xref: /llvm-project/llvm/test/CodeGen/AMDGPU/spill-scavenge-offset.ll (revision 6548b6354d1d990e1c98736f5e7c3de876bedc8e)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn -mcpu=verde -enable-misched=0 -post-RA-scheduler=0 -amdgpu-spill-sgpr-to-vgpr=0 < %s | FileCheck -check-prefixes=CHECK,GFX6 %s
3; RUN: llc -sgpr-regalloc=basic -vgpr-regalloc=basic -mtriple=amdgcn -mcpu=tonga -enable-misched=0 -post-RA-scheduler=0 -amdgpu-spill-sgpr-to-vgpr=0 < %s | FileCheck --check-prefix=CHECK %s
4; RUN: llc -mtriple=amdgcn -mattr=-xnack,+enable-flat-scratch -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -amdgpu-spill-sgpr-to-vgpr=0 < %s | FileCheck -check-prefixes=CHECK,GFX9-FLATSCR,FLATSCR %s
5; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -enable-misched=0 -post-RA-scheduler=0 -amdgpu-spill-sgpr-to-vgpr=0 -mattr=+enable-flat-scratch < %s | FileCheck -check-prefixes=CHECK,GFX10-FLATSCR,FLATSCR %s
6;
7; There is something about Tonga that causes this test to spend a lot of time
8; in the default register allocator.
9
10
11; When the offset of VGPR spills into scratch space gets too large, an additional SGPR
12; is used to calculate the scratch load/store address. Make sure that this
13; mechanism works even when many spills happen.
14
15; Just test that it compiles successfully.
16define amdgpu_kernel void @test(ptr addrspace(1) %out, ptr addrspace(1) %in) {
17; GFX6-LABEL: test:
18; GFX6:       ; %bb.0: ; %entry
19; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
20; GFX6-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, -1, 0
21; GFX6-NEXT:    v_mbcnt_hi_u32_b32_e32 v0, -1, v0
22; GFX6-NEXT:    v_lshlrev_b32_e32 v5, 13, v0
23; GFX6-NEXT:    s_mov_b32 s6, 0
24; GFX6-NEXT:    v_mov_b32_e32 v6, 0
25; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
26; GFX6-NEXT:    v_add_i32_e32 v7, vcc, s2, v5
27; GFX6-NEXT:    v_mov_b32_e32 v0, s3
28; GFX6-NEXT:    s_mov_b32 s7, 0xf000
29; GFX6-NEXT:    s_mov_b64 s[4:5], s[2:3]
30; GFX6-NEXT:    v_addc_u32_e32 v8, vcc, 0, v0, vcc
31; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64
32; GFX6-NEXT:    s_mov_b32 s40, SCRATCH_RSRC_DWORD0
33; GFX6-NEXT:    s_mov_b32 s41, SCRATCH_RSRC_DWORD1
34; GFX6-NEXT:    s_mov_b32 s42, -1
35; GFX6-NEXT:    s_mov_b32 s43, 0xe8f000
36; GFX6-NEXT:    s_add_u32 s40, s40, s11
37; GFX6-NEXT:    s_addc_u32 s41, s41, 0
38; GFX6-NEXT:    s_mov_b32 s2, 0x3fd00
39; GFX6-NEXT:    s_mov_b64 s[8:9], 0x100
40; GFX6-NEXT:    s_mov_b64 s[10:11], s[6:7]
41; GFX6-NEXT:    s_mov_b64 s[12:13], 0x180
42; GFX6-NEXT:    s_mov_b64 s[14:15], s[6:7]
43; GFX6-NEXT:    s_mov_b64 s[16:17], 0x200
44; GFX6-NEXT:    s_mov_b64 s[18:19], s[6:7]
45; GFX6-NEXT:    s_mov_b64 s[20:21], 0x280
46; GFX6-NEXT:    s_mov_b64 s[22:23], s[6:7]
47; GFX6-NEXT:    s_mov_b64 s[24:25], 0x300
48; GFX6-NEXT:    s_mov_b64 s[26:27], s[6:7]
49; GFX6-NEXT:    s_mov_b64 s[28:29], 0x380
50; GFX6-NEXT:    s_mov_b64 s[30:31], s[6:7]
51; GFX6-NEXT:    s_mov_b64 s[36:37], 0x400
52; GFX6-NEXT:    s_mov_b64 s[38:39], s[6:7]
53; GFX6-NEXT:    s_mov_b32 s33, 0x4f900
54; GFX6-NEXT:    s_waitcnt vmcnt(0)
55; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4 ; 4-byte Folded Spill
56; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:8 ; 4-byte Folded Spill
57; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:12 ; 4-byte Folded Spill
58; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:16 ; 4-byte Folded Spill
59; GFX6-NEXT:    s_waitcnt expcnt(0)
60; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:16
61; GFX6-NEXT:    s_waitcnt vmcnt(0)
62; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:20 ; 4-byte Folded Spill
63; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:24 ; 4-byte Folded Spill
64; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:28 ; 4-byte Folded Spill
65; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:32 ; 4-byte Folded Spill
66; GFX6-NEXT:    s_waitcnt expcnt(0)
67; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:32
68; GFX6-NEXT:    s_waitcnt vmcnt(0)
69; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:36 ; 4-byte Folded Spill
70; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:40 ; 4-byte Folded Spill
71; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:44 ; 4-byte Folded Spill
72; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:48 ; 4-byte Folded Spill
73; GFX6-NEXT:    s_waitcnt expcnt(0)
74; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:48
75; GFX6-NEXT:    s_waitcnt vmcnt(0)
76; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:52 ; 4-byte Folded Spill
77; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:56 ; 4-byte Folded Spill
78; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:60 ; 4-byte Folded Spill
79; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:64 ; 4-byte Folded Spill
80; GFX6-NEXT:    s_waitcnt expcnt(0)
81; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:64
82; GFX6-NEXT:    s_waitcnt vmcnt(0)
83; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:68 ; 4-byte Folded Spill
84; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:72 ; 4-byte Folded Spill
85; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:76 ; 4-byte Folded Spill
86; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:80 ; 4-byte Folded Spill
87; GFX6-NEXT:    s_waitcnt expcnt(0)
88; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:80
89; GFX6-NEXT:    s_waitcnt vmcnt(0)
90; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:84 ; 4-byte Folded Spill
91; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:88 ; 4-byte Folded Spill
92; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:92 ; 4-byte Folded Spill
93; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:96 ; 4-byte Folded Spill
94; GFX6-NEXT:    s_waitcnt expcnt(0)
95; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:96
96; GFX6-NEXT:    s_waitcnt vmcnt(0)
97; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:100 ; 4-byte Folded Spill
98; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:104 ; 4-byte Folded Spill
99; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:108 ; 4-byte Folded Spill
100; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:112 ; 4-byte Folded Spill
101; GFX6-NEXT:    s_waitcnt expcnt(0)
102; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:112
103; GFX6-NEXT:    s_waitcnt vmcnt(0)
104; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:116 ; 4-byte Folded Spill
105; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:120 ; 4-byte Folded Spill
106; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:124 ; 4-byte Folded Spill
107; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:128 ; 4-byte Folded Spill
108; GFX6-NEXT:    s_waitcnt expcnt(0)
109; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:128
110; GFX6-NEXT:    s_waitcnt vmcnt(0)
111; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:132 ; 4-byte Folded Spill
112; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:136 ; 4-byte Folded Spill
113; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:140 ; 4-byte Folded Spill
114; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:144 ; 4-byte Folded Spill
115; GFX6-NEXT:    s_waitcnt expcnt(0)
116; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:144
117; GFX6-NEXT:    s_waitcnt vmcnt(0)
118; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:148 ; 4-byte Folded Spill
119; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:152 ; 4-byte Folded Spill
120; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:156 ; 4-byte Folded Spill
121; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:160 ; 4-byte Folded Spill
122; GFX6-NEXT:    s_waitcnt expcnt(0)
123; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:160
124; GFX6-NEXT:    s_waitcnt vmcnt(0)
125; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:164 ; 4-byte Folded Spill
126; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:168 ; 4-byte Folded Spill
127; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:172 ; 4-byte Folded Spill
128; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:176 ; 4-byte Folded Spill
129; GFX6-NEXT:    s_waitcnt expcnt(0)
130; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:176
131; GFX6-NEXT:    s_waitcnt vmcnt(0)
132; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:180 ; 4-byte Folded Spill
133; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:184 ; 4-byte Folded Spill
134; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:188 ; 4-byte Folded Spill
135; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:192 ; 4-byte Folded Spill
136; GFX6-NEXT:    s_waitcnt expcnt(0)
137; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:192
138; GFX6-NEXT:    s_waitcnt vmcnt(0)
139; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:196 ; 4-byte Folded Spill
140; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:200 ; 4-byte Folded Spill
141; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:204 ; 4-byte Folded Spill
142; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:208 ; 4-byte Folded Spill
143; GFX6-NEXT:    s_waitcnt expcnt(0)
144; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:208
145; GFX6-NEXT:    s_waitcnt vmcnt(0)
146; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:212 ; 4-byte Folded Spill
147; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:216 ; 4-byte Folded Spill
148; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:220 ; 4-byte Folded Spill
149; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:224 ; 4-byte Folded Spill
150; GFX6-NEXT:    s_waitcnt expcnt(0)
151; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:224
152; GFX6-NEXT:    s_waitcnt vmcnt(0)
153; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:228 ; 4-byte Folded Spill
154; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:232 ; 4-byte Folded Spill
155; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:236 ; 4-byte Folded Spill
156; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:240 ; 4-byte Folded Spill
157; GFX6-NEXT:    s_waitcnt expcnt(0)
158; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:240
159; GFX6-NEXT:    s_waitcnt vmcnt(0)
160; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:244 ; 4-byte Folded Spill
161; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:248 ; 4-byte Folded Spill
162; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:252 ; 4-byte Folded Spill
163; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:256 ; 4-byte Folded Spill
164; GFX6-NEXT:    s_waitcnt expcnt(0)
165; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:256
166; GFX6-NEXT:    s_waitcnt vmcnt(0)
167; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:260 ; 4-byte Folded Spill
168; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:264 ; 4-byte Folded Spill
169; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:268 ; 4-byte Folded Spill
170; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:272 ; 4-byte Folded Spill
171; GFX6-NEXT:    s_waitcnt expcnt(0)
172; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:272
173; GFX6-NEXT:    s_waitcnt vmcnt(0)
174; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:276 ; 4-byte Folded Spill
175; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:280 ; 4-byte Folded Spill
176; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:284 ; 4-byte Folded Spill
177; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:288 ; 4-byte Folded Spill
178; GFX6-NEXT:    s_waitcnt expcnt(0)
179; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:288
180; GFX6-NEXT:    s_waitcnt vmcnt(0)
181; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:292 ; 4-byte Folded Spill
182; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:296 ; 4-byte Folded Spill
183; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:300 ; 4-byte Folded Spill
184; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:304 ; 4-byte Folded Spill
185; GFX6-NEXT:    s_waitcnt expcnt(0)
186; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:304
187; GFX6-NEXT:    s_waitcnt vmcnt(0)
188; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:308 ; 4-byte Folded Spill
189; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:312 ; 4-byte Folded Spill
190; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:316 ; 4-byte Folded Spill
191; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:320 ; 4-byte Folded Spill
192; GFX6-NEXT:    s_waitcnt expcnt(0)
193; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:320
194; GFX6-NEXT:    s_waitcnt vmcnt(0)
195; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:324 ; 4-byte Folded Spill
196; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:328 ; 4-byte Folded Spill
197; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:332 ; 4-byte Folded Spill
198; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:336 ; 4-byte Folded Spill
199; GFX6-NEXT:    s_waitcnt expcnt(0)
200; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:336
201; GFX6-NEXT:    s_waitcnt vmcnt(0)
202; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:340 ; 4-byte Folded Spill
203; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:344 ; 4-byte Folded Spill
204; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:348 ; 4-byte Folded Spill
205; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:352 ; 4-byte Folded Spill
206; GFX6-NEXT:    s_waitcnt expcnt(0)
207; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:352
208; GFX6-NEXT:    s_waitcnt vmcnt(0)
209; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:356 ; 4-byte Folded Spill
210; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:360 ; 4-byte Folded Spill
211; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:364 ; 4-byte Folded Spill
212; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:368 ; 4-byte Folded Spill
213; GFX6-NEXT:    s_waitcnt expcnt(0)
214; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:368
215; GFX6-NEXT:    s_waitcnt vmcnt(0)
216; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:372 ; 4-byte Folded Spill
217; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:376 ; 4-byte Folded Spill
218; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:380 ; 4-byte Folded Spill
219; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:384 ; 4-byte Folded Spill
220; GFX6-NEXT:    s_waitcnt expcnt(0)
221; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:384
222; GFX6-NEXT:    s_waitcnt vmcnt(0)
223; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:388 ; 4-byte Folded Spill
224; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:392 ; 4-byte Folded Spill
225; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:396 ; 4-byte Folded Spill
226; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:400 ; 4-byte Folded Spill
227; GFX6-NEXT:    s_waitcnt expcnt(0)
228; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:400
229; GFX6-NEXT:    s_waitcnt vmcnt(0)
230; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:404 ; 4-byte Folded Spill
231; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:408 ; 4-byte Folded Spill
232; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:412 ; 4-byte Folded Spill
233; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:416 ; 4-byte Folded Spill
234; GFX6-NEXT:    s_waitcnt expcnt(0)
235; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:416
236; GFX6-NEXT:    s_waitcnt vmcnt(0)
237; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:420 ; 4-byte Folded Spill
238; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:424 ; 4-byte Folded Spill
239; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:428 ; 4-byte Folded Spill
240; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:432 ; 4-byte Folded Spill
241; GFX6-NEXT:    s_waitcnt expcnt(0)
242; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:432
243; GFX6-NEXT:    s_waitcnt vmcnt(0)
244; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:436 ; 4-byte Folded Spill
245; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:440 ; 4-byte Folded Spill
246; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:444 ; 4-byte Folded Spill
247; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:448 ; 4-byte Folded Spill
248; GFX6-NEXT:    s_waitcnt expcnt(0)
249; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:448
250; GFX6-NEXT:    s_waitcnt vmcnt(0)
251; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:452 ; 4-byte Folded Spill
252; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:456 ; 4-byte Folded Spill
253; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:460 ; 4-byte Folded Spill
254; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:464 ; 4-byte Folded Spill
255; GFX6-NEXT:    s_waitcnt expcnt(0)
256; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:464
257; GFX6-NEXT:    s_waitcnt vmcnt(0)
258; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:468 ; 4-byte Folded Spill
259; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:472 ; 4-byte Folded Spill
260; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:476 ; 4-byte Folded Spill
261; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:480 ; 4-byte Folded Spill
262; GFX6-NEXT:    s_waitcnt expcnt(0)
263; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:480
264; GFX6-NEXT:    s_waitcnt vmcnt(0)
265; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:484 ; 4-byte Folded Spill
266; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:488 ; 4-byte Folded Spill
267; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:492 ; 4-byte Folded Spill
268; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:496 ; 4-byte Folded Spill
269; GFX6-NEXT:    s_waitcnt expcnt(0)
270; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:496
271; GFX6-NEXT:    s_waitcnt vmcnt(0)
272; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:500 ; 4-byte Folded Spill
273; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:504 ; 4-byte Folded Spill
274; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:508 ; 4-byte Folded Spill
275; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:512 ; 4-byte Folded Spill
276; GFX6-NEXT:    s_waitcnt expcnt(0)
277; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:512
278; GFX6-NEXT:    s_waitcnt vmcnt(0)
279; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:516 ; 4-byte Folded Spill
280; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:520 ; 4-byte Folded Spill
281; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:524 ; 4-byte Folded Spill
282; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:528 ; 4-byte Folded Spill
283; GFX6-NEXT:    s_waitcnt expcnt(0)
284; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:528
285; GFX6-NEXT:    s_waitcnt vmcnt(0)
286; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:532 ; 4-byte Folded Spill
287; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:536 ; 4-byte Folded Spill
288; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:540 ; 4-byte Folded Spill
289; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:544 ; 4-byte Folded Spill
290; GFX6-NEXT:    s_waitcnt expcnt(0)
291; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:544
292; GFX6-NEXT:    s_waitcnt vmcnt(0)
293; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:548 ; 4-byte Folded Spill
294; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:552 ; 4-byte Folded Spill
295; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:556 ; 4-byte Folded Spill
296; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:560 ; 4-byte Folded Spill
297; GFX6-NEXT:    s_waitcnt expcnt(0)
298; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:560
299; GFX6-NEXT:    s_waitcnt vmcnt(0)
300; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:564 ; 4-byte Folded Spill
301; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:568 ; 4-byte Folded Spill
302; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:572 ; 4-byte Folded Spill
303; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:576 ; 4-byte Folded Spill
304; GFX6-NEXT:    s_waitcnt expcnt(0)
305; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:576
306; GFX6-NEXT:    s_waitcnt vmcnt(0)
307; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:580 ; 4-byte Folded Spill
308; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:584 ; 4-byte Folded Spill
309; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:588 ; 4-byte Folded Spill
310; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:592 ; 4-byte Folded Spill
311; GFX6-NEXT:    s_waitcnt expcnt(0)
312; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:592
313; GFX6-NEXT:    s_waitcnt vmcnt(0)
314; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:596 ; 4-byte Folded Spill
315; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:600 ; 4-byte Folded Spill
316; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:604 ; 4-byte Folded Spill
317; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:608 ; 4-byte Folded Spill
318; GFX6-NEXT:    s_waitcnt expcnt(0)
319; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:608
320; GFX6-NEXT:    s_waitcnt vmcnt(0)
321; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:612 ; 4-byte Folded Spill
322; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:616 ; 4-byte Folded Spill
323; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:620 ; 4-byte Folded Spill
324; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:624 ; 4-byte Folded Spill
325; GFX6-NEXT:    s_waitcnt expcnt(0)
326; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:624
327; GFX6-NEXT:    s_waitcnt vmcnt(0)
328; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:628 ; 4-byte Folded Spill
329; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:632 ; 4-byte Folded Spill
330; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:636 ; 4-byte Folded Spill
331; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:640 ; 4-byte Folded Spill
332; GFX6-NEXT:    s_waitcnt expcnt(0)
333; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:640
334; GFX6-NEXT:    s_waitcnt vmcnt(0)
335; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:644 ; 4-byte Folded Spill
336; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:648 ; 4-byte Folded Spill
337; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:652 ; 4-byte Folded Spill
338; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:656 ; 4-byte Folded Spill
339; GFX6-NEXT:    s_waitcnt expcnt(0)
340; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:656
341; GFX6-NEXT:    s_waitcnt vmcnt(0)
342; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:660 ; 4-byte Folded Spill
343; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:664 ; 4-byte Folded Spill
344; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:668 ; 4-byte Folded Spill
345; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:672 ; 4-byte Folded Spill
346; GFX6-NEXT:    s_waitcnt expcnt(0)
347; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:672
348; GFX6-NEXT:    s_waitcnt vmcnt(0)
349; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:676 ; 4-byte Folded Spill
350; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:680 ; 4-byte Folded Spill
351; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:684 ; 4-byte Folded Spill
352; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:688 ; 4-byte Folded Spill
353; GFX6-NEXT:    s_waitcnt expcnt(0)
354; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:688
355; GFX6-NEXT:    s_waitcnt vmcnt(0)
356; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:692 ; 4-byte Folded Spill
357; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:696 ; 4-byte Folded Spill
358; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:700 ; 4-byte Folded Spill
359; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:704 ; 4-byte Folded Spill
360; GFX6-NEXT:    s_waitcnt expcnt(0)
361; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:704
362; GFX6-NEXT:    s_waitcnt vmcnt(0)
363; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:708 ; 4-byte Folded Spill
364; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:712 ; 4-byte Folded Spill
365; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:716 ; 4-byte Folded Spill
366; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:720 ; 4-byte Folded Spill
367; GFX6-NEXT:    s_waitcnt expcnt(0)
368; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:720
369; GFX6-NEXT:    s_waitcnt vmcnt(0)
370; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:724 ; 4-byte Folded Spill
371; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:728 ; 4-byte Folded Spill
372; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:732 ; 4-byte Folded Spill
373; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:736 ; 4-byte Folded Spill
374; GFX6-NEXT:    s_waitcnt expcnt(0)
375; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:736
376; GFX6-NEXT:    s_waitcnt vmcnt(0)
377; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:740 ; 4-byte Folded Spill
378; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:744 ; 4-byte Folded Spill
379; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:748 ; 4-byte Folded Spill
380; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:752 ; 4-byte Folded Spill
381; GFX6-NEXT:    s_waitcnt expcnt(0)
382; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:752
383; GFX6-NEXT:    s_waitcnt vmcnt(0)
384; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:756 ; 4-byte Folded Spill
385; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:760 ; 4-byte Folded Spill
386; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:764 ; 4-byte Folded Spill
387; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:768 ; 4-byte Folded Spill
388; GFX6-NEXT:    s_waitcnt expcnt(0)
389; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:768
390; GFX6-NEXT:    s_waitcnt vmcnt(0)
391; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:772 ; 4-byte Folded Spill
392; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:776 ; 4-byte Folded Spill
393; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:780 ; 4-byte Folded Spill
394; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:784 ; 4-byte Folded Spill
395; GFX6-NEXT:    s_waitcnt expcnt(0)
396; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:784
397; GFX6-NEXT:    s_waitcnt vmcnt(0)
398; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:788 ; 4-byte Folded Spill
399; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:792 ; 4-byte Folded Spill
400; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:796 ; 4-byte Folded Spill
401; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:800 ; 4-byte Folded Spill
402; GFX6-NEXT:    s_waitcnt expcnt(0)
403; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:800
404; GFX6-NEXT:    s_waitcnt vmcnt(0)
405; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:804 ; 4-byte Folded Spill
406; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:808 ; 4-byte Folded Spill
407; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:812 ; 4-byte Folded Spill
408; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:816 ; 4-byte Folded Spill
409; GFX6-NEXT:    s_waitcnt expcnt(0)
410; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:816
411; GFX6-NEXT:    s_waitcnt vmcnt(0)
412; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:820 ; 4-byte Folded Spill
413; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:824 ; 4-byte Folded Spill
414; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:828 ; 4-byte Folded Spill
415; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:832 ; 4-byte Folded Spill
416; GFX6-NEXT:    s_waitcnt expcnt(0)
417; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:832
418; GFX6-NEXT:    s_waitcnt vmcnt(0)
419; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:836 ; 4-byte Folded Spill
420; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:840 ; 4-byte Folded Spill
421; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:844 ; 4-byte Folded Spill
422; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:848 ; 4-byte Folded Spill
423; GFX6-NEXT:    s_waitcnt expcnt(0)
424; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:848
425; GFX6-NEXT:    s_waitcnt vmcnt(0)
426; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:852 ; 4-byte Folded Spill
427; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:856 ; 4-byte Folded Spill
428; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:860 ; 4-byte Folded Spill
429; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:864 ; 4-byte Folded Spill
430; GFX6-NEXT:    s_waitcnt expcnt(0)
431; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:864
432; GFX6-NEXT:    s_waitcnt vmcnt(0)
433; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:868 ; 4-byte Folded Spill
434; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:872 ; 4-byte Folded Spill
435; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:876 ; 4-byte Folded Spill
436; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:880 ; 4-byte Folded Spill
437; GFX6-NEXT:    s_waitcnt expcnt(0)
438; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:880
439; GFX6-NEXT:    s_waitcnt vmcnt(0)
440; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:884 ; 4-byte Folded Spill
441; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:888 ; 4-byte Folded Spill
442; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:892 ; 4-byte Folded Spill
443; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:896 ; 4-byte Folded Spill
444; GFX6-NEXT:    s_waitcnt expcnt(0)
445; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:896
446; GFX6-NEXT:    s_waitcnt vmcnt(0)
447; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:900 ; 4-byte Folded Spill
448; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:904 ; 4-byte Folded Spill
449; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:908 ; 4-byte Folded Spill
450; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:912 ; 4-byte Folded Spill
451; GFX6-NEXT:    s_waitcnt expcnt(0)
452; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:912
453; GFX6-NEXT:    s_waitcnt vmcnt(0)
454; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:916 ; 4-byte Folded Spill
455; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:920 ; 4-byte Folded Spill
456; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:924 ; 4-byte Folded Spill
457; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:928 ; 4-byte Folded Spill
458; GFX6-NEXT:    s_waitcnt expcnt(0)
459; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:928
460; GFX6-NEXT:    s_waitcnt vmcnt(0)
461; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:932 ; 4-byte Folded Spill
462; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:936 ; 4-byte Folded Spill
463; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:940 ; 4-byte Folded Spill
464; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:944 ; 4-byte Folded Spill
465; GFX6-NEXT:    s_waitcnt expcnt(0)
466; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:944
467; GFX6-NEXT:    s_waitcnt vmcnt(0)
468; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:948 ; 4-byte Folded Spill
469; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:952 ; 4-byte Folded Spill
470; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:956 ; 4-byte Folded Spill
471; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:960 ; 4-byte Folded Spill
472; GFX6-NEXT:    s_waitcnt expcnt(0)
473; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:960
474; GFX6-NEXT:    s_waitcnt vmcnt(0)
475; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:964 ; 4-byte Folded Spill
476; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:968 ; 4-byte Folded Spill
477; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:972 ; 4-byte Folded Spill
478; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:976 ; 4-byte Folded Spill
479; GFX6-NEXT:    s_waitcnt expcnt(0)
480; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:976
481; GFX6-NEXT:    s_waitcnt vmcnt(0)
482; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:980 ; 4-byte Folded Spill
483; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:984 ; 4-byte Folded Spill
484; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:988 ; 4-byte Folded Spill
485; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:992 ; 4-byte Folded Spill
486; GFX6-NEXT:    s_waitcnt expcnt(0)
487; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:992
488; GFX6-NEXT:    s_waitcnt vmcnt(0)
489; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:996 ; 4-byte Folded Spill
490; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1000 ; 4-byte Folded Spill
491; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1004 ; 4-byte Folded Spill
492; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1008 ; 4-byte Folded Spill
493; GFX6-NEXT:    s_waitcnt expcnt(0)
494; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1008
495; GFX6-NEXT:    s_waitcnt vmcnt(0)
496; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1012 ; 4-byte Folded Spill
497; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1016 ; 4-byte Folded Spill
498; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1020 ; 4-byte Folded Spill
499; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1024 ; 4-byte Folded Spill
500; GFX6-NEXT:    s_waitcnt expcnt(0)
501; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1024
502; GFX6-NEXT:    s_waitcnt vmcnt(0)
503; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1028 ; 4-byte Folded Spill
504; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1032 ; 4-byte Folded Spill
505; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1036 ; 4-byte Folded Spill
506; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1040 ; 4-byte Folded Spill
507; GFX6-NEXT:    s_waitcnt expcnt(0)
508; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1040
509; GFX6-NEXT:    s_waitcnt vmcnt(0)
510; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1044 ; 4-byte Folded Spill
511; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1048 ; 4-byte Folded Spill
512; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1052 ; 4-byte Folded Spill
513; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1056 ; 4-byte Folded Spill
514; GFX6-NEXT:    s_waitcnt expcnt(0)
515; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1056
516; GFX6-NEXT:    s_waitcnt vmcnt(0)
517; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1060 ; 4-byte Folded Spill
518; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1064 ; 4-byte Folded Spill
519; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1068 ; 4-byte Folded Spill
520; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1072 ; 4-byte Folded Spill
521; GFX6-NEXT:    s_waitcnt expcnt(0)
522; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1072
523; GFX6-NEXT:    s_waitcnt vmcnt(0)
524; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1076 ; 4-byte Folded Spill
525; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1080 ; 4-byte Folded Spill
526; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1084 ; 4-byte Folded Spill
527; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1088 ; 4-byte Folded Spill
528; GFX6-NEXT:    s_waitcnt expcnt(0)
529; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1088
530; GFX6-NEXT:    s_waitcnt vmcnt(0)
531; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1092 ; 4-byte Folded Spill
532; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1096 ; 4-byte Folded Spill
533; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1100 ; 4-byte Folded Spill
534; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1104 ; 4-byte Folded Spill
535; GFX6-NEXT:    s_waitcnt expcnt(0)
536; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1104
537; GFX6-NEXT:    s_waitcnt vmcnt(0)
538; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1108 ; 4-byte Folded Spill
539; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1112 ; 4-byte Folded Spill
540; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1116 ; 4-byte Folded Spill
541; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1120 ; 4-byte Folded Spill
542; GFX6-NEXT:    s_waitcnt expcnt(0)
543; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1120
544; GFX6-NEXT:    s_waitcnt vmcnt(0)
545; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1124 ; 4-byte Folded Spill
546; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1128 ; 4-byte Folded Spill
547; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1132 ; 4-byte Folded Spill
548; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1136 ; 4-byte Folded Spill
549; GFX6-NEXT:    s_waitcnt expcnt(0)
550; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1136
551; GFX6-NEXT:    s_waitcnt vmcnt(0)
552; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1140 ; 4-byte Folded Spill
553; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1144 ; 4-byte Folded Spill
554; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1148 ; 4-byte Folded Spill
555; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1152 ; 4-byte Folded Spill
556; GFX6-NEXT:    s_waitcnt expcnt(0)
557; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1152
558; GFX6-NEXT:    s_waitcnt vmcnt(0)
559; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1156 ; 4-byte Folded Spill
560; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1160 ; 4-byte Folded Spill
561; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1164 ; 4-byte Folded Spill
562; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1168 ; 4-byte Folded Spill
563; GFX6-NEXT:    s_waitcnt expcnt(0)
564; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1168
565; GFX6-NEXT:    s_waitcnt vmcnt(0)
566; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1172 ; 4-byte Folded Spill
567; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1176 ; 4-byte Folded Spill
568; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1180 ; 4-byte Folded Spill
569; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1184 ; 4-byte Folded Spill
570; GFX6-NEXT:    s_waitcnt expcnt(0)
571; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1184
572; GFX6-NEXT:    s_waitcnt vmcnt(0)
573; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1188 ; 4-byte Folded Spill
574; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1192 ; 4-byte Folded Spill
575; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1196 ; 4-byte Folded Spill
576; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1200 ; 4-byte Folded Spill
577; GFX6-NEXT:    s_waitcnt expcnt(0)
578; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1200
579; GFX6-NEXT:    s_waitcnt vmcnt(0)
580; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1204 ; 4-byte Folded Spill
581; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1208 ; 4-byte Folded Spill
582; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1212 ; 4-byte Folded Spill
583; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1216 ; 4-byte Folded Spill
584; GFX6-NEXT:    s_waitcnt expcnt(0)
585; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1216
586; GFX6-NEXT:    s_waitcnt vmcnt(0)
587; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1220 ; 4-byte Folded Spill
588; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1224 ; 4-byte Folded Spill
589; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1228 ; 4-byte Folded Spill
590; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1232 ; 4-byte Folded Spill
591; GFX6-NEXT:    s_waitcnt expcnt(0)
592; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1232
593; GFX6-NEXT:    s_waitcnt vmcnt(0)
594; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1236 ; 4-byte Folded Spill
595; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1240 ; 4-byte Folded Spill
596; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1244 ; 4-byte Folded Spill
597; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1248 ; 4-byte Folded Spill
598; GFX6-NEXT:    s_waitcnt expcnt(0)
599; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1248
600; GFX6-NEXT:    s_waitcnt vmcnt(0)
601; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1252 ; 4-byte Folded Spill
602; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1256 ; 4-byte Folded Spill
603; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1260 ; 4-byte Folded Spill
604; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1264 ; 4-byte Folded Spill
605; GFX6-NEXT:    s_waitcnt expcnt(0)
606; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1264
607; GFX6-NEXT:    s_waitcnt vmcnt(0)
608; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1268 ; 4-byte Folded Spill
609; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1272 ; 4-byte Folded Spill
610; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1276 ; 4-byte Folded Spill
611; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1280 ; 4-byte Folded Spill
612; GFX6-NEXT:    s_waitcnt expcnt(0)
613; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1280
614; GFX6-NEXT:    s_waitcnt vmcnt(0)
615; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1284 ; 4-byte Folded Spill
616; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1288 ; 4-byte Folded Spill
617; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1292 ; 4-byte Folded Spill
618; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1296 ; 4-byte Folded Spill
619; GFX6-NEXT:    s_waitcnt expcnt(0)
620; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1296
621; GFX6-NEXT:    s_waitcnt vmcnt(0)
622; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1300 ; 4-byte Folded Spill
623; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1304 ; 4-byte Folded Spill
624; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1308 ; 4-byte Folded Spill
625; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1312 ; 4-byte Folded Spill
626; GFX6-NEXT:    s_waitcnt expcnt(0)
627; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1312
628; GFX6-NEXT:    s_waitcnt vmcnt(0)
629; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1316 ; 4-byte Folded Spill
630; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1320 ; 4-byte Folded Spill
631; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1324 ; 4-byte Folded Spill
632; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1328 ; 4-byte Folded Spill
633; GFX6-NEXT:    s_waitcnt expcnt(0)
634; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1328
635; GFX6-NEXT:    s_waitcnt vmcnt(0)
636; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1332 ; 4-byte Folded Spill
637; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1336 ; 4-byte Folded Spill
638; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1340 ; 4-byte Folded Spill
639; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1344 ; 4-byte Folded Spill
640; GFX6-NEXT:    s_waitcnt expcnt(0)
641; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1344
642; GFX6-NEXT:    s_waitcnt vmcnt(0)
643; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1348 ; 4-byte Folded Spill
644; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1352 ; 4-byte Folded Spill
645; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1356 ; 4-byte Folded Spill
646; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1360 ; 4-byte Folded Spill
647; GFX6-NEXT:    s_waitcnt expcnt(0)
648; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1360
649; GFX6-NEXT:    s_waitcnt vmcnt(0)
650; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1364 ; 4-byte Folded Spill
651; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1368 ; 4-byte Folded Spill
652; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1372 ; 4-byte Folded Spill
653; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1376 ; 4-byte Folded Spill
654; GFX6-NEXT:    s_waitcnt expcnt(0)
655; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1376
656; GFX6-NEXT:    s_waitcnt vmcnt(0)
657; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1380 ; 4-byte Folded Spill
658; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1384 ; 4-byte Folded Spill
659; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1388 ; 4-byte Folded Spill
660; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1392 ; 4-byte Folded Spill
661; GFX6-NEXT:    s_waitcnt expcnt(0)
662; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1392
663; GFX6-NEXT:    s_waitcnt vmcnt(0)
664; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1396 ; 4-byte Folded Spill
665; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1400 ; 4-byte Folded Spill
666; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1404 ; 4-byte Folded Spill
667; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1408 ; 4-byte Folded Spill
668; GFX6-NEXT:    s_waitcnt expcnt(0)
669; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1408
670; GFX6-NEXT:    s_waitcnt vmcnt(0)
671; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1412 ; 4-byte Folded Spill
672; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1416 ; 4-byte Folded Spill
673; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1420 ; 4-byte Folded Spill
674; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1424 ; 4-byte Folded Spill
675; GFX6-NEXT:    s_waitcnt expcnt(0)
676; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1424
677; GFX6-NEXT:    s_waitcnt vmcnt(0)
678; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1428 ; 4-byte Folded Spill
679; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1432 ; 4-byte Folded Spill
680; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1436 ; 4-byte Folded Spill
681; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1440 ; 4-byte Folded Spill
682; GFX6-NEXT:    s_waitcnt expcnt(0)
683; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1440
684; GFX6-NEXT:    s_waitcnt vmcnt(0)
685; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1444 ; 4-byte Folded Spill
686; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1448 ; 4-byte Folded Spill
687; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1452 ; 4-byte Folded Spill
688; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1456 ; 4-byte Folded Spill
689; GFX6-NEXT:    s_waitcnt expcnt(0)
690; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1456
691; GFX6-NEXT:    s_waitcnt vmcnt(0)
692; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1460 ; 4-byte Folded Spill
693; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1464 ; 4-byte Folded Spill
694; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1468 ; 4-byte Folded Spill
695; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1472 ; 4-byte Folded Spill
696; GFX6-NEXT:    s_waitcnt expcnt(0)
697; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1472
698; GFX6-NEXT:    s_waitcnt vmcnt(0)
699; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1476 ; 4-byte Folded Spill
700; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1480 ; 4-byte Folded Spill
701; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1484 ; 4-byte Folded Spill
702; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1488 ; 4-byte Folded Spill
703; GFX6-NEXT:    s_waitcnt expcnt(0)
704; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1488
705; GFX6-NEXT:    s_waitcnt vmcnt(0)
706; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1492 ; 4-byte Folded Spill
707; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1496 ; 4-byte Folded Spill
708; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1500 ; 4-byte Folded Spill
709; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1504 ; 4-byte Folded Spill
710; GFX6-NEXT:    s_waitcnt expcnt(0)
711; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1504
712; GFX6-NEXT:    s_waitcnt vmcnt(0)
713; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1508 ; 4-byte Folded Spill
714; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1512 ; 4-byte Folded Spill
715; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1516 ; 4-byte Folded Spill
716; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1520 ; 4-byte Folded Spill
717; GFX6-NEXT:    s_waitcnt expcnt(0)
718; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1520
719; GFX6-NEXT:    s_waitcnt vmcnt(0)
720; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1524 ; 4-byte Folded Spill
721; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1528 ; 4-byte Folded Spill
722; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1532 ; 4-byte Folded Spill
723; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1536 ; 4-byte Folded Spill
724; GFX6-NEXT:    s_waitcnt expcnt(0)
725; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1536
726; GFX6-NEXT:    s_waitcnt vmcnt(0)
727; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1540 ; 4-byte Folded Spill
728; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1544 ; 4-byte Folded Spill
729; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1548 ; 4-byte Folded Spill
730; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1552 ; 4-byte Folded Spill
731; GFX6-NEXT:    s_waitcnt expcnt(0)
732; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1552
733; GFX6-NEXT:    s_waitcnt vmcnt(0)
734; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1556 ; 4-byte Folded Spill
735; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1560 ; 4-byte Folded Spill
736; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1564 ; 4-byte Folded Spill
737; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1568 ; 4-byte Folded Spill
738; GFX6-NEXT:    s_waitcnt expcnt(0)
739; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1568
740; GFX6-NEXT:    s_waitcnt vmcnt(0)
741; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1572 ; 4-byte Folded Spill
742; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1576 ; 4-byte Folded Spill
743; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1580 ; 4-byte Folded Spill
744; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1584 ; 4-byte Folded Spill
745; GFX6-NEXT:    s_waitcnt expcnt(0)
746; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1584
747; GFX6-NEXT:    s_waitcnt vmcnt(0)
748; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1588 ; 4-byte Folded Spill
749; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1592 ; 4-byte Folded Spill
750; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1596 ; 4-byte Folded Spill
751; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1600 ; 4-byte Folded Spill
752; GFX6-NEXT:    s_waitcnt expcnt(0)
753; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1600
754; GFX6-NEXT:    s_waitcnt vmcnt(0)
755; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1604 ; 4-byte Folded Spill
756; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1608 ; 4-byte Folded Spill
757; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1612 ; 4-byte Folded Spill
758; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1616 ; 4-byte Folded Spill
759; GFX6-NEXT:    s_waitcnt expcnt(0)
760; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1616
761; GFX6-NEXT:    s_waitcnt vmcnt(0)
762; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1620 ; 4-byte Folded Spill
763; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1624 ; 4-byte Folded Spill
764; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1628 ; 4-byte Folded Spill
765; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1632 ; 4-byte Folded Spill
766; GFX6-NEXT:    s_waitcnt expcnt(0)
767; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1632
768; GFX6-NEXT:    s_waitcnt vmcnt(0)
769; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1636 ; 4-byte Folded Spill
770; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1640 ; 4-byte Folded Spill
771; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1644 ; 4-byte Folded Spill
772; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1648 ; 4-byte Folded Spill
773; GFX6-NEXT:    s_waitcnt expcnt(0)
774; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1648
775; GFX6-NEXT:    s_waitcnt vmcnt(0)
776; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1652 ; 4-byte Folded Spill
777; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1656 ; 4-byte Folded Spill
778; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1660 ; 4-byte Folded Spill
779; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1664 ; 4-byte Folded Spill
780; GFX6-NEXT:    s_waitcnt expcnt(0)
781; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1664
782; GFX6-NEXT:    s_waitcnt vmcnt(0)
783; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1668 ; 4-byte Folded Spill
784; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1672 ; 4-byte Folded Spill
785; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1676 ; 4-byte Folded Spill
786; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1680 ; 4-byte Folded Spill
787; GFX6-NEXT:    s_waitcnt expcnt(0)
788; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1680
789; GFX6-NEXT:    s_waitcnt vmcnt(0)
790; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1684 ; 4-byte Folded Spill
791; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1688 ; 4-byte Folded Spill
792; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1692 ; 4-byte Folded Spill
793; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1696 ; 4-byte Folded Spill
794; GFX6-NEXT:    s_waitcnt expcnt(0)
795; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1696
796; GFX6-NEXT:    s_waitcnt vmcnt(0)
797; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1700 ; 4-byte Folded Spill
798; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1704 ; 4-byte Folded Spill
799; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1708 ; 4-byte Folded Spill
800; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1712 ; 4-byte Folded Spill
801; GFX6-NEXT:    s_waitcnt expcnt(0)
802; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1712
803; GFX6-NEXT:    s_waitcnt vmcnt(0)
804; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1716 ; 4-byte Folded Spill
805; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1720 ; 4-byte Folded Spill
806; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1724 ; 4-byte Folded Spill
807; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1728 ; 4-byte Folded Spill
808; GFX6-NEXT:    s_waitcnt expcnt(0)
809; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1728
810; GFX6-NEXT:    s_waitcnt vmcnt(0)
811; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1732 ; 4-byte Folded Spill
812; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1736 ; 4-byte Folded Spill
813; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1740 ; 4-byte Folded Spill
814; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1744 ; 4-byte Folded Spill
815; GFX6-NEXT:    s_waitcnt expcnt(0)
816; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1744
817; GFX6-NEXT:    s_waitcnt vmcnt(0)
818; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1748 ; 4-byte Folded Spill
819; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1752 ; 4-byte Folded Spill
820; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1756 ; 4-byte Folded Spill
821; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1760 ; 4-byte Folded Spill
822; GFX6-NEXT:    s_waitcnt expcnt(0)
823; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1760
824; GFX6-NEXT:    s_waitcnt vmcnt(0)
825; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1764 ; 4-byte Folded Spill
826; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1768 ; 4-byte Folded Spill
827; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1772 ; 4-byte Folded Spill
828; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1776 ; 4-byte Folded Spill
829; GFX6-NEXT:    s_waitcnt expcnt(0)
830; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1776
831; GFX6-NEXT:    s_waitcnt vmcnt(0)
832; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1780 ; 4-byte Folded Spill
833; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1784 ; 4-byte Folded Spill
834; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1788 ; 4-byte Folded Spill
835; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1792 ; 4-byte Folded Spill
836; GFX6-NEXT:    s_waitcnt expcnt(0)
837; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1792
838; GFX6-NEXT:    s_waitcnt vmcnt(0)
839; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1796 ; 4-byte Folded Spill
840; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1800 ; 4-byte Folded Spill
841; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1804 ; 4-byte Folded Spill
842; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1808 ; 4-byte Folded Spill
843; GFX6-NEXT:    s_waitcnt expcnt(0)
844; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1808
845; GFX6-NEXT:    s_waitcnt vmcnt(0)
846; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1812 ; 4-byte Folded Spill
847; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1816 ; 4-byte Folded Spill
848; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1820 ; 4-byte Folded Spill
849; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1824 ; 4-byte Folded Spill
850; GFX6-NEXT:    s_waitcnt expcnt(0)
851; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1824
852; GFX6-NEXT:    s_waitcnt vmcnt(0)
853; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1828 ; 4-byte Folded Spill
854; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1832 ; 4-byte Folded Spill
855; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1836 ; 4-byte Folded Spill
856; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1840 ; 4-byte Folded Spill
857; GFX6-NEXT:    s_waitcnt expcnt(0)
858; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1840
859; GFX6-NEXT:    s_waitcnt vmcnt(0)
860; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1844 ; 4-byte Folded Spill
861; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1848 ; 4-byte Folded Spill
862; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1852 ; 4-byte Folded Spill
863; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1856 ; 4-byte Folded Spill
864; GFX6-NEXT:    s_waitcnt expcnt(0)
865; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1856
866; GFX6-NEXT:    s_waitcnt vmcnt(0)
867; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1860 ; 4-byte Folded Spill
868; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1864 ; 4-byte Folded Spill
869; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1868 ; 4-byte Folded Spill
870; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1872 ; 4-byte Folded Spill
871; GFX6-NEXT:    s_waitcnt expcnt(0)
872; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1872
873; GFX6-NEXT:    s_waitcnt vmcnt(0)
874; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1876 ; 4-byte Folded Spill
875; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1880 ; 4-byte Folded Spill
876; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1884 ; 4-byte Folded Spill
877; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1888 ; 4-byte Folded Spill
878; GFX6-NEXT:    s_waitcnt expcnt(0)
879; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1888
880; GFX6-NEXT:    s_waitcnt vmcnt(0)
881; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1892 ; 4-byte Folded Spill
882; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1896 ; 4-byte Folded Spill
883; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1900 ; 4-byte Folded Spill
884; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1904 ; 4-byte Folded Spill
885; GFX6-NEXT:    s_waitcnt expcnt(0)
886; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1904
887; GFX6-NEXT:    s_waitcnt vmcnt(0)
888; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1908 ; 4-byte Folded Spill
889; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1912 ; 4-byte Folded Spill
890; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1916 ; 4-byte Folded Spill
891; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1920 ; 4-byte Folded Spill
892; GFX6-NEXT:    s_waitcnt expcnt(0)
893; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1920
894; GFX6-NEXT:    s_waitcnt vmcnt(0)
895; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1924 ; 4-byte Folded Spill
896; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1928 ; 4-byte Folded Spill
897; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1932 ; 4-byte Folded Spill
898; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1936 ; 4-byte Folded Spill
899; GFX6-NEXT:    s_waitcnt expcnt(0)
900; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1936
901; GFX6-NEXT:    s_waitcnt vmcnt(0)
902; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1940 ; 4-byte Folded Spill
903; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1944 ; 4-byte Folded Spill
904; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1948 ; 4-byte Folded Spill
905; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1952 ; 4-byte Folded Spill
906; GFX6-NEXT:    s_waitcnt expcnt(0)
907; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1952
908; GFX6-NEXT:    s_waitcnt vmcnt(0)
909; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1956 ; 4-byte Folded Spill
910; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1960 ; 4-byte Folded Spill
911; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1964 ; 4-byte Folded Spill
912; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1968 ; 4-byte Folded Spill
913; GFX6-NEXT:    s_waitcnt expcnt(0)
914; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1968
915; GFX6-NEXT:    s_waitcnt vmcnt(0)
916; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1972 ; 4-byte Folded Spill
917; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1976 ; 4-byte Folded Spill
918; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1980 ; 4-byte Folded Spill
919; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:1984 ; 4-byte Folded Spill
920; GFX6-NEXT:    s_waitcnt expcnt(0)
921; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:1984
922; GFX6-NEXT:    s_waitcnt vmcnt(0)
923; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:1988 ; 4-byte Folded Spill
924; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:1992 ; 4-byte Folded Spill
925; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:1996 ; 4-byte Folded Spill
926; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2000 ; 4-byte Folded Spill
927; GFX6-NEXT:    s_waitcnt expcnt(0)
928; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2000
929; GFX6-NEXT:    s_waitcnt vmcnt(0)
930; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2004 ; 4-byte Folded Spill
931; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2008 ; 4-byte Folded Spill
932; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2012 ; 4-byte Folded Spill
933; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2016 ; 4-byte Folded Spill
934; GFX6-NEXT:    s_waitcnt expcnt(0)
935; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2016
936; GFX6-NEXT:    s_waitcnt vmcnt(0)
937; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2020 ; 4-byte Folded Spill
938; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2024 ; 4-byte Folded Spill
939; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2028 ; 4-byte Folded Spill
940; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2032 ; 4-byte Folded Spill
941; GFX6-NEXT:    s_waitcnt expcnt(0)
942; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2032
943; GFX6-NEXT:    s_waitcnt vmcnt(0)
944; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2036 ; 4-byte Folded Spill
945; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2040 ; 4-byte Folded Spill
946; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2044 ; 4-byte Folded Spill
947; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2048 ; 4-byte Folded Spill
948; GFX6-NEXT:    s_waitcnt expcnt(0)
949; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2048
950; GFX6-NEXT:    s_waitcnt vmcnt(0)
951; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2052 ; 4-byte Folded Spill
952; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2056 ; 4-byte Folded Spill
953; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2060 ; 4-byte Folded Spill
954; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2064 ; 4-byte Folded Spill
955; GFX6-NEXT:    s_waitcnt expcnt(0)
956; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2064
957; GFX6-NEXT:    s_waitcnt vmcnt(0)
958; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2068 ; 4-byte Folded Spill
959; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2072 ; 4-byte Folded Spill
960; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2076 ; 4-byte Folded Spill
961; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2080 ; 4-byte Folded Spill
962; GFX6-NEXT:    s_waitcnt expcnt(0)
963; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2080
964; GFX6-NEXT:    s_waitcnt vmcnt(0)
965; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2084 ; 4-byte Folded Spill
966; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2088 ; 4-byte Folded Spill
967; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2092 ; 4-byte Folded Spill
968; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2096 ; 4-byte Folded Spill
969; GFX6-NEXT:    s_waitcnt expcnt(0)
970; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2096
971; GFX6-NEXT:    s_waitcnt vmcnt(0)
972; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2100 ; 4-byte Folded Spill
973; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2104 ; 4-byte Folded Spill
974; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2108 ; 4-byte Folded Spill
975; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2112 ; 4-byte Folded Spill
976; GFX6-NEXT:    s_waitcnt expcnt(0)
977; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2112
978; GFX6-NEXT:    s_waitcnt vmcnt(0)
979; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2116 ; 4-byte Folded Spill
980; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2120 ; 4-byte Folded Spill
981; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2124 ; 4-byte Folded Spill
982; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2128 ; 4-byte Folded Spill
983; GFX6-NEXT:    s_waitcnt expcnt(0)
984; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2128
985; GFX6-NEXT:    s_waitcnt vmcnt(0)
986; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2132 ; 4-byte Folded Spill
987; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2136 ; 4-byte Folded Spill
988; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2140 ; 4-byte Folded Spill
989; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2144 ; 4-byte Folded Spill
990; GFX6-NEXT:    s_waitcnt expcnt(0)
991; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2144
992; GFX6-NEXT:    s_waitcnt vmcnt(0)
993; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2148 ; 4-byte Folded Spill
994; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2152 ; 4-byte Folded Spill
995; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2156 ; 4-byte Folded Spill
996; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2160 ; 4-byte Folded Spill
997; GFX6-NEXT:    s_waitcnt expcnt(0)
998; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2160
999; GFX6-NEXT:    s_waitcnt vmcnt(0)
1000; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2164 ; 4-byte Folded Spill
1001; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2168 ; 4-byte Folded Spill
1002; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2172 ; 4-byte Folded Spill
1003; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2176 ; 4-byte Folded Spill
1004; GFX6-NEXT:    s_waitcnt expcnt(0)
1005; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2176
1006; GFX6-NEXT:    s_waitcnt vmcnt(0)
1007; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2180 ; 4-byte Folded Spill
1008; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2184 ; 4-byte Folded Spill
1009; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2188 ; 4-byte Folded Spill
1010; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2192 ; 4-byte Folded Spill
1011; GFX6-NEXT:    s_waitcnt expcnt(0)
1012; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2192
1013; GFX6-NEXT:    s_waitcnt vmcnt(0)
1014; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2196 ; 4-byte Folded Spill
1015; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2200 ; 4-byte Folded Spill
1016; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2204 ; 4-byte Folded Spill
1017; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2208 ; 4-byte Folded Spill
1018; GFX6-NEXT:    s_waitcnt expcnt(0)
1019; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2208
1020; GFX6-NEXT:    s_waitcnt vmcnt(0)
1021; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2212 ; 4-byte Folded Spill
1022; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2216 ; 4-byte Folded Spill
1023; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2220 ; 4-byte Folded Spill
1024; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2224 ; 4-byte Folded Spill
1025; GFX6-NEXT:    s_waitcnt expcnt(0)
1026; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2224
1027; GFX6-NEXT:    s_waitcnt vmcnt(0)
1028; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2228 ; 4-byte Folded Spill
1029; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2232 ; 4-byte Folded Spill
1030; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2236 ; 4-byte Folded Spill
1031; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2240 ; 4-byte Folded Spill
1032; GFX6-NEXT:    s_waitcnt expcnt(0)
1033; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2240
1034; GFX6-NEXT:    s_waitcnt vmcnt(0)
1035; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2244 ; 4-byte Folded Spill
1036; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2248 ; 4-byte Folded Spill
1037; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2252 ; 4-byte Folded Spill
1038; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2256 ; 4-byte Folded Spill
1039; GFX6-NEXT:    s_waitcnt expcnt(0)
1040; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2256
1041; GFX6-NEXT:    s_waitcnt vmcnt(0)
1042; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2260 ; 4-byte Folded Spill
1043; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2264 ; 4-byte Folded Spill
1044; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2268 ; 4-byte Folded Spill
1045; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2272 ; 4-byte Folded Spill
1046; GFX6-NEXT:    s_waitcnt expcnt(0)
1047; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2272
1048; GFX6-NEXT:    s_waitcnt vmcnt(0)
1049; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2276 ; 4-byte Folded Spill
1050; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2280 ; 4-byte Folded Spill
1051; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2284 ; 4-byte Folded Spill
1052; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2288 ; 4-byte Folded Spill
1053; GFX6-NEXT:    s_waitcnt expcnt(0)
1054; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2288
1055; GFX6-NEXT:    s_waitcnt vmcnt(0)
1056; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2292 ; 4-byte Folded Spill
1057; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2296 ; 4-byte Folded Spill
1058; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2300 ; 4-byte Folded Spill
1059; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2304 ; 4-byte Folded Spill
1060; GFX6-NEXT:    s_waitcnt expcnt(0)
1061; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2304
1062; GFX6-NEXT:    s_waitcnt vmcnt(0)
1063; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2308 ; 4-byte Folded Spill
1064; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2312 ; 4-byte Folded Spill
1065; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2316 ; 4-byte Folded Spill
1066; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2320 ; 4-byte Folded Spill
1067; GFX6-NEXT:    s_waitcnt expcnt(0)
1068; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2320
1069; GFX6-NEXT:    s_waitcnt vmcnt(0)
1070; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2324 ; 4-byte Folded Spill
1071; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2328 ; 4-byte Folded Spill
1072; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2332 ; 4-byte Folded Spill
1073; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2336 ; 4-byte Folded Spill
1074; GFX6-NEXT:    s_waitcnt expcnt(0)
1075; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2336
1076; GFX6-NEXT:    s_waitcnt vmcnt(0)
1077; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2340 ; 4-byte Folded Spill
1078; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2344 ; 4-byte Folded Spill
1079; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2348 ; 4-byte Folded Spill
1080; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2352 ; 4-byte Folded Spill
1081; GFX6-NEXT:    s_waitcnt expcnt(0)
1082; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2352
1083; GFX6-NEXT:    s_waitcnt vmcnt(0)
1084; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2356 ; 4-byte Folded Spill
1085; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2360 ; 4-byte Folded Spill
1086; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2364 ; 4-byte Folded Spill
1087; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2368 ; 4-byte Folded Spill
1088; GFX6-NEXT:    s_waitcnt expcnt(0)
1089; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2368
1090; GFX6-NEXT:    s_waitcnt vmcnt(0)
1091; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2372 ; 4-byte Folded Spill
1092; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2376 ; 4-byte Folded Spill
1093; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2380 ; 4-byte Folded Spill
1094; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2384 ; 4-byte Folded Spill
1095; GFX6-NEXT:    s_waitcnt expcnt(0)
1096; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2384
1097; GFX6-NEXT:    s_waitcnt vmcnt(0)
1098; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2388 ; 4-byte Folded Spill
1099; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2392 ; 4-byte Folded Spill
1100; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2396 ; 4-byte Folded Spill
1101; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2400 ; 4-byte Folded Spill
1102; GFX6-NEXT:    s_waitcnt expcnt(0)
1103; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2400
1104; GFX6-NEXT:    s_waitcnt vmcnt(0)
1105; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2404 ; 4-byte Folded Spill
1106; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2408 ; 4-byte Folded Spill
1107; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2412 ; 4-byte Folded Spill
1108; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2416 ; 4-byte Folded Spill
1109; GFX6-NEXT:    s_waitcnt expcnt(0)
1110; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2416
1111; GFX6-NEXT:    s_waitcnt vmcnt(0)
1112; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2420 ; 4-byte Folded Spill
1113; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2424 ; 4-byte Folded Spill
1114; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2428 ; 4-byte Folded Spill
1115; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2432 ; 4-byte Folded Spill
1116; GFX6-NEXT:    s_waitcnt expcnt(0)
1117; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2432
1118; GFX6-NEXT:    s_waitcnt vmcnt(0)
1119; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2436 ; 4-byte Folded Spill
1120; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2440 ; 4-byte Folded Spill
1121; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2444 ; 4-byte Folded Spill
1122; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2448 ; 4-byte Folded Spill
1123; GFX6-NEXT:    s_waitcnt expcnt(0)
1124; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2448
1125; GFX6-NEXT:    s_waitcnt vmcnt(0)
1126; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2452 ; 4-byte Folded Spill
1127; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2456 ; 4-byte Folded Spill
1128; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2460 ; 4-byte Folded Spill
1129; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2464 ; 4-byte Folded Spill
1130; GFX6-NEXT:    s_waitcnt expcnt(0)
1131; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2464
1132; GFX6-NEXT:    s_waitcnt vmcnt(0)
1133; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2468 ; 4-byte Folded Spill
1134; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2472 ; 4-byte Folded Spill
1135; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2476 ; 4-byte Folded Spill
1136; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2480 ; 4-byte Folded Spill
1137; GFX6-NEXT:    s_waitcnt expcnt(0)
1138; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2480
1139; GFX6-NEXT:    s_waitcnt vmcnt(0)
1140; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2484 ; 4-byte Folded Spill
1141; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2488 ; 4-byte Folded Spill
1142; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2492 ; 4-byte Folded Spill
1143; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2496 ; 4-byte Folded Spill
1144; GFX6-NEXT:    s_waitcnt expcnt(0)
1145; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2496
1146; GFX6-NEXT:    s_waitcnt vmcnt(0)
1147; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2500 ; 4-byte Folded Spill
1148; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2504 ; 4-byte Folded Spill
1149; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2508 ; 4-byte Folded Spill
1150; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2512 ; 4-byte Folded Spill
1151; GFX6-NEXT:    s_waitcnt expcnt(0)
1152; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2512
1153; GFX6-NEXT:    s_waitcnt vmcnt(0)
1154; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2516 ; 4-byte Folded Spill
1155; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2520 ; 4-byte Folded Spill
1156; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2524 ; 4-byte Folded Spill
1157; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2528 ; 4-byte Folded Spill
1158; GFX6-NEXT:    s_waitcnt expcnt(0)
1159; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2528
1160; GFX6-NEXT:    s_waitcnt vmcnt(0)
1161; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2532 ; 4-byte Folded Spill
1162; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2536 ; 4-byte Folded Spill
1163; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2540 ; 4-byte Folded Spill
1164; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2544 ; 4-byte Folded Spill
1165; GFX6-NEXT:    s_waitcnt expcnt(0)
1166; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2544
1167; GFX6-NEXT:    s_waitcnt vmcnt(0)
1168; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2548 ; 4-byte Folded Spill
1169; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2552 ; 4-byte Folded Spill
1170; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2556 ; 4-byte Folded Spill
1171; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2560 ; 4-byte Folded Spill
1172; GFX6-NEXT:    s_waitcnt expcnt(0)
1173; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2560
1174; GFX6-NEXT:    s_waitcnt vmcnt(0)
1175; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2564 ; 4-byte Folded Spill
1176; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2568 ; 4-byte Folded Spill
1177; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2572 ; 4-byte Folded Spill
1178; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2576 ; 4-byte Folded Spill
1179; GFX6-NEXT:    s_waitcnt expcnt(0)
1180; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2576
1181; GFX6-NEXT:    s_waitcnt vmcnt(0)
1182; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2580 ; 4-byte Folded Spill
1183; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2584 ; 4-byte Folded Spill
1184; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2588 ; 4-byte Folded Spill
1185; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2592 ; 4-byte Folded Spill
1186; GFX6-NEXT:    s_waitcnt expcnt(0)
1187; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2592
1188; GFX6-NEXT:    s_waitcnt vmcnt(0)
1189; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2596 ; 4-byte Folded Spill
1190; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2600 ; 4-byte Folded Spill
1191; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2604 ; 4-byte Folded Spill
1192; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2608 ; 4-byte Folded Spill
1193; GFX6-NEXT:    s_waitcnt expcnt(0)
1194; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2608
1195; GFX6-NEXT:    s_waitcnt vmcnt(0)
1196; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2612 ; 4-byte Folded Spill
1197; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2616 ; 4-byte Folded Spill
1198; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2620 ; 4-byte Folded Spill
1199; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2624 ; 4-byte Folded Spill
1200; GFX6-NEXT:    s_waitcnt expcnt(0)
1201; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2624
1202; GFX6-NEXT:    s_waitcnt vmcnt(0)
1203; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2628 ; 4-byte Folded Spill
1204; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2632 ; 4-byte Folded Spill
1205; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2636 ; 4-byte Folded Spill
1206; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2640 ; 4-byte Folded Spill
1207; GFX6-NEXT:    s_waitcnt expcnt(0)
1208; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2640
1209; GFX6-NEXT:    s_waitcnt vmcnt(0)
1210; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2644 ; 4-byte Folded Spill
1211; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2648 ; 4-byte Folded Spill
1212; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2652 ; 4-byte Folded Spill
1213; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2656 ; 4-byte Folded Spill
1214; GFX6-NEXT:    s_waitcnt expcnt(0)
1215; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2656
1216; GFX6-NEXT:    s_waitcnt vmcnt(0)
1217; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2660 ; 4-byte Folded Spill
1218; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2664 ; 4-byte Folded Spill
1219; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2668 ; 4-byte Folded Spill
1220; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2672 ; 4-byte Folded Spill
1221; GFX6-NEXT:    s_waitcnt expcnt(0)
1222; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2672
1223; GFX6-NEXT:    s_waitcnt vmcnt(0)
1224; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2676 ; 4-byte Folded Spill
1225; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2680 ; 4-byte Folded Spill
1226; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2684 ; 4-byte Folded Spill
1227; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2688 ; 4-byte Folded Spill
1228; GFX6-NEXT:    s_waitcnt expcnt(0)
1229; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2688
1230; GFX6-NEXT:    s_waitcnt vmcnt(0)
1231; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2692 ; 4-byte Folded Spill
1232; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2696 ; 4-byte Folded Spill
1233; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2700 ; 4-byte Folded Spill
1234; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2704 ; 4-byte Folded Spill
1235; GFX6-NEXT:    s_waitcnt expcnt(0)
1236; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2704
1237; GFX6-NEXT:    s_waitcnt vmcnt(0)
1238; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2708 ; 4-byte Folded Spill
1239; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2712 ; 4-byte Folded Spill
1240; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2716 ; 4-byte Folded Spill
1241; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2720 ; 4-byte Folded Spill
1242; GFX6-NEXT:    s_waitcnt expcnt(0)
1243; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2720
1244; GFX6-NEXT:    s_waitcnt vmcnt(0)
1245; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2724 ; 4-byte Folded Spill
1246; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2728 ; 4-byte Folded Spill
1247; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2732 ; 4-byte Folded Spill
1248; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2736 ; 4-byte Folded Spill
1249; GFX6-NEXT:    s_waitcnt expcnt(0)
1250; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2736
1251; GFX6-NEXT:    s_waitcnt vmcnt(0)
1252; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2740 ; 4-byte Folded Spill
1253; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2744 ; 4-byte Folded Spill
1254; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2748 ; 4-byte Folded Spill
1255; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2752 ; 4-byte Folded Spill
1256; GFX6-NEXT:    s_waitcnt expcnt(0)
1257; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2752
1258; GFX6-NEXT:    s_waitcnt vmcnt(0)
1259; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2756 ; 4-byte Folded Spill
1260; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2760 ; 4-byte Folded Spill
1261; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2764 ; 4-byte Folded Spill
1262; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2768 ; 4-byte Folded Spill
1263; GFX6-NEXT:    s_waitcnt expcnt(0)
1264; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2768
1265; GFX6-NEXT:    s_waitcnt vmcnt(0)
1266; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2772 ; 4-byte Folded Spill
1267; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2776 ; 4-byte Folded Spill
1268; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2780 ; 4-byte Folded Spill
1269; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2784 ; 4-byte Folded Spill
1270; GFX6-NEXT:    s_waitcnt expcnt(0)
1271; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2784
1272; GFX6-NEXT:    s_waitcnt vmcnt(0)
1273; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2788 ; 4-byte Folded Spill
1274; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2792 ; 4-byte Folded Spill
1275; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2796 ; 4-byte Folded Spill
1276; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2800 ; 4-byte Folded Spill
1277; GFX6-NEXT:    s_waitcnt expcnt(0)
1278; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2800
1279; GFX6-NEXT:    s_waitcnt vmcnt(0)
1280; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2804 ; 4-byte Folded Spill
1281; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2808 ; 4-byte Folded Spill
1282; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2812 ; 4-byte Folded Spill
1283; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2816 ; 4-byte Folded Spill
1284; GFX6-NEXT:    s_waitcnt expcnt(0)
1285; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2816
1286; GFX6-NEXT:    s_waitcnt vmcnt(0)
1287; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2820 ; 4-byte Folded Spill
1288; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2824 ; 4-byte Folded Spill
1289; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2828 ; 4-byte Folded Spill
1290; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2832 ; 4-byte Folded Spill
1291; GFX6-NEXT:    s_waitcnt expcnt(0)
1292; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2832
1293; GFX6-NEXT:    s_waitcnt vmcnt(0)
1294; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2836 ; 4-byte Folded Spill
1295; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2840 ; 4-byte Folded Spill
1296; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2844 ; 4-byte Folded Spill
1297; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2848 ; 4-byte Folded Spill
1298; GFX6-NEXT:    s_waitcnt expcnt(0)
1299; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2848
1300; GFX6-NEXT:    s_waitcnt vmcnt(0)
1301; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2852 ; 4-byte Folded Spill
1302; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2856 ; 4-byte Folded Spill
1303; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2860 ; 4-byte Folded Spill
1304; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2864 ; 4-byte Folded Spill
1305; GFX6-NEXT:    s_waitcnt expcnt(0)
1306; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2864
1307; GFX6-NEXT:    s_waitcnt vmcnt(0)
1308; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2868 ; 4-byte Folded Spill
1309; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2872 ; 4-byte Folded Spill
1310; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2876 ; 4-byte Folded Spill
1311; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2880 ; 4-byte Folded Spill
1312; GFX6-NEXT:    s_waitcnt expcnt(0)
1313; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2880
1314; GFX6-NEXT:    s_waitcnt vmcnt(0)
1315; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2884 ; 4-byte Folded Spill
1316; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2888 ; 4-byte Folded Spill
1317; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2892 ; 4-byte Folded Spill
1318; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2896 ; 4-byte Folded Spill
1319; GFX6-NEXT:    s_waitcnt expcnt(0)
1320; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2896
1321; GFX6-NEXT:    s_waitcnt vmcnt(0)
1322; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2900 ; 4-byte Folded Spill
1323; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2904 ; 4-byte Folded Spill
1324; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2908 ; 4-byte Folded Spill
1325; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2912 ; 4-byte Folded Spill
1326; GFX6-NEXT:    s_waitcnt expcnt(0)
1327; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2912
1328; GFX6-NEXT:    s_waitcnt vmcnt(0)
1329; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2916 ; 4-byte Folded Spill
1330; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2920 ; 4-byte Folded Spill
1331; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2924 ; 4-byte Folded Spill
1332; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2928 ; 4-byte Folded Spill
1333; GFX6-NEXT:    s_waitcnt expcnt(0)
1334; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2928
1335; GFX6-NEXT:    s_waitcnt vmcnt(0)
1336; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2932 ; 4-byte Folded Spill
1337; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2936 ; 4-byte Folded Spill
1338; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2940 ; 4-byte Folded Spill
1339; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2944 ; 4-byte Folded Spill
1340; GFX6-NEXT:    s_waitcnt expcnt(0)
1341; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2944
1342; GFX6-NEXT:    s_waitcnt vmcnt(0)
1343; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2948 ; 4-byte Folded Spill
1344; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2952 ; 4-byte Folded Spill
1345; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2956 ; 4-byte Folded Spill
1346; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2960 ; 4-byte Folded Spill
1347; GFX6-NEXT:    s_waitcnt expcnt(0)
1348; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2960
1349; GFX6-NEXT:    s_waitcnt vmcnt(0)
1350; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2964 ; 4-byte Folded Spill
1351; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2968 ; 4-byte Folded Spill
1352; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2972 ; 4-byte Folded Spill
1353; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2976 ; 4-byte Folded Spill
1354; GFX6-NEXT:    s_waitcnt expcnt(0)
1355; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2976
1356; GFX6-NEXT:    s_waitcnt vmcnt(0)
1357; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2980 ; 4-byte Folded Spill
1358; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:2984 ; 4-byte Folded Spill
1359; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:2988 ; 4-byte Folded Spill
1360; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:2992 ; 4-byte Folded Spill
1361; GFX6-NEXT:    s_waitcnt expcnt(0)
1362; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:2992
1363; GFX6-NEXT:    s_waitcnt vmcnt(0)
1364; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:2996 ; 4-byte Folded Spill
1365; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3000 ; 4-byte Folded Spill
1366; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3004 ; 4-byte Folded Spill
1367; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3008 ; 4-byte Folded Spill
1368; GFX6-NEXT:    s_waitcnt expcnt(0)
1369; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3008
1370; GFX6-NEXT:    s_waitcnt vmcnt(0)
1371; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3012 ; 4-byte Folded Spill
1372; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3016 ; 4-byte Folded Spill
1373; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3020 ; 4-byte Folded Spill
1374; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3024 ; 4-byte Folded Spill
1375; GFX6-NEXT:    s_waitcnt expcnt(0)
1376; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3024
1377; GFX6-NEXT:    s_waitcnt vmcnt(0)
1378; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3028 ; 4-byte Folded Spill
1379; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3032 ; 4-byte Folded Spill
1380; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3036 ; 4-byte Folded Spill
1381; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3040 ; 4-byte Folded Spill
1382; GFX6-NEXT:    s_waitcnt expcnt(0)
1383; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3040
1384; GFX6-NEXT:    s_waitcnt vmcnt(0)
1385; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3044 ; 4-byte Folded Spill
1386; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3048 ; 4-byte Folded Spill
1387; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3052 ; 4-byte Folded Spill
1388; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3056 ; 4-byte Folded Spill
1389; GFX6-NEXT:    s_waitcnt expcnt(0)
1390; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3056
1391; GFX6-NEXT:    s_waitcnt vmcnt(0)
1392; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3060 ; 4-byte Folded Spill
1393; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3064 ; 4-byte Folded Spill
1394; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3068 ; 4-byte Folded Spill
1395; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3072 ; 4-byte Folded Spill
1396; GFX6-NEXT:    s_waitcnt expcnt(0)
1397; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3072
1398; GFX6-NEXT:    s_waitcnt vmcnt(0)
1399; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3076 ; 4-byte Folded Spill
1400; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3080 ; 4-byte Folded Spill
1401; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3084 ; 4-byte Folded Spill
1402; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3088 ; 4-byte Folded Spill
1403; GFX6-NEXT:    s_waitcnt expcnt(0)
1404; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3088
1405; GFX6-NEXT:    s_waitcnt vmcnt(0)
1406; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3092 ; 4-byte Folded Spill
1407; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3096 ; 4-byte Folded Spill
1408; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3100 ; 4-byte Folded Spill
1409; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3104 ; 4-byte Folded Spill
1410; GFX6-NEXT:    s_waitcnt expcnt(0)
1411; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3104
1412; GFX6-NEXT:    s_waitcnt vmcnt(0)
1413; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3108 ; 4-byte Folded Spill
1414; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3112 ; 4-byte Folded Spill
1415; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3116 ; 4-byte Folded Spill
1416; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3120 ; 4-byte Folded Spill
1417; GFX6-NEXT:    s_waitcnt expcnt(0)
1418; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3120
1419; GFX6-NEXT:    s_waitcnt vmcnt(0)
1420; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3124 ; 4-byte Folded Spill
1421; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3128 ; 4-byte Folded Spill
1422; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3132 ; 4-byte Folded Spill
1423; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3136 ; 4-byte Folded Spill
1424; GFX6-NEXT:    s_waitcnt expcnt(0)
1425; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3136
1426; GFX6-NEXT:    s_waitcnt vmcnt(0)
1427; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3140 ; 4-byte Folded Spill
1428; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3144 ; 4-byte Folded Spill
1429; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3148 ; 4-byte Folded Spill
1430; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3152 ; 4-byte Folded Spill
1431; GFX6-NEXT:    s_waitcnt expcnt(0)
1432; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3152
1433; GFX6-NEXT:    s_waitcnt vmcnt(0)
1434; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3156 ; 4-byte Folded Spill
1435; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3160 ; 4-byte Folded Spill
1436; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3164 ; 4-byte Folded Spill
1437; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3168 ; 4-byte Folded Spill
1438; GFX6-NEXT:    s_waitcnt expcnt(0)
1439; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3168
1440; GFX6-NEXT:    s_waitcnt vmcnt(0)
1441; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3172 ; 4-byte Folded Spill
1442; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3176 ; 4-byte Folded Spill
1443; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3180 ; 4-byte Folded Spill
1444; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3184 ; 4-byte Folded Spill
1445; GFX6-NEXT:    s_waitcnt expcnt(0)
1446; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3184
1447; GFX6-NEXT:    s_waitcnt vmcnt(0)
1448; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3188 ; 4-byte Folded Spill
1449; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3192 ; 4-byte Folded Spill
1450; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3196 ; 4-byte Folded Spill
1451; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3200 ; 4-byte Folded Spill
1452; GFX6-NEXT:    s_waitcnt expcnt(0)
1453; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3200
1454; GFX6-NEXT:    s_waitcnt vmcnt(0)
1455; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3204 ; 4-byte Folded Spill
1456; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3208 ; 4-byte Folded Spill
1457; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3212 ; 4-byte Folded Spill
1458; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3216 ; 4-byte Folded Spill
1459; GFX6-NEXT:    s_waitcnt expcnt(0)
1460; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3216
1461; GFX6-NEXT:    s_waitcnt vmcnt(0)
1462; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3220 ; 4-byte Folded Spill
1463; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3224 ; 4-byte Folded Spill
1464; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3228 ; 4-byte Folded Spill
1465; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3232 ; 4-byte Folded Spill
1466; GFX6-NEXT:    s_waitcnt expcnt(0)
1467; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3232
1468; GFX6-NEXT:    s_waitcnt vmcnt(0)
1469; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3236 ; 4-byte Folded Spill
1470; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3240 ; 4-byte Folded Spill
1471; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3244 ; 4-byte Folded Spill
1472; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3248 ; 4-byte Folded Spill
1473; GFX6-NEXT:    s_waitcnt expcnt(0)
1474; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3248
1475; GFX6-NEXT:    s_waitcnt vmcnt(0)
1476; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3252 ; 4-byte Folded Spill
1477; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3256 ; 4-byte Folded Spill
1478; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3260 ; 4-byte Folded Spill
1479; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3264 ; 4-byte Folded Spill
1480; GFX6-NEXT:    s_waitcnt expcnt(0)
1481; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3264
1482; GFX6-NEXT:    s_waitcnt vmcnt(0)
1483; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3268 ; 4-byte Folded Spill
1484; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3272 ; 4-byte Folded Spill
1485; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3276 ; 4-byte Folded Spill
1486; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3280 ; 4-byte Folded Spill
1487; GFX6-NEXT:    s_waitcnt expcnt(0)
1488; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3280
1489; GFX6-NEXT:    s_waitcnt vmcnt(0)
1490; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3284 ; 4-byte Folded Spill
1491; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3288 ; 4-byte Folded Spill
1492; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3292 ; 4-byte Folded Spill
1493; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3296 ; 4-byte Folded Spill
1494; GFX6-NEXT:    s_waitcnt expcnt(0)
1495; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3296
1496; GFX6-NEXT:    s_waitcnt vmcnt(0)
1497; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3300 ; 4-byte Folded Spill
1498; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3304 ; 4-byte Folded Spill
1499; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3308 ; 4-byte Folded Spill
1500; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3312 ; 4-byte Folded Spill
1501; GFX6-NEXT:    s_waitcnt expcnt(0)
1502; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3312
1503; GFX6-NEXT:    s_waitcnt vmcnt(0)
1504; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3316 ; 4-byte Folded Spill
1505; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3320 ; 4-byte Folded Spill
1506; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3324 ; 4-byte Folded Spill
1507; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3328 ; 4-byte Folded Spill
1508; GFX6-NEXT:    s_waitcnt expcnt(0)
1509; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3328
1510; GFX6-NEXT:    s_waitcnt vmcnt(0)
1511; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3332 ; 4-byte Folded Spill
1512; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3336 ; 4-byte Folded Spill
1513; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3340 ; 4-byte Folded Spill
1514; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3344 ; 4-byte Folded Spill
1515; GFX6-NEXT:    s_waitcnt expcnt(0)
1516; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3344
1517; GFX6-NEXT:    s_waitcnt vmcnt(0)
1518; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3348 ; 4-byte Folded Spill
1519; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3352 ; 4-byte Folded Spill
1520; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3356 ; 4-byte Folded Spill
1521; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3360 ; 4-byte Folded Spill
1522; GFX6-NEXT:    s_waitcnt expcnt(0)
1523; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3360
1524; GFX6-NEXT:    s_waitcnt vmcnt(0)
1525; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3364 ; 4-byte Folded Spill
1526; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3368 ; 4-byte Folded Spill
1527; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3372 ; 4-byte Folded Spill
1528; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3376 ; 4-byte Folded Spill
1529; GFX6-NEXT:    s_waitcnt expcnt(0)
1530; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3376
1531; GFX6-NEXT:    s_waitcnt vmcnt(0)
1532; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3380 ; 4-byte Folded Spill
1533; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3384 ; 4-byte Folded Spill
1534; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3388 ; 4-byte Folded Spill
1535; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3392 ; 4-byte Folded Spill
1536; GFX6-NEXT:    s_waitcnt expcnt(0)
1537; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3392
1538; GFX6-NEXT:    s_waitcnt vmcnt(0)
1539; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3396 ; 4-byte Folded Spill
1540; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3400 ; 4-byte Folded Spill
1541; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3404 ; 4-byte Folded Spill
1542; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3408 ; 4-byte Folded Spill
1543; GFX6-NEXT:    s_waitcnt expcnt(0)
1544; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3408
1545; GFX6-NEXT:    s_waitcnt vmcnt(0)
1546; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3412 ; 4-byte Folded Spill
1547; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3416 ; 4-byte Folded Spill
1548; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3420 ; 4-byte Folded Spill
1549; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3424 ; 4-byte Folded Spill
1550; GFX6-NEXT:    s_waitcnt expcnt(0)
1551; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3424
1552; GFX6-NEXT:    s_waitcnt vmcnt(0)
1553; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3428 ; 4-byte Folded Spill
1554; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3432 ; 4-byte Folded Spill
1555; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3436 ; 4-byte Folded Spill
1556; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3440 ; 4-byte Folded Spill
1557; GFX6-NEXT:    s_waitcnt expcnt(0)
1558; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3440
1559; GFX6-NEXT:    s_waitcnt vmcnt(0)
1560; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3444 ; 4-byte Folded Spill
1561; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3448 ; 4-byte Folded Spill
1562; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3452 ; 4-byte Folded Spill
1563; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3456 ; 4-byte Folded Spill
1564; GFX6-NEXT:    s_waitcnt expcnt(0)
1565; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3456
1566; GFX6-NEXT:    s_waitcnt vmcnt(0)
1567; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3460 ; 4-byte Folded Spill
1568; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3464 ; 4-byte Folded Spill
1569; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3468 ; 4-byte Folded Spill
1570; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3472 ; 4-byte Folded Spill
1571; GFX6-NEXT:    s_waitcnt expcnt(0)
1572; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3472
1573; GFX6-NEXT:    s_waitcnt vmcnt(0)
1574; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3476 ; 4-byte Folded Spill
1575; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3480 ; 4-byte Folded Spill
1576; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3484 ; 4-byte Folded Spill
1577; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3488 ; 4-byte Folded Spill
1578; GFX6-NEXT:    s_waitcnt expcnt(0)
1579; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3488
1580; GFX6-NEXT:    s_waitcnt vmcnt(0)
1581; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3492 ; 4-byte Folded Spill
1582; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3496 ; 4-byte Folded Spill
1583; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3500 ; 4-byte Folded Spill
1584; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3504 ; 4-byte Folded Spill
1585; GFX6-NEXT:    s_waitcnt expcnt(0)
1586; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3504
1587; GFX6-NEXT:    s_waitcnt vmcnt(0)
1588; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3508 ; 4-byte Folded Spill
1589; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3512 ; 4-byte Folded Spill
1590; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3516 ; 4-byte Folded Spill
1591; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3520 ; 4-byte Folded Spill
1592; GFX6-NEXT:    s_waitcnt expcnt(0)
1593; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3520
1594; GFX6-NEXT:    s_waitcnt vmcnt(0)
1595; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3524 ; 4-byte Folded Spill
1596; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3528 ; 4-byte Folded Spill
1597; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3532 ; 4-byte Folded Spill
1598; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3536 ; 4-byte Folded Spill
1599; GFX6-NEXT:    s_waitcnt expcnt(0)
1600; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3536
1601; GFX6-NEXT:    s_waitcnt vmcnt(0)
1602; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3540 ; 4-byte Folded Spill
1603; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3544 ; 4-byte Folded Spill
1604; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3548 ; 4-byte Folded Spill
1605; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3552 ; 4-byte Folded Spill
1606; GFX6-NEXT:    s_waitcnt expcnt(0)
1607; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3552
1608; GFX6-NEXT:    s_waitcnt vmcnt(0)
1609; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3556 ; 4-byte Folded Spill
1610; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3560 ; 4-byte Folded Spill
1611; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3564 ; 4-byte Folded Spill
1612; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3568 ; 4-byte Folded Spill
1613; GFX6-NEXT:    s_waitcnt expcnt(0)
1614; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3568
1615; GFX6-NEXT:    s_waitcnt vmcnt(0)
1616; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3572 ; 4-byte Folded Spill
1617; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3576 ; 4-byte Folded Spill
1618; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3580 ; 4-byte Folded Spill
1619; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3584 ; 4-byte Folded Spill
1620; GFX6-NEXT:    s_waitcnt expcnt(0)
1621; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3584
1622; GFX6-NEXT:    s_waitcnt vmcnt(0)
1623; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3588 ; 4-byte Folded Spill
1624; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3592 ; 4-byte Folded Spill
1625; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3596 ; 4-byte Folded Spill
1626; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3600 ; 4-byte Folded Spill
1627; GFX6-NEXT:    s_waitcnt expcnt(0)
1628; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3600
1629; GFX6-NEXT:    s_waitcnt vmcnt(0)
1630; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3604 ; 4-byte Folded Spill
1631; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3608 ; 4-byte Folded Spill
1632; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3612 ; 4-byte Folded Spill
1633; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3616 ; 4-byte Folded Spill
1634; GFX6-NEXT:    s_waitcnt expcnt(0)
1635; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3616
1636; GFX6-NEXT:    s_waitcnt vmcnt(0)
1637; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3620 ; 4-byte Folded Spill
1638; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3624 ; 4-byte Folded Spill
1639; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3628 ; 4-byte Folded Spill
1640; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3632 ; 4-byte Folded Spill
1641; GFX6-NEXT:    s_waitcnt expcnt(0)
1642; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3632
1643; GFX6-NEXT:    s_waitcnt vmcnt(0)
1644; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3636 ; 4-byte Folded Spill
1645; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3640 ; 4-byte Folded Spill
1646; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3644 ; 4-byte Folded Spill
1647; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3648 ; 4-byte Folded Spill
1648; GFX6-NEXT:    s_waitcnt expcnt(0)
1649; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3648
1650; GFX6-NEXT:    s_waitcnt vmcnt(0)
1651; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3652 ; 4-byte Folded Spill
1652; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3656 ; 4-byte Folded Spill
1653; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3660 ; 4-byte Folded Spill
1654; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3664 ; 4-byte Folded Spill
1655; GFX6-NEXT:    s_waitcnt expcnt(0)
1656; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3664
1657; GFX6-NEXT:    s_waitcnt vmcnt(0)
1658; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3668 ; 4-byte Folded Spill
1659; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3672 ; 4-byte Folded Spill
1660; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3676 ; 4-byte Folded Spill
1661; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3680 ; 4-byte Folded Spill
1662; GFX6-NEXT:    s_waitcnt expcnt(0)
1663; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3680
1664; GFX6-NEXT:    s_waitcnt vmcnt(0)
1665; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3684 ; 4-byte Folded Spill
1666; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3688 ; 4-byte Folded Spill
1667; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3692 ; 4-byte Folded Spill
1668; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3696 ; 4-byte Folded Spill
1669; GFX6-NEXT:    s_waitcnt expcnt(0)
1670; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3696
1671; GFX6-NEXT:    s_waitcnt vmcnt(0)
1672; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3700 ; 4-byte Folded Spill
1673; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3704 ; 4-byte Folded Spill
1674; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3708 ; 4-byte Folded Spill
1675; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3712 ; 4-byte Folded Spill
1676; GFX6-NEXT:    s_waitcnt expcnt(0)
1677; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3712
1678; GFX6-NEXT:    s_waitcnt vmcnt(0)
1679; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3716 ; 4-byte Folded Spill
1680; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3720 ; 4-byte Folded Spill
1681; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3724 ; 4-byte Folded Spill
1682; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3728 ; 4-byte Folded Spill
1683; GFX6-NEXT:    s_waitcnt expcnt(0)
1684; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3728
1685; GFX6-NEXT:    s_waitcnt vmcnt(0)
1686; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3732 ; 4-byte Folded Spill
1687; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3736 ; 4-byte Folded Spill
1688; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3740 ; 4-byte Folded Spill
1689; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3744 ; 4-byte Folded Spill
1690; GFX6-NEXT:    s_waitcnt expcnt(0)
1691; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3744
1692; GFX6-NEXT:    s_waitcnt vmcnt(0)
1693; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3748 ; 4-byte Folded Spill
1694; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3752 ; 4-byte Folded Spill
1695; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3756 ; 4-byte Folded Spill
1696; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3760 ; 4-byte Folded Spill
1697; GFX6-NEXT:    s_waitcnt expcnt(0)
1698; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3760
1699; GFX6-NEXT:    s_waitcnt vmcnt(0)
1700; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3764 ; 4-byte Folded Spill
1701; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3768 ; 4-byte Folded Spill
1702; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3772 ; 4-byte Folded Spill
1703; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3776 ; 4-byte Folded Spill
1704; GFX6-NEXT:    s_waitcnt expcnt(0)
1705; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3776
1706; GFX6-NEXT:    s_waitcnt vmcnt(0)
1707; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3780 ; 4-byte Folded Spill
1708; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3784 ; 4-byte Folded Spill
1709; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3788 ; 4-byte Folded Spill
1710; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3792 ; 4-byte Folded Spill
1711; GFX6-NEXT:    s_waitcnt expcnt(0)
1712; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3792
1713; GFX6-NEXT:    s_waitcnt vmcnt(0)
1714; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3796 ; 4-byte Folded Spill
1715; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3800 ; 4-byte Folded Spill
1716; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3804 ; 4-byte Folded Spill
1717; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3808 ; 4-byte Folded Spill
1718; GFX6-NEXT:    s_waitcnt expcnt(0)
1719; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3808
1720; GFX6-NEXT:    s_waitcnt vmcnt(0)
1721; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3812 ; 4-byte Folded Spill
1722; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3816 ; 4-byte Folded Spill
1723; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3820 ; 4-byte Folded Spill
1724; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3824 ; 4-byte Folded Spill
1725; GFX6-NEXT:    s_waitcnt expcnt(0)
1726; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3824
1727; GFX6-NEXT:    s_waitcnt vmcnt(0)
1728; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3828 ; 4-byte Folded Spill
1729; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3832 ; 4-byte Folded Spill
1730; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3836 ; 4-byte Folded Spill
1731; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3840 ; 4-byte Folded Spill
1732; GFX6-NEXT:    s_waitcnt expcnt(0)
1733; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3840
1734; GFX6-NEXT:    s_waitcnt vmcnt(0)
1735; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3844 ; 4-byte Folded Spill
1736; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3848 ; 4-byte Folded Spill
1737; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3852 ; 4-byte Folded Spill
1738; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3856 ; 4-byte Folded Spill
1739; GFX6-NEXT:    s_waitcnt expcnt(0)
1740; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3856
1741; GFX6-NEXT:    s_waitcnt vmcnt(0)
1742; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3860 ; 4-byte Folded Spill
1743; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3864 ; 4-byte Folded Spill
1744; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3868 ; 4-byte Folded Spill
1745; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3872 ; 4-byte Folded Spill
1746; GFX6-NEXT:    s_waitcnt expcnt(0)
1747; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3872
1748; GFX6-NEXT:    s_waitcnt vmcnt(0)
1749; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3876 ; 4-byte Folded Spill
1750; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3880 ; 4-byte Folded Spill
1751; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3884 ; 4-byte Folded Spill
1752; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3888 ; 4-byte Folded Spill
1753; GFX6-NEXT:    s_waitcnt expcnt(0)
1754; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3888
1755; GFX6-NEXT:    s_waitcnt vmcnt(0)
1756; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3892 ; 4-byte Folded Spill
1757; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3896 ; 4-byte Folded Spill
1758; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3900 ; 4-byte Folded Spill
1759; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3904 ; 4-byte Folded Spill
1760; GFX6-NEXT:    s_waitcnt expcnt(0)
1761; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3904
1762; GFX6-NEXT:    s_waitcnt vmcnt(0)
1763; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3908 ; 4-byte Folded Spill
1764; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3912 ; 4-byte Folded Spill
1765; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3916 ; 4-byte Folded Spill
1766; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3920 ; 4-byte Folded Spill
1767; GFX6-NEXT:    s_waitcnt expcnt(0)
1768; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3920
1769; GFX6-NEXT:    s_waitcnt vmcnt(0)
1770; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3924 ; 4-byte Folded Spill
1771; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3928 ; 4-byte Folded Spill
1772; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3932 ; 4-byte Folded Spill
1773; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3936 ; 4-byte Folded Spill
1774; GFX6-NEXT:    s_waitcnt expcnt(0)
1775; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3936
1776; GFX6-NEXT:    s_waitcnt vmcnt(0)
1777; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3940 ; 4-byte Folded Spill
1778; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3944 ; 4-byte Folded Spill
1779; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3948 ; 4-byte Folded Spill
1780; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3952 ; 4-byte Folded Spill
1781; GFX6-NEXT:    s_waitcnt expcnt(0)
1782; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3952
1783; GFX6-NEXT:    s_waitcnt vmcnt(0)
1784; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3956 ; 4-byte Folded Spill
1785; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3960 ; 4-byte Folded Spill
1786; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3964 ; 4-byte Folded Spill
1787; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3968 ; 4-byte Folded Spill
1788; GFX6-NEXT:    s_waitcnt expcnt(0)
1789; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3968
1790; GFX6-NEXT:    s_waitcnt vmcnt(0)
1791; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3972 ; 4-byte Folded Spill
1792; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3976 ; 4-byte Folded Spill
1793; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3980 ; 4-byte Folded Spill
1794; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:3984 ; 4-byte Folded Spill
1795; GFX6-NEXT:    s_waitcnt expcnt(0)
1796; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:3984
