 %s1.elt2 to i32
2200  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
2201  %conv4 = sext i16 %s2.elt2 to i32
2202  %mul2 = mul i32 %conv4, %conv3
2203
2204  %s3 = load i32, i32 addrspace(1)* %dst, align 4
2205  %add0 = add i32 %mul1, %s3
2206
2207  %add1 = add i32 %mul2, %add0
2208  %add2 = add i32 %add1, %mul1
2209
2210  store i32 %add2, i32 addrspace(1)* %dst, align 4
2211  ret void
2212}
2213
2214define amdgpu_kernel void @udot2_MultipleUses_mul2(<2 x i16> addrspace(1)* %src1,
2215; GFX7-LABEL: udot2_MultipleUses_mul2:
2216; GFX7:       ; %bb.0: ; %entry
2217; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
2218; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
2219; GFX7-NEXT:    s_mov_b32 s8, 0xffff
2220; GFX7-NEXT:    s_mov_b32 s3, 0xf000
2221; GFX7-NEXT:    s_mov_b32 s2, -1
2222; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2223; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
2224; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
2225; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2226; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
2227; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
2228; GFX7-NEXT:    s_and_b32 s4, s4, s8
2229; GFX7-NEXT:    s_and_b32 s5, s5, s8
2230; GFX7-NEXT:    s_load_dword s8, s[0:1], 0x0
2231; GFX7-NEXT:    v_mov_b32_e32 v0, s6
2232; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2233; GFX7-NEXT:    v_mov_b32_e32 v1, s8
2234; GFX7-NEXT:    v_mad_u32_u24 v1, s7, v0, v1
2235; GFX7-NEXT:    v_mad_u32_u24 v0, s7, v0, v1
2236; GFX7-NEXT:    v_mov_b32_e32 v1, s4
2237; GFX7-NEXT:    v_mad_u32_u24 v0, s5, v1, v0
2238; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
2239; GFX7-NEXT:    s_endpgm
2240;
2241; GFX8-LABEL: udot2_MultipleUses_mul2:
2242; GFX8:       ; %bb.0: ; %entry
2243; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2244; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2245; GFX8-NEXT:    s_mov_b32 s2, 0xffff
2246; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2247; GFX8-NEXT:    s_load_dword s3, s[4:5], 0x0
2248; GFX8-NEXT:    s_load_dword s4, s[6:7], 0x0
2249; GFX8-NEXT:    s_load_dword s5, s[0:1], 0x0
2250; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2251; GFX8-NEXT:    s_and_b32 s6, s3, s2
2252; GFX8-NEXT:    s_lshr_b32 s3, s3, 16
2253; GFX8-NEXT:    s_and_b32 s2, s4, s2
2254; GFX8-NEXT:    s_lshr_b32 s4, s4, 16
2255; GFX8-NEXT:    v_mov_b32_e32 v0, s5
2256; GFX8-NEXT:    v_mov_b32_e32 v1, s3
2257; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
2258; GFX8-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
2259; GFX8-NEXT:    v_mov_b32_e32 v1, s6
2260; GFX8-NEXT:    v_mad_u32_u24 v2, s2, v1, v0
2261; GFX8-NEXT:    v_mov_b32_e32 v0, s0
2262; GFX8-NEXT:    v_mov_b32_e32 v1, s1
2263; GFX8-NEXT:    flat_store_dword v[0:1], v2
2264; GFX8-NEXT:    s_endpgm
2265;
2266; GFX9-NODL-LABEL: udot2_MultipleUses_mul2:
2267; GFX9-NODL:       ; %bb.0: ; %entry
2268; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2269; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2270; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
2271; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
2272; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2273; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
2274; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
2275; GFX9-NODL-NEXT:    s_load_dword s5, s[0:1], 0x0
2276; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2277; GFX9-NODL-NEXT:    s_and_b32 s6, s3, s2
2278; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
2279; GFX9-NODL-NEXT:    s_and_b32 s2, s4, s2
2280; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
2281; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s3
2282; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
2283; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, s4, v1, v2
2284; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
