1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GFX6 %s
3; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
4; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
5
6define amdgpu_ps <4 x float> @load_2d_v4f32_xyzw(<8 x i32> inreg %rsrc, i32 %s, i32 %t) {
7; GFX6-LABEL: load_2d_v4f32_xyzw:
8; GFX6:       ; %bb.0:
9; GFX6-NEXT:    s_mov_b32 s0, s2
10; GFX6-NEXT:    s_mov_b32 s1, s3
11; GFX6-NEXT:    s_mov_b32 s2, s4
12; GFX6-NEXT:    s_mov_b32 s3, s5
13; GFX6-NEXT:    s_mov_b32 s4, s6
14; GFX6-NEXT:    s_mov_b32 s5, s7
15; GFX6-NEXT:    s_mov_b32 s6, s8
16; GFX6-NEXT:    s_mov_b32 s7, s9
17; GFX6-NEXT:    image_load v[0:3], v[0:1], s[0:7] dmask:0xf unorm
18; GFX6-NEXT:    s_waitcnt vmcnt(0)
19; GFX6-NEXT:    ; return to shader part epilog
20;
21; GFX10PLUS-LABEL: load_2d_v4f32_xyzw:
22; GFX10PLUS:       ; %bb.0:
23; GFX10PLUS-NEXT:    s_mov_b32 s0, s2
24; GFX10PLUS-NEXT:    s_mov_b32 s1, s3
25; GFX10PLUS-NEXT:    s_mov_b32 s2, s4
26; GFX10PLUS-NEXT:    s_mov_b32 s3, s5
27; GFX10PLUS-NEXT:    s_mov_b32 s4, s6
28; GFX10PLUS-NEXT:    s_mov_b32 s5, s7
29; GFX10PLUS-NEXT:    s_mov_b32 s6, s8
30; GFX10PLUS-NEXT:    s_mov_b32 s7, s9
31; GFX10PLUS-NEXT:    image_load v[0:3], v[0:1], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
32; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
33; GFX10PLUS-NEXT:    ; return to shader part epilog
34  %v = call <4 x float> @llvm.amdgcn.image.load.2d.v4f32.i32(i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 0, i32 0)
35  ret <4 x float> %v
36}
37
38define amdgpu_ps <4 x float> @load_2d_v4f32_xyzw_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t) {
39; GFX6-LABEL: load_2d_v4f32_xyzw_tfe:
40; GFX6:       ; %bb.0:
41; GFX6-NEXT:    v_mov_b32_e32 v5, v0
42; GFX6-NEXT:    v_mov_b32_e32 v0, 0
43; GFX6-NEXT:    s_mov_b32 s0, s2
44; GFX6-NEXT:    s_mov_b32 s1, s3
45; GFX6-NEXT:    s_mov_b32 s2, s4
46; GFX6-NEXT:    s_mov_b32 s3, s5
47; GFX6-NEXT:    s_mov_b32 s4, s6
48; GFX6-NEXT:    s_mov_b32 s5, s7
49; GFX6-NEXT:    s_mov_b32 s6, s8
50; GFX6-NEXT:    s_mov_b32 s7, s9
51; GFX6-NEXT:    v_mov_b32_e32 v6, v1
52; GFX6-NEXT:    v_mov_b32_e32 v1, v0
53; GFX6-NEXT:    v_mov_b32_e32 v2, v0
54; GFX6-NEXT:    v_mov_b32_e32 v3, v0
55; GFX6-NEXT:    v_mov_b32_e32 v4, v0
56; GFX6-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf unorm tfe
57; GFX6-NEXT:    s_mov_b32 s8, s10
58; GFX6-NEXT:    s_mov_b32 s9, s11
59; GFX6-NEXT:    s_mov_b32 s10, -1
60; GFX6-NEXT:    s_mov_b32 s11, 0xf000
61; GFX6-NEXT:    s_waitcnt vmcnt(0)
62; GFX6-NEXT:    buffer_store_dword v4, off, s[8:11], 0
63; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
64; GFX6-NEXT:    ; return to shader part epilog
65;
66; GFX10-LABEL: load_2d_v4f32_xyzw_tfe:
67; GFX10:       ; %bb.0:
68; GFX10-NEXT:    v_mov_b32_e32 v7, 0
69; GFX10-NEXT:    v_mov_b32_e32 v5, v0
70; GFX10-NEXT:    v_mov_b32_e32 v6, v1
71; GFX10-NEXT:    s_mov_b32 s0, s2
72; GFX10-NEXT:    s_mov_b32 s1, s3
73; GFX10-NEXT:    v_mov_b32_e32 v8, v7
74; GFX10-NEXT:    v_mov_b32_e32 v9, v7
75; GFX10-NEXT:    v_mov_b32_e32 v10, v7
76; GFX10-NEXT:    v_mov_b32_e32 v11, v7