1797; GFX6-NEXT:    s_waitcnt vmcnt(0)
1798; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:3988 ; 4-byte Folded Spill
1799; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:3992 ; 4-byte Folded Spill
1800; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:3996 ; 4-byte Folded Spill
1801; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4000 ; 4-byte Folded Spill
1802; GFX6-NEXT:    s_waitcnt expcnt(0)
1803; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4000
1804; GFX6-NEXT:    s_waitcnt vmcnt(0)
1805; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4004 ; 4-byte Folded Spill
1806; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:4008 ; 4-byte Folded Spill
1807; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:4012 ; 4-byte Folded Spill
1808; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4016 ; 4-byte Folded Spill
1809; GFX6-NEXT:    s_waitcnt expcnt(0)
1810; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4016
1811; GFX6-NEXT:    s_waitcnt vmcnt(0)
1812; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4020 ; 4-byte Folded Spill
1813; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:4024 ; 4-byte Folded Spill
1814; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:4028 ; 4-byte Folded Spill
1815; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4032 ; 4-byte Folded Spill
1816; GFX6-NEXT:    s_waitcnt expcnt(0)
1817; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4032
1818; GFX6-NEXT:    s_waitcnt vmcnt(0)
1819; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4036 ; 4-byte Folded Spill
1820; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:4040 ; 4-byte Folded Spill
1821; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:4044 ; 4-byte Folded Spill
1822; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4048 ; 4-byte Folded Spill
1823; GFX6-NEXT:    s_waitcnt expcnt(0)
1824; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4048
1825; GFX6-NEXT:    s_waitcnt vmcnt(0)
1826; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4052 ; 4-byte Folded Spill
1827; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:4056 ; 4-byte Folded Spill
1828; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:4060 ; 4-byte Folded Spill
1829; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4064 ; 4-byte Folded Spill
1830; GFX6-NEXT:    s_waitcnt expcnt(0)
1831; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4064
1832; GFX6-NEXT:    s_waitcnt vmcnt(0)
1833; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], 0 offset:4068 ; 4-byte Folded Spill
1834; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0 offset:4072 ; 4-byte Folded Spill
1835; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], 0 offset:4076 ; 4-byte Folded Spill
1836; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], 0 offset:4080 ; 4-byte Folded Spill
1837; GFX6-NEXT:    s_waitcnt expcnt(0)
1838; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:4080
1839; GFX6-NEXT:    s_mov_b64 s[4:5], 0x80
1840; GFX6-NEXT:    s_waitcnt vmcnt(0)
1841; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1842; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1843; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1844; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1845; GFX6-NEXT:    s_waitcnt expcnt(0)
1846; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:3968
1847; GFX6-NEXT:    s_mov_b32 s2, 0x40100
1848; GFX6-NEXT:    s_waitcnt vmcnt(0)
1849; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1850; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1851; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1852; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1853; GFX6-NEXT:    s_waitcnt expcnt(0)
1854; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:3984
1855; GFX6-NEXT:    s_mov_b32 s2, 0x40500
1856; GFX6-NEXT:    s_waitcnt vmcnt(0)
1857; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1858; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1859; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1860; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1861; GFX6-NEXT:    s_waitcnt expcnt(0)
1862; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4000
1863; GFX6-NEXT:    s_mov_b32 s2, 0x40900
1864; GFX6-NEXT:    s_waitcnt vmcnt(0)
1865; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1866; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1867; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1868; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1869; GFX6-NEXT:    s_waitcnt expcnt(0)
1870; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4016
1871; GFX6-NEXT:    s_mov_b32 s2, 0x40d00
1872; GFX6-NEXT:    s_waitcnt vmcnt(0)
1873; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1874; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1875; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1876; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1877; GFX6-NEXT:    s_waitcnt expcnt(0)
1878; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4032
1879; GFX6-NEXT:    s_mov_b32 s2, 0x41100
1880; GFX6-NEXT:    s_waitcnt vmcnt(0)
1881; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1882; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1883; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1884; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1885; GFX6-NEXT:    s_waitcnt expcnt(0)
1886; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4048
1887; GFX6-NEXT:    s_mov_b32 s2, 0x41500
1888; GFX6-NEXT:    s_waitcnt vmcnt(0)
1889; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1890; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1891; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1892; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1893; GFX6-NEXT:    s_waitcnt expcnt(0)
1894; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4064
1895; GFX6-NEXT:    s_mov_b32 s2, 0x41900
1896; GFX6-NEXT:    s_waitcnt vmcnt(0)
1897; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1898; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1899; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1900; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1901; GFX6-NEXT:    s_waitcnt expcnt(0)
1902; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4080
1903; GFX6-NEXT:    s_mov_b32 s2, 0x41d00
1904; GFX6-NEXT:    s_waitcnt vmcnt(0)
1905; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1906; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1907; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1908; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1909; GFX6-NEXT:    s_waitcnt expcnt(0)
1910; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:3968
1911; GFX6-NEXT:    s_mov_b32 s2, 0x42100
1912; GFX6-NEXT:    s_waitcnt vmcnt(0)
1913; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1914; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1915; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1916; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1917; GFX6-NEXT:    s_waitcnt expcnt(0)
1918; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:3984
1919; GFX6-NEXT:    s_mov_b32 s2, 0x42500
1920; GFX6-NEXT:    s_waitcnt vmcnt(0)
1921; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1922; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1923; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1924; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1925; GFX6-NEXT:    s_waitcnt expcnt(0)
1926; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4000
1927; GFX6-NEXT:    s_mov_b32 s2, 0x42900
1928; GFX6-NEXT:    s_waitcnt vmcnt(0)
1929; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1930; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1931; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1932; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1933; GFX6-NEXT:    s_waitcnt expcnt(0)
1934; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4016
1935; GFX6-NEXT:    s_mov_b32 s2, 0x42d00
1936; GFX6-NEXT:    s_waitcnt vmcnt(0)
1937; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1938; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1939; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1940; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1941; GFX6-NEXT:    s_waitcnt expcnt(0)
1942; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4032
1943; GFX6-NEXT:    s_mov_b32 s2, 0x43100
1944; GFX6-NEXT:    s_waitcnt vmcnt(0)
1945; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1946; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1947; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1948; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1949; GFX6-NEXT:    s_waitcnt expcnt(0)
1950; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4048
1951; GFX6-NEXT:    s_mov_b32 s2, 0x43500
1952; GFX6-NEXT:    s_waitcnt vmcnt(0)
1953; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1954; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1955; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1956; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1957; GFX6-NEXT:    s_waitcnt expcnt(0)
1958; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4064
1959; GFX6-NEXT:    s_mov_b32 s2, 0x43900
1960; GFX6-NEXT:    s_waitcnt vmcnt(0)
1961; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1962; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1963; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1964; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1965; GFX6-NEXT:    s_waitcnt expcnt(0)
1966; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4080
1967; GFX6-NEXT:    s_mov_b32 s2, 0x43d00
1968; GFX6-NEXT:    s_waitcnt vmcnt(0)
1969; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1970; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1971; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1972; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1973; GFX6-NEXT:    s_waitcnt expcnt(0)
1974; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:3968
1975; GFX6-NEXT:    s_mov_b32 s2, 0x44100
1976; GFX6-NEXT:    s_waitcnt vmcnt(0)
1977; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1978; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1979; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1980; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1981; GFX6-NEXT:    s_waitcnt expcnt(0)
1982; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:3984
1983; GFX6-NEXT:    s_mov_b32 s2, 0x44500
1984; GFX6-NEXT:    s_waitcnt vmcnt(0)
1985; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1986; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1987; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1988; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1989; GFX6-NEXT:    s_waitcnt expcnt(0)
1990; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4000
1991; GFX6-NEXT:    s_mov_b32 s2, 0x44900
1992; GFX6-NEXT:    s_waitcnt vmcnt(0)
1993; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
1994; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
1995; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
1996; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
1997; GFX6-NEXT:    s_waitcnt expcnt(0)
1998; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4016
1999; GFX6-NEXT:    s_mov_b32 s2, 0x44d00
2000; GFX6-NEXT:    s_waitcnt vmcnt(0)
2001; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2002; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2003; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2004; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2005; GFX6-NEXT:    s_waitcnt expcnt(0)
2006; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4032
2007; GFX6-NEXT:    s_mov_b32 s2, 0x45100
2008; GFX6-NEXT:    s_waitcnt vmcnt(0)
2009; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2010; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2011; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2012; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2013; GFX6-NEXT:    s_waitcnt expcnt(0)
2014; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4048
2015; GFX6-NEXT:    s_mov_b32 s2, 0x45500
2016; GFX6-NEXT:    s_waitcnt vmcnt(0)
2017; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2018; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2019; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2020; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2021; GFX6-NEXT:    s_waitcnt expcnt(0)
2022; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4064
2023; GFX6-NEXT:    s_mov_b32 s2, 0x45900
2024; GFX6-NEXT:    s_waitcnt vmcnt(0)
2025; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2026; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2027; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2028; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2029; GFX6-NEXT:    s_waitcnt expcnt(0)
2030; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4080
2031; GFX6-NEXT:    s_mov_b32 s2, 0x45d00
2032; GFX6-NEXT:    s_waitcnt vmcnt(0)
2033; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2034; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2035; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2036; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2037; GFX6-NEXT:    s_waitcnt expcnt(0)
2038; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:3968
2039; GFX6-NEXT:    s_mov_b32 s2, 0x46100
2040; GFX6-NEXT:    s_waitcnt vmcnt(0)
2041; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2042; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2043; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2044; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2045; GFX6-NEXT:    s_waitcnt expcnt(0)
2046; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:3984
2047; GFX6-NEXT:    s_mov_b32 s2, 0x46500
2048; GFX6-NEXT:    s_waitcnt vmcnt(0)
2049; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2050; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2051; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2052; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2053; GFX6-NEXT:    s_waitcnt expcnt(0)
2054; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4000
2055; GFX6-NEXT:    s_mov_b32 s2, 0x46900
2056; GFX6-NEXT:    s_waitcnt vmcnt(0)
2057; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2058; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2059; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2060; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2061; GFX6-NEXT:    s_waitcnt expcnt(0)
2062; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4016
2063; GFX6-NEXT:    s_mov_b32 s2, 0x46d00
2064; GFX6-NEXT:    s_waitcnt vmcnt(0)
2065; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2066; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2067; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2068; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2069; GFX6-NEXT:    s_waitcnt expcnt(0)
2070; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4032
2071; GFX6-NEXT:    s_mov_b32 s2, 0x47100
2072; GFX6-NEXT:    s_waitcnt vmcnt(0)
2073; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2074; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2075; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2076; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2077; GFX6-NEXT:    s_waitcnt expcnt(0)
2078; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4048
2079; GFX6-NEXT:    s_mov_b32 s2, 0x47500
2080; GFX6-NEXT:    s_waitcnt vmcnt(0)
2081; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2082; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2083; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2084; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2085; GFX6-NEXT:    s_waitcnt expcnt(0)
2086; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4064
2087; GFX6-NEXT:    s_mov_b32 s2, 0x47900
2088; GFX6-NEXT:    s_waitcnt vmcnt(0)
2089; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2090; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2091; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2092; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2093; GFX6-NEXT:    s_waitcnt expcnt(0)
2094; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4080
2095; GFX6-NEXT:    s_mov_b32 s2, 0x47d00
2096; GFX6-NEXT:    s_waitcnt vmcnt(0)
2097; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2098; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2099; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2100; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2101; GFX6-NEXT:    s_waitcnt expcnt(0)
2102; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:3968
2103; GFX6-NEXT:    s_mov_b32 s2, 0x48100
2104; GFX6-NEXT:    s_waitcnt vmcnt(0)
2105; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2106; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2107; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2108; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2109; GFX6-NEXT:    s_waitcnt expcnt(0)
2110; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:3984
2111; GFX6-NEXT:    s_mov_b32 s2, 0x48500
2112; GFX6-NEXT:    s_waitcnt vmcnt(0)
2113; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2114; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2115; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2116; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2117; GFX6-NEXT:    s_waitcnt expcnt(0)
2118; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4000
2119; GFX6-NEXT:    s_mov_b32 s2, 0x48900
2120; GFX6-NEXT:    s_waitcnt vmcnt(0)
2121; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2122; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2123; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2124; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2125; GFX6-NEXT:    s_waitcnt expcnt(0)
2126; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4016
2127; GFX6-NEXT:    s_mov_b32 s2, 0x48d00
2128; GFX6-NEXT:    s_waitcnt vmcnt(0)
2129; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2130; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2131; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2132; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2133; GFX6-NEXT:    s_waitcnt expcnt(0)
2134; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4032
2135; GFX6-NEXT:    s_mov_b32 s2, 0x49100
2136; GFX6-NEXT:    s_waitcnt vmcnt(0)
2137; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2138; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2139; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2140; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2141; GFX6-NEXT:    s_waitcnt expcnt(0)
2142; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4048
2143; GFX6-NEXT:    s_mov_b32 s2, 0x49500
2144; GFX6-NEXT:    s_waitcnt vmcnt(0)
2145; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2146; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2147; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2148; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2149; GFX6-NEXT:    s_waitcnt expcnt(0)
2150; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4064
2151; GFX6-NEXT:    s_mov_b32 s2, 0x49900
2152; GFX6-NEXT:    s_waitcnt vmcnt(0)
2153; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2154; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2155; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2156; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2157; GFX6-NEXT:    s_waitcnt expcnt(0)
2158; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4080
2159; GFX6-NEXT:    s_mov_b32 s2, 0x49d00
2160; GFX6-NEXT:    s_waitcnt vmcnt(0)
2161; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2162; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2163; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2164; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2165; GFX6-NEXT:    s_waitcnt expcnt(0)
2166; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:3968
2167; GFX6-NEXT:    s_mov_b32 s2, 0x4a100
2168; GFX6-NEXT:    s_waitcnt vmcnt(0)
2169; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2170; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2171; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2172; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2173; GFX6-NEXT:    s_waitcnt expcnt(0)
2174; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:3984
2175; GFX6-NEXT:    s_mov_b32 s2, 0x4a500
2176; GFX6-NEXT:    s_waitcnt vmcnt(0)
2177; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2178; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2179; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2180; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2181; GFX6-NEXT:    s_waitcnt expcnt(0)
2182; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4000
2183; GFX6-NEXT:    s_mov_b32 s2, 0x4a900
2184; GFX6-NEXT:    s_waitcnt vmcnt(0)
2185; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2186; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2187; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2188; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2189; GFX6-NEXT:    s_waitcnt expcnt(0)
2190; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4016
2191; GFX6-NEXT:    s_mov_b32 s2, 0x4ad00
2192; GFX6-NEXT:    s_waitcnt vmcnt(0)
2193; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2194; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2195; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2196; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2197; GFX6-NEXT:    s_waitcnt expcnt(0)
2198; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4032
2199; GFX6-NEXT:    s_mov_b32 s2, 0x4b100
2200; GFX6-NEXT:    s_waitcnt vmcnt(0)
2201; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2202; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2203; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2204; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2205; GFX6-NEXT:    s_waitcnt expcnt(0)
2206; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4048
2207; GFX6-NEXT:    s_mov_b32 s2, 0x4b500
2208; GFX6-NEXT:    s_waitcnt vmcnt(0)
2209; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2210; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2211; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2212; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2213; GFX6-NEXT:    s_waitcnt expcnt(0)
2214; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4064
2215; GFX6-NEXT:    s_mov_b32 s2, 0x4b900
2216; GFX6-NEXT:    s_waitcnt vmcnt(0)
2217; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2218; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2219; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2220; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2221; GFX6-NEXT:    s_waitcnt expcnt(0)
2222; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4080
2223; GFX6-NEXT:    s_mov_b32 s2, 0x4bd00
2224; GFX6-NEXT:    s_waitcnt vmcnt(0)
2225; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2226; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2227; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2228; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2229; GFX6-NEXT:    s_waitcnt expcnt(0)
2230; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:3968
2231; GFX6-NEXT:    s_mov_b32 s2, 0x4c100
2232; GFX6-NEXT:    s_waitcnt vmcnt(0)
2233; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2234; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2235; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2236; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2237; GFX6-NEXT:    s_waitcnt expcnt(0)
2238; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:3984
2239; GFX6-NEXT:    s_mov_b32 s2, 0x4c500
2240; GFX6-NEXT:    s_waitcnt vmcnt(0)
2241; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2242; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2243; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2244; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2245; GFX6-NEXT:    s_waitcnt expcnt(0)
2246; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4000
2247; GFX6-NEXT:    s_mov_b32 s2, 0x4c900
2248; GFX6-NEXT:    s_waitcnt vmcnt(0)
2249; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2250; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2251; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2252; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2253; GFX6-NEXT:    s_waitcnt expcnt(0)
2254; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4016
2255; GFX6-NEXT:    s_mov_b32 s2, 0x4cd00
2256; GFX6-NEXT:    s_waitcnt vmcnt(0)
2257; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2258; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2259; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2260; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2261; GFX6-NEXT:    s_waitcnt expcnt(0)
2262; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4032
2263; GFX6-NEXT:    s_mov_b32 s2, 0x4d100
2264; GFX6-NEXT:    s_waitcnt vmcnt(0)
2265; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2266; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2267; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2268; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2269; GFX6-NEXT:    s_waitcnt expcnt(0)
2270; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4048
2271; GFX6-NEXT:    s_mov_b32 s2, 0x4d500
2272; GFX6-NEXT:    s_waitcnt vmcnt(0)
2273; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2274; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2275; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2276; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2277; GFX6-NEXT:    s_waitcnt expcnt(0)
2278; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4064
2279; GFX6-NEXT:    s_mov_b32 s2, 0x4d900
2280; GFX6-NEXT:    s_waitcnt vmcnt(0)
2281; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2282; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2283; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2284; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2285; GFX6-NEXT:    s_waitcnt expcnt(0)
2286; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4080
2287; GFX6-NEXT:    s_mov_b32 s2, 0x4dd00
2288; GFX6-NEXT:    s_waitcnt vmcnt(0)
2289; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2290; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2291; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2292; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2293; GFX6-NEXT:    s_waitcnt expcnt(0)
2294; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:3968
2295; GFX6-NEXT:    s_mov_b32 s2, 0x4e100
2296; GFX6-NEXT:    s_waitcnt vmcnt(0)
2297; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2298; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2299; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2300; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2301; GFX6-NEXT:    s_waitcnt expcnt(0)
2302; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:3984
2303; GFX6-NEXT:    s_mov_b32 s2, 0x4e500
2304; GFX6-NEXT:    s_waitcnt vmcnt(0)
2305; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2306; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2307; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2308; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2309; GFX6-NEXT:    s_waitcnt expcnt(0)
2310; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4000
2311; GFX6-NEXT:    s_mov_b32 s2, 0x4e900
2312; GFX6-NEXT:    s_waitcnt vmcnt(0)
2313; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2314; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2315; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2316; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2317; GFX6-NEXT:    s_waitcnt expcnt(0)
2318; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4016
2319; GFX6-NEXT:    s_mov_b32 s2, 0x4ed00
2320; GFX6-NEXT:    s_waitcnt vmcnt(0)
2321; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2322; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2323; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2324; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2325; GFX6-NEXT:    s_waitcnt expcnt(0)
2326; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4032
2327; GFX6-NEXT:    s_mov_b32 s2, 0x4f100
2328; GFX6-NEXT:    s_waitcnt vmcnt(0)
2329; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2330; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2331; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2332; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2333; GFX6-NEXT:    s_waitcnt expcnt(0)
2334; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4048
2335; GFX6-NEXT:    s_mov_b32 s2, 0x4f500
2336; GFX6-NEXT:    s_waitcnt vmcnt(0)
2337; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2338; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2339; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2340; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2341; GFX6-NEXT:    s_waitcnt expcnt(0)
2342; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4064
2343; GFX6-NEXT:    s_mov_b32 s2, 0x4f900
2344; GFX6-NEXT:    s_waitcnt vmcnt(0)
2345; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
2346; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
2347; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
2348; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
2349; GFX6-NEXT:    s_waitcnt expcnt(0)
2350; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4080
2351; GFX6-NEXT:    ;;#ASMSTART
2352; GFX6-NEXT:    ;;#ASMEND
2353; GFX6-NEXT:    v_add_i32_e32 v7, vcc, s0, v5
2354; GFX6-NEXT:    v_mov_b32_e32 v4, s1
2355; GFX6-NEXT:    v_addc_u32_e32 v8, vcc, 0, v4, vcc
2356; GFX6-NEXT:    ;;#ASMSTART
2357; GFX6-NEXT:    ;;#ASMEND
2358; GFX6-NEXT:    ;;#ASMSTART
2359; GFX6-NEXT:    ;;#ASMEND
2360; GFX6-NEXT:    ;;#ASMSTART
2361; GFX6-NEXT:    ;;#ASMEND
2362; GFX6-NEXT:    ;;#ASMSTART
2363; GFX6-NEXT:    ;;#ASMEND
2364; GFX6-NEXT:    ;;#ASMSTART
2365; GFX6-NEXT:    ;;#ASMEND
2366; GFX6-NEXT:    ;;#ASMSTART
2367; GFX6-NEXT:    ;;#ASMEND
2368; GFX6-NEXT:    s_mov_b64 s[2:3], s[6:7]
2369; GFX6-NEXT:    s_waitcnt vmcnt(0)
2370; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4080
2371; GFX6-NEXT:    s_waitcnt expcnt(0)
2372; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2373; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2374; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2375; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2376; GFX6-NEXT:    s_mov_b32 s33, 0x4f500
2377; GFX6-NEXT:    s_waitcnt vmcnt(0)
2378; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4064
2379; GFX6-NEXT:    s_waitcnt expcnt(0)
2380; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2381; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2382; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2383; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2384; GFX6-NEXT:    s_mov_b32 s33, 0x4f100
2385; GFX6-NEXT:    s_waitcnt vmcnt(0)
2386; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4048
2387; GFX6-NEXT:    s_waitcnt expcnt(0)
2388; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2389; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2390; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2391; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2392; GFX6-NEXT:    s_mov_b32 s33, 0x4ed00
2393; GFX6-NEXT:    s_waitcnt vmcnt(0)
2394; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4032
2395; GFX6-NEXT:    s_waitcnt expcnt(0)
2396; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2397; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2398; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2399; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2400; GFX6-NEXT:    s_mov_b32 s33, 0x4e900
2401; GFX6-NEXT:    s_waitcnt vmcnt(0)
2402; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4016
2403; GFX6-NEXT:    s_waitcnt expcnt(0)
2404; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2405; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2406; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2407; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2408; GFX6-NEXT:    s_mov_b32 s33, 0x4e500
2409; GFX6-NEXT:    s_waitcnt vmcnt(0)
2410; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:4000
2411; GFX6-NEXT:    s_waitcnt expcnt(0)
2412; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2413; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2414; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2415; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2416; GFX6-NEXT:    s_mov_b32 s33, 0x4e100
2417; GFX6-NEXT:    s_waitcnt vmcnt(0)
2418; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:3984
2419; GFX6-NEXT:    s_waitcnt expcnt(0)
2420; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2421; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2422; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2423; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2424; GFX6-NEXT:    s_mov_b32 s33, 0x4dd00
2425; GFX6-NEXT:    s_waitcnt vmcnt(0)
2426; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[36:39], 0 addr64 offset:3968
2427; GFX6-NEXT:    s_waitcnt expcnt(0)
2428; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2429; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2430; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2431; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2432; GFX6-NEXT:    s_mov_b32 s33, 0x4d900
2433; GFX6-NEXT:    s_waitcnt vmcnt(0)
2434; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4080
2435; GFX6-NEXT:    s_waitcnt expcnt(0)
2436; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2437; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2438; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2439; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2440; GFX6-NEXT:    s_mov_b32 s33, 0x4d500
2441; GFX6-NEXT:    s_waitcnt vmcnt(0)
2442; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4064
2443; GFX6-NEXT:    s_waitcnt expcnt(0)
2444; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2445; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2446; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2447; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2448; GFX6-NEXT:    s_mov_b32 s33, 0x4d100
2449; GFX6-NEXT:    s_waitcnt vmcnt(0)
2450; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4048
2451; GFX6-NEXT:    s_waitcnt expcnt(0)
2452; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2453; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2454; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2455; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2456; GFX6-NEXT:    s_mov_b32 s33, 0x4cd00
2457; GFX6-NEXT:    s_waitcnt vmcnt(0)
2458; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4032
2459; GFX6-NEXT:    s_waitcnt expcnt(0)
2460; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2461; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2462; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2463; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2464; GFX6-NEXT:    s_mov_b32 s33, 0x4c900
2465; GFX6-NEXT:    s_waitcnt vmcnt(0)
2466; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4016
2467; GFX6-NEXT:    s_waitcnt expcnt(0)
2468; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2469; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2470; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2471; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2472; GFX6-NEXT:    s_mov_b32 s33, 0x4c500
2473; GFX6-NEXT:    s_waitcnt vmcnt(0)
2474; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:4000
2475; GFX6-NEXT:    s_waitcnt expcnt(0)
2476; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2477; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2478; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2479; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2480; GFX6-NEXT:    s_mov_b32 s33, 0x4c100
2481; GFX6-NEXT:    s_waitcnt vmcnt(0)
2482; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:3984
2483; GFX6-NEXT:    s_waitcnt expcnt(0)
2484; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s33 ; 4-byte Folded Reload
2485; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s33 offset:4 ; 4-byte Folded Reload
2486; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s33 offset:8 ; 4-byte Folded Reload
2487; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s33 offset:12 ; 4-byte Folded Reload
2488; GFX6-NEXT:    s_waitcnt vmcnt(0)
2489; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[28:31], 0 addr64 offset:3968
2490; GFX6-NEXT:    s_mov_b32 s28, 0x4bd00
2491; GFX6-NEXT:    s_waitcnt expcnt(0)
2492; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2493; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2494; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2495; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2496; GFX6-NEXT:    s_mov_b32 s28, 0x4b900
2497; GFX6-NEXT:    s_waitcnt vmcnt(0)
2498; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4080
2499; GFX6-NEXT:    s_waitcnt expcnt(0)
2500; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2501; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2502; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2503; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2504; GFX6-NEXT:    s_mov_b32 s28, 0x4b500
2505; GFX6-NEXT:    s_waitcnt vmcnt(0)
2506; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4064
2507; GFX6-NEXT:    s_waitcnt expcnt(0)
2508; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2509; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2510; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2511; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2512; GFX6-NEXT:    s_mov_b32 s28, 0x4b100
2513; GFX6-NEXT:    s_waitcnt vmcnt(0)
2514; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4048
2515; GFX6-NEXT:    s_waitcnt expcnt(0)
2516; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2517; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2518; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2519; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2520; GFX6-NEXT:    s_mov_b32 s28, 0x4ad00
2521; GFX6-NEXT:    s_waitcnt vmcnt(0)
2522; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4032
2523; GFX6-NEXT:    s_waitcnt expcnt(0)
2524; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2525; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2526; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2527; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2528; GFX6-NEXT:    s_mov_b32 s28, 0x4a900
2529; GFX6-NEXT:    s_waitcnt vmcnt(0)
2530; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4016
2531; GFX6-NEXT:    s_waitcnt expcnt(0)
2532; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2533; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2534; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2535; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2536; GFX6-NEXT:    s_mov_b32 s28, 0x4a500
2537; GFX6-NEXT:    s_waitcnt vmcnt(0)
2538; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:4000
2539; GFX6-NEXT:    s_waitcnt expcnt(0)
2540; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2541; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2542; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2543; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2544; GFX6-NEXT:    s_mov_b32 s28, 0x4a100
2545; GFX6-NEXT:    s_waitcnt vmcnt(0)
2546; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:3984
2547; GFX6-NEXT:    s_waitcnt expcnt(0)
2548; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s28 ; 4-byte Folded Reload
2549; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s28 offset:4 ; 4-byte Folded Reload
2550; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s28 offset:8 ; 4-byte Folded Reload
2551; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s28 offset:12 ; 4-byte Folded Reload
2552; GFX6-NEXT:    s_waitcnt vmcnt(0)
2553; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[24:27], 0 addr64 offset:3968
2554; GFX6-NEXT:    s_mov_b32 s24, 0x49d00
2555; GFX6-NEXT:    s_waitcnt expcnt(0)
2556; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2557; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2558; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2559; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2560; GFX6-NEXT:    s_mov_b32 s24, 0x49900
2561; GFX6-NEXT:    s_waitcnt vmcnt(0)
2562; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4080
2563; GFX6-NEXT:    s_waitcnt expcnt(0)
2564; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2565; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2566; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2567; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2568; GFX6-NEXT:    s_mov_b32 s24, 0x49500
2569; GFX6-NEXT:    s_waitcnt vmcnt(0)
2570; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4064
2571; GFX6-NEXT:    s_waitcnt expcnt(0)
2572; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2573; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2574; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2575; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2576; GFX6-NEXT:    s_mov_b32 s24, 0x49100
2577; GFX6-NEXT:    s_waitcnt vmcnt(0)
2578; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4048
2579; GFX6-NEXT:    s_waitcnt expcnt(0)
2580; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2581; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2582; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2583; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2584; GFX6-NEXT:    s_mov_b32 s24, 0x48d00
2585; GFX6-NEXT:    s_waitcnt vmcnt(0)
2586; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4032
2587; GFX6-NEXT:    s_waitcnt expcnt(0)
2588; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2589; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2590; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2591; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2592; GFX6-NEXT:    s_mov_b32 s24, 0x48900
2593; GFX6-NEXT:    s_waitcnt vmcnt(0)
2594; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4016
2595; GFX6-NEXT:    s_waitcnt expcnt(0)
2596; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2597; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2598; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2599; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2600; GFX6-NEXT:    s_mov_b32 s24, 0x48500
2601; GFX6-NEXT:    s_waitcnt vmcnt(0)
2602; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:4000
2603; GFX6-NEXT:    s_waitcnt expcnt(0)
2604; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2605; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2606; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2607; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2608; GFX6-NEXT:    s_mov_b32 s24, 0x48100
2609; GFX6-NEXT:    s_waitcnt vmcnt(0)
2610; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:3984
2611; GFX6-NEXT:    s_waitcnt expcnt(0)
2612; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s24 ; 4-byte Folded Reload
2613; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s24 offset:4 ; 4-byte Folded Reload
2614; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s24 offset:8 ; 4-byte Folded Reload
2615; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s24 offset:12 ; 4-byte Folded Reload
2616; GFX6-NEXT:    s_waitcnt vmcnt(0)
2617; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[20:23], 0 addr64 offset:3968
2618; GFX6-NEXT:    s_mov_b32 s20, 0x47d00
2619; GFX6-NEXT:    s_waitcnt expcnt(0)
2620; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2621; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2622; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2623; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2624; GFX6-NEXT:    s_mov_b32 s20, 0x47900
2625; GFX6-NEXT:    s_waitcnt vmcnt(0)
2626; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4080
2627; GFX6-NEXT:    s_waitcnt expcnt(0)
2628; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2629; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2630; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2631; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2632; GFX6-NEXT:    s_mov_b32 s20, 0x47500
2633; GFX6-NEXT:    s_waitcnt vmcnt(0)
2634; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4064
2635; GFX6-NEXT:    s_waitcnt expcnt(0)
2636; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2637; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2638; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2639; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2640; GFX6-NEXT:    s_mov_b32 s20, 0x47100
2641; GFX6-NEXT:    s_waitcnt vmcnt(0)
2642; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4048
2643; GFX6-NEXT:    s_waitcnt expcnt(0)
2644; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2645; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2646; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2647; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2648; GFX6-NEXT:    s_mov_b32 s20, 0x46d00
2649; GFX6-NEXT:    s_waitcnt vmcnt(0)
2650; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4032
2651; GFX6-NEXT:    s_waitcnt expcnt(0)
2652; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2653; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2654; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2655; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2656; GFX6-NEXT:    s_mov_b32 s20, 0x46900
2657; GFX6-NEXT:    s_waitcnt vmcnt(0)
2658; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4016
2659; GFX6-NEXT:    s_waitcnt expcnt(0)
2660; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2661; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2662; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2663; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2664; GFX6-NEXT:    s_mov_b32 s20, 0x46500
2665; GFX6-NEXT:    s_waitcnt vmcnt(0)
2666; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:4000
2667; GFX6-NEXT:    s_waitcnt expcnt(0)
2668; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2669; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2670; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2671; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2672; GFX6-NEXT:    s_mov_b32 s20, 0x46100
2673; GFX6-NEXT:    s_waitcnt vmcnt(0)
2674; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:3984
2675; GFX6-NEXT:    s_waitcnt expcnt(0)
2676; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s20 ; 4-byte Folded Reload
2677; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s20 offset:4 ; 4-byte Folded Reload
2678; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s20 offset:8 ; 4-byte Folded Reload
2679; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s20 offset:12 ; 4-byte Folded Reload
2680; GFX6-NEXT:    s_waitcnt vmcnt(0)
2681; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[16:19], 0 addr64 offset:3968
2682; GFX6-NEXT:    s_mov_b32 s16, 0x45d00
2683; GFX6-NEXT:    s_waitcnt expcnt(0)
2684; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2685; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2686; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2687; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2688; GFX6-NEXT:    s_mov_b32 s16, 0x45900
2689; GFX6-NEXT:    s_waitcnt vmcnt(0)
2690; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4080
2691; GFX6-NEXT:    s_waitcnt expcnt(0)
2692; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2693; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2694; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2695; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2696; GFX6-NEXT:    s_mov_b32 s16, 0x45500
2697; GFX6-NEXT:    s_waitcnt vmcnt(0)
2698; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4064
2699; GFX6-NEXT:    s_waitcnt expcnt(0)
2700; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2701; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2702; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2703; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2704; GFX6-NEXT:    s_mov_b32 s16, 0x45100
2705; GFX6-NEXT:    s_waitcnt vmcnt(0)
2706; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4048
2707; GFX6-NEXT:    s_waitcnt expcnt(0)
2708; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2709; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2710; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2711; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2712; GFX6-NEXT:    s_mov_b32 s16, 0x44d00
2713; GFX6-NEXT:    s_waitcnt vmcnt(0)
2714; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4032
2715; GFX6-NEXT:    s_waitcnt expcnt(0)
2716; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2717; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2718; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2719; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2720; GFX6-NEXT:    s_mov_b32 s16, 0x44900
2721; GFX6-NEXT:    s_waitcnt vmcnt(0)
2722; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4016
2723; GFX6-NEXT:    s_waitcnt expcnt(0)
2724; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2725; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2726; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2727; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2728; GFX6-NEXT:    s_mov_b32 s16, 0x44500
2729; GFX6-NEXT:    s_waitcnt vmcnt(0)
2730; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:4000
2731; GFX6-NEXT:    s_waitcnt expcnt(0)
2732; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2733; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2734; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2735; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2736; GFX6-NEXT:    s_mov_b32 s16, 0x44100
2737; GFX6-NEXT:    s_waitcnt vmcnt(0)
2738; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:3984
2739; GFX6-NEXT:    s_waitcnt expcnt(0)
2740; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s16 ; 4-byte Folded Reload
2741; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s16 offset:4 ; 4-byte Folded Reload
2742; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s16 offset:8 ; 4-byte Folded Reload
2743; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s16 offset:12 ; 4-byte Folded Reload
2744; GFX6-NEXT:    s_waitcnt vmcnt(0)
2745; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[12:15], 0 addr64 offset:3968
2746; GFX6-NEXT:    s_mov_b32 s12, 0x43d00
2747; GFX6-NEXT:    s_waitcnt expcnt(0)
2748; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2749; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2750; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2751; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2752; GFX6-NEXT:    s_mov_b32 s12, 0x43900
2753; GFX6-NEXT:    s_waitcnt vmcnt(0)
2754; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4080
2755; GFX6-NEXT:    s_waitcnt expcnt(0)
2756; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2757; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2758; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2759; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2760; GFX6-NEXT:    s_mov_b32 s12, 0x43500
2761; GFX6-NEXT:    s_waitcnt vmcnt(0)
2762; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4064
2763; GFX6-NEXT:    s_waitcnt expcnt(0)
2764; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2765; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2766; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2767; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2768; GFX6-NEXT:    s_mov_b32 s12, 0x43100
2769; GFX6-NEXT:    s_waitcnt vmcnt(0)
2770; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4048
2771; GFX6-NEXT:    s_waitcnt expcnt(0)
2772; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2773; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2774; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2775; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2776; GFX6-NEXT:    s_mov_b32 s12, 0x42d00
2777; GFX6-NEXT:    s_waitcnt vmcnt(0)
2778; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4032
2779; GFX6-NEXT:    s_waitcnt expcnt(0)
2780; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2781; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2782; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2783; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2784; GFX6-NEXT:    s_mov_b32 s12, 0x42900
2785; GFX6-NEXT:    s_waitcnt vmcnt(0)
2786; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4016
2787; GFX6-NEXT:    s_waitcnt expcnt(0)
2788; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2789; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2790; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2791; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2792; GFX6-NEXT:    s_mov_b32 s12, 0x42500
2793; GFX6-NEXT:    s_waitcnt vmcnt(0)
2794; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:4000
2795; GFX6-NEXT:    s_waitcnt expcnt(0)
2796; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2797; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2798; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2799; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2800; GFX6-NEXT:    s_mov_b32 s12, 0x42100
2801; GFX6-NEXT:    s_waitcnt vmcnt(0)
2802; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:3984
2803; GFX6-NEXT:    s_waitcnt expcnt(0)
2804; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s12 ; 4-byte Folded Reload
2805; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s12 offset:4 ; 4-byte Folded Reload
2806; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s12 offset:8 ; 4-byte Folded Reload
2807; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s12 offset:12 ; 4-byte Folded Reload
2808; GFX6-NEXT:    s_waitcnt vmcnt(0)
2809; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[8:11], 0 addr64 offset:3968
2810; GFX6-NEXT:    s_mov_b32 s8, 0x41d00
2811; GFX6-NEXT:    s_waitcnt expcnt(0)
2812; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2813; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2814; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2815; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2816; GFX6-NEXT:    s_mov_b32 s8, 0x41900
2817; GFX6-NEXT:    s_waitcnt vmcnt(0)
2818; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4080
2819; GFX6-NEXT:    s_waitcnt expcnt(0)
2820; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2821; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2822; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2823; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2824; GFX6-NEXT:    s_mov_b32 s8, 0x41500
2825; GFX6-NEXT:    s_waitcnt vmcnt(0)
2826; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4064
2827; GFX6-NEXT:    s_waitcnt expcnt(0)
2828; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2829; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2830; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2831; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2832; GFX6-NEXT:    s_mov_b32 s8, 0x41100
2833; GFX6-NEXT:    s_waitcnt vmcnt(0)
2834; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4048
2835; GFX6-NEXT:    s_waitcnt expcnt(0)
2836; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2837; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2838; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2839; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2840; GFX6-NEXT:    s_mov_b32 s8, 0x40d00
2841; GFX6-NEXT:    s_waitcnt vmcnt(0)
2842; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4032
2843; GFX6-NEXT:    s_waitcnt expcnt(0)
2844; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2845; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2846; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2847; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2848; GFX6-NEXT:    s_mov_b32 s8, 0x40900
2849; GFX6-NEXT:    s_waitcnt vmcnt(0)
2850; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4016
2851; GFX6-NEXT:    s_waitcnt expcnt(0)
2852; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2853; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2854; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2855; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2856; GFX6-NEXT:    s_mov_b32 s8, 0x40500
2857; GFX6-NEXT:    s_waitcnt vmcnt(0)
2858; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:4000
2859; GFX6-NEXT:    s_waitcnt expcnt(0)
2860; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2861; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2862; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2863; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2864; GFX6-NEXT:    s_mov_b32 s8, 0x40100
2865; GFX6-NEXT:    s_waitcnt vmcnt(0)
2866; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:3984
2867; GFX6-NEXT:    s_waitcnt expcnt(0)
2868; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s8 ; 4-byte Folded Reload
2869; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s8 offset:4 ; 4-byte Folded Reload
2870; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s8 offset:8 ; 4-byte Folded Reload
2871; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s8 offset:12 ; 4-byte Folded Reload
2872; GFX6-NEXT:    s_waitcnt vmcnt(0)
2873; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[7:8], s[4:7], 0 addr64 offset:3968
2874; GFX6-NEXT:    s_mov_b32 s4, 0x3fd00
2875; GFX6-NEXT:    s_waitcnt expcnt(0)
2876; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], s4 ; 4-byte Folded Reload
2877; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], s4 offset:4 ; 4-byte Folded Reload
2878; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], s4 offset:8 ; 4-byte Folded Reload
2879; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], s4 offset:12 ; 4-byte Folded Reload
2880; GFX6-NEXT:    s_waitcnt vmcnt(0)
2881; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4080
2882; GFX6-NEXT:    s_waitcnt expcnt(0)
2883; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4068 ; 4-byte Folded Reload
2884; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:4072 ; 4-byte Folded Reload
2885; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:4076 ; 4-byte Folded Reload
2886; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4080 ; 4-byte Folded Reload
2887; GFX6-NEXT:    s_waitcnt vmcnt(0)
2888; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4064
2889; GFX6-NEXT:    s_waitcnt expcnt(0)
2890; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4052 ; 4-byte Folded Reload
2891; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:4056 ; 4-byte Folded Reload
2892; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:4060 ; 4-byte Folded Reload
2893; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4064 ; 4-byte Folded Reload
2894; GFX6-NEXT:    s_waitcnt vmcnt(0)
2895; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4048
2896; GFX6-NEXT:    s_waitcnt expcnt(0)
2897; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4036 ; 4-byte Folded Reload
2898; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:4040 ; 4-byte Folded Reload
2899; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:4044 ; 4-byte Folded Reload
2900; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4048 ; 4-byte Folded Reload
2901; GFX6-NEXT:    s_waitcnt vmcnt(0)
2902; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4032
2903; GFX6-NEXT:    s_waitcnt expcnt(0)
2904; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4020 ; 4-byte Folded Reload
2905; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:4024 ; 4-byte Folded Reload
2906; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:4028 ; 4-byte Folded Reload
2907; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4032 ; 4-byte Folded Reload
2908; GFX6-NEXT:    s_waitcnt vmcnt(0)
2909; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4016
2910; GFX6-NEXT:    s_waitcnt expcnt(0)
2911; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4004 ; 4-byte Folded Reload
2912; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:4008 ; 4-byte Folded Reload
2913; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:4012 ; 4-byte Folded Reload
2914; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4016 ; 4-byte Folded Reload
2915; GFX6-NEXT:    s_waitcnt vmcnt(0)
2916; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:4000
2917; GFX6-NEXT:    s_waitcnt expcnt(0)
2918; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3988 ; 4-byte Folded Reload
2919; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3992 ; 4-byte Folded Reload
2920; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3996 ; 4-byte Folded Reload