2285; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s6
2286; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
2287; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
2288; GFX9-NODL-NEXT:    s_endpgm
2289;
2290; GFX9-DL-LABEL: udot2_MultipleUses_mul2:
2291; GFX9-DL:       ; %bb.0: ; %entry
2292; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2293; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2294; GFX9-DL-NEXT:    s_mov_b32 s2, 0xffff
2295; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2296; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2297; GFX9-DL-NEXT:    s_load_dword s3, s[4:5], 0x0
2298; GFX9-DL-NEXT:    s_load_dword s4, s[6:7], 0x0
2299; GFX9-DL-NEXT:    s_load_dword s5, s[0:1], 0x0
2300; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2301; GFX9-DL-NEXT:    s_and_b32 s6, s3, s2
2302; GFX9-DL-NEXT:    s_lshr_b32 s3, s3, 16
2303; GFX9-DL-NEXT:    s_and_b32 s2, s4, s2
2304; GFX9-DL-NEXT:    s_lshr_b32 s4, s4, 16
2305; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s3
2306; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
2307; GFX9-DL-NEXT:    v_mad_u32_u24 v2, s4, v1, v2
2308; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s4, v1, v2
2309; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s6
2310; GFX9-DL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
2311; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2312; GFX9-DL-NEXT:    s_endpgm
2313;
2314; GFX10-DL-LABEL: udot2_MultipleUses_mul2:
2315; GFX10-DL:       ; %bb.0: ; %entry
2316; GFX10-DL-NEXT:    s_clause 0x1
2317; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
2318; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
2319; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
2320; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2321; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2322; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
2323; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
2324; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
2325; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2326; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
2327; GFX10-DL-NEXT:    s_lshr_b32 s2, s0, 16
2328; GFX10-DL-NEXT:    s_lshr_b32 s3, s1, 16
2329; GFX10-DL-NEXT:    s_mov_b32 s6, 0xffff
2330; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s2, v0
2331; GFX10-DL-NEXT:    s_and_b32 s0, s0, s6
2332; GFX10-DL-NEXT:    s_and_b32 s1, s1, s6
2333; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s3, s2, v0
2334; GFX10-DL-NEXT:    v_mad_u32_u24 v0, s1, s0, v0
2335; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
2336; GFX10-DL-NEXT:    s_endpgm
2337                                                   <2 x i16> addrspace(1)* %src2,
2338                                                   i32 addrspace(1)* nocapture %dst) {
2339entry:
2340  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
2341  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
2342
2343  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
2344  %conv = zext i16 %s1.elt1 to i32
2345  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
2346  %conv2 = zext i16 %s2.elt1 to i32
2347  %mul1 = mul i32 %conv2, %conv
2348
2349  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
2350  %conv3 = zext i16 %s1.elt2 to i32
2351  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
2352  %conv4 = zext i16 %s2.elt2 to i32
2353  %mul2 = mul i32 %conv4, %conv3
2354
2355  %s3 = load i32, i32 addrspace(1)* %dst, align 4
2356  %add0 = add i32 %mul2, %s3
2357
2358  %add1 = add i32 %mul2, %add0
2359  %add2 = add i32 %add1, %mul1
2360
2361  store i32 %add2, i32 addrspace(1)* %dst, align 4
2362  ret void
2363}
2364
2365define amdgpu_kernel void @idot2_MultipleUses_mul2(<2 x i16> addrspace(1)* %src1,
2366; GFX7-LABEL: idot2_MultipleUses_mul2:
2367; GFX7:       ; %bb.0: ; %entry