77; GFX10-NEXT:    s_mov_b32 s2, s4
78; GFX10-NEXT:    s_mov_b32 s3, s5
79; GFX10-NEXT:    s_mov_b32 s4, s6
80; GFX10-NEXT:    s_mov_b32 s5, s7
81; GFX10-NEXT:    s_mov_b32 s6, s8
82; GFX10-NEXT:    s_mov_b32 s7, s9
83; GFX10-NEXT:    v_mov_b32_e32 v0, v7
84; GFX10-NEXT:    v_mov_b32_e32 v1, v8
85; GFX10-NEXT:    v_mov_b32_e32 v2, v9
86; GFX10-NEXT:    v_mov_b32_e32 v3, v10
87; GFX10-NEXT:    v_mov_b32_e32 v4, v11
88; GFX10-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe
89; GFX10-NEXT:    s_waitcnt vmcnt(0)
90; GFX10-NEXT:    global_store_dword v7, v4, s[10:11]
91; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
92; GFX10-NEXT:    ; return to shader part epilog
93;
94; GFX11-LABEL: load_2d_v4f32_xyzw_tfe:
95; GFX11:       ; %bb.0:
96; GFX11-NEXT:    v_mov_b32_e32 v7, 0
97; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
98; GFX11-NEXT:    s_mov_b32 s0, s2
99; GFX11-NEXT:    s_mov_b32 s1, s3
100; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
101; GFX11-NEXT:    v_mov_b32_e32 v9, v7
102; GFX11-NEXT:    v_mov_b32_e32 v11, v7
103; GFX11-NEXT:    v_mov_b32_e32 v10, v7
104; GFX11-NEXT:    v_mov_b32_e32 v8, v7
105; GFX11-NEXT:    s_mov_b32 s2, s4
106; GFX11-NEXT:    s_mov_b32 s3, s5
107; GFX11-NEXT:    s_mov_b32 s4, s6
108; GFX11-NEXT:    s_mov_b32 s5, s7
109; GFX11-NEXT:    s_mov_b32 s6, s8
110; GFX11-NEXT:    s_mov_b32 s7, s9
111; GFX11-NEXT:    v_mov_b32_e32 v0, v7
112; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
113; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
114; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe
115; GFX11-NEXT:    s_waitcnt vmcnt(0)
116; GFX11-NEXT:    global_store_b32 v7, v4, s[10:11]
117; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
118; GFX11-NEXT:    ; return to shader part epilog
119  %v = call { <4 x float>, i32 } @llvm.amdgcn.image.load.2d.sl_v4f32i32s.i32(i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 1, i32 0)
120  %v.vec = extractvalue { <4 x float>, i32 } %v, 0
121  %v.err = extractvalue { <4 x float>, i32 } %v, 1
122  store i32 %v.err, i32 addrspace(1)* %out, align 4
123  ret <4 x float> %v.vec
124}
125
126define amdgpu_ps <4 x float> @load_2d_v4f32_xyzw_tfe_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t) {
127; GFX6-LABEL: load_2d_v4f32_xyzw_tfe_lwe:
128; GFX6:       ; %bb.0:
129; GFX6-NEXT:    v_mov_b32_e32 v5, v0
130; GFX6-NEXT:    v_mov_b32_e32 v0, 0
131; GFX6-NEXT:    s_mov_b32 s0, s2
132; GFX6-NEXT:    s_mov_b32 s1, s3
133; GFX6-NEXT:    s_mov_b32 s2, s4
134; GFX6-NEXT:    s_mov_b32 s3, s5
135; GFX6-NEXT:    s_mov_b32 s4, s6
136; GFX6-NEXT:    s_mov_b32 s5, s7
137; GFX6-NEXT:    s_mov_b32 s6, s8
138; GFX6-NEXT:    s_mov_b32 s7, s9
139; GFX6-NEXT:    v_mov_b32_e32 v6, v1
140; GFX6-NEXT:    v_mov_b32_e32 v1, v0
141; GFX6-NEXT:    v_mov_b32_e32 v2, v0
142; GFX6-NEXT:    v_mov_b32_e32 v3, v0
143; GFX6-NEXT:    v_mov_b32_e32 v4, v0
144; GFX6-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf unorm tfe lwe
145; GFX6-NEXT:    s_mov_b32 s8, s10
146; GFX6-NEXT:    s_mov_b32 s9, s11
147; GFX6-NEXT:    s_mov_b32 s10, -1
148; GFX6-NEXT:    s_mov_b32 s11, 0xf000