2921; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:4000 ; 4-byte Folded Reload
2922; GFX6-NEXT:    s_waitcnt vmcnt(0)
2923; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3984
2924; GFX6-NEXT:    s_waitcnt expcnt(0)
2925; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3972 ; 4-byte Folded Reload
2926; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3976 ; 4-byte Folded Reload
2927; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3980 ; 4-byte Folded Reload
2928; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3984 ; 4-byte Folded Reload
2929; GFX6-NEXT:    s_waitcnt vmcnt(0)
2930; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3968
2931; GFX6-NEXT:    s_waitcnt expcnt(0)
2932; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3956 ; 4-byte Folded Reload
2933; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3960 ; 4-byte Folded Reload
2934; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3964 ; 4-byte Folded Reload
2935; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3968 ; 4-byte Folded Reload
2936; GFX6-NEXT:    s_waitcnt vmcnt(0)
2937; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3952
2938; GFX6-NEXT:    s_waitcnt expcnt(0)
2939; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3940 ; 4-byte Folded Reload
2940; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3944 ; 4-byte Folded Reload
2941; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3948 ; 4-byte Folded Reload
2942; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3952 ; 4-byte Folded Reload
2943; GFX6-NEXT:    s_waitcnt vmcnt(0)
2944; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3936
2945; GFX6-NEXT:    s_waitcnt expcnt(0)
2946; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3924 ; 4-byte Folded Reload
2947; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3928 ; 4-byte Folded Reload
2948; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3932 ; 4-byte Folded Reload
2949; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3936 ; 4-byte Folded Reload
2950; GFX6-NEXT:    s_waitcnt vmcnt(0)
2951; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3920
2952; GFX6-NEXT:    s_waitcnt expcnt(0)
2953; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3908 ; 4-byte Folded Reload
2954; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3912 ; 4-byte Folded Reload
2955; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3916 ; 4-byte Folded Reload
2956; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3920 ; 4-byte Folded Reload
2957; GFX6-NEXT:    s_waitcnt vmcnt(0)
2958; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3904
2959; GFX6-NEXT:    s_waitcnt expcnt(0)
2960; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3892 ; 4-byte Folded Reload
2961; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3896 ; 4-byte Folded Reload
2962; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3900 ; 4-byte Folded Reload
2963; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3904 ; 4-byte Folded Reload
2964; GFX6-NEXT:    s_waitcnt vmcnt(0)
2965; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3888
2966; GFX6-NEXT:    s_waitcnt expcnt(0)
2967; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3876 ; 4-byte Folded Reload
2968; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3880 ; 4-byte Folded Reload
2969; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3884 ; 4-byte Folded Reload
2970; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3888 ; 4-byte Folded Reload
2971; GFX6-NEXT:    s_waitcnt vmcnt(0)
2972; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3872
2973; GFX6-NEXT:    s_waitcnt expcnt(0)
2974; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3860 ; 4-byte Folded Reload
2975; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3864 ; 4-byte Folded Reload
2976; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3868 ; 4-byte Folded Reload
2977; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3872 ; 4-byte Folded Reload
2978; GFX6-NEXT:    s_waitcnt vmcnt(0)
2979; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3856
2980; GFX6-NEXT:    s_waitcnt expcnt(0)
2981; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3844 ; 4-byte Folded Reload
2982; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3848 ; 4-byte Folded Reload
2983; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3852 ; 4-byte Folded Reload
2984; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3856 ; 4-byte Folded Reload
2985; GFX6-NEXT:    s_waitcnt vmcnt(0)
2986; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3840
2987; GFX6-NEXT:    s_waitcnt expcnt(0)
2988; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3828 ; 4-byte Folded Reload
2989; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3832 ; 4-byte Folded Reload
2990; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3836 ; 4-byte Folded Reload
2991; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3840 ; 4-byte Folded Reload
2992; GFX6-NEXT:    s_waitcnt vmcnt(0)
2993; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3824
2994; GFX6-NEXT:    s_waitcnt expcnt(0)
2995; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3812 ; 4-byte Folded Reload
2996; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3816 ; 4-byte Folded Reload
2997; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3820 ; 4-byte Folded Reload
2998; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3824 ; 4-byte Folded Reload
2999; GFX6-NEXT:    s_waitcnt vmcnt(0)
3000; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3808
3001; GFX6-NEXT:    s_waitcnt expcnt(0)
3002; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3796 ; 4-byte Folded Reload
3003; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3800 ; 4-byte Folded Reload
3004; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3804 ; 4-byte Folded Reload
3005; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3808 ; 4-byte Folded Reload
3006; GFX6-NEXT:    s_waitcnt vmcnt(0)
3007; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3792
3008; GFX6-NEXT:    s_waitcnt expcnt(0)
3009; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3780 ; 4-byte Folded Reload
3010; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3784 ; 4-byte Folded Reload
3011; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3788 ; 4-byte Folded Reload
3012; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3792 ; 4-byte Folded Reload
3013; GFX6-NEXT:    s_waitcnt vmcnt(0)
3014; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3776
3015; GFX6-NEXT:    s_waitcnt expcnt(0)
3016; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3764 ; 4-byte Folded Reload
3017; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3768 ; 4-byte Folded Reload
3018; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3772 ; 4-byte Folded Reload
3019; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3776 ; 4-byte Folded Reload
3020; GFX6-NEXT:    s_waitcnt vmcnt(0)
3021; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3760
3022; GFX6-NEXT:    s_waitcnt expcnt(0)
3023; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3748 ; 4-byte Folded Reload
3024; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3752 ; 4-byte Folded Reload
3025; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3756 ; 4-byte Folded Reload
3026; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3760 ; 4-byte Folded Reload
3027; GFX6-NEXT:    s_waitcnt vmcnt(0)
3028; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3744
3029; GFX6-NEXT:    s_waitcnt expcnt(0)
3030; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3732 ; 4-byte Folded Reload
3031; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3736 ; 4-byte Folded Reload
3032; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3740 ; 4-byte Folded Reload
3033; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3744 ; 4-byte Folded Reload
3034; GFX6-NEXT:    s_waitcnt vmcnt(0)
3035; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3728
3036; GFX6-NEXT:    s_waitcnt expcnt(0)
3037; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3716 ; 4-byte Folded Reload
3038; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3720 ; 4-byte Folded Reload
3039; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3724 ; 4-byte Folded Reload
3040; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3728 ; 4-byte Folded Reload
3041; GFX6-NEXT:    s_waitcnt vmcnt(0)
3042; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3712
3043; GFX6-NEXT:    s_waitcnt expcnt(0)
3044; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3700 ; 4-byte Folded Reload
3045; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3704 ; 4-byte Folded Reload
3046; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3708 ; 4-byte Folded Reload
3047; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3712 ; 4-byte Folded Reload
3048; GFX6-NEXT:    s_waitcnt vmcnt(0)
3049; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3696
3050; GFX6-NEXT:    s_waitcnt expcnt(0)
3051; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3684 ; 4-byte Folded Reload
3052; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3688 ; 4-byte Folded Reload
3053; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3692 ; 4-byte Folded Reload
3054; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3696 ; 4-byte Folded Reload
3055; GFX6-NEXT:    s_waitcnt vmcnt(0)
3056; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3680
3057; GFX6-NEXT:    s_waitcnt expcnt(0)
3058; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3668 ; 4-byte Folded Reload
3059; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3672 ; 4-byte Folded Reload
3060; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3676 ; 4-byte Folded Reload
3061; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3680 ; 4-byte Folded Reload
3062; GFX6-NEXT:    s_waitcnt vmcnt(0)
3063; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3664
3064; GFX6-NEXT:    s_waitcnt expcnt(0)
3065; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3652 ; 4-byte Folded Reload
3066; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3656 ; 4-byte Folded Reload
3067; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3660 ; 4-byte Folded Reload
3068; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3664 ; 4-byte Folded Reload
3069; GFX6-NEXT:    s_waitcnt vmcnt(0)
3070; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3648
3071; GFX6-NEXT:    s_waitcnt expcnt(0)
3072; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3636 ; 4-byte Folded Reload
3073; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3640 ; 4-byte Folded Reload
3074; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3644 ; 4-byte Folded Reload
3075; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3648 ; 4-byte Folded Reload
3076; GFX6-NEXT:    s_waitcnt vmcnt(0)
3077; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3632
3078; GFX6-NEXT:    s_waitcnt expcnt(0)
3079; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3620 ; 4-byte Folded Reload
3080; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3624 ; 4-byte Folded Reload
3081; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3628 ; 4-byte Folded Reload
3082; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3632 ; 4-byte Folded Reload
3083; GFX6-NEXT:    s_waitcnt vmcnt(0)
3084; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3616
3085; GFX6-NEXT:    s_waitcnt expcnt(0)
3086; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3604 ; 4-byte Folded Reload
3087; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3608 ; 4-byte Folded Reload
3088; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3612 ; 4-byte Folded Reload
3089; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3616 ; 4-byte Folded Reload
3090; GFX6-NEXT:    s_waitcnt vmcnt(0)
3091; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3600
3092; GFX6-NEXT:    s_waitcnt expcnt(0)
3093; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3588 ; 4-byte Folded Reload
3094; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3592 ; 4-byte Folded Reload
3095; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3596 ; 4-byte Folded Reload
3096; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3600 ; 4-byte Folded Reload
3097; GFX6-NEXT:    s_waitcnt vmcnt(0)
3098; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3584
3099; GFX6-NEXT:    s_waitcnt expcnt(0)
3100; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3572 ; 4-byte Folded Reload
3101; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3576 ; 4-byte Folded Reload
3102; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3580 ; 4-byte Folded Reload
3103; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3584 ; 4-byte Folded Reload
3104; GFX6-NEXT:    s_waitcnt vmcnt(0)
3105; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3568
3106; GFX6-NEXT:    s_waitcnt expcnt(0)
3107; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3556 ; 4-byte Folded Reload
3108; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3560 ; 4-byte Folded Reload
3109; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3564 ; 4-byte Folded Reload
3110; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3568 ; 4-byte Folded Reload
3111; GFX6-NEXT:    s_waitcnt vmcnt(0)
3112; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3552
3113; GFX6-NEXT:    s_waitcnt expcnt(0)
3114; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3540 ; 4-byte Folded Reload
3115; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3544 ; 4-byte Folded Reload
3116; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3548 ; 4-byte Folded Reload
3117; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3552 ; 4-byte Folded Reload
3118; GFX6-NEXT:    s_waitcnt vmcnt(0)
3119; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3536
3120; GFX6-NEXT:    s_waitcnt expcnt(0)
3121; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3524 ; 4-byte Folded Reload
3122; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3528 ; 4-byte Folded Reload
3123; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3532 ; 4-byte Folded Reload
3124; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3536 ; 4-byte Folded Reload
3125; GFX6-NEXT:    s_waitcnt vmcnt(0)
3126; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3520
3127; GFX6-NEXT:    s_waitcnt expcnt(0)
3128; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3508 ; 4-byte Folded Reload
3129; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3512 ; 4-byte Folded Reload
3130; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3516 ; 4-byte Folded Reload
3131; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3520 ; 4-byte Folded Reload
3132; GFX6-NEXT:    s_waitcnt vmcnt(0)
3133; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3504
3134; GFX6-NEXT:    s_waitcnt expcnt(0)
3135; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3492 ; 4-byte Folded Reload
3136; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3496 ; 4-byte Folded Reload
3137; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3500 ; 4-byte Folded Reload
3138; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3504 ; 4-byte Folded Reload
3139; GFX6-NEXT:    s_waitcnt vmcnt(0)
3140; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3488
3141; GFX6-NEXT:    s_waitcnt expcnt(0)
3142; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3476 ; 4-byte Folded Reload
3143; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3480 ; 4-byte Folded Reload
3144; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3484 ; 4-byte Folded Reload
3145; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3488 ; 4-byte Folded Reload
3146; GFX6-NEXT:    s_waitcnt vmcnt(0)
3147; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3472
3148; GFX6-NEXT:    s_waitcnt expcnt(0)
3149; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3460 ; 4-byte Folded Reload
3150; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3464 ; 4-byte Folded Reload
3151; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3468 ; 4-byte Folded Reload
3152; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3472 ; 4-byte Folded Reload
3153; GFX6-NEXT:    s_waitcnt vmcnt(0)
3154; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3456
3155; GFX6-NEXT:    s_waitcnt expcnt(0)
3156; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3444 ; 4-byte Folded Reload
3157; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3448 ; 4-byte Folded Reload
3158; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3452 ; 4-byte Folded Reload
3159; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3456 ; 4-byte Folded Reload
3160; GFX6-NEXT:    s_waitcnt vmcnt(0)
3161; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3440
3162; GFX6-NEXT:    s_waitcnt expcnt(0)
3163; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3428 ; 4-byte Folded Reload
3164; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3432 ; 4-byte Folded Reload
3165; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3436 ; 4-byte Folded Reload
3166; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3440 ; 4-byte Folded Reload
3167; GFX6-NEXT:    s_waitcnt vmcnt(0)
3168; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3424
3169; GFX6-NEXT:    s_waitcnt expcnt(0)
3170; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3412 ; 4-byte Folded Reload
3171; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3416 ; 4-byte Folded Reload
3172; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3420 ; 4-byte Folded Reload
3173; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3424 ; 4-byte Folded Reload
3174; GFX6-NEXT:    s_waitcnt vmcnt(0)
3175; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3408
3176; GFX6-NEXT:    s_waitcnt expcnt(0)
3177; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3396 ; 4-byte Folded Reload
3178; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3400 ; 4-byte Folded Reload
3179; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3404 ; 4-byte Folded Reload
3180; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3408 ; 4-byte Folded Reload
3181; GFX6-NEXT:    s_waitcnt vmcnt(0)
3182; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3392
3183; GFX6-NEXT:    s_waitcnt expcnt(0)
3184; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3380 ; 4-byte Folded Reload
3185; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3384 ; 4-byte Folded Reload
3186; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3388 ; 4-byte Folded Reload
3187; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3392 ; 4-byte Folded Reload
3188; GFX6-NEXT:    s_waitcnt vmcnt(0)
3189; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3376
3190; GFX6-NEXT:    s_waitcnt expcnt(0)
3191; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3364 ; 4-byte Folded Reload
3192; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3368 ; 4-byte Folded Reload
3193; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3372 ; 4-byte Folded Reload
3194; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3376 ; 4-byte Folded Reload
3195; GFX6-NEXT:    s_waitcnt vmcnt(0)
3196; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3360
3197; GFX6-NEXT:    s_waitcnt expcnt(0)
3198; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3348 ; 4-byte Folded Reload
3199; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3352 ; 4-byte Folded Reload
3200; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3356 ; 4-byte Folded Reload
3201; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3360 ; 4-byte Folded Reload
3202; GFX6-NEXT:    s_waitcnt vmcnt(0)
3203; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3344
3204; GFX6-NEXT:    s_waitcnt expcnt(0)
3205; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3332 ; 4-byte Folded Reload
3206; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3336 ; 4-byte Folded Reload
3207; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3340 ; 4-byte Folded Reload
3208; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3344 ; 4-byte Folded Reload
3209; GFX6-NEXT:    s_waitcnt vmcnt(0)
3210; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3328
3211; GFX6-NEXT:    s_waitcnt expcnt(0)
3212; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3316 ; 4-byte Folded Reload
3213; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3320 ; 4-byte Folded Reload
3214; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3324 ; 4-byte Folded Reload
3215; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3328 ; 4-byte Folded Reload
3216; GFX6-NEXT:    s_waitcnt vmcnt(0)
3217; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3312
3218; GFX6-NEXT:    s_waitcnt expcnt(0)
3219; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3300 ; 4-byte Folded Reload
3220; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3304 ; 4-byte Folded Reload
3221; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3308 ; 4-byte Folded Reload
3222; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3312 ; 4-byte Folded Reload
3223; GFX6-NEXT:    s_waitcnt vmcnt(0)
3224; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3296
3225; GFX6-NEXT:    s_waitcnt expcnt(0)
3226; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3284 ; 4-byte Folded Reload
3227; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3288 ; 4-byte Folded Reload
3228; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3292 ; 4-byte Folded Reload
3229; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3296 ; 4-byte Folded Reload
3230; GFX6-NEXT:    s_waitcnt vmcnt(0)
3231; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3280
3232; GFX6-NEXT:    s_waitcnt expcnt(0)
3233; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3268 ; 4-byte Folded Reload
3234; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3272 ; 4-byte Folded Reload
3235; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3276 ; 4-byte Folded Reload
3236; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3280 ; 4-byte Folded Reload
3237; GFX6-NEXT:    s_waitcnt vmcnt(0)
3238; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3264
3239; GFX6-NEXT:    s_waitcnt expcnt(0)
3240; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3252 ; 4-byte Folded Reload
3241; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3256 ; 4-byte Folded Reload
3242; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3260 ; 4-byte Folded Reload
3243; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3264 ; 4-byte Folded Reload
3244; GFX6-NEXT:    s_waitcnt vmcnt(0)
3245; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3248
3246; GFX6-NEXT:    s_waitcnt expcnt(0)
3247; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3236 ; 4-byte Folded Reload
3248; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3240 ; 4-byte Folded Reload
3249; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3244 ; 4-byte Folded Reload
3250; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3248 ; 4-byte Folded Reload
3251; GFX6-NEXT:    s_waitcnt vmcnt(0)
3252; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3232
3253; GFX6-NEXT:    s_waitcnt expcnt(0)
3254; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3220 ; 4-byte Folded Reload
3255; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3224 ; 4-byte Folded Reload
3256; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3228 ; 4-byte Folded Reload
3257; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3232 ; 4-byte Folded Reload
3258; GFX6-NEXT:    s_waitcnt vmcnt(0)
3259; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3216
3260; GFX6-NEXT:    s_waitcnt expcnt(0)
3261; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3204 ; 4-byte Folded Reload
3262; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3208 ; 4-byte Folded Reload
3263; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3212 ; 4-byte Folded Reload
3264; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3216 ; 4-byte Folded Reload
3265; GFX6-NEXT:    s_waitcnt vmcnt(0)
3266; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3200
3267; GFX6-NEXT:    s_waitcnt expcnt(0)
3268; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3188 ; 4-byte Folded Reload
3269; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3192 ; 4-byte Folded Reload
3270; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3196 ; 4-byte Folded Reload
3271; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3200 ; 4-byte Folded Reload
3272; GFX6-NEXT:    s_waitcnt vmcnt(0)
3273; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3184
3274; GFX6-NEXT:    s_waitcnt expcnt(0)
3275; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3172 ; 4-byte Folded Reload
3276; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3176 ; 4-byte Folded Reload
3277; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3180 ; 4-byte Folded Reload
3278; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3184 ; 4-byte Folded Reload
3279; GFX6-NEXT:    s_waitcnt vmcnt(0)
3280; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3168
3281; GFX6-NEXT:    s_waitcnt expcnt(0)
3282; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3156 ; 4-byte Folded Reload
3283; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3160 ; 4-byte Folded Reload
3284; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3164 ; 4-byte Folded Reload
3285; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3168 ; 4-byte Folded Reload
3286; GFX6-NEXT:    s_waitcnt vmcnt(0)
3287; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3152
3288; GFX6-NEXT:    s_waitcnt expcnt(0)
3289; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3140 ; 4-byte Folded Reload
3290; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3144 ; 4-byte Folded Reload
3291; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3148 ; 4-byte Folded Reload
3292; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3152 ; 4-byte Folded Reload
3293; GFX6-NEXT:    s_waitcnt vmcnt(0)
3294; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3136
3295; GFX6-NEXT:    s_waitcnt expcnt(0)
3296; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3124 ; 4-byte Folded Reload
3297; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3128 ; 4-byte Folded Reload
3298; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3132 ; 4-byte Folded Reload
3299; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3136 ; 4-byte Folded Reload
3300; GFX6-NEXT:    s_waitcnt vmcnt(0)
3301; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3120
3302; GFX6-NEXT:    s_waitcnt expcnt(0)
3303; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3108 ; 4-byte Folded Reload
3304; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3112 ; 4-byte Folded Reload
3305; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3116 ; 4-byte Folded Reload
3306; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3120 ; 4-byte Folded Reload
3307; GFX6-NEXT:    s_waitcnt vmcnt(0)
3308; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3104
3309; GFX6-NEXT:    s_waitcnt expcnt(0)
3310; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3092 ; 4-byte Folded Reload
3311; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3096 ; 4-byte Folded Reload
3312; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3100 ; 4-byte Folded Reload
3313; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3104 ; 4-byte Folded Reload
3314; GFX6-NEXT:    s_waitcnt vmcnt(0)
3315; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3088
3316; GFX6-NEXT:    s_waitcnt expcnt(0)
3317; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3076 ; 4-byte Folded Reload
3318; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3080 ; 4-byte Folded Reload
3319; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3084 ; 4-byte Folded Reload
3320; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3088 ; 4-byte Folded Reload
3321; GFX6-NEXT:    s_waitcnt vmcnt(0)
3322; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3072
3323; GFX6-NEXT:    s_waitcnt expcnt(0)
3324; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3060 ; 4-byte Folded Reload
3325; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3064 ; 4-byte Folded Reload
3326; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3068 ; 4-byte Folded Reload
3327; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3072 ; 4-byte Folded Reload
3328; GFX6-NEXT:    s_waitcnt vmcnt(0)
3329; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3056
3330; GFX6-NEXT:    s_waitcnt expcnt(0)
3331; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3044 ; 4-byte Folded Reload
3332; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3048 ; 4-byte Folded Reload
3333; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3052 ; 4-byte Folded Reload
3334; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3056 ; 4-byte Folded Reload
3335; GFX6-NEXT:    s_waitcnt vmcnt(0)
3336; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3040
3337; GFX6-NEXT:    s_waitcnt expcnt(0)
3338; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3028 ; 4-byte Folded Reload
3339; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3032 ; 4-byte Folded Reload
3340; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3036 ; 4-byte Folded Reload
3341; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3040 ; 4-byte Folded Reload
3342; GFX6-NEXT:    s_waitcnt vmcnt(0)
3343; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3024
3344; GFX6-NEXT:    s_waitcnt expcnt(0)
3345; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:3012 ; 4-byte Folded Reload
3346; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3016 ; 4-byte Folded Reload
3347; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3020 ; 4-byte Folded Reload
3348; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3024 ; 4-byte Folded Reload
3349; GFX6-NEXT:    s_waitcnt vmcnt(0)
3350; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:3008
3351; GFX6-NEXT:    s_waitcnt expcnt(0)
3352; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2996 ; 4-byte Folded Reload
3353; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:3000 ; 4-byte Folded Reload
3354; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:3004 ; 4-byte Folded Reload
3355; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:3008 ; 4-byte Folded Reload
3356; GFX6-NEXT:    s_waitcnt vmcnt(0)
3357; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2992
3358; GFX6-NEXT:    s_waitcnt expcnt(0)
3359; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2980 ; 4-byte Folded Reload
3360; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2984 ; 4-byte Folded Reload
3361; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2988 ; 4-byte Folded Reload
3362; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2992 ; 4-byte Folded Reload
3363; GFX6-NEXT:    s_waitcnt vmcnt(0)
3364; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2976
3365; GFX6-NEXT:    s_waitcnt expcnt(0)
3366; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2964 ; 4-byte Folded Reload
3367; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2968 ; 4-byte Folded Reload
3368; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2972 ; 4-byte Folded Reload
3369; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2976 ; 4-byte Folded Reload
3370; GFX6-NEXT:    s_waitcnt vmcnt(0)
3371; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2960
3372; GFX6-NEXT:    s_waitcnt expcnt(0)
3373; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2948 ; 4-byte Folded Reload
3374; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2952 ; 4-byte Folded Reload
3375; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2956 ; 4-byte Folded Reload
3376; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2960 ; 4-byte Folded Reload
3377; GFX6-NEXT:    s_waitcnt vmcnt(0)
3378; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2944
3379; GFX6-NEXT:    s_waitcnt expcnt(0)
3380; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2932 ; 4-byte Folded Reload
3381; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2936 ; 4-byte Folded Reload
3382; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2940 ; 4-byte Folded Reload
3383; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2944 ; 4-byte Folded Reload
3384; GFX6-NEXT:    s_waitcnt vmcnt(0)
3385; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2928
3386; GFX6-NEXT:    s_waitcnt expcnt(0)
3387; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2916 ; 4-byte Folded Reload
3388; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2920 ; 4-byte Folded Reload
3389; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2924 ; 4-byte Folded Reload
3390; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2928 ; 4-byte Folded Reload
3391; GFX6-NEXT:    s_waitcnt vmcnt(0)
3392; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2912
3393; GFX6-NEXT:    s_waitcnt expcnt(0)
3394; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2900 ; 4-byte Folded Reload
3395; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2904 ; 4-byte Folded Reload
3396; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2908 ; 4-byte Folded Reload
3397; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2912 ; 4-byte Folded Reload
3398; GFX6-NEXT:    s_waitcnt vmcnt(0)
3399; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2896
3400; GFX6-NEXT:    s_waitcnt expcnt(0)
3401; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2884 ; 4-byte Folded Reload
3402; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2888 ; 4-byte Folded Reload
3403; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2892 ; 4-byte Folded Reload
3404; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2896 ; 4-byte Folded Reload
3405; GFX6-NEXT:    s_waitcnt vmcnt(0)
3406; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2880
3407; GFX6-NEXT:    s_waitcnt expcnt(0)
3408; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2868 ; 4-byte Folded Reload
3409; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2872 ; 4-byte Folded Reload
3410; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2876 ; 4-byte Folded Reload
3411; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2880 ; 4-byte Folded Reload
3412; GFX6-NEXT:    s_waitcnt vmcnt(0)
3413; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2864
3414; GFX6-NEXT:    s_waitcnt expcnt(0)
3415; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2852 ; 4-byte Folded Reload
3416; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2856 ; 4-byte Folded Reload
3417; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2860 ; 4-byte Folded Reload
3418; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2864 ; 4-byte Folded Reload
3419; GFX6-NEXT:    s_waitcnt vmcnt(0)
3420; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2848
3421; GFX6-NEXT:    s_waitcnt expcnt(0)
3422; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2836 ; 4-byte Folded Reload
3423; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2840 ; 4-byte Folded Reload
3424; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2844 ; 4-byte Folded Reload
3425; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2848 ; 4-byte Folded Reload
3426; GFX6-NEXT:    s_waitcnt vmcnt(0)
3427; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2832
3428; GFX6-NEXT:    s_waitcnt expcnt(0)
3429; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2820 ; 4-byte Folded Reload
3430; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2824 ; 4-byte Folded Reload
3431; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2828 ; 4-byte Folded Reload
3432; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2832 ; 4-byte Folded Reload
3433; GFX6-NEXT:    s_waitcnt vmcnt(0)
3434; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2816
3435; GFX6-NEXT:    s_waitcnt expcnt(0)
3436; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2804 ; 4-byte Folded Reload
3437; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2808 ; 4-byte Folded Reload
3438; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2812 ; 4-byte Folded Reload
3439; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2816 ; 4-byte Folded Reload
3440; GFX6-NEXT:    s_waitcnt vmcnt(0)
3441; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2800
3442; GFX6-NEXT:    s_waitcnt expcnt(0)
3443; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2788 ; 4-byte Folded Reload
3444; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2792 ; 4-byte Folded Reload
3445; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2796 ; 4-byte Folded Reload
3446; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2800 ; 4-byte Folded Reload
3447; GFX6-NEXT:    s_waitcnt vmcnt(0)
3448; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2784
3449; GFX6-NEXT:    s_waitcnt expcnt(0)
3450; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2772 ; 4-byte Folded Reload
3451; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2776 ; 4-byte Folded Reload
3452; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2780 ; 4-byte Folded Reload
3453; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2784 ; 4-byte Folded Reload
3454; GFX6-NEXT:    s_waitcnt vmcnt(0)
3455; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2768
3456; GFX6-NEXT:    s_waitcnt expcnt(0)
3457; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2756 ; 4-byte Folded Reload
3458; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2760 ; 4-byte Folded Reload
3459; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2764 ; 4-byte Folded Reload
3460; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2768 ; 4-byte Folded Reload
3461; GFX6-NEXT:    s_waitcnt vmcnt(0)
3462; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2752
3463; GFX6-NEXT:    s_waitcnt expcnt(0)
3464; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2740 ; 4-byte Folded Reload
3465; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2744 ; 4-byte Folded Reload
3466; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2748 ; 4-byte Folded Reload
3467; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2752 ; 4-byte Folded Reload
3468; GFX6-NEXT:    s_waitcnt vmcnt(0)
3469; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2736
3470; GFX6-NEXT:    s_waitcnt expcnt(0)
3471; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2724 ; 4-byte Folded Reload
3472; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2728 ; 4-byte Folded Reload
3473; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2732 ; 4-byte Folded Reload
3474; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2736 ; 4-byte Folded Reload
3475; GFX6-NEXT:    s_waitcnt vmcnt(0)
3476; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2720
3477; GFX6-NEXT:    s_waitcnt expcnt(0)
3478; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2708 ; 4-byte Folded Reload
3479; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2712 ; 4-byte Folded Reload
3480; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2716 ; 4-byte Folded Reload
3481; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2720 ; 4-byte Folded Reload
3482; GFX6-NEXT:    s_waitcnt vmcnt(0)
3483; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2704
3484; GFX6-NEXT:    s_waitcnt expcnt(0)
3485; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2692 ; 4-byte Folded Reload
3486; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2696 ; 4-byte Folded Reload
3487; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2700 ; 4-byte Folded Reload
3488; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2704 ; 4-byte Folded Reload
3489; GFX6-NEXT:    s_waitcnt vmcnt(0)
3490; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2688
3491; GFX6-NEXT:    s_waitcnt expcnt(0)
3492; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2676 ; 4-byte Folded Reload
3493; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2680 ; 4-byte Folded Reload
3494; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2684 ; 4-byte Folded Reload
3495; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2688 ; 4-byte Folded Reload
3496; GFX6-NEXT:    s_waitcnt vmcnt(0)
3497; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2672
3498; GFX6-NEXT:    s_waitcnt expcnt(0)
3499; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2660 ; 4-byte Folded Reload
3500; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2664 ; 4-byte Folded Reload
3501; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2668 ; 4-byte Folded Reload
3502; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2672 ; 4-byte Folded Reload
3503; GFX6-NEXT:    s_waitcnt vmcnt(0)
3504; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2656
3505; GFX6-NEXT:    s_waitcnt expcnt(0)
3506; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2644 ; 4-byte Folded Reload
3507; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2648 ; 4-byte Folded Reload
3508; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2652 ; 4-byte Folded Reload
3509; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2656 ; 4-byte Folded Reload
3510; GFX6-NEXT:    s_waitcnt vmcnt(0)
3511; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2640
3512; GFX6-NEXT:    s_waitcnt expcnt(0)
3513; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2628 ; 4-byte Folded Reload
3514; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2632 ; 4-byte Folded Reload
3515; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2636 ; 4-byte Folded Reload
3516; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2640 ; 4-byte Folded Reload
3517; GFX6-NEXT:    s_waitcnt vmcnt(0)
3518; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2624
3519; GFX6-NEXT:    s_waitcnt expcnt(0)
3520; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2612 ; 4-byte Folded Reload
3521; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2616 ; 4-byte Folded Reload
3522; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2620 ; 4-byte Folded Reload
3523; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2624 ; 4-byte Folded Reload
3524; GFX6-NEXT:    s_waitcnt vmcnt(0)
3525; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2608
3526; GFX6-NEXT:    s_waitcnt expcnt(0)
3527; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2596 ; 4-byte Folded Reload
3528; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2600 ; 4-byte Folded Reload
3529; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2604 ; 4-byte Folded Reload
3530; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2608 ; 4-byte Folded Reload
3531; GFX6-NEXT:    s_waitcnt vmcnt(0)
3532; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2592
3533; GFX6-NEXT:    s_waitcnt expcnt(0)
3534; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2580 ; 4-byte Folded Reload
3535; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2584 ; 4-byte Folded Reload
3536; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2588 ; 4-byte Folded Reload
3537; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2592 ; 4-byte Folded Reload
3538; GFX6-NEXT:    s_waitcnt vmcnt(0)
3539; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2576
3540; GFX6-NEXT:    s_waitcnt expcnt(0)
3541; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2564 ; 4-byte Folded Reload
3542; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2568 ; 4-byte Folded Reload
3543; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2572 ; 4-byte Folded Reload
3544; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2576 ; 4-byte Folded Reload
3545; GFX6-NEXT:    s_waitcnt vmcnt(0)
3546; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2560
3547; GFX6-NEXT:    s_waitcnt expcnt(0)
3548; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2548 ; 4-byte Folded Reload
3549; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2552 ; 4-byte Folded Reload
3550; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2556 ; 4-byte Folded Reload
3551; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2560 ; 4-byte Folded Reload
3552; GFX6-NEXT:    s_waitcnt vmcnt(0)
3553; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2544
3554; GFX6-NEXT:    s_waitcnt expcnt(0)
3555; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2532 ; 4-byte Folded Reload
3556; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2536 ; 4-byte Folded Reload
3557; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2540 ; 4-byte Folded Reload
3558; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2544 ; 4-byte Folded Reload
3559; GFX6-NEXT:    s_waitcnt vmcnt(0)
3560; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2528
3561; GFX6-NEXT:    s_waitcnt expcnt(0)
3562; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2516 ; 4-byte Folded Reload
3563; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2520 ; 4-byte Folded Reload
3564; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2524 ; 4-byte Folded Reload
3565; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2528 ; 4-byte Folded Reload
3566; GFX6-NEXT:    s_waitcnt vmcnt(0)
3567; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2512
3568; GFX6-NEXT:    s_waitcnt expcnt(0)
3569; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2500 ; 4-byte Folded Reload
3570; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2504 ; 4-byte Folded Reload
3571; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2508 ; 4-byte Folded Reload
3572; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2512 ; 4-byte Folded Reload
3573; GFX6-NEXT:    s_waitcnt vmcnt(0)
3574; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2496
3575; GFX6-NEXT:    s_waitcnt expcnt(0)
3576; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2484 ; 4-byte Folded Reload
3577; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2488 ; 4-byte Folded Reload
3578; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2492 ; 4-byte Folded Reload
3579; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2496 ; 4-byte Folded Reload
3580; GFX6-NEXT:    s_waitcnt vmcnt(0)
3581; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2480
3582; GFX6-NEXT:    s_waitcnt expcnt(0)
3583; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2468 ; 4-byte Folded Reload
3584; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2472 ; 4-byte Folded Reload
3585; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2476 ; 4-byte Folded Reload
3586; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2480 ; 4-byte Folded Reload
3587; GFX6-NEXT:    s_waitcnt vmcnt(0)
3588; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2464
3589; GFX6-NEXT:    s_waitcnt expcnt(0)
3590; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2452 ; 4-byte Folded Reload
3591; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2456 ; 4-byte Folded Reload
3592; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2460 ; 4-byte Folded Reload
3593; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2464 ; 4-byte Folded Reload
3594; GFX6-NEXT:    s_waitcnt vmcnt(0)
3595; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2448
3596; GFX6-NEXT:    s_waitcnt expcnt(0)
3597; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2436 ; 4-byte Folded Reload
3598; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2440 ; 4-byte Folded Reload
3599; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2444 ; 4-byte Folded Reload
3600; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2448 ; 4-byte Folded Reload
3601; GFX6-NEXT:    s_waitcnt vmcnt(0)
3602; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2432
3603; GFX6-NEXT:    s_waitcnt expcnt(0)
3604; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2420 ; 4-byte Folded Reload
3605; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2424 ; 4-byte Folded Reload
3606; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2428 ; 4-byte Folded Reload
3607; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2432 ; 4-byte Folded Reload
3608; GFX6-NEXT:    s_waitcnt vmcnt(0)
3609; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2416
3610; GFX6-NEXT:    s_waitcnt expcnt(0)
3611; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2404 ; 4-byte Folded Reload
3612; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2408 ; 4-byte Folded Reload
3613; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2412 ; 4-byte Folded Reload
3614; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2416 ; 4-byte Folded Reload
3615; GFX6-NEXT:    s_waitcnt vmcnt(0)
3616; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2400
3617; GFX6-NEXT:    s_waitcnt expcnt(0)
3618; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2388 ; 4-byte Folded Reload
3619; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2392 ; 4-byte Folded Reload
3620; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2396 ; 4-byte Folded Reload
3621; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2400 ; 4-byte Folded Reload
3622; GFX6-NEXT:    s_waitcnt vmcnt(0)
3623; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2384
3624; GFX6-NEXT:    s_waitcnt expcnt(0)
3625; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2372 ; 4-byte Folded Reload
3626; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2376 ; 4-byte Folded Reload
3627; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2380 ; 4-byte Folded Reload
3628; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2384 ; 4-byte Folded Reload
3629; GFX6-NEXT:    s_waitcnt vmcnt(0)
3630; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2368
3631; GFX6-NEXT:    s_waitcnt expcnt(0)
3632; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2356 ; 4-byte Folded Reload
3633; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2360 ; 4-byte Folded Reload
3634; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2364 ; 4-byte Folded Reload
3635; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2368 ; 4-byte Folded Reload
3636; GFX6-NEXT:    s_waitcnt vmcnt(0)
3637; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2352
3638; GFX6-NEXT:    s_waitcnt expcnt(0)
3639; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2340 ; 4-byte Folded Reload
3640; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2344 ; 4-byte Folded Reload
3641; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2348 ; 4-byte Folded Reload
3642; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2352 ; 4-byte Folded Reload
3643; GFX6-NEXT:    s_waitcnt vmcnt(0)
3644; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2336
3645; GFX6-NEXT:    s_waitcnt expcnt(0)
3646; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2324 ; 4-byte Folded Reload
3647; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2328 ; 4-byte Folded Reload
3648; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2332 ; 4-byte Folded Reload
3649; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2336 ; 4-byte Folded Reload
3650; GFX6-NEXT:    s_waitcnt vmcnt(0)
3651; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2320
3652; GFX6-NEXT:    s_waitcnt expcnt(0)
3653; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2308 ; 4-byte Folded Reload
3654; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2312 ; 4-byte Folded Reload
3655; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2316 ; 4-byte Folded Reload
3656; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2320 ; 4-byte Folded Reload
3657; GFX6-NEXT:    s_waitcnt vmcnt(0)
3658; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2304
3659; GFX6-NEXT:    s_waitcnt expcnt(0)
3660; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2292 ; 4-byte Folded Reload
3661; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2296 ; 4-byte Folded Reload
3662; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2300 ; 4-byte Folded Reload
3663; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2304 ; 4-byte Folded Reload
3664; GFX6-NEXT:    s_waitcnt vmcnt(0)
3665; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2288
3666; GFX6-NEXT:    s_waitcnt expcnt(0)
3667; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2276 ; 4-byte Folded Reload
3668; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2280 ; 4-byte Folded Reload
3669; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2284 ; 4-byte Folded Reload
3670; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2288 ; 4-byte Folded Reload
3671; GFX6-NEXT:    s_waitcnt vmcnt(0)
3672; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2272
3673; GFX6-NEXT:    s_waitcnt expcnt(0)
3674; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2260 ; 4-byte Folded Reload
3675; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2264 ; 4-byte Folded Reload
3676; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2268 ; 4-byte Folded Reload
3677; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2272 ; 4-byte Folded Reload
3678; GFX6-NEXT:    s_waitcnt vmcnt(0)
3679; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2256
3680; GFX6-NEXT:    s_waitcnt expcnt(0)
3681; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2244 ; 4-byte Folded Reload
3682; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2248 ; 4-byte Folded Reload
3683; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2252 ; 4-byte Folded Reload
3684; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2256 ; 4-byte Folded Reload
3685; GFX6-NEXT:    s_waitcnt vmcnt(0)
3686; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2240
3687; GFX6-NEXT:    s_waitcnt expcnt(0)
3688; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2228 ; 4-byte Folded Reload
3689; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2232 ; 4-byte Folded Reload
3690; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2236 ; 4-byte Folded Reload
3691; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2240 ; 4-byte Folded Reload
3692; GFX6-NEXT:    s_waitcnt vmcnt(0)
3693; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2224
3694; GFX6-NEXT:    s_waitcnt expcnt(0)
3695; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2212 ; 4-byte Folded Reload
3696; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2216 ; 4-byte Folded Reload
3697; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2220 ; 4-byte Folded Reload
3698; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2224 ; 4-byte Folded Reload
3699; GFX6-NEXT:    s_waitcnt vmcnt(0)
3700; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2208
3701; GFX6-NEXT:    s_waitcnt expcnt(0)
3702; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2196 ; 4-byte Folded Reload
3703; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2200 ; 4-byte Folded Reload
3704; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2204 ; 4-byte Folded Reload
3705; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2208 ; 4-byte Folded Reload
3706; GFX6-NEXT:    s_waitcnt vmcnt(0)
3707; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2192
3708; GFX6-NEXT:    s_waitcnt expcnt(0)
3709; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2180 ; 4-byte Folded Reload
3710; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2184 ; 4-byte Folded Reload
3711; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2188 ; 4-byte Folded Reload
3712; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2192 ; 4-byte Folded Reload
3713; GFX6-NEXT:    s_waitcnt vmcnt(0)
3714; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2176
3715; GFX6-NEXT:    s_waitcnt expcnt(0)
3716; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2164 ; 4-byte Folded Reload
3717; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2168 ; 4-byte Folded Reload
3718; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2172 ; 4-byte Folded Reload
3719; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2176 ; 4-byte Folded Reload
3720; GFX6-NEXT:    s_waitcnt vmcnt(0)
3721; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2160
3722; GFX6-NEXT:    s_waitcnt expcnt(0)
3723; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2148 ; 4-byte Folded Reload
3724; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2152 ; 4-byte Folded Reload
3725; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2156 ; 4-byte Folded Reload
3726; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2160 ; 4-byte Folded Reload
3727; GFX6-NEXT:    s_waitcnt vmcnt(0)
3728; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2144
3729; GFX6-NEXT:    s_waitcnt expcnt(0)
3730; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2132 ; 4-byte Folded Reload
3731; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2136 ; 4-byte Folded Reload
3732; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2140 ; 4-byte Folded Reload
3733; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2144 ; 4-byte Folded Reload
3734; GFX6-NEXT:    s_waitcnt vmcnt(0)
3735; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2128
3736; GFX6-NEXT:    s_waitcnt expcnt(0)
3737; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2116 ; 4-byte Folded Reload
3738; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2120 ; 4-byte Folded Reload
3739; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2124 ; 4-byte Folded Reload
3740; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2128 ; 4-byte Folded Reload
3741; GFX6-NEXT:    s_waitcnt vmcnt(0)
3742; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2112
3743; GFX6-NEXT:    s_waitcnt expcnt(0)
3744; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2100 ; 4-byte Folded Reload
3745; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2104 ; 4-byte Folded Reload
3746; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2108 ; 4-byte Folded Reload
3747; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2112 ; 4-byte Folded Reload
3748; GFX6-NEXT:    s_waitcnt vmcnt(0)
3749; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2096
3750; GFX6-NEXT:    s_waitcnt expcnt(0)
3751; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2084 ; 4-byte Folded Reload
3752; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2088 ; 4-byte Folded Reload
3753; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2092 ; 4-byte Folded Reload
3754; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2096 ; 4-byte Folded Reload
3755; GFX6-NEXT:    s_waitcnt vmcnt(0)
3756; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2080
3757; GFX6-NEXT:    s_waitcnt expcnt(0)
3758; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2068 ; 4-byte Folded Reload
3759; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2072 ; 4-byte Folded Reload
3760; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2076 ; 4-byte Folded Reload
3761; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2080 ; 4-byte Folded Reload
3762; GFX6-NEXT:    s_waitcnt vmcnt(0)
3763; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2064
3764; GFX6-NEXT:    s_waitcnt expcnt(0)
3765; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2052 ; 4-byte Folded Reload
3766; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2056 ; 4-byte Folded Reload
3767; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2060 ; 4-byte Folded Reload
3768; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2064 ; 4-byte Folded Reload
3769; GFX6-NEXT:    s_waitcnt vmcnt(0)
3770; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2048
3771; GFX6-NEXT:    s_waitcnt expcnt(0)
3772; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2036 ; 4-byte Folded Reload
3773; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2040 ; 4-byte Folded Reload
3774; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2044 ; 4-byte Folded Reload
3775; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2048 ; 4-byte Folded Reload
3776; GFX6-NEXT:    s_waitcnt vmcnt(0)
3777; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2032
3778; GFX6-NEXT:    s_waitcnt expcnt(0)
3779; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2020 ; 4-byte Folded Reload
3780; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2024 ; 4-byte Folded Reload
3781; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2028 ; 4-byte Folded Reload
3782; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2032 ; 4-byte Folded Reload
3783; GFX6-NEXT:    s_waitcnt vmcnt(0)
3784; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2016
3785; GFX6-NEXT:    s_waitcnt expcnt(0)
3786; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:2004 ; 4-byte Folded Reload
3787; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:2008 ; 4-byte Folded Reload
3788; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:2012 ; 4-byte Folded Reload
3789; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2016 ; 4-byte Folded Reload
3790; GFX6-NEXT:    s_waitcnt vmcnt(0)
3791; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:2000
3792; GFX6-NEXT:    s_waitcnt expcnt(0)
3793; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1988 ; 4-byte Folded Reload
3794; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1992 ; 4-byte Folded Reload
3795; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1996 ; 4-byte Folded Reload
3796; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:2000 ; 4-byte Folded Reload
3797; GFX6-NEXT:    s_waitcnt vmcnt(0)
3798; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1984
3799; GFX6-NEXT:    s_waitcnt expcnt(0)
3800; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1972 ; 4-byte Folded Reload