2368; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
2369; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
2370; GFX7-NEXT:    s_mov_b32 s3, 0xf000
2371; GFX7-NEXT:    s_mov_b32 s2, -1
2372; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2373; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
2374; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
2375; GFX7-NEXT:    s_load_dword s6, s[0:1], 0x0
2376; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2377; GFX7-NEXT:    s_sext_i32_i16 s7, s4
2378; GFX7-NEXT:    s_ashr_i32 s4, s4, 16
2379; GFX7-NEXT:    s_sext_i32_i16 s8, s5
2380; GFX7-NEXT:    s_ashr_i32 s5, s5, 16
2381; GFX7-NEXT:    v_mov_b32_e32 v0, s4
2382; GFX7-NEXT:    v_mov_b32_e32 v1, s6
2383; GFX7-NEXT:    v_mad_i32_i24 v1, s5, v0, v1
2384; GFX7-NEXT:    v_mad_i32_i24 v0, s5, v0, v1
2385; GFX7-NEXT:    v_mov_b32_e32 v1, s7
2386; GFX7-NEXT:    v_mad_i32_i24 v0, s8, v1, v0
2387; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
2388; GFX7-NEXT:    s_endpgm
2389;
2390; GFX8-LABEL: idot2_MultipleUses_mul2:
2391; GFX8:       ; %bb.0: ; %entry
2392; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2393; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2394; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2395; GFX8-NEXT:    s_load_dword s2, s[4:5], 0x0
2396; GFX8-NEXT:    s_load_dword s3, s[6:7], 0x0
2397; GFX8-NEXT:    s_load_dword s4, s[0:1], 0x0
2398; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2399; GFX8-NEXT:    s_sext_i32_i16 s5, s2
2400; GFX8-NEXT:    s_ashr_i32 s2, s2, 16
2401; GFX8-NEXT:    s_sext_i32_i16 s6, s3
2402; GFX8-NEXT:    s_ashr_i32 s3, s3, 16
2403; GFX8-NEXT:    v_mov_b32_e32 v0, s4
2404; GFX8-NEXT:    v_mov_b32_e32 v1, s2
2405; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v1, v0
2406; GFX8-NEXT:    v_mov_b32_e32 v2, s5
2407; GFX8-NEXT:    v_mad_i32_i24 v0, s3, v1, v0
2408; GFX8-NEXT:    v_mad_i32_i24 v2, s6, v2, v0
2409; GFX8-NEXT:    v_mov_b32_e32 v0, s0
2410; GFX8-NEXT:    v_mov_b32_e32 v1, s1
2411; GFX8-NEXT:    flat_store_dword v[0:1], v2
2412; GFX8-NEXT:    s_endpgm
2413;
2414; GFX9-NODL-LABEL: idot2_MultipleUses_mul2:
2415; GFX9-NODL:       ; %bb.0: ; %entry
2416; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2417; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2418; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
2419; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2420; GFX9-NODL-NEXT:    s_load_dword s2, s[4:5], 0x0
2421; GFX9-NODL-NEXT:    s_load_dword s3, s[6:7], 0x0
2422; GFX9-NODL-NEXT:    s_load_dword s4, s[0:1], 0x0
2423; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2424; GFX9-NODL-NEXT:    s_sext_i32_i16 s5, s2
2425; GFX9-NODL-NEXT:    s_ashr_i32 s2, s2, 16
2426; GFX9-NODL-NEXT:    s_sext_i32_i16 s6, s3
2427; GFX9-NODL-NEXT:    s_ashr_i32 s3, s3, 16
2428; GFX9-NODL-NEXT:    v_mov_b32_e32 v1, s4
2429; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s2
2430; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
2431; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
2432; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
2433; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2434; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
2435; GFX9-NODL-NEXT:    s_endpgm
2436;
2437; GFX9-DL-LABEL: idot2_MultipleUses_mul2:
2438; GFX9-DL:       ; %bb.0: ; %entry
2439; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2440; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2441; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2442; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2443; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
2444; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
2445; GFX9-DL-NEXT:    s_load_dword s4, s[0:1], 0x0
2446; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2447; GFX9-DL-NEXT:    s_sext_i32_i16 s5, s2
2448; GFX9-DL-NEXT:    s_ashr_i32 s2, s2, 16