149; GFX6-NEXT:    s_waitcnt vmcnt(0)
150; GFX6-NEXT:    buffer_store_dword v4, off, s[8:11], 0
151; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
152; GFX6-NEXT:    ; return to shader part epilog
153;
154; GFX10-LABEL: load_2d_v4f32_xyzw_tfe_lwe:
155; GFX10:       ; %bb.0:
156; GFX10-NEXT:    v_mov_b32_e32 v7, 0
157; GFX10-NEXT:    v_mov_b32_e32 v5, v0
158; GFX10-NEXT:    v_mov_b32_e32 v6, v1
159; GFX10-NEXT:    s_mov_b32 s0, s2
160; GFX10-NEXT:    s_mov_b32 s1, s3
161; GFX10-NEXT:    v_mov_b32_e32 v8, v7
162; GFX10-NEXT:    v_mov_b32_e32 v9, v7
163; GFX10-NEXT:    v_mov_b32_e32 v10, v7
164; GFX10-NEXT:    v_mov_b32_e32 v11, v7
165; GFX10-NEXT:    s_mov_b32 s2, s4
166; GFX10-NEXT:    s_mov_b32 s3, s5
167; GFX10-NEXT:    s_mov_b32 s4, s6
168; GFX10-NEXT:    s_mov_b32 s5, s7
169; GFX10-NEXT:    s_mov_b32 s6, s8
170; GFX10-NEXT:    s_mov_b32 s7, s9
171; GFX10-NEXT:    v_mov_b32_e32 v0, v7
172; GFX10-NEXT:    v_mov_b32_e32 v1, v8
173; GFX10-NEXT:    v_mov_b32_e32 v2, v9
174; GFX10-NEXT:    v_mov_b32_e32 v3, v10
175; GFX10-NEXT:    v_mov_b32_e32 v4, v11
176; GFX10-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe lwe
177; GFX10-NEXT:    s_waitcnt vmcnt(0)
178; GFX10-NEXT:    global_store_dword v7, v4, s[10:11]
179; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
180; GFX10-NEXT:    ; return to shader part epilog
181;
182; GFX11-LABEL: load_2d_v4f32_xyzw_tfe_lwe:
183; GFX11:       ; %bb.0:
184; GFX11-NEXT:    v_mov_b32_e32 v7, 0
185; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
186; GFX11-NEXT:    s_mov_b32 s0, s2
187; GFX11-NEXT:    s_mov_b32 s1, s3
188; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
189; GFX11-NEXT:    v_mov_b32_e32 v9, v7
190; GFX11-NEXT:    v_mov_b32_e32 v11, v7
191; GFX11-NEXT:    v_mov_b32_e32 v10, v7
192; GFX11-NEXT:    v_mov_b32_e32 v8, v7
193; GFX11-NEXT:    s_mov_b32 s2, s4
194; GFX11-NEXT:    s_mov_b32 s3, s5
195; GFX11-NEXT:    s_mov_b32 s4, s6
196; GFX11-NEXT:    s_mov_b32 s5, s7
197; GFX11-NEXT:    s_mov_b32 s6, s8
198; GFX11-NEXT:    s_mov_b32 s7, s9
199; GFX11-NEXT:    v_mov_b32_e32 v0, v7
200; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
201; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
202; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe lwe
203; GFX11-NEXT:    s_waitcnt vmcnt(0)
204; GFX11-NEXT:    global_store_b32 v7, v4, s[10:11]
205; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
206; GFX11-NEXT:    ; return to shader part epilog
207  %v = call { <4 x float>, i32 } @llvm.amdgcn.image.load.2d.sl_v4f32i32s.i32(i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 3, i32 0)
208  %v.vec = extractvalue { <4 x float>, i32 } %v, 0
209  %v.err = extractvalue { <4 x float>, i32 } %v, 1
210  store i32 %v.err, i32 addrspace(1)* %out, align 4
211  ret <4 x float> %v.vec
212}
213
214declare <4 x float> @llvm.amdgcn.image.load.2d.v4f32.i32(i32 immarg, i32, i32, <8 x i32>, i32 immarg, i32 immarg) #0
215declare { <4 x float>, i32 } @llvm.amdgcn.image.load.2d.sl_v4f32i32s.i32(i32 immarg, i32, i32, <8 x i32>, i32 immarg, i32 immarg) #0
216
217attributes #0 = { nounwind readonly }
218