3801; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1976 ; 4-byte Folded Reload
3802; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1980 ; 4-byte Folded Reload
3803; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1984 ; 4-byte Folded Reload
3804; GFX6-NEXT:    s_waitcnt vmcnt(0)
3805; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1968
3806; GFX6-NEXT:    s_waitcnt expcnt(0)
3807; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1956 ; 4-byte Folded Reload
3808; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1960 ; 4-byte Folded Reload
3809; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1964 ; 4-byte Folded Reload
3810; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1968 ; 4-byte Folded Reload
3811; GFX6-NEXT:    s_waitcnt vmcnt(0)
3812; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1952
3813; GFX6-NEXT:    s_waitcnt expcnt(0)
3814; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1940 ; 4-byte Folded Reload
3815; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1944 ; 4-byte Folded Reload
3816; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1948 ; 4-byte Folded Reload
3817; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1952 ; 4-byte Folded Reload
3818; GFX6-NEXT:    s_waitcnt vmcnt(0)
3819; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1936
3820; GFX6-NEXT:    s_waitcnt expcnt(0)
3821; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1924 ; 4-byte Folded Reload
3822; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1928 ; 4-byte Folded Reload
3823; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1932 ; 4-byte Folded Reload
3824; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1936 ; 4-byte Folded Reload
3825; GFX6-NEXT:    s_waitcnt vmcnt(0)
3826; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1920
3827; GFX6-NEXT:    s_waitcnt expcnt(0)
3828; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1908 ; 4-byte Folded Reload
3829; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1912 ; 4-byte Folded Reload
3830; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1916 ; 4-byte Folded Reload
3831; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1920 ; 4-byte Folded Reload
3832; GFX6-NEXT:    s_waitcnt vmcnt(0)
3833; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1904
3834; GFX6-NEXT:    s_waitcnt expcnt(0)
3835; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1892 ; 4-byte Folded Reload
3836; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1896 ; 4-byte Folded Reload
3837; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1900 ; 4-byte Folded Reload
3838; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1904 ; 4-byte Folded Reload
3839; GFX6-NEXT:    s_waitcnt vmcnt(0)
3840; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1888
3841; GFX6-NEXT:    s_waitcnt expcnt(0)
3842; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1876 ; 4-byte Folded Reload
3843; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1880 ; 4-byte Folded Reload
3844; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1884 ; 4-byte Folded Reload
3845; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1888 ; 4-byte Folded Reload
3846; GFX6-NEXT:    s_waitcnt vmcnt(0)
3847; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1872
3848; GFX6-NEXT:    s_waitcnt expcnt(0)
3849; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1860 ; 4-byte Folded Reload
3850; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1864 ; 4-byte Folded Reload
3851; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1868 ; 4-byte Folded Reload
3852; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1872 ; 4-byte Folded Reload
3853; GFX6-NEXT:    s_waitcnt vmcnt(0)
3854; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1856
3855; GFX6-NEXT:    s_waitcnt expcnt(0)
3856; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1844 ; 4-byte Folded Reload
3857; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1848 ; 4-byte Folded Reload
3858; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1852 ; 4-byte Folded Reload
3859; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1856 ; 4-byte Folded Reload
3860; GFX6-NEXT:    s_waitcnt vmcnt(0)
3861; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1840
3862; GFX6-NEXT:    s_waitcnt expcnt(0)
3863; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1828 ; 4-byte Folded Reload
3864; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1832 ; 4-byte Folded Reload
3865; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1836 ; 4-byte Folded Reload
3866; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1840 ; 4-byte Folded Reload
3867; GFX6-NEXT:    s_waitcnt vmcnt(0)
3868; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1824
3869; GFX6-NEXT:    s_waitcnt expcnt(0)
3870; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1812 ; 4-byte Folded Reload
3871; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1816 ; 4-byte Folded Reload
3872; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1820 ; 4-byte Folded Reload
3873; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1824 ; 4-byte Folded Reload
3874; GFX6-NEXT:    s_waitcnt vmcnt(0)
3875; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1808
3876; GFX6-NEXT:    s_waitcnt expcnt(0)
3877; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1796 ; 4-byte Folded Reload
3878; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1800 ; 4-byte Folded Reload
3879; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1804 ; 4-byte Folded Reload
3880; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1808 ; 4-byte Folded Reload
3881; GFX6-NEXT:    s_waitcnt vmcnt(0)
3882; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1792
3883; GFX6-NEXT:    s_waitcnt expcnt(0)
3884; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1780 ; 4-byte Folded Reload
3885; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1784 ; 4-byte Folded Reload
3886; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1788 ; 4-byte Folded Reload
3887; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1792 ; 4-byte Folded Reload
3888; GFX6-NEXT:    s_waitcnt vmcnt(0)
3889; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1776
3890; GFX6-NEXT:    s_waitcnt expcnt(0)
3891; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1764 ; 4-byte Folded Reload
3892; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1768 ; 4-byte Folded Reload
3893; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1772 ; 4-byte Folded Reload
3894; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1776 ; 4-byte Folded Reload
3895; GFX6-NEXT:    s_waitcnt vmcnt(0)
3896; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1760
3897; GFX6-NEXT:    s_waitcnt expcnt(0)
3898; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1748 ; 4-byte Folded Reload
3899; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1752 ; 4-byte Folded Reload
3900; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1756 ; 4-byte Folded Reload
3901; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1760 ; 4-byte Folded Reload
3902; GFX6-NEXT:    s_waitcnt vmcnt(0)
3903; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1744
3904; GFX6-NEXT:    s_waitcnt expcnt(0)
3905; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1732 ; 4-byte Folded Reload
3906; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1736 ; 4-byte Folded Reload
3907; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1740 ; 4-byte Folded Reload
3908; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1744 ; 4-byte Folded Reload
3909; GFX6-NEXT:    s_waitcnt vmcnt(0)
3910; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1728
3911; GFX6-NEXT:    s_waitcnt expcnt(0)
3912; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1716 ; 4-byte Folded Reload
3913; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1720 ; 4-byte Folded Reload
3914; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1724 ; 4-byte Folded Reload
3915; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1728 ; 4-byte Folded Reload
3916; GFX6-NEXT:    s_waitcnt vmcnt(0)
3917; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1712
3918; GFX6-NEXT:    s_waitcnt expcnt(0)
3919; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1700 ; 4-byte Folded Reload
3920; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1704 ; 4-byte Folded Reload
3921; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1708 ; 4-byte Folded Reload
3922; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1712 ; 4-byte Folded Reload
3923; GFX6-NEXT:    s_waitcnt vmcnt(0)
3924; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1696
3925; GFX6-NEXT:    s_waitcnt expcnt(0)
3926; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1684 ; 4-byte Folded Reload
3927; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1688 ; 4-byte Folded Reload
3928; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1692 ; 4-byte Folded Reload
3929; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1696 ; 4-byte Folded Reload
3930; GFX6-NEXT:    s_waitcnt vmcnt(0)
3931; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1680
3932; GFX6-NEXT:    s_waitcnt expcnt(0)
3933; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1668 ; 4-byte Folded Reload
3934; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1672 ; 4-byte Folded Reload
3935; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1676 ; 4-byte Folded Reload
3936; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1680 ; 4-byte Folded Reload
3937; GFX6-NEXT:    s_waitcnt vmcnt(0)
3938; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1664
3939; GFX6-NEXT:    s_waitcnt expcnt(0)
3940; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1652 ; 4-byte Folded Reload
3941; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1656 ; 4-byte Folded Reload
3942; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1660 ; 4-byte Folded Reload
3943; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1664 ; 4-byte Folded Reload
3944; GFX6-NEXT:    s_waitcnt vmcnt(0)
3945; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1648
3946; GFX6-NEXT:    s_waitcnt expcnt(0)
3947; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1636 ; 4-byte Folded Reload
3948; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1640 ; 4-byte Folded Reload
3949; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1644 ; 4-byte Folded Reload
3950; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1648 ; 4-byte Folded Reload
3951; GFX6-NEXT:    s_waitcnt vmcnt(0)
3952; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1632
3953; GFX6-NEXT:    s_waitcnt expcnt(0)
3954; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1620 ; 4-byte Folded Reload
3955; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1624 ; 4-byte Folded Reload
3956; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1628 ; 4-byte Folded Reload
3957; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1632 ; 4-byte Folded Reload
3958; GFX6-NEXT:    s_waitcnt vmcnt(0)
3959; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1616
3960; GFX6-NEXT:    s_waitcnt expcnt(0)
3961; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1604 ; 4-byte Folded Reload
3962; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1608 ; 4-byte Folded Reload
3963; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1612 ; 4-byte Folded Reload
3964; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1616 ; 4-byte Folded Reload
3965; GFX6-NEXT:    s_waitcnt vmcnt(0)
3966; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1600
3967; GFX6-NEXT:    s_waitcnt expcnt(0)
3968; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1588 ; 4-byte Folded Reload
3969; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1592 ; 4-byte Folded Reload
3970; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1596 ; 4-byte Folded Reload
3971; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1600 ; 4-byte Folded Reload
3972; GFX6-NEXT:    s_waitcnt vmcnt(0)
3973; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1584
3974; GFX6-NEXT:    s_waitcnt expcnt(0)
3975; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1572 ; 4-byte Folded Reload
3976; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1576 ; 4-byte Folded Reload
3977; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1580 ; 4-byte Folded Reload
3978; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1584 ; 4-byte Folded Reload
3979; GFX6-NEXT:    s_waitcnt vmcnt(0)
3980; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1568
3981; GFX6-NEXT:    s_waitcnt expcnt(0)
3982; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1556 ; 4-byte Folded Reload
3983; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1560 ; 4-byte Folded Reload
3984; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1564 ; 4-byte Folded Reload
3985; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1568 ; 4-byte Folded Reload
3986; GFX6-NEXT:    s_waitcnt vmcnt(0)
3987; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1552
3988; GFX6-NEXT:    s_waitcnt expcnt(0)
3989; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1540 ; 4-byte Folded Reload
3990; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1544 ; 4-byte Folded Reload
3991; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1548 ; 4-byte Folded Reload
3992; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1552 ; 4-byte Folded Reload
3993; GFX6-NEXT:    s_waitcnt vmcnt(0)
3994; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1536
3995; GFX6-NEXT:    s_waitcnt expcnt(0)
3996; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1524 ; 4-byte Folded Reload
3997; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1528 ; 4-byte Folded Reload
3998; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1532 ; 4-byte Folded Reload
3999; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1536 ; 4-byte Folded Reload
4000; GFX6-NEXT:    s_waitcnt vmcnt(0)
4001; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1520
4002; GFX6-NEXT:    s_waitcnt expcnt(0)
4003; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1508 ; 4-byte Folded Reload
4004; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1512 ; 4-byte Folded Reload
4005; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1516 ; 4-byte Folded Reload
4006; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1520 ; 4-byte Folded Reload
4007; GFX6-NEXT:    s_waitcnt vmcnt(0)
4008; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1504
4009; GFX6-NEXT:    s_waitcnt expcnt(0)
4010; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1492 ; 4-byte Folded Reload
4011; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1496 ; 4-byte Folded Reload
4012; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1500 ; 4-byte Folded Reload
4013; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1504 ; 4-byte Folded Reload
4014; GFX6-NEXT:    s_waitcnt vmcnt(0)
4015; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1488
4016; GFX6-NEXT:    s_waitcnt expcnt(0)
4017; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1476 ; 4-byte Folded Reload
4018; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1480 ; 4-byte Folded Reload
4019; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1484 ; 4-byte Folded Reload
4020; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1488 ; 4-byte Folded Reload
4021; GFX6-NEXT:    s_waitcnt vmcnt(0)
4022; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1472
4023; GFX6-NEXT:    s_waitcnt expcnt(0)
4024; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1460 ; 4-byte Folded Reload
4025; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1464 ; 4-byte Folded Reload
4026; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1468 ; 4-byte Folded Reload
4027; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1472 ; 4-byte Folded Reload
4028; GFX6-NEXT:    s_waitcnt vmcnt(0)
4029; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1456
4030; GFX6-NEXT:    s_waitcnt expcnt(0)
4031; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1444 ; 4-byte Folded Reload
4032; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1448 ; 4-byte Folded Reload
4033; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1452 ; 4-byte Folded Reload
4034; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1456 ; 4-byte Folded Reload
4035; GFX6-NEXT:    s_waitcnt vmcnt(0)
4036; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1440
4037; GFX6-NEXT:    s_waitcnt expcnt(0)
4038; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1428 ; 4-byte Folded Reload
4039; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1432 ; 4-byte Folded Reload
4040; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1436 ; 4-byte Folded Reload
4041; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1440 ; 4-byte Folded Reload
4042; GFX6-NEXT:    s_waitcnt vmcnt(0)
4043; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1424
4044; GFX6-NEXT:    s_waitcnt expcnt(0)
4045; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1412 ; 4-byte Folded Reload
4046; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1416 ; 4-byte Folded Reload
4047; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1420 ; 4-byte Folded Reload
4048; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1424 ; 4-byte Folded Reload
4049; GFX6-NEXT:    s_waitcnt vmcnt(0)
4050; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1408
4051; GFX6-NEXT:    s_waitcnt expcnt(0)
4052; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1396 ; 4-byte Folded Reload
4053; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1400 ; 4-byte Folded Reload
4054; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1404 ; 4-byte Folded Reload
4055; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1408 ; 4-byte Folded Reload
4056; GFX6-NEXT:    s_waitcnt vmcnt(0)
4057; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1392
4058; GFX6-NEXT:    s_waitcnt expcnt(0)
4059; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1380 ; 4-byte Folded Reload
4060; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1384 ; 4-byte Folded Reload
4061; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1388 ; 4-byte Folded Reload
4062; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1392 ; 4-byte Folded Reload
4063; GFX6-NEXT:    s_waitcnt vmcnt(0)
4064; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1376
4065; GFX6-NEXT:    s_waitcnt expcnt(0)
4066; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1364 ; 4-byte Folded Reload
4067; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1368 ; 4-byte Folded Reload
4068; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1372 ; 4-byte Folded Reload
4069; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1376 ; 4-byte Folded Reload
4070; GFX6-NEXT:    s_waitcnt vmcnt(0)
4071; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1360
4072; GFX6-NEXT:    s_waitcnt expcnt(0)
4073; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1348 ; 4-byte Folded Reload
4074; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1352 ; 4-byte Folded Reload
4075; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1356 ; 4-byte Folded Reload
4076; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1360 ; 4-byte Folded Reload
4077; GFX6-NEXT:    s_waitcnt vmcnt(0)
4078; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1344
4079; GFX6-NEXT:    s_waitcnt expcnt(0)
4080; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1332 ; 4-byte Folded Reload
4081; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1336 ; 4-byte Folded Reload
4082; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1340 ; 4-byte Folded Reload
4083; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1344 ; 4-byte Folded Reload
4084; GFX6-NEXT:    s_waitcnt vmcnt(0)
4085; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1328
4086; GFX6-NEXT:    s_waitcnt expcnt(0)
4087; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1316 ; 4-byte Folded Reload
4088; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1320 ; 4-byte Folded Reload
4089; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1324 ; 4-byte Folded Reload
4090; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1328 ; 4-byte Folded Reload
4091; GFX6-NEXT:    s_waitcnt vmcnt(0)
4092; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1312
4093; GFX6-NEXT:    s_waitcnt expcnt(0)
4094; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1300 ; 4-byte Folded Reload
4095; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1304 ; 4-byte Folded Reload
4096; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1308 ; 4-byte Folded Reload
4097; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1312 ; 4-byte Folded Reload
4098; GFX6-NEXT:    s_waitcnt vmcnt(0)
4099; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1296
4100; GFX6-NEXT:    s_waitcnt expcnt(0)
4101; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1284 ; 4-byte Folded Reload
4102; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1288 ; 4-byte Folded Reload
4103; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1292 ; 4-byte Folded Reload
4104; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1296 ; 4-byte Folded Reload
4105; GFX6-NEXT:    s_waitcnt vmcnt(0)
4106; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1280
4107; GFX6-NEXT:    s_waitcnt expcnt(0)
4108; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1268 ; 4-byte Folded Reload
4109; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1272 ; 4-byte Folded Reload
4110; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1276 ; 4-byte Folded Reload
4111; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1280 ; 4-byte Folded Reload
4112; GFX6-NEXT:    s_waitcnt vmcnt(0)
4113; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1264
4114; GFX6-NEXT:    s_waitcnt expcnt(0)
4115; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1252 ; 4-byte Folded Reload
4116; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1256 ; 4-byte Folded Reload
4117; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1260 ; 4-byte Folded Reload
4118; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1264 ; 4-byte Folded Reload
4119; GFX6-NEXT:    s_waitcnt vmcnt(0)
4120; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1248
4121; GFX6-NEXT:    s_waitcnt expcnt(0)
4122; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1236 ; 4-byte Folded Reload
4123; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1240 ; 4-byte Folded Reload
4124; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1244 ; 4-byte Folded Reload
4125; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1248 ; 4-byte Folded Reload
4126; GFX6-NEXT:    s_waitcnt vmcnt(0)
4127; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1232
4128; GFX6-NEXT:    s_waitcnt expcnt(0)
4129; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1220 ; 4-byte Folded Reload
4130; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1224 ; 4-byte Folded Reload
4131; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1228 ; 4-byte Folded Reload
4132; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1232 ; 4-byte Folded Reload
4133; GFX6-NEXT:    s_waitcnt vmcnt(0)
4134; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1216
4135; GFX6-NEXT:    s_waitcnt expcnt(0)
4136; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1204 ; 4-byte Folded Reload
4137; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1208 ; 4-byte Folded Reload
4138; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1212 ; 4-byte Folded Reload
4139; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1216 ; 4-byte Folded Reload
4140; GFX6-NEXT:    s_waitcnt vmcnt(0)
4141; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1200
4142; GFX6-NEXT:    s_waitcnt expcnt(0)
4143; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1188 ; 4-byte Folded Reload
4144; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1192 ; 4-byte Folded Reload
4145; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1196 ; 4-byte Folded Reload
4146; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1200 ; 4-byte Folded Reload
4147; GFX6-NEXT:    s_waitcnt vmcnt(0)
4148; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1184
4149; GFX6-NEXT:    s_waitcnt expcnt(0)
4150; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1172 ; 4-byte Folded Reload
4151; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1176 ; 4-byte Folded Reload
4152; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1180 ; 4-byte Folded Reload
4153; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1184 ; 4-byte Folded Reload
4154; GFX6-NEXT:    s_waitcnt vmcnt(0)
4155; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1168
4156; GFX6-NEXT:    s_waitcnt expcnt(0)
4157; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1156 ; 4-byte Folded Reload
4158; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1160 ; 4-byte Folded Reload
4159; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1164 ; 4-byte Folded Reload
4160; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1168 ; 4-byte Folded Reload
4161; GFX6-NEXT:    s_waitcnt vmcnt(0)
4162; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1152
4163; GFX6-NEXT:    s_waitcnt expcnt(0)
4164; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1140 ; 4-byte Folded Reload
4165; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1144 ; 4-byte Folded Reload
4166; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1148 ; 4-byte Folded Reload
4167; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1152 ; 4-byte Folded Reload
4168; GFX6-NEXT:    s_waitcnt vmcnt(0)
4169; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1136
4170; GFX6-NEXT:    s_waitcnt expcnt(0)
4171; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1124 ; 4-byte Folded Reload
4172; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1128 ; 4-byte Folded Reload
4173; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1132 ; 4-byte Folded Reload
4174; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1136 ; 4-byte Folded Reload
4175; GFX6-NEXT:    s_waitcnt vmcnt(0)
4176; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1120
4177; GFX6-NEXT:    s_waitcnt expcnt(0)
4178; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1108 ; 4-byte Folded Reload
4179; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1112 ; 4-byte Folded Reload
4180; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1116 ; 4-byte Folded Reload
4181; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1120 ; 4-byte Folded Reload
4182; GFX6-NEXT:    s_waitcnt vmcnt(0)
4183; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1104
4184; GFX6-NEXT:    s_waitcnt expcnt(0)
4185; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1092 ; 4-byte Folded Reload
4186; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1096 ; 4-byte Folded Reload
4187; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1100 ; 4-byte Folded Reload
4188; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1104 ; 4-byte Folded Reload
4189; GFX6-NEXT:    s_waitcnt vmcnt(0)
4190; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1088
4191; GFX6-NEXT:    s_waitcnt expcnt(0)
4192; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1076 ; 4-byte Folded Reload
4193; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1080 ; 4-byte Folded Reload
4194; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1084 ; 4-byte Folded Reload
4195; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1088 ; 4-byte Folded Reload
4196; GFX6-NEXT:    s_waitcnt vmcnt(0)
4197; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1072
4198; GFX6-NEXT:    s_waitcnt expcnt(0)
4199; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1060 ; 4-byte Folded Reload
4200; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1064 ; 4-byte Folded Reload
4201; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1068 ; 4-byte Folded Reload
4202; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1072 ; 4-byte Folded Reload
4203; GFX6-NEXT:    s_waitcnt vmcnt(0)
4204; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1056
4205; GFX6-NEXT:    s_waitcnt expcnt(0)
4206; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1044 ; 4-byte Folded Reload
4207; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1048 ; 4-byte Folded Reload
4208; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1052 ; 4-byte Folded Reload
4209; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1056 ; 4-byte Folded Reload
4210; GFX6-NEXT:    s_waitcnt vmcnt(0)
4211; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1040
4212; GFX6-NEXT:    s_waitcnt expcnt(0)
4213; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1028 ; 4-byte Folded Reload
4214; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1032 ; 4-byte Folded Reload
4215; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1036 ; 4-byte Folded Reload
4216; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1040 ; 4-byte Folded Reload
4217; GFX6-NEXT:    s_waitcnt vmcnt(0)
4218; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1024
4219; GFX6-NEXT:    s_waitcnt expcnt(0)
4220; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:1012 ; 4-byte Folded Reload
4221; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1016 ; 4-byte Folded Reload
4222; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1020 ; 4-byte Folded Reload
4223; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1024 ; 4-byte Folded Reload
4224; GFX6-NEXT:    s_waitcnt vmcnt(0)
4225; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:1008
4226; GFX6-NEXT:    s_waitcnt expcnt(0)
4227; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:996 ; 4-byte Folded Reload
4228; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:1000 ; 4-byte Folded Reload
4229; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:1004 ; 4-byte Folded Reload
4230; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:1008 ; 4-byte Folded Reload
4231; GFX6-NEXT:    s_waitcnt vmcnt(0)
4232; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:992
4233; GFX6-NEXT:    s_waitcnt expcnt(0)
4234; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:980 ; 4-byte Folded Reload
4235; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:984 ; 4-byte Folded Reload
4236; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:988 ; 4-byte Folded Reload
4237; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:992 ; 4-byte Folded Reload
4238; GFX6-NEXT:    s_waitcnt vmcnt(0)
4239; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:976
4240; GFX6-NEXT:    s_waitcnt expcnt(0)
4241; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:964 ; 4-byte Folded Reload
4242; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:968 ; 4-byte Folded Reload
4243; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:972 ; 4-byte Folded Reload
4244; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:976 ; 4-byte Folded Reload
4245; GFX6-NEXT:    s_waitcnt vmcnt(0)
4246; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:960
4247; GFX6-NEXT:    s_waitcnt expcnt(0)
4248; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:948 ; 4-byte Folded Reload
4249; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:952 ; 4-byte Folded Reload
4250; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:956 ; 4-byte Folded Reload
4251; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:960 ; 4-byte Folded Reload
4252; GFX6-NEXT:    s_waitcnt vmcnt(0)
4253; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:944
4254; GFX6-NEXT:    s_waitcnt expcnt(0)
4255; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:932 ; 4-byte Folded Reload
4256; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:936 ; 4-byte Folded Reload
4257; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:940 ; 4-byte Folded Reload
4258; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:944 ; 4-byte Folded Reload
4259; GFX6-NEXT:    s_waitcnt vmcnt(0)
4260; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:928
4261; GFX6-NEXT:    s_waitcnt expcnt(0)
4262; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:916 ; 4-byte Folded Reload
4263; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:920 ; 4-byte Folded Reload
4264; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:924 ; 4-byte Folded Reload
4265; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:928 ; 4-byte Folded Reload
4266; GFX6-NEXT:    s_waitcnt vmcnt(0)
4267; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:912
4268; GFX6-NEXT:    s_waitcnt expcnt(0)
4269; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:900 ; 4-byte Folded Reload
4270; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:904 ; 4-byte Folded Reload
4271; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:908 ; 4-byte Folded Reload
4272; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:912 ; 4-byte Folded Reload
4273; GFX6-NEXT:    s_waitcnt vmcnt(0)
4274; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:896
4275; GFX6-NEXT:    s_waitcnt expcnt(0)
4276; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:884 ; 4-byte Folded Reload
4277; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:888 ; 4-byte Folded Reload
4278; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:892 ; 4-byte Folded Reload
4279; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:896 ; 4-byte Folded Reload
4280; GFX6-NEXT:    s_waitcnt vmcnt(0)
4281; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:880
4282; GFX6-NEXT:    s_waitcnt expcnt(0)
4283; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:868 ; 4-byte Folded Reload
4284; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:872 ; 4-byte Folded Reload
4285; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:876 ; 4-byte Folded Reload
4286; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:880 ; 4-byte Folded Reload
4287; GFX6-NEXT:    s_waitcnt vmcnt(0)
4288; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:864
4289; GFX6-NEXT:    s_waitcnt expcnt(0)
4290; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:852 ; 4-byte Folded Reload
4291; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:856 ; 4-byte Folded Reload
4292; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:860 ; 4-byte Folded Reload
4293; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:864 ; 4-byte Folded Reload
4294; GFX6-NEXT:    s_waitcnt vmcnt(0)
4295; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:848
4296; GFX6-NEXT:    s_waitcnt expcnt(0)
4297; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:836 ; 4-byte Folded Reload
4298; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:840 ; 4-byte Folded Reload
4299; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:844 ; 4-byte Folded Reload
4300; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:848 ; 4-byte Folded Reload
4301; GFX6-NEXT:    s_waitcnt vmcnt(0)
4302; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:832
4303; GFX6-NEXT:    s_waitcnt expcnt(0)
4304; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:820 ; 4-byte Folded Reload
4305; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:824 ; 4-byte Folded Reload
4306; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:828 ; 4-byte Folded Reload
4307; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:832 ; 4-byte Folded Reload
4308; GFX6-NEXT:    s_waitcnt vmcnt(0)
4309; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:816
4310; GFX6-NEXT:    s_waitcnt expcnt(0)
4311; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:804 ; 4-byte Folded Reload
4312; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:808 ; 4-byte Folded Reload
4313; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:812 ; 4-byte Folded Reload
4314; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:816 ; 4-byte Folded Reload
4315; GFX6-NEXT:    s_waitcnt vmcnt(0)
4316; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:800
4317; GFX6-NEXT:    s_waitcnt expcnt(0)
4318; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:788 ; 4-byte Folded Reload
4319; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:792 ; 4-byte Folded Reload
4320; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:796 ; 4-byte Folded Reload
4321; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:800 ; 4-byte Folded Reload
4322; GFX6-NEXT:    s_waitcnt vmcnt(0)
4323; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:784
4324; GFX6-NEXT:    s_waitcnt expcnt(0)
4325; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:772 ; 4-byte Folded Reload
4326; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:776 ; 4-byte Folded Reload
4327; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:780 ; 4-byte Folded Reload
4328; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:784 ; 4-byte Folded Reload
4329; GFX6-NEXT:    s_waitcnt vmcnt(0)
4330; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:768
4331; GFX6-NEXT:    s_waitcnt expcnt(0)
4332; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:756 ; 4-byte Folded Reload
4333; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:760 ; 4-byte Folded Reload
4334; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:764 ; 4-byte Folded Reload
4335; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:768 ; 4-byte Folded Reload
4336; GFX6-NEXT:    s_waitcnt vmcnt(0)
4337; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:752
4338; GFX6-NEXT:    s_waitcnt expcnt(0)
4339; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:740 ; 4-byte Folded Reload
4340; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:744 ; 4-byte Folded Reload
4341; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:748 ; 4-byte Folded Reload
4342; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:752 ; 4-byte Folded Reload
4343; GFX6-NEXT:    s_waitcnt vmcnt(0)
4344; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:736
4345; GFX6-NEXT:    s_waitcnt expcnt(0)
4346; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:724 ; 4-byte Folded Reload
4347; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:728 ; 4-byte Folded Reload
4348; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:732 ; 4-byte Folded Reload
4349; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:736 ; 4-byte Folded Reload
4350; GFX6-NEXT:    s_waitcnt vmcnt(0)
4351; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:720
4352; GFX6-NEXT:    s_waitcnt expcnt(0)
4353; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:708 ; 4-byte Folded Reload
4354; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:712 ; 4-byte Folded Reload
4355; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:716 ; 4-byte Folded Reload
4356; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:720 ; 4-byte Folded Reload
4357; GFX6-NEXT:    s_waitcnt vmcnt(0)
4358; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:704
4359; GFX6-NEXT:    s_waitcnt expcnt(0)
4360; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:692 ; 4-byte Folded Reload
4361; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:696 ; 4-byte Folded Reload
4362; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:700 ; 4-byte Folded Reload
4363; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:704 ; 4-byte Folded Reload
4364; GFX6-NEXT:    s_waitcnt vmcnt(0)
4365; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:688
4366; GFX6-NEXT:    s_waitcnt expcnt(0)
4367; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:676 ; 4-byte Folded Reload
4368; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:680 ; 4-byte Folded Reload
4369; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:684 ; 4-byte Folded Reload
4370; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:688 ; 4-byte Folded Reload
4371; GFX6-NEXT:    s_waitcnt vmcnt(0)
4372; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:672
4373; GFX6-NEXT:    s_waitcnt expcnt(0)
4374; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:660 ; 4-byte Folded Reload
4375; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:664 ; 4-byte Folded Reload
4376; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:668 ; 4-byte Folded Reload
4377; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:672 ; 4-byte Folded Reload
4378; GFX6-NEXT:    s_waitcnt vmcnt(0)
4379; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:656
4380; GFX6-NEXT:    s_waitcnt expcnt(0)
4381; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:644 ; 4-byte Folded Reload
4382; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:648 ; 4-byte Folded Reload
4383; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:652 ; 4-byte Folded Reload
4384; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:656 ; 4-byte Folded Reload
4385; GFX6-NEXT:    s_waitcnt vmcnt(0)
4386; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:640
4387; GFX6-NEXT:    s_waitcnt expcnt(0)
4388; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:628 ; 4-byte Folded Reload
4389; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:632 ; 4-byte Folded Reload
4390; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:636 ; 4-byte Folded Reload
4391; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:640 ; 4-byte Folded Reload
4392; GFX6-NEXT:    s_waitcnt vmcnt(0)
4393; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:624
4394; GFX6-NEXT:    s_waitcnt expcnt(0)
4395; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:612 ; 4-byte Folded Reload
4396; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:616 ; 4-byte Folded Reload
4397; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:620 ; 4-byte Folded Reload
4398; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:624 ; 4-byte Folded Reload
4399; GFX6-NEXT:    s_waitcnt vmcnt(0)
4400; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:608
4401; GFX6-NEXT:    s_waitcnt expcnt(0)
4402; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:596 ; 4-byte Folded Reload
4403; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:600 ; 4-byte Folded Reload
4404; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:604 ; 4-byte Folded Reload
4405; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:608 ; 4-byte Folded Reload
4406; GFX6-NEXT:    s_waitcnt vmcnt(0)
4407; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:592
4408; GFX6-NEXT:    s_waitcnt expcnt(0)
4409; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:580 ; 4-byte Folded Reload
4410; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:584 ; 4-byte Folded Reload
4411; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:588 ; 4-byte Folded Reload
4412; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:592 ; 4-byte Folded Reload
4413; GFX6-NEXT:    s_waitcnt vmcnt(0)
4414; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:576
4415; GFX6-NEXT:    s_waitcnt expcnt(0)
4416; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:564 ; 4-byte Folded Reload
4417; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:568 ; 4-byte Folded Reload
4418; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:572 ; 4-byte Folded Reload
4419; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:576 ; 4-byte Folded Reload
4420; GFX6-NEXT:    s_waitcnt vmcnt(0)
4421; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:560
4422; GFX6-NEXT:    s_waitcnt expcnt(0)
4423; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:548 ; 4-byte Folded Reload
4424; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:552 ; 4-byte Folded Reload
4425; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:556 ; 4-byte Folded Reload
4426; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:560 ; 4-byte Folded Reload
4427; GFX6-NEXT:    s_waitcnt vmcnt(0)
4428; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:544
4429; GFX6-NEXT:    s_waitcnt expcnt(0)
4430; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:532 ; 4-byte Folded Reload
4431; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:536 ; 4-byte Folded Reload
4432; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:540 ; 4-byte Folded Reload
4433; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:544 ; 4-byte Folded Reload
4434; GFX6-NEXT:    s_waitcnt vmcnt(0)
4435; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:528
4436; GFX6-NEXT:    s_waitcnt expcnt(0)
4437; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:516 ; 4-byte Folded Reload
4438; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:520 ; 4-byte Folded Reload
4439; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:524 ; 4-byte Folded Reload
4440; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:528 ; 4-byte Folded Reload
4441; GFX6-NEXT:    s_waitcnt vmcnt(0)
4442; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:512
4443; GFX6-NEXT:    s_waitcnt expcnt(0)
4444; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:500 ; 4-byte Folded Reload
4445; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:504 ; 4-byte Folded Reload
4446; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:508 ; 4-byte Folded Reload
4447; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:512 ; 4-byte Folded Reload
4448; GFX6-NEXT:    s_waitcnt vmcnt(0)
4449; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:496
4450; GFX6-NEXT:    s_waitcnt expcnt(0)
4451; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:484 ; 4-byte Folded Reload
4452; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:488 ; 4-byte Folded Reload
4453; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:492 ; 4-byte Folded Reload
4454; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:496 ; 4-byte Folded Reload
4455; GFX6-NEXT:    s_waitcnt vmcnt(0)
4456; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:480
4457; GFX6-NEXT:    s_waitcnt expcnt(0)
4458; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:468 ; 4-byte Folded Reload
4459; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:472 ; 4-byte Folded Reload
4460; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:476 ; 4-byte Folded Reload
4461; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:480 ; 4-byte Folded Reload
4462; GFX6-NEXT:    s_waitcnt vmcnt(0)
4463; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:464
4464; GFX6-NEXT:    s_waitcnt expcnt(0)
4465; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:452 ; 4-byte Folded Reload
4466; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:456 ; 4-byte Folded Reload
4467; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:460 ; 4-byte Folded Reload
4468; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:464 ; 4-byte Folded Reload
4469; GFX6-NEXT:    s_waitcnt vmcnt(0)
4470; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:448
4471; GFX6-NEXT:    s_waitcnt expcnt(0)
4472; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:436 ; 4-byte Folded Reload
4473; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:440 ; 4-byte Folded Reload
4474; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:444 ; 4-byte Folded Reload
4475; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:448 ; 4-byte Folded Reload
4476; GFX6-NEXT:    s_waitcnt vmcnt(0)
4477; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:432
4478; GFX6-NEXT:    s_waitcnt expcnt(0)
4479; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:420 ; 4-byte Folded Reload
4480; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:424 ; 4-byte Folded Reload
4481; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:428 ; 4-byte Folded Reload
4482; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:432 ; 4-byte Folded Reload
4483; GFX6-NEXT:    s_waitcnt vmcnt(0)
4484; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:416
4485; GFX6-NEXT:    s_waitcnt expcnt(0)
4486; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:404 ; 4-byte Folded Reload
4487; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:408 ; 4-byte Folded Reload
4488; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:412 ; 4-byte Folded Reload
4489; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:416 ; 4-byte Folded Reload
4490; GFX6-NEXT:    s_waitcnt vmcnt(0)
4491; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:400
4492; GFX6-NEXT:    s_waitcnt expcnt(0)
4493; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:388 ; 4-byte Folded Reload
4494; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:392 ; 4-byte Folded Reload
4495; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:396 ; 4-byte Folded Reload
4496; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:400 ; 4-byte Folded Reload
4497; GFX6-NEXT:    s_waitcnt vmcnt(0)
4498; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:384
4499; GFX6-NEXT:    s_waitcnt expcnt(0)
4500; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:372 ; 4-byte Folded Reload
4501; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:376 ; 4-byte Folded Reload
4502; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:380 ; 4-byte Folded Reload
4503; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:384 ; 4-byte Folded Reload
4504; GFX6-NEXT:    s_waitcnt vmcnt(0)
4505; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:368
4506; GFX6-NEXT:    s_waitcnt expcnt(0)
4507; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:356 ; 4-byte Folded Reload
4508; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:360 ; 4-byte Folded Reload
4509; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:364 ; 4-byte Folded Reload
4510; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:368 ; 4-byte Folded Reload
4511; GFX6-NEXT:    s_waitcnt vmcnt(0)
4512; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:352
4513; GFX6-NEXT:    s_waitcnt expcnt(0)
4514; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:340 ; 4-byte Folded Reload
4515; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:344 ; 4-byte Folded Reload
4516; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:348 ; 4-byte Folded Reload
4517; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:352 ; 4-byte Folded Reload
4518; GFX6-NEXT:    s_waitcnt vmcnt(0)
4519; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:336
4520; GFX6-NEXT:    s_waitcnt expcnt(0)
4521; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:324 ; 4-byte Folded Reload
4522; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:328 ; 4-byte Folded Reload
4523; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:332 ; 4-byte Folded Reload
4524; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:336 ; 4-byte Folded Reload
4525; GFX6-NEXT:    s_waitcnt vmcnt(0)
4526; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:320
4527; GFX6-NEXT:    s_waitcnt expcnt(0)
4528; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:308 ; 4-byte Folded Reload
4529; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:312 ; 4-byte Folded Reload
4530; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:316 ; 4-byte Folded Reload
4531; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:320 ; 4-byte Folded Reload
4532; GFX6-NEXT:    s_waitcnt vmcnt(0)
4533; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:304
4534; GFX6-NEXT:    s_waitcnt expcnt(0)
4535; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:292 ; 4-byte Folded Reload
4536; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:296 ; 4-byte Folded Reload
4537; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:300 ; 4-byte Folded Reload
4538; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:304 ; 4-byte Folded Reload
4539; GFX6-NEXT:    s_waitcnt vmcnt(0)
4540; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:288
4541; GFX6-NEXT:    s_waitcnt expcnt(0)
4542; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:276 ; 4-byte Folded Reload
4543; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:280 ; 4-byte Folded Reload
4544; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:284 ; 4-byte Folded Reload
4545; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:288 ; 4-byte Folded Reload
4546; GFX6-NEXT:    s_waitcnt vmcnt(0)
4547; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:272
4548; GFX6-NEXT:    s_waitcnt expcnt(0)
4549; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:260 ; 4-byte Folded Reload
4550; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:264 ; 4-byte Folded Reload
4551; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:268 ; 4-byte Folded Reload
4552; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:272 ; 4-byte Folded Reload
4553; GFX6-NEXT:    s_waitcnt vmcnt(0)
4554; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:256
4555; GFX6-NEXT:    s_waitcnt expcnt(0)
4556; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:244 ; 4-byte Folded Reload
4557; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:248 ; 4-byte Folded Reload
4558; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:252 ; 4-byte Folded Reload
4559; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:256 ; 4-byte Folded Reload
4560; GFX6-NEXT:    s_waitcnt vmcnt(0)
4561; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:240
4562; GFX6-NEXT:    s_waitcnt expcnt(0)
4563; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:228 ; 4-byte Folded Reload
4564; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:232 ; 4-byte Folded Reload
4565; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:236 ; 4-byte Folded Reload
4566; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:240 ; 4-byte Folded Reload
4567; GFX6-NEXT:    s_waitcnt vmcnt(0)
4568; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:224
4569; GFX6-NEXT:    s_waitcnt expcnt(0)
4570; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:212 ; 4-byte Folded Reload
4571; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:216 ; 4-byte Folded Reload
4572; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:220 ; 4-byte Folded Reload
4573; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:224 ; 4-byte Folded Reload
4574; GFX6-NEXT:    s_waitcnt vmcnt(0)
4575; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:208
4576; GFX6-NEXT:    s_waitcnt expcnt(0)
4577; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:196 ; 4-byte Folded Reload
4578; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:200 ; 4-byte Folded Reload
4579; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:204 ; 4-byte Folded Reload
4580; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:208 ; 4-byte Folded Reload
4581; GFX6-NEXT:    s_waitcnt vmcnt(0)
4582; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:192
4583; GFX6-NEXT:    s_waitcnt expcnt(0)
4584; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:180 ; 4-byte Folded Reload
4585; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:184 ; 4-byte Folded Reload
4586; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:188 ; 4-byte Folded Reload
4587; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:192 ; 4-byte Folded Reload
4588; GFX6-NEXT:    s_waitcnt vmcnt(0)
4589; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:176
4590; GFX6-NEXT:    s_waitcnt expcnt(0)
4591; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:164 ; 4-byte Folded Reload
4592; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:168 ; 4-byte Folded Reload
4593; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:172 ; 4-byte Folded Reload
4594; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:176 ; 4-byte Folded Reload
4595; GFX6-NEXT:    s_waitcnt vmcnt(0)
4596; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:160
4597; GFX6-NEXT:    s_waitcnt expcnt(0)
4598; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:148 ; 4-byte Folded Reload
4599; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:152 ; 4-byte Folded Reload
4600; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:156 ; 4-byte Folded Reload
4601; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:160 ; 4-byte Folded Reload
4602; GFX6-NEXT:    s_waitcnt vmcnt(0)
4603; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:144
4604; GFX6-NEXT:    s_waitcnt expcnt(0)
4605; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:132 ; 4-byte Folded Reload
4606; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:136 ; 4-byte Folded Reload
4607; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:140 ; 4-byte Folded Reload
4608; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:144 ; 4-byte Folded Reload
4609; GFX6-NEXT:    s_waitcnt vmcnt(0)
4610; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:128
4611; GFX6-NEXT:    s_waitcnt expcnt(0)
4612; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:116 ; 4-byte Folded Reload
4613; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:120 ; 4-byte Folded Reload
4614; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:124 ; 4-byte Folded Reload
4615; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:128 ; 4-byte Folded Reload
4616; GFX6-NEXT:    s_waitcnt vmcnt(0)
4617; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:112
4618; GFX6-NEXT:    s_waitcnt expcnt(0)
4619; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:100 ; 4-byte Folded Reload
4620; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:104 ; 4-byte Folded Reload
4621; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:108 ; 4-byte Folded Reload
4622; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:112 ; 4-byte Folded Reload
4623; GFX6-NEXT:    s_waitcnt vmcnt(0)
4624; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:96
4625; GFX6-NEXT:    s_waitcnt expcnt(0)
4626; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:84 ; 4-byte Folded Reload
4627; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:88 ; 4-byte Folded Reload
4628; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:92 ; 4-byte Folded Reload
4629; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:96 ; 4-byte Folded Reload
4630; GFX6-NEXT:    s_waitcnt vmcnt(0)
4631; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:80
4632; GFX6-NEXT:    s_waitcnt expcnt(0)
4633; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:68 ; 4-byte Folded Reload
4634; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:72 ; 4-byte Folded Reload
4635; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:76 ; 4-byte Folded Reload
4636; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:80 ; 4-byte Folded Reload
4637; GFX6-NEXT:    s_waitcnt vmcnt(0)
4638; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:64
4639; GFX6-NEXT:    s_waitcnt expcnt(0)
4640; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:52 ; 4-byte Folded Reload
4641; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:56 ; 4-byte Folded Reload
4642; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:60 ; 4-byte Folded Reload
4643; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:64 ; 4-byte Folded Reload
4644; GFX6-NEXT:    s_waitcnt vmcnt(0)
4645; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:48
4646; GFX6-NEXT:    s_waitcnt expcnt(0)
4647; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:36 ; 4-byte Folded Reload
4648; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:40 ; 4-byte Folded Reload
4649; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:44 ; 4-byte Folded Reload
4650; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:48 ; 4-byte Folded Reload
4651; GFX6-NEXT:    s_waitcnt vmcnt(0)
4652; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:32
4653; GFX6-NEXT:    s_waitcnt expcnt(0)
4654; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:20 ; 4-byte Folded Reload
4655; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:24 ; 4-byte Folded Reload
4656; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:28 ; 4-byte Folded Reload
4657; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:32 ; 4-byte Folded Reload
4658; GFX6-NEXT:    s_waitcnt vmcnt(0)
4659; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64 offset:16
4660; GFX6-NEXT:    s_waitcnt expcnt(0)
4661; GFX6-NEXT:    buffer_load_dword v0, off, s[40:43], 0 offset:4 ; 4-byte Folded Reload
4662; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0 offset:8 ; 4-byte Folded Reload
4663; GFX6-NEXT:    buffer_load_dword v2, off, s[40:43], 0 offset:12 ; 4-byte Folded Reload
4664; GFX6-NEXT:    buffer_load_dword v3, off, s[40:43], 0 offset:16 ; 4-byte Folded Reload
4665; GFX6-NEXT:    s_waitcnt vmcnt(0)
4666; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[0:3], 0 addr64
4667; GFX6-NEXT:    s_endpgm
4668;
4669; GFX9-FLATSCR-LABEL: test:
4670; GFX9-FLATSCR:       ; %bb.0: ; %entry
4671; GFX9-FLATSCR-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
4672; GFX9-FLATSCR-NEXT:    v_mbcnt_lo_u32_b32 v0, -1, 0
4673; GFX9-FLATSCR-NEXT:    v_mbcnt_hi_u32_b32 v0, -1, v0
4674; GFX9-FLATSCR-NEXT:    v_lshlrev_b32_e32 v5, 13, v0
4675; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x80
4676; GFX9-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
4677; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v2, vcc, s2, v5
4678; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v0, s3
4679; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v0, vcc
4680; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4681; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4682; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4683; GFX9-FLATSCR-NEXT:    s_add_u32 flat_scratch_lo, s8, s13