2449; GFX9-DL-NEXT:    s_sext_i32_i16 s6, s3
2450; GFX9-DL-NEXT:    s_ashr_i32 s3, s3, 16
2451; GFX9-DL-NEXT:    v_mov_b32_e32 v1, s4
2452; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s2
2453; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
2454; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s3, v2, v1
2455; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s5
2456; GFX9-DL-NEXT:    v_mad_i32_i24 v1, s6, v2, v1
2457; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2458; GFX9-DL-NEXT:    s_endpgm
2459;
2460; GFX10-DL-LABEL: idot2_MultipleUses_mul2:
2461; GFX10-DL:       ; %bb.0: ; %entry
2462; GFX10-DL-NEXT:    s_clause 0x1
2463; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
2464; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
2465; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
2466; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2467; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2468; GFX10-DL-NEXT:    s_load_dword s6, s[4:5], 0x0
2469; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
2470; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
2471; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2472; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s6
2473; GFX10-DL-NEXT:    s_ashr_i32 s2, s0, 16
2474; GFX10-DL-NEXT:    s_ashr_i32 s3, s1, 16
2475; GFX10-DL-NEXT:    s_sext_i32_i16 s0, s0
2476; GFX10-DL-NEXT:    s_sext_i32_i16 s1, s1
2477; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
2478; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s3, s2, v0
2479; GFX10-DL-NEXT:    v_mad_i32_i24 v0, s1, s0, v0
2480; GFX10-DL-NEXT:    global_store_dword v1, v0, s[4:5]
2481; GFX10-DL-NEXT:    s_endpgm
2482                                                   <2 x i16> addrspace(1)* %src2,
2483                                                   i32 addrspace(1)* nocapture %dst) {
2484entry:
2485  %vec1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
2486  %vec2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
2487
2488  %s1.elt1 = extractelement <2 x i16> %vec1, i64 0
2489  %conv = sext i16 %s1.elt1 to i32
2490  %s2.elt1 = extractelement <2 x i16> %vec2, i64 0
2491  %conv2 = sext i16 %s2.elt1 to i32
2492  %mul1 = mul i32 %conv2, %conv
2493
2494  %s1.elt2 = extractelement <2 x i16> %vec1, i64 1
2495  %conv3 = sext i16 %s1.elt2 to i32
2496  %s2.elt2 = extractelement <2 x i16> %vec2, i64 1
2497  %conv4 = sext i16 %s2.elt2 to i32
2498  %mul2 = mul i32 %conv4, %conv3
2499
2500  %s3 = load i32, i32 addrspace(1)* %dst, align 4
2501  %add0 = add i32 %mul2, %s3
2502
2503  %add1 = add i32 %mul2, %add0
2504  %add2 = add i32 %add1, %mul1
2505
2506  store i32 %add2, i32 addrspace(1)* %dst, align 4
2507  ret void
2508}
2509
2510define amdgpu_kernel void @udot2_acc16(<2 x i16> addrspace(1)* %src1,
2511; GFX7-LABEL: udot2_acc16:
2512; GFX7:       ; %bb.0: ; %entry
2513; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
2514; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
2515; GFX7-NEXT:    s_mov_b32 s3, 0xf000
2516; GFX7-NEXT:    s_mov_b32 s2, -1
2517; GFX7-NEXT:    s_mov_b32 s8, 0xffff
2518; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2519; GFX7-NEXT:    buffer_load_ushort v0, off, s[0:3], 0
2520; GFX7-NEXT:    s_load_dword s4, s[4:5], 0x0
2521; GFX7-NEXT:    s_load_dword s5, s[6:7], 0x0
2522; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2523; GFX7-NEXT:    s_lshr_b32 s6, s4, 16
2524; GFX7-NEXT:    s_lshr_b32 s7, s5, 16
2525; GFX7-NEXT:    v_mov_b32_e32 v1, s7
2526; GFX7-NEXT:    s_and_b32 s5, s5, s8
2527; GFX7-NEXT:    s_and_b32 s4, s4, s8
2528; GFX7-NEXT:    s_waitcnt vmcnt(0)
2529; GFX7-NEXT:    v_mad_u32_u24 v0, s6, v1, v0
2530; GFX7-NEXT:    v_mov_b32_e32 v1, s5
2531; GFX7-NEXT:    v_mad_u32_u24 v0, s4, v1, v0
2532; GFX7-NEXT:    buffer_store_short v0, off, s[0:3], 0
2533; GFX7-NEXT:    s_endpgm
2534;
2535; GFX8-LABEL: udot2_acc16:
2536; GFX8:       ; %bb.0: ; %entry
2537; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2538; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2539; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2540; GFX8-NEXT:    v_mov_b32_e32 v0, s0
2541; GFX8-NEXT:    v_mov_b32_e32 v1, s1
2542; GFX8-NEXT:    flat_load_ushort v2, v[0:1]
2543; GFX8-NEXT:    s_load_dword s1, s[4:5], 0x0
2544; GFX8-NEXT:    s_load_dword s2, s[6:7], 0x0
2545; GFX8-NEXT:    s_mov_b32 s0, 0xffff
2546; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2547; GFX8-NEXT:    s_and_b32 s3, s2, s0
2548; GFX8-NEXT:    s_lshr_b32 s2, s2, 16
2549; GFX8-NEXT:    s_and_b32 s0, s1, s0
2550; GFX8-NEXT:    s_lshr_b32 s1, s1, 16
2551; GFX8-NEXT:    v_mov_b32_e32 v3, s2
2552; GFX8-NEXT:    s_waitcnt vmcnt(0)
2553; GFX8-NEXT:    v_mad_u32_u24 v2, s1, v3, v2
2554; GFX8-NEXT:    v_mov_b32_e32 v3, s3
2555; GFX8-NEXT:    v_mad_u32_u24 v2, s0, v3, v2
2556; GFX8-NEXT:    flat_store_short v[0:1], v2
2557; GFX8-NEXT:    s_endpgm
2558;
2559; GFX9-NODL-LABEL: udot2_acc16:
2560; GFX9-NODL:       ; %bb.0: ; %entry
2561; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2562; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2563; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
2564; GFX9-NODL-NEXT:    s_mov_b32 s2, 0xffff
2565; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2566; GFX9-NODL-NEXT:    global_load_ushort v1, v0, s[0:1]
2567; GFX9-NODL-NEXT:    s_load_dword s3, s[4:5], 0x0
2568; GFX9-NODL-NEXT:    s_load_dword s4, s[6:7], 0x0
2569; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2570; GFX9-NODL-NEXT:    s_and_b32 s5, s4, s2
2571; GFX9-NODL-NEXT:    s_lshr_b32 s4, s4, 16
2572; GFX9-NODL-NEXT:    s_and_b32 s2, s3, s2
2573; GFX9-NODL-NEXT:    s_lshr_b32 s3, s3, 16
2574; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s4
2575; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
2576; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s3, v2, v1
2577; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s5
2578; GFX9-NODL-NEXT:    v_mad_u32_u24 v1, s2, v2, v1
2579; GFX9-NODL-NEXT:    global_store_short v0, v1, s[0:1]
2580; GFX9-NODL-NEXT:    s_endpgm
2581;
2582; GFX9-DL-LABEL: udot2_acc16:
2583; GFX9-DL:       ; %bb.0: ; %entry
2584; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2585; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2586; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2587; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2588; GFX9-DL-NEXT:    s_load_dword s2, s[4:5], 0x0
2589; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
2590; GFX9-DL-NEXT:    global_load_ushort v1, v0, s[0:1]
2591; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2592; GFX9-DL-NEXT:    v_mov_b32_e32 v2, s3
2593; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
2594; GFX9-DL-NEXT:    v_dot2_u32_u16 v1, s2, v2, v1
2595; GFX9-DL-NEXT:    global_store_short v0, v1, s[0:1]
2596; GFX9-DL-NEXT:    s_endpgm
2597;
2598; GFX10-DL-LABEL: udot2_acc16:
2599; GFX10-DL:       ; %bb.0: ; %entry
2600; GFX10-DL-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x34
2601; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
2602; GFX10-DL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
2603; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2604; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2605; GFX10-DL-NEXT:    global_load_ushort v1, v0, s[4:5]
2606; GFX10-DL-NEXT:    s_load_dword s0, s[0:1], 0x0
2607; GFX10-DL-NEXT:    s_load_dword s1, s[2:3], 0x0
2608; GFX10-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
2609; GFX10-DL-NEXT:    v_dot2_u32_u16 v1, s0, s1, v1
2610; GFX10-DL-NEXT:    global_store_short v0, v1, s[4:5]
2611; GFX10-DL-NEXT:    s_endpgm
2612                                       <2 x i16> addrspace(1)* %src2,
2613                                       i16 addrspace(1)* nocapture %dst) {
2614entry:
2615  %v1 = load <2 x i16>, <2 x i16> addrspace(1)* %src1
2616  %v2 = load <2 x i16>, <2 x i16> addrspace(1)* %src2
2617
2618  %v1e1 = extractelement <2 x i16> %v1, i64 0
2619  %v2e1 = extractelement <2 x i16> %v2, i64 0
2620  %mul1 = mul i16 %v1e1, %v2e1
2621
2622  %v1e2 = extractelement <2 x i16> %v1, i64 1
2623  %v2e2 = extractelement <2 x i16> %v2, i64 1
2624  %mul2 = mul i16 %v1e2, %v2e2
2625
2626  %s2 = load i16, i16 addrspace(1)* %dst, align 2
2627  %add1 = add i16 %mul2, %s2
2628  %add2 = add i16 %add1, %mul1
2629  store i16 %add2, i16 addrspace(1)* %dst, align 2
2630  ret void
2631}
2632
2633define amdgpu_kernel void @notsdot2_sext8(<2 x i8> addrspace(1)* %src1,
2634; GFX7-LABEL: notsdot2_sext8:
2635; GFX7:       ; %bb.0: ; %entry
2636; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
2637; GFX7-NEXT:    s_load_dwordx2 s[8:9], s[0:1], 0xd
2638; GFX7-NEXT:    s_mov_b32 s3, 0xf000
2639; GFX7-NEXT:    s_mov_b32 s2, -1
2640; GFX7-NEXT:    s_mov_b32 s10, s2
2641; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2642; GFX7-NEXT:    s_mov_b32 s0, s4
2643; GFX7-NEXT:    s_mov_b32 s1, s5
2644; GFX7-NEXT:    s_mov_b32 s4, s6
2645; GFX7-NEXT:    s_mov_b32 s5, s7
2646; GFX7-NEXT:    s_mov_b32 s6, s2
2647; GFX7-NEXT:    s_mov_b32 s7, s3
2648; GFX7-NEXT:    buffer_load_ushort v0, off, s[0:3], 0
2649; GFX7-NEXT:    buffer_load_ushort v1, off, s[4:7], 0
2650; GFX7-NEXT:    s_load_dword s0, s[8:9], 0x0
2651; GFX7-NEXT:    s_mov_b32 s11, s3
2652; GFX7-NEXT:    s_waitcnt vmcnt(1)
2653; GFX7-NEXT:    v_bfe_i32 v2, v0, 0, 8
2654; GFX7-NEXT:    s_waitcnt vmcnt(0)
2655; GFX7-NEXT:    v_bfe_i32 v3, v1, 0, 8
2656; GFX7-NEXT:    v_bfe_i32 v0, v0, 8, 8
2657; GFX7-NEXT:    v_bfe_i32 v1, v1, 8, 8
2658; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
2659; GFX7-NEXT:    v_mad_i32_i24 v0, v1, v0, s0
2660; GFX7-NEXT:    v_mad_i32_i24 v0, v3, v2, v0
2661; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
2662; GFX7-NEXT:    s_endpgm
2663;
2664; GFX8-LABEL: notsdot2_sext8:
2665; GFX8:       ; %bb.0: ; %entry
2666; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2667; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2668; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2669; GFX8-NEXT:    v_mov_b32_e32 v0, s4
2670; GFX8-NEXT:    v_mov_b32_e32 v1, s5
2671; GFX8-NEXT:    v_mov_b32_e32 v2, s6
2672; GFX8-NEXT:    v_mov_b32_e32 v3, s7
2673; GFX8-NEXT:    flat_load_ushort v0, v[0:1]
2674; GFX8-NEXT:    flat_load_ushort v1, v[2:3]
2675; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
2676; GFX8-NEXT:    s_waitcnt vmcnt(1)
2677; GFX8-NEXT:    v_bfe_i32 v2, v0, 0, 8
2678; GFX8-NEXT:    v_lshrrev_b16_e32 v0, 8, v0
2679; GFX8-NEXT:    s_waitcnt vmcnt(0)
2680; GFX8-NEXT:    v_bfe_i32 v3, v1, 0, 8
2681; GFX8-NEXT:    v_lshrrev_b16_e32 v1, 8, v1
2682; GFX8-NEXT:    v_bfe_i32 v0, v0, 0, 8
2683; GFX8-NEXT:    v_bfe_i32 v1, v1, 0, 8
2684; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
2685; GFX8-NEXT:    v_mad_i32_i24 v0, v1, v0, s2
2686; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v2, v0
2687; GFX8-NEXT:    v_mov_b32_e32 v0, s0
2688; GFX8-NEXT:    v_mov_b32_e32 v1, s1
2689; GFX8-NEXT:    flat_store_dword v[0:1], v2
2690; GFX8-NEXT:    s_endpgm
2691;
2692; GFX9-NODL-LABEL: notsdot2_sext8:
2693; GFX9-NODL:       ; %bb.0: ; %entry
2694; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2695; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2696; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
2697; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2698; GFX9-NODL-NEXT:    global_load_ushort v1, v0, s[4:5]
2699; GFX9-NODL-NEXT:    global_load_ushort v2, v0, s[6:7]
2700; GFX9-NODL-NEXT:    s_load_dword s2, s[0:1], 0x0
2701; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
2702; GFX9-NODL-NEXT:    v_bfe_i32 v3, v1, 0, 8
2703; GFX9-NODL-NEXT:    v_lshrrev_b16_e32 v1, 8, v1
2704; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