4684; GFX9-FLATSCR-NEXT:    s_addc_u32 flat_scratch_hi, s9, 0
4685; GFX9-FLATSCR-NEXT:    s_mov_b32 s4, 4
4686; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4687; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4688; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4689; GFX9-FLATSCR-NEXT:    s_mov_b32 s4, 20
4690; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4691; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4692; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4693; GFX9-FLATSCR-NEXT:    s_mov_b32 s4, 36
4694; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4695; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4696; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4697; GFX9-FLATSCR-NEXT:    s_mov_b32 s4, 52
4698; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4699; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4700; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4701; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x44
4702; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4703; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4704; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4705; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x54
4706; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4707; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4708; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4709; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x64
4710; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4711; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4712; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4713; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x74
4714; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4715; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4716; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x100
4717; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4718; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4719; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4720; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x84
4721; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4722; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4723; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4724; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x94
4725; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4726; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4727; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4728; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa4
4729; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4730; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4731; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4732; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb4
4733; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4734; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4735; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4736; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc4
4737; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4738; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4739; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4740; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd4
4741; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4742; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4743; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4744; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe4
4745; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4746; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4747; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4748; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf4
4749; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4750; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4751; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x180
4752; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4753; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4754; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4755; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x104
4756; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4757; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4758; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4759; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x114
4760; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4761; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4762; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4763; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x124
4764; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4765; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4766; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4767; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x134
4768; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4769; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4770; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4771; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x144
4772; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4773; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4774; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4775; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x154
4776; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4777; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4778; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4779; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x164
4780; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4781; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4782; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4783; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x174
4784; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4785; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4786; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x200
4787; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4788; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4789; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4790; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x184
4791; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4792; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4793; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4794; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x194
4795; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4796; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4797; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4798; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1a4
4799; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4800; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4801; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4802; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1b4
4803; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4804; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4805; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4806; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1c4
4807; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4808; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4809; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4810; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1d4
4811; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4812; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4813; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4814; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1e4
4815; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4816; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4817; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4818; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1f4
4819; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4820; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4821; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x280
4822; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4823; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4824; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4825; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x204
4826; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4827; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4828; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4829; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x214
4830; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4831; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4832; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4833; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x224
4834; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4835; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4836; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4837; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x234
4838; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4839; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4840; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4841; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x244
4842; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4843; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4844; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4845; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x254
4846; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4847; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4848; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4849; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x264
4850; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4851; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4852; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4853; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x274
4854; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4855; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4856; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x300
4857; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4858; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4859; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4860; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x284
4861; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4862; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4863; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4864; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x294
4865; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4866; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4867; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4868; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2a4
4869; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4870; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4871; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4872; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2b4
4873; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4874; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4875; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4876; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2c4
4877; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4878; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4879; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4880; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2d4
4881; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4882; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4883; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4884; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2e4
4885; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4886; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4887; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4888; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x2f4
4889; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4890; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4891; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x380
4892; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4893; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4894; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4895; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x304
4896; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4897; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4898; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4899; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x314
4900; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4901; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4902; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4903; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x324
4904; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4905; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4906; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4907; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x334
4908; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4909; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4910; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4911; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x344
4912; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4913; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4914; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4915; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x354
4916; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4917; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4918; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4919; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x364
4920; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4921; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4922; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4080
4923; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x374
4924; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4925; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4926; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x400
4927; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, s4, v2
4928; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v3, vcc
4929; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3968
4930; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x384
4931; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4932; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4933; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:3984
4934; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x394
4935; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4936; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4937; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4000
4938; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3a4
4939; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4940; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4941; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4016
4942; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3b4
4943; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4944; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4945; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4032
4946; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3c4
4947; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4948; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4949; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4048
4950; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3d4
4951; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4952; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4953; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:4064
4954; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:4080
4955; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3e4
4956; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
4957; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
4958; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x3f4
4959; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
4960; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4961; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3]
4962; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x404
4963; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v6, s1
4964; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4965; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4966; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:16
4967; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x414
4968; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4969; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4970; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:32
4971; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x424
4972; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4973; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4974; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:48
4975; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x434
4976; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4977; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4978; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:64
4979; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x444
4980; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4981; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4982; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:80
4983; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x454
4984; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4985; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4986; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:96
4987; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x464
4988; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4989; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4990; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:112
4991; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x474
4992; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4993; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4994; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:128
4995; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x484
4996; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
4997; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
4998; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:144
4999; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x494
5000; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5001; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5002; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:160
5003; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4a4
5004; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5005; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5006; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:176
5007; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4b4
5008; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5009; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5010; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:192
5011; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4c4
5012; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5013; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5014; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:208
5015; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4d4
5016; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5017; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5018; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:224
5019; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4e4
5020; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5021; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5022; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:240
5023; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x4f4
5024; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5025; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5026; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:256
5027; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x504
5028; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5029; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5030; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:272
5031; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x514
5032; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5033; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5034; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:288
5035; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x524
5036; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5037; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5038; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:304
5039; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x534
5040; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5041; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5042; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:320
5043; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x544
5044; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5045; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5046; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:336
5047; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x554
5048; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5049; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5050; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:352
5051; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x564
5052; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5053; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5054; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:368
5055; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x574
5056; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5057; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5058; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:384
5059; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x584
5060; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5061; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5062; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:400
5063; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x594
5064; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5065; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5066; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:416
5067; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5a4
5068; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5069; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5070; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:432
5071; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5b4
5072; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5073; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5074; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:448
5075; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5c4
5076; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5077; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5078; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:464
5079; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5d4
5080; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5081; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5082; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:480
5083; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5e4
5084; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5085; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5086; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:496
5087; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x5f4
5088; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5089; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5090; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:512
5091; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x604
5092; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5093; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5094; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:528
5095; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x614
5096; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5097; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5098; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:544
5099; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x624
5100; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5101; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5102; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:560
5103; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x634
5104; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5105; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5106; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:576
5107; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x644
5108; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5109; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5110; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:592
5111; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x654
5112; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5113; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5114; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:608
5115; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x664
5116; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5117; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5118; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:624
5119; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x674
5120; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5121; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5122; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:640
5123; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x684
5124; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5125; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5126; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:656
5127; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x694
5128; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5129; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5130; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:672
5131; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6a4
5132; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5133; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5134; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:688
5135; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6b4
5136; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5137; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5138; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:704
5139; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6c4
5140; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5141; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5142; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:720
5143; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6d4
5144; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5145; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5146; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:736
5147; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6e4
5148; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5149; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5150; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:752
5151; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x6f4
5152; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5153; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5154; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:768
5155; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x704
5156; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5157; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5158; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:784
5159; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x714
5160; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5161; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5162; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:800
5163; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x724
5164; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5165; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5166; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:816
5167; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x734
5168; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5169; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5170; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:832
5171; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x744
5172; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5173; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5174; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:848
5175; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x754
5176; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5177; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5178; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:864
5179; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x764
5180; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5181; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5182; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:880
5183; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x774
5184; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5185; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5186; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:896
5187; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x784
5188; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5189; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5190; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:912
5191; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x794
5192; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5193; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5194; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:928
5195; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7a4
5196; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5197; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5198; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:944
5199; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7b4
5200; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5201; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5202; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:960
5203; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7c4
5204; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5205; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5206; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:976
5207; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7d4
5208; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5209; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5210; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:992
5211; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7e4
5212; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5213; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5214; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1008
5215; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x7f4
5216; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5217; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5218; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1024
5219; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x804
5220; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5221; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5222; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1040
5223; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x814
5224; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5225; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5226; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1056
5227; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x824
5228; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5229; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5230; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1072
5231; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x834
5232; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5233; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5234; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1088
5235; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x844
5236; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5237; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5238; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1104
5239; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x854
5240; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5241; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5242; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1120
5243; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x864
5244; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5245; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5246; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1136
5247; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x874
5248; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5249; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5250; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1152
5251; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x884
5252; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5253; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5254; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1168
5255; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x894
5256; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5257; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5258; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1184
5259; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8a4
5260; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5261; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5262; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1200
5263; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8b4
5264; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5265; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5266; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1216
5267; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8c4
5268; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5269; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5270; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1232
5271; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8d4
5272; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5273; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5274; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1248
5275; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8e4
5276; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5277; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5278; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1264
5279; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x8f4
5280; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5281; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5282; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1280
5283; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x904
5284; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5285; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5286; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1296
5287; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x914
5288; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5289; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5290; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1312
5291; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x924
5292; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5293; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5294; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1328
5295; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x934
5296; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5297; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5298; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1344
5299; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x944
5300; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5301; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5302; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1360
5303; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x954
5304; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5305; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5306; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1376
5307; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x964
5308; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5309; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5310; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1392
5311; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x974
5312; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5313; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5314; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1408
5315; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x984
5316; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5317; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5318; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1424
5319; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x994
5320; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5321; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5322; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1440
5323; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9a4
5324; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5325; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5326; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1456
5327; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9b4
5328; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5329; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5330; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1472
5331; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9c4
5332; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5333; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5334; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1488
5335; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9d4
5336; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5337; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5338; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1504
5339; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9e4
5340; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5341; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5342; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1520
5343; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x9f4
5344; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5345; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5346; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1536
5347; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa04
5348; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5349; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5350; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1552
5351; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa14
5352; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5353; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5354; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1568
5355; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa24
5356; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5357; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5358; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1584
5359; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa34
5360; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5361; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5362; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1600
5363; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa44
5364; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5365; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5366; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1616
5367; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa54
5368; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5369; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5370; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1632
5371; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa64
5372; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5373; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5374; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1648
5375; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa74
5376; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5377; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5378; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1664
5379; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa84
5380; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5381; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5382; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1680
5383; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xa94
5384; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5385; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5386; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1696
5387; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xaa4
5388; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5389; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5390; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1712
5391; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xab4
5392; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5393; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5394; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1728
5395; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xac4
5396; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5397; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5398; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1744
5399; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xad4
5400; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5401; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5402; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1760
5403; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xae4
5404; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5405; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5406; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1776
5407; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xaf4
5408; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5409; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5410; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1792
5411; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb04
5412; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5413; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5414; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1808
5415; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb14
5416; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5417; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5418; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1824
5419; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb24
5420; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5421; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5422; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1840
5423; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb34
5424; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5425; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5426; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1856
5427; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb44
5428; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5429; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5430; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1872
5431; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb54
5432; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5433; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5434; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1888
5435; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb64
5436; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5437; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5438; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1904
5439; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb74
5440; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5441; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5442; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1920
5443; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb84
5444; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5445; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5446; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1936
5447; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xb94
5448; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5449; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5450; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1952
5451; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xba4
5452; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5453; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5454; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1968
5455; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xbb4
5456; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5457; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5458; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1984
5459; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xbc4
5460; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5461; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5462; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2000
5463; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xbd4
5464; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5465; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5466; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2016
5467; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xbe4
5468; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5469; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5470; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2032
5471; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xbf4
5472; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5473; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5474; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2048
5475; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc04
5476; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5477; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5478; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2064
5479; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc14
5480; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5481; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5482; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2080
5483; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc24
5484; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5485; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5486; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2096
5487; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc34
5488; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5489; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5490; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2112
5491; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc44
5492; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5493; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5494; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2128
5495; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc54
5496; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5497; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5498; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2144
5499; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc64
5500; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5501; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5502; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2160
5503; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc74
5504; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5505; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5506; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2176
5507; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc84
5508; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5509; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5510; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2192
5511; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xc94
5512; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5513; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5514; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2208
5515; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xca4
5516; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5517; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5518; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2224
5519; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xcb4
5520; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5521; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5522; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2240
5523; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xcc4
5524; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5525; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5526; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2256
5527; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xcd4
5528; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5529; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5530; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2272
5531; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xce4
5532; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5533; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5534; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2288
5535; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xcf4
5536; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5537; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5538; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2304
5539; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd04
5540; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5541; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5542; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2320
5543; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd14
5544; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5545; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5546; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2336
5547; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd24
5548; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5549; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5550; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2352
5551; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd34
5552; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5553; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5554; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2368
5555; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd44
5556; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5557; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5558; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2384
5559; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd54
5560; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5561; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5562; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2400
5563; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd64
5564; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5565; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5566; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2416
5567; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd74
5568; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5569; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5570; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2432
5571; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd84
5572; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5573; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5574; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2448
5575; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xd94
5576; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5577; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5578; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2464
5579; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xda4
5580; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5581; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5582; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2480
5583; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xdb4
5584; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5585; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5586; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2496
5587; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xdc4
5588; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5589; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5590; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2512
5591; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xdd4
5592; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5593; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5594; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2528
5595; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xde4
5596; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5597; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5598; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2544
5599; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xdf4
5600; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5601; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5602; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2560
5603; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe04
5604; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5605; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5606; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2576
5607; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe14
5608; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5609; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5610; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2592
5611; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe24
5612; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5613; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5614; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2608
5615; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe34
5616; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5617; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5618; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2624
5619; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe44
5620; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5621; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5622; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2640
5623; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe54
5624; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5625; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5626; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2656
5627; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe64
5628; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5629; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5630; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2672
5631; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe74
5632; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5633; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5634; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2688
5635; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe84
5636; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5637; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5638; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2704
5639; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xe94
5640; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5641; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5642; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2720
5643; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xea4
5644; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5645; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5646; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2736
5647; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xeb4
5648; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5649; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5650; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2752
5651; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xec4
5652; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5653; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5654; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2768
5655; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xed4
5656; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5657; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5658; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2784
5659; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xee4
5660; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5661; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5662; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2800
5663; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xef4
5664; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5665; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5666; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2816
5667; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf04
5668; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5669; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5670; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2832
5671; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf14
5672; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5673; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5674; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2848
5675; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf24
5676; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5677; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5678; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2864
5679; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf34
5680; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5681; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5682; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2880
5683; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf44
5684; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5685; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5686; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2896
5687; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf54
5688; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5689; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5690; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2912
5691; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf64
5692; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5693; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5694; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2928
5695; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf74
5696; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5697; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5698; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2944
5699; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf84
5700; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5701; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5702; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2960
5703; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xf94
5704; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5705; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5706; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2976
5707; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xfa4
5708; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5709; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5710; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2992
5711; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xfb4
5712; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5713; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5714; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3008
5715; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xfc4
5716; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5717; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5718; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3024
5719; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xfd4
5720; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5721; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5722; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3040
5723; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xfe4
5724; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5725; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5726; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3056
5727; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0xff4
5728; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5729; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5730; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3072
5731; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1004
5732; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5733; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5734; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3088
5735; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1014
5736; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5737; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5738; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3104
5739; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1024
5740; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5741; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5742; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3120
5743; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1034
5744; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5745; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5746; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3136
5747; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1044
5748; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5749; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5750; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3152
5751; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1054
5752; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5753; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5754; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3168
5755; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1064
5756; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5757; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5758; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3184
5759; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1074
5760; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5761; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5762; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3200
5763; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1084
5764; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5765; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5766; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3216
5767; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1094
5768; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5769; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5770; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3232
5771; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10a4
5772; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5773; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5774; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3248
5775; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10b4
5776; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5777; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5778; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3264
5779; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10c4
5780; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5781; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5782; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3280
5783; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10d4
5784; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5785; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5786; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3296
5787; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10e4
5788; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5789; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5790; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3312
5791; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x10f4
5792; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5793; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5794; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3328
5795; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1104
5796; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5797; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5798; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3344
5799; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1114
5800; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5801; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5802; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3360
5803; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1124
5804; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5805; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5806; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3376
5807; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1134
5808; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5809; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5810; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3392
5811; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1144
5812; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5813; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5814; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3408
5815; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1154
5816; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5817; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5818; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3424
5819; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1164
5820; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5821; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5822; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3440
5823; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1174
5824; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5825; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5826; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3456
5827; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1184
5828; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5829; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5830; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3472
5831; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1194
5832; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5833; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5834; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3488
5835; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11a4
5836; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5837; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5838; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3504
5839; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11b4
5840; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5841; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5842; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3520
5843; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11c4
5844; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5845; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5846; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3536
5847; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11d4
5848; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5849; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5850; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3552
5851; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11e4
5852; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5853; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5854; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3568
5855; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x11f4
5856; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5857; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5858; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3584
5859; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1204
5860; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5861; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5862; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3600
5863; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1214
5864; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5865; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5866; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3616
5867; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1224
5868; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5869; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5870; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3632
5871; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1234
5872; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5873; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5874; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3648
5875; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1244
5876; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5877; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5878; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3664
5879; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1254
5880; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5881; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5882; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3680
5883; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1264
5884; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5885; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5886; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3696
5887; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1274
5888; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5889; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5890; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3712
5891; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1284
5892; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5893; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5894; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3728
5895; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1294
5896; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5897; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5898; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3744
5899; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12a4
5900; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5901; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5902; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3760
5903; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12b4
5904; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5905; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5906; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3776
5907; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12c4
5908; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5909; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5910; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3792
5911; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12d4
5912; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5913; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5914; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3808
5915; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12e4
5916; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5917; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5918; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3824
5919; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x12f4
5920; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5921; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5922; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3840
5923; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1304
5924; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5925; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5926; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3856
5927; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1314
5928; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5929; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5930; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3872
5931; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1324
5932; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5933; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5934; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3888
5935; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1334
5936; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5937; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5938; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3904
5939; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1344
5940; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5941; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5942; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3920
5943; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1354
5944; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5945; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5946; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3936
5947; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1364
5948; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5949; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5950; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3952
5951; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1374
5952; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5953; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5954; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3968
5955; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1384
5956; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5957; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5958; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:3984
5959; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x1394
5960; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5961; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5962; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4000
5963; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x13a4
5964; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5965; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5966; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4016
5967; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x13b4
5968; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5969; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5970; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4032
5971; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x13c4
5972; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5973; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5974; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4048
5975; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x13d4
5976; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5977; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5978; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4064
5979; GFX9-FLATSCR-NEXT:    s_movk_i32 s4, 0x13e4
5980; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
5981; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
5982; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:4080
5983; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x13e4
5984; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5985; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5986; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5987; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5988; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5989; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5990; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5991; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5992; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5993; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5994; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5995; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5996; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
5997; GFX9-FLATSCR-NEXT:    ;;#ASMEND
5998; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v4, vcc, s0, v5
5999; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
6000; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6001; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4080
6002; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6003; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x13d4
6004; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6005; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4064
6006; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6007; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x13c4
6008; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6009; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4048
6010; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6011; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x13b4
6012; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6013; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4032
6014; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6015; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x13a4
6016; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6017; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4016
6018; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6019; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1394
6020; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6021; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:4000
6022; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6023; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1384
6024; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6025; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3984
6026; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6027; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1374
6028; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6029; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3968
6030; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6031; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1364
6032; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6033; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3952
6034; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6035; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1354
6036; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6037; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3936
6038; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6039; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1344
6040; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6041; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3920
6042; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6043; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1334
6044; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6045; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3904
6046; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6047; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1324
6048; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6049; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3888
6050; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6051; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1314
6052; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6053; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3872
6054; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6055; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1304
6056; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6057; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3856
6058; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6059; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12f4
6060; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6061; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3840
6062; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6063; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12e4
6064; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6065; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3824
6066; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6067; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12d4
6068; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6069; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3808
6070; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6071; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12c4
6072; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6073; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3792
6074; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6075; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12b4
6076; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6077; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3776
6078; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6079; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x12a4
6080; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6081; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3760
6082; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6083; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1294
6084; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6085; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3744
6086; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6087; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1284
6088; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6089; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3728
6090; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6091; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1274
6092; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6093; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3712
6094; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6095; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1264
6096; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6097; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3696
6098; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6099; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1254
6100; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6101; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3680
6102; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6103; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1244
6104; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6105; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3664
6106; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6107; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1234
6108; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6109; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3648
6110; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6111; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1224
6112; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6113; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3632
6114; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6115; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1214
6116; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6117; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3616
6118; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6119; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1204
6120; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6121; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3600
6122; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6123; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11f4
6124; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6125; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3584
6126; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6127; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11e4
6128; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6129; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3568
6130; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6131; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11d4
6132; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6133; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3552
6134; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6135; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11c4
6136; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6137; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3536
6138; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6139; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11b4
6140; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6141; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3520
6142; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6143; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x11a4
6144; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6145; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3504
6146; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6147; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1194
6148; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6149; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3488
6150; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6151; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1184
6152; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6153; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3472
6154; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6155; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1174
6156; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6157; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3456
6158; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6159; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1164
6160; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6161; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3440
6162; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6163; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1154
6164; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6165; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3424
6166; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6167; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1144
6168; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6169; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3408
6170; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6171; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1134
6172; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6173; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3392
6174; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6175; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1124
6176; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6177; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3376
6178; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6179; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1114
6180; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6181; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3360
6182; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6183; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1104
6184; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6185; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3344
6186; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6187; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10f4
6188; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6189; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3328
6190; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6191; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10e4
6192; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6193; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3312
6194; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6195; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10d4
6196; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6197; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3296
6198; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6199; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10c4
6200; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6201; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3280
6202; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6203; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10b4
6204; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6205; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3264
6206; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6207; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x10a4
6208; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6209; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3248
6210; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6211; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1094
6212; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6213; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3232
6214; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6215; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1084
6216; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6217; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3216
6218; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6219; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1074
6220; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6221; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3200
6222; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6223; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1064
6224; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6225; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3184
6226; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6227; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1054
6228; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6229; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3168
6230; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6231; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1044
6232; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6233; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3152
6234; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6235; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1034
6236; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6237; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3136
6238; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6239; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1024
6240; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6241; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3120
6242; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6243; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1014
6244; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6245; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3104
6246; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6247; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x1004
6248; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6249; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3088
6250; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6251; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xff4
6252; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6253; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3072
6254; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6255; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xfe4
6256; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6257; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3056
6258; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6259; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xfd4
6260; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6261; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3040
6262; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6263; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xfc4
6264; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6265; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3024
6266; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6267; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xfb4
6268; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6269; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:3008
6270; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6271; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xfa4
6272; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6273; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2992
6274; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6275; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf94
6276; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6277; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2976
6278; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6279; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf84
6280; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6281; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2960
6282; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6283; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf74
6284; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6285; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2944
6286; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6287; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf64
6288; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6289; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2928
6290; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6291; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf54
6292; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6293; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2912
6294; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6295; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf44
6296; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6297; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2896
6298; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6299; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf34
6300; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6301; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2880
6302; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6303; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf24
6304; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6305; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2864
6306; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6307; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf14
6308; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6309; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2848
6310; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6311; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xf04
6312; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6313; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2832
6314; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6315; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xef4
6316; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6317; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2816
6318; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6319; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xee4
6320; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6321; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2800
6322; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6323; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xed4
6324; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6325; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2784
6326; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6327; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xec4
6328; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6329; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2768
6330; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6331; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xeb4
6332; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6333; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2752
6334; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6335; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xea4
6336; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6337; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2736
6338; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6339; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe94
6340; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6341; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2720
6342; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6343; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe84
6344; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6345; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2704
6346; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6347; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe74
6348; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6349; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2688
6350; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6351; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe64
6352; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6353; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2672
6354; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6355; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe54
6356; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6357; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2656
6358; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6359; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe44
6360; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6361; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2640
6362; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6363; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe34
6364; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6365; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2624
6366; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6367; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe24
6368; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6369; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2608
6370; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6371; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe14
6372; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6373; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2592
6374; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6375; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xe04
6376; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6377; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2576
6378; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6379; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xdf4
6380; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6381; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2560
6382; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6383; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xde4
6384; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6385; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2544
6386; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6387; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xdd4
6388; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6389; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2528
6390; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6391; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xdc4
6392; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6393; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2512
6394; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6395; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xdb4
6396; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6397; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2496
6398; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6399; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xda4
6400; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6401; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2480
6402; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6403; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd94
6404; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6405; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2464
6406; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6407; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd84
6408; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6409; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2448
6410; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6411; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd74
6412; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6413; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2432
6414; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6415; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd64
6416; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6417; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2416
6418; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6419; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd54
6420; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6421; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2400
6422; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6423; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd44
6424; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6425; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2384
6426; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6427; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd34
6428; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6429; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2368
6430; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6431; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd24
6432; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6433; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2352
6434; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6435; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd14
6436; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6437; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2336
6438; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6439; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xd04
6440; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6441; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2320
6442; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6443; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xcf4
6444; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6445; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2304
6446; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6447; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xce4
6448; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6449; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2288
6450; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6451; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xcd4
6452; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6453; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2272
6454; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6455; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xcc4
6456; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6457; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2256
6458; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6459; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xcb4
6460; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6461; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2240
6462; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6463; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xca4
6464; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6465; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2224
6466; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6467; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc94
6468; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6469; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2208
6470; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6471; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc84
6472; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6473; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2192
6474; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6475; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc74
6476; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6477; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2176
6478; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6479; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc64
6480; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6481; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2160
6482; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6483; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc54
6484; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6485; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2144
6486; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6487; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc44
6488; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6489; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2128
6490; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6491; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc34
6492; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6493; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2112
6494; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6495; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc24
6496; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6497; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2096
6498; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6499; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc14
6500; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6501; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2080
6502; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6503; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xc04
6504; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6505; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2064
6506; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6507; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xbf4
6508; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6509; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2048
6510; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6511; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xbe4
6512; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6513; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2032
6514; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6515; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xbd4
6516; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6517; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2016
6518; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6519; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xbc4
6520; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6521; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2000
6522; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6523; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xbb4
6524; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6525; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1984
6526; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6527; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xba4
6528; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6529; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1968
6530; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6531; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb94
6532; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6533; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1952
6534; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6535; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb84
6536; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6537; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1936
6538; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6539; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb74
6540; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6541; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1920
6542; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6543; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb64
6544; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6545; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1904
6546; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6547; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb54
6548; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6549; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1888
6550; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6551; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb44
6552; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6553; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1872
6554; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6555; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb34
6556; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6557; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1856
6558; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6559; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb24
6560; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6561; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1840
6562; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6563; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb14
6564; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6565; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1824
6566; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6567; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xb04
6568; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6569; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1808
6570; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6571; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xaf4
6572; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6573; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1792
6574; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6575; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xae4
6576; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6577; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1776
6578; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6579; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xad4
6580; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6581; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1760
6582; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6583; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xac4
6584; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6585; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1744
6586; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6587; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xab4
6588; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6589; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1728
6590; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6591; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xaa4
6592; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6593; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1712
6594; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6595; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa94
6596; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6597; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1696
6598; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6599; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa84
6600; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6601; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1680
6602; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6603; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa74
6604; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6605; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1664
6606; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6607; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa64
6608; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6609; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1648
6610; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6611; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa54
6612; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6613; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1632
6614; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6615; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa44
6616; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6617; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1616
6618; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6619; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa34
6620; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6621; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1600
6622; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6623; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa24
6624; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6625; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1584
6626; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6627; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa14
6628; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6629; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1568
6630; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6631; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0xa04
6632; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6633; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1552
6634; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6635; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9f4
6636; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6637; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1536
6638; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6639; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9e4
6640; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6641; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1520
6642; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6643; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9d4
6644; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6645; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1504
6646; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6647; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9c4
6648; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6649; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1488
6650; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6651; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9b4
6652; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6653; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1472
6654; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6655; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x9a4
6656; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6657; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1456
6658; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6659; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x994
6660; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6661; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1440
6662; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6663; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x984
6664; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6665; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1424
6666; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6667; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x974
6668; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6669; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1408
6670; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6671; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x964
6672; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6673; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1392
6674; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6675; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x954
6676; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6677; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1376
6678; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6679; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x944
6680; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6681; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1360
6682; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6683; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x934
6684; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6685; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1344
6686; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6687; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x924
6688; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6689; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1328
6690; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6691; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x914
6692; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6693; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1312
6694; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6695; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x904
6696; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6697; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1296
6698; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6699; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8f4
6700; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6701; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1280
6702; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6703; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8e4
6704; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6705; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1264
6706; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6707; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8d4