2705; GFX9-NODL-NEXT:    v_bfe_i32 v4, v2, 0, 8
2706; GFX9-NODL-NEXT:    v_lshrrev_b16_e32 v2, 8, v2
2707; GFX9-NODL-NEXT:    v_bfe_i32 v1, v1, 0, 8
2708; GFX9-NODL-NEXT:    v_bfe_i32 v2, v2, 0, 8
2709; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
2710; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, v2, v1, s2
2711; GFX9-NODL-NEXT:    v_mad_i32_i24 v1, v4, v3, v1
2712; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
2713; GFX9-NODL-NEXT:    s_endpgm
2714;
2715; GFX9-DL-LABEL: notsdot2_sext8:
2716; GFX9-DL:       ; %bb.0: ; %entry
2717; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2718; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2719; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
2720; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2721; GFX9-DL-NEXT:    global_load_ushort v1, v0, s[4:5]
2722; GFX9-DL-NEXT:    global_load_ushort v2, v0, s[6:7]
2723; GFX9-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
2724; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
2725; GFX9-DL-NEXT:    v_bfe_i32 v3, v1, 0, 8
2726; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v1, 8, v1
2727; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
2728; GFX9-DL-NEXT:    v_bfe_i32 v4, v2, 0, 8
2729; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v2, 8, v2
2730; GFX9-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
2731; GFX9-DL-NEXT:    v_bfe_i32 v2, v2, 0, 8
2732; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
2733; GFX9-DL-NEXT:    v_mad_i32_i24 v1, v2, v1, s2
2734; GFX9-DL-NEXT:    v_mad_i32_i24 v1, v4, v3, v1
2735; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2736; GFX9-DL-NEXT:    s_endpgm
2737;
2738; GFX10-DL-LABEL: notsdot2_sext8:
2739; GFX10-DL:       ; %bb.0: ; %entry
2740; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
2741; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
2742; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
2743; GFX10-DL-NEXT:    ; implicit-def: $vcc_hi
2744; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2745; GFX10-DL-NEXT:    s_clause 0x1
2746; GFX10-DL-NEXT:    global_load_ushort v1, v0, s[4:5]
2747; GFX10-DL-NEXT:    global_load_ushort v2, v0, s[6:7]
2748; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
2749; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
2750; GFX10-DL-NEXT:    v_lshrrev_b16_e64 v3, 8, v1
2751; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
2752; GFX10-DL-NEXT:    v_lshrrev_b16_e64 v4, 8, v2
2753; GFX10-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
2754; GFX10-DL-NEXT:    v_bfe_i32 v2, v2, 0, 8
2755; GFX10-DL-NEXT:    v_bfe_i32 v3, v3, 0, 8
2756; GFX10-DL-NEXT:    v_bfe_i32 v4, v4, 0, 8
2757; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
2758; GFX10-DL-NEXT:    v_mad_i32_i24 v3, v4, v3, s2
2759; GFX10-DL-NEXT:    v_mad_i32_i24 v1, v2, v1, v3
2760; GFX10-DL-NEXT:    global_store_dword v0, v1, s[0:1]
2761; GFX10-DL-NEXT:    s_endpgm
2762                                          <2 x i8> addrspace(1)* %src2,
2763                                          i32 addrspace(1)* nocapture %dst) {
2764entry:
2765  %vec1 = load <2 x i8>, <2 x i8> addrspace(1)* %src1
2766  %vec2 = load <2 x i8>, <2 x i8> addrspace(1)* %src2
2767
2768  %s1.elt1 = extractelement <2 x i8> %vec1, i64 0
2769  %conv = sext i8 %s1.elt1 to i32
2770  %s2.elt1 = extractelement <2 x i8> %vec2, i64 0
2771  %conv2 = sext i8 %s2.elt1 to i32
2772  %mul1 = mul nuw i32 %conv2, %conv
2773
2774  %s1.elt2 = extractelement <2 x i8> %vec1, i64 1
2775  %conv3 = sext i8 %s1.elt2 to i32
2776  %s2.elt2 = extractelement <2 x i8> %vec2, i64 1
2777  %conv4 = sext i8 %s2.elt2 to i32
2778  %mul2 = mul nuw i32 %conv4, %conv3
2779
2780  %s3 = load i32, i32 addrspace(1)* %dst, align 4
2781  %add = add i32 %mul2, %s3
2782  %add6 = add i32 %add, %mul1
2783  store i32 %add6, i32 addrspace(1)* %dst, align 4
2784  ret void
2785}
2786