6708; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6709; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1248
6710; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6711; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8c4
6712; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6713; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1232
6714; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6715; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8b4
6716; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6717; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1216
6718; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6719; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x8a4
6720; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6721; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1200
6722; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6723; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x894
6724; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6725; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1184
6726; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6727; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x884
6728; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6729; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1168
6730; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6731; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x874
6732; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6733; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1152
6734; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6735; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x864
6736; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6737; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1136
6738; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6739; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x854
6740; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6741; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1120
6742; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6743; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x844
6744; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6745; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1104
6746; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6747; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x834
6748; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6749; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1088
6750; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6751; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x824
6752; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6753; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1072
6754; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6755; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x814
6756; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6757; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1056
6758; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6759; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x804
6760; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6761; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1040
6762; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6763; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7f4
6764; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6765; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1024
6766; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6767; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7e4
6768; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6769; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1008
6770; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6771; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7d4
6772; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6773; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:992
6774; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6775; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7c4
6776; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6777; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:976
6778; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6779; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7b4
6780; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6781; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:960
6782; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6783; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x7a4
6784; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6785; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:944
6786; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6787; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x794
6788; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6789; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:928
6790; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6791; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x784
6792; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6793; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:912
6794; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6795; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x774
6796; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6797; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:896
6798; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6799; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x764
6800; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6801; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:880
6802; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6803; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x754
6804; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6805; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:864
6806; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6807; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x744
6808; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6809; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:848
6810; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6811; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x734
6812; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6813; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:832
6814; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6815; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x724
6816; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6817; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:816
6818; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6819; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x714
6820; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6821; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:800
6822; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6823; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x704
6824; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6825; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:784
6826; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6827; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6f4
6828; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6829; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:768
6830; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6831; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6e4
6832; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6833; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:752
6834; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6835; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6d4
6836; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6837; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:736
6838; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6839; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6c4
6840; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6841; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:720
6842; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6843; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6b4
6844; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6845; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:704
6846; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6847; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x6a4
6848; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6849; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:688
6850; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6851; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x694
6852; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6853; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:672
6854; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6855; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x684
6856; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6857; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:656
6858; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6859; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x674
6860; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6861; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:640
6862; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6863; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x664
6864; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6865; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:624
6866; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6867; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x654
6868; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6869; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:608
6870; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6871; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x644
6872; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6873; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:592
6874; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6875; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x634
6876; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6877; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:576
6878; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6879; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x624
6880; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6881; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:560
6882; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6883; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x614
6884; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6885; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:544
6886; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6887; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x604
6888; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6889; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:528
6890; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6891; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5f4
6892; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6893; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:512
6894; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6895; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5e4
6896; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6897; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:496
6898; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6899; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5d4
6900; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6901; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:480
6902; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6903; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5c4
6904; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6905; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:464
6906; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6907; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5b4
6908; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6909; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:448
6910; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6911; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x5a4
6912; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6913; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:432
6914; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6915; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x594
6916; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6917; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:416
6918; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6919; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x584
6920; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6921; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:400
6922; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6923; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x574
6924; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6925; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:384
6926; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6927; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x564
6928; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6929; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:368
6930; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6931; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x554
6932; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6933; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:352
6934; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6935; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x544
6936; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6937; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:336
6938; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6939; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x534
6940; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6941; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:320
6942; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6943; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x524
6944; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6945; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:304
6946; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6947; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x514
6948; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6949; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:288
6950; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6951; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x504
6952; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6953; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:272
6954; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6955; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4f4
6956; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6957; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:256
6958; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6959; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4e4
6960; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6961; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:240
6962; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6963; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4d4
6964; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6965; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:224
6966; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6967; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4c4
6968; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6969; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:208
6970; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6971; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4b4
6972; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6973; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:192
6974; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6975; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x4a4
6976; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6977; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:176
6978; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6979; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x494
6980; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6981; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:160
6982; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6983; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x484
6984; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6985; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:144
6986; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6987; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x474
6988; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6989; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:128
6990; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6991; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x464
6992; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6993; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:112
6994; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6995; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x454
6996; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
6997; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:96
6998; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
6999; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x444
7000; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7001; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:80
7002; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
7003; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x434
7004; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7005; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:64
7006; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
7007; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x424
7008; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7009; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:48
7010; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
7011; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x414
7012; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7013; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:32
7014; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
7015; GFX9-FLATSCR-NEXT:    s_movk_i32 s2, 0x404
7016; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7017; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:16
7018; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
7019; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7020; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1]
7021; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3f4
7022; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7023; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x400, v4
7024; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7025; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3e4
7026; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7027; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7028; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7029; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3d4
7030; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7031; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7032; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7033; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3c4
7034; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7035; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7036; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7037; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3b4
7038; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7039; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7040; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7041; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x3a4
7042; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7043; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7044; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7045; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x394
7046; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7047; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7048; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7049; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x384
7050; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7051; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7052; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7053; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x374
7054; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7055; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7056; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7057; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x380, v4
7058; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7059; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x364
7060; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7061; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7062; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7063; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x354
7064; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7065; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7066; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7067; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x344
7068; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7069; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7070; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7071; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x334
7072; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7073; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7074; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7075; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x324
7076; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7077; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7078; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7079; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x314
7080; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7081; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7082; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7083; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x304
7084; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7085; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7086; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7087; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2f4
7088; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7089; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7090; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7091; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x300, v4
7092; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7093; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2e4
7094; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7095; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7096; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7097; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2d4
7098; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7099; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7100; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7101; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2c4
7102; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7103; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7104; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7105; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2b4
7106; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7107; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7108; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7109; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2a4
7110; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7111; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7112; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7113; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x294
7114; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7115; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7116; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7117; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x284
7118; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7119; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7120; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7121; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x274
7122; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7123; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7124; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7125; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x280, v4
7126; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7127; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x264
7128; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7129; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7130; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7131; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x254
7132; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7133; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7134; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7135; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x244
7136; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7137; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7138; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7139; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x234
7140; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7141; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7142; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7143; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x224
7144; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7145; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7146; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7147; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x214
7148; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7149; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7150; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7151; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x204
7152; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7153; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7154; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7155; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1f4
7156; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7157; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7158; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7159; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x200, v4
7160; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7161; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1e4
7162; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7163; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7164; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7165; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1d4
7166; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7167; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7168; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7169; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1c4
7170; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7171; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7172; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7173; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1b4
7174; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7175; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7176; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7177; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x1a4
7178; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7179; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7180; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7181; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x194
7182; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7183; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7184; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7185; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x184
7186; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7187; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7188; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7189; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x174
7190; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7191; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7192; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7193; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x180, v4
7194; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7195; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x164
7196; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7197; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7198; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7199; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x154
7200; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7201; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7202; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7203; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x144
7204; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7205; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7206; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7207; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x134
7208; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7209; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7210; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7211; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x124
7212; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7213; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7214; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7215; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x114
7216; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7217; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7218; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7219; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x104
7220; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7221; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7222; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7223; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xf4
7224; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7225; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7226; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7227; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x100, v4
7228; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7229; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xe4
7230; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7231; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4080
7232; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7233; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xd4
7234; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7235; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4064
7236; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7237; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xc4
7238; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7239; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4048
7240; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7241; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xb4
7242; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7243; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4032
7244; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7245; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0xa4
7246; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7247; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4016
7248; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7249; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x94
7250; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7251; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:4000
7252; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7253; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x84
7254; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7255; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3984
7256; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
7257; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x74
7258; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7259; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[7:10], off offset:3968
7260; GFX9-FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, 0x80, v4
7261; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7262; GFX9-FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v6, vcc
7263; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x64
7264; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7265; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4080
7266; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7267; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x54
7268; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7269; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4064
7270; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7271; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x44
7272; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7273; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4048
7274; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7275; GFX9-FLATSCR-NEXT:    s_mov_b32 s0, 52
7276; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7277; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4032
7278; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7279; GFX9-FLATSCR-NEXT:    s_mov_b32 s0, 36
7280; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7281; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4016
7282; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7283; GFX9-FLATSCR-NEXT:    s_mov_b32 s0, 20
7284; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7285; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:4000
7286; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7287; GFX9-FLATSCR-NEXT:    s_mov_b32 s0, 4
7288; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7289; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:3984
7290; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[2:5], off, s0 ; 16-byte Folded Reload
7291; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7292; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[2:5], off offset:3968
7293; GFX9-FLATSCR-NEXT:    s_endpgm
7294;
7295; GFX10-FLATSCR-LABEL: test:
7296; GFX10-FLATSCR:       ; %bb.0: ; %entry
7297; GFX10-FLATSCR-NEXT:    s_add_u32 s8, s8, s13
7298; GFX10-FLATSCR-NEXT:    s_addc_u32 s9, s9, 0
7299; GFX10-FLATSCR-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s8
7300; GFX10-FLATSCR-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s9
7301; GFX10-FLATSCR-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
7302; GFX10-FLATSCR-NEXT:    v_mbcnt_lo_u32_b32 v0, -1, 0
7303; GFX10-FLATSCR-NEXT:    v_mbcnt_hi_u32_b32 v0, -1, v0
7304; GFX10-FLATSCR-NEXT:    v_lshlrev_b32_e32 v5, 13, v0
7305; GFX10-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
7306; GFX10-FLATSCR-NEXT:    v_add_co_u32 v4, s4, s2, v5
7307; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e64 v22, null, s3, 0, s4
7308; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x804
7309; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x80, v4
7310; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v22, vcc_lo
7311; GFX10-FLATSCR-NEXT:    v_add_co_u32 v2, vcc_lo, 0x100, v4
7312; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v22, vcc_lo
7313; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7314; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7315; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:4 ; 16-byte Folded Spill
7316; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7317; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7318; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:20 ; 16-byte Folded Spill
7319; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7320; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7321; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:36 ; 16-byte Folded Spill
7322; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7323; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7324; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:52 ; 16-byte Folded Spill
7325; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7326; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7327; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:68 ; 16-byte Folded Spill
7328; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7329; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7330; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:84 ; 16-byte Folded Spill
7331; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7332; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7333; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:100 ; 16-byte Folded Spill
7334; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2032
7335; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7336; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:116 ; 16-byte Folded Spill
7337; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1920
7338; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7339; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:132 ; 16-byte Folded Spill
7340; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1936
7341; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7342; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:148 ; 16-byte Folded Spill
7343; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1952
7344; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7345; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:164 ; 16-byte Folded Spill
7346; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1968
7347; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7348; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:180 ; 16-byte Folded Spill
7349; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:1984
7350; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7351; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:196 ; 16-byte Folded Spill
7352; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:2000
7353; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7354; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:212 ; 16-byte Folded Spill
7355; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:2016
7356; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7357; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:228 ; 16-byte Folded Spill
7358; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[2:3], off offset:2032
7359; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7360; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], off offset:244 ; 16-byte Folded Spill
7361; GFX10-FLATSCR-NEXT:    v_add_co_u32 v6, vcc_lo, 0x180, v4
7362; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v7, vcc_lo, 0, v22, vcc_lo
7363; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:1920
7364; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7365; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:260 ; 16-byte Folded Spill
7366; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:1936
7367; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7368; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:276 ; 16-byte Folded Spill
7369; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:1952
7370; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7371; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:292 ; 16-byte Folded Spill
7372; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:1968
7373; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7374; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:308 ; 16-byte Folded Spill
7375; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:1984
7376; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7377; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:324 ; 16-byte Folded Spill
7378; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:2000
7379; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7380; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:340 ; 16-byte Folded Spill
7381; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:2016
7382; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7383; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:356 ; 16-byte Folded Spill
7384; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[8:11], v[6:7], off offset:2032
7385; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7386; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[8:11], off offset:372 ; 16-byte Folded Spill
7387; GFX10-FLATSCR-NEXT:    v_add_co_u32 v8, vcc_lo, 0x200, v4
7388; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v9, vcc_lo, 0, v22, vcc_lo
7389; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:1920
7390; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7391; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:388 ; 16-byte Folded Spill
7392; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:1936
7393; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7394; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:404 ; 16-byte Folded Spill
7395; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:1952
7396; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7397; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:420 ; 16-byte Folded Spill
7398; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:1968
7399; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7400; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:436 ; 16-byte Folded Spill
7401; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:1984
7402; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7403; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:452 ; 16-byte Folded Spill
7404; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:2000
7405; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7406; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:468 ; 16-byte Folded Spill
7407; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:2016
7408; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7409; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:484 ; 16-byte Folded Spill
7410; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[10:13], v[8:9], off offset:2032
7411; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7412; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[10:13], off offset:500 ; 16-byte Folded Spill
7413; GFX10-FLATSCR-NEXT:    v_add_co_u32 v10, vcc_lo, 0x280, v4
7414; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v11, vcc_lo, 0, v22, vcc_lo
7415; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:1920
7416; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7417; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:516 ; 16-byte Folded Spill
7418; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:1936
7419; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7420; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:532 ; 16-byte Folded Spill
7421; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:1952
7422; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7423; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:548 ; 16-byte Folded Spill
7424; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:1968
7425; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7426; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:564 ; 16-byte Folded Spill
7427; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:1984
7428; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7429; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:580 ; 16-byte Folded Spill
7430; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:2000
7431; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7432; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:596 ; 16-byte Folded Spill
7433; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:2016
7434; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7435; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:612 ; 16-byte Folded Spill
7436; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[12:15], v[10:11], off offset:2032
7437; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7438; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[12:15], off offset:628 ; 16-byte Folded Spill
7439; GFX10-FLATSCR-NEXT:    v_add_co_u32 v12, vcc_lo, 0x300, v4
7440; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v13, vcc_lo, 0, v22, vcc_lo
7441; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:1920
7442; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7443; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:644 ; 16-byte Folded Spill
7444; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:1936
7445; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7446; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:660 ; 16-byte Folded Spill
7447; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:1952
7448; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7449; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:676 ; 16-byte Folded Spill
7450; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:1968
7451; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7452; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:692 ; 16-byte Folded Spill
7453; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:1984
7454; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7455; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:708 ; 16-byte Folded Spill
7456; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:2000
7457; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7458; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:724 ; 16-byte Folded Spill
7459; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:2016
7460; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7461; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:740 ; 16-byte Folded Spill
7462; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[14:17], v[12:13], off offset:2032
7463; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7464; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[14:17], off offset:756 ; 16-byte Folded Spill
7465; GFX10-FLATSCR-NEXT:    v_add_co_u32 v14, vcc_lo, 0x380, v4
7466; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v15, vcc_lo, 0, v22, vcc_lo
7467; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:1920
7468; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7469; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:772 ; 16-byte Folded Spill
7470; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:1936
7471; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7472; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:788 ; 16-byte Folded Spill
7473; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:1952
7474; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7475; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:804 ; 16-byte Folded Spill
7476; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:1968
7477; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7478; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:820 ; 16-byte Folded Spill
7479; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:1984
7480; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7481; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:836 ; 16-byte Folded Spill
7482; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:2000
7483; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7484; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:852 ; 16-byte Folded Spill
7485; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:2016
7486; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7487; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:868 ; 16-byte Folded Spill
7488; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[16:19], v[14:15], off offset:2032
7489; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7490; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[16:19], off offset:884 ; 16-byte Folded Spill
7491; GFX10-FLATSCR-NEXT:    v_add_co_u32 v16, vcc_lo, 0x400, v4
7492; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v17, vcc_lo, 0, v22, vcc_lo
7493; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:1920
7494; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7495; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:900 ; 16-byte Folded Spill
7496; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:1936
7497; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7498; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:916 ; 16-byte Folded Spill
7499; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:1952
7500; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7501; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:932 ; 16-byte Folded Spill
7502; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:1968
7503; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7504; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:948 ; 16-byte Folded Spill
7505; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:1984
7506; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7507; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:964 ; 16-byte Folded Spill
7508; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:2000
7509; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7510; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:980 ; 16-byte Folded Spill
7511; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:2016
7512; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7513; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:996 ; 16-byte Folded Spill
7514; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[18:21], v[16:17], off offset:2032
7515; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7516; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[18:21], off offset:1012 ; 16-byte Folded Spill
7517; GFX10-FLATSCR-NEXT:    v_add_co_u32 v18, vcc_lo, 0x480, v4
7518; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v19, vcc_lo, 0, v22, vcc_lo
7519; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x500, v4
7520; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7521; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:1920
7522; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7523; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1028 ; 16-byte Folded Spill
7524; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:1936
7525; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7526; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1044 ; 16-byte Folded Spill
7527; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:1952
7528; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7529; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1060 ; 16-byte Folded Spill
7530; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:1968
7531; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7532; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1076 ; 16-byte Folded Spill
7533; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:1984
7534; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7535; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1092 ; 16-byte Folded Spill
7536; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:2000
7537; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7538; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1108 ; 16-byte Folded Spill
7539; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:2016
7540; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7541; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1124 ; 16-byte Folded Spill
7542; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[18:19], off offset:2032
7543; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7544; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1140 ; 16-byte Folded Spill
7545; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1920
7546; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7547; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1156 ; 16-byte Folded Spill
7548; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1936
7549; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7550; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1172 ; 16-byte Folded Spill
7551; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1952
7552; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7553; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1188 ; 16-byte Folded Spill
7554; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1968
7555; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7556; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1204 ; 16-byte Folded Spill
7557; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1984
7558; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7559; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1220 ; 16-byte Folded Spill
7560; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2000
7561; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7562; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1236 ; 16-byte Folded Spill
7563; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2016
7564; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7565; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1252 ; 16-byte Folded Spill
7566; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2032
7567; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x580, v4
7568; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7569; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7570; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1268 ; 16-byte Folded Spill
7571; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1920
7572; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7573; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1284 ; 16-byte Folded Spill
7574; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1936
7575; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7576; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1300 ; 16-byte Folded Spill
7577; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1952
7578; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7579; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1316 ; 16-byte Folded Spill
7580; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1968
7581; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7582; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1332 ; 16-byte Folded Spill
7583; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1984
7584; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7585; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1348 ; 16-byte Folded Spill
7586; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2000
7587; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7588; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1364 ; 16-byte Folded Spill
7589; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2016
7590; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7591; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1380 ; 16-byte Folded Spill
7592; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2032
7593; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x600, v4
7594; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7595; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7596; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1396 ; 16-byte Folded Spill
7597; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1920
7598; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7599; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1412 ; 16-byte Folded Spill
7600; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1936
7601; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7602; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1428 ; 16-byte Folded Spill
7603; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1952
7604; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7605; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1444 ; 16-byte Folded Spill
7606; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1968
7607; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7608; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1460 ; 16-byte Folded Spill
7609; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1984
7610; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7611; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1476 ; 16-byte Folded Spill
7612; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2000
7613; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7614; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1492 ; 16-byte Folded Spill
7615; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2016
7616; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7617; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1508 ; 16-byte Folded Spill
7618; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2032
7619; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x680, v4
7620; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7621; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7622; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1524 ; 16-byte Folded Spill
7623; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1920
7624; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7625; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1540 ; 16-byte Folded Spill
7626; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1936
7627; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7628; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1556 ; 16-byte Folded Spill
7629; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1952
7630; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7631; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1572 ; 16-byte Folded Spill
7632; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1968
7633; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7634; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1588 ; 16-byte Folded Spill
7635; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1984
7636; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7637; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1604 ; 16-byte Folded Spill
7638; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2000
7639; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7640; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1620 ; 16-byte Folded Spill
7641; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2016
7642; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7643; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1636 ; 16-byte Folded Spill
7644; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2032
7645; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x700, v4
7646; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7647; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7648; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1652 ; 16-byte Folded Spill
7649; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1920
7650; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7651; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1668 ; 16-byte Folded Spill
7652; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1936
7653; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7654; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1684 ; 16-byte Folded Spill
7655; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1952
7656; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7657; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1700 ; 16-byte Folded Spill
7658; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1968
7659; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7660; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1716 ; 16-byte Folded Spill
7661; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:1984
7662; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7663; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1732 ; 16-byte Folded Spill
7664; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2000
7665; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7666; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1748 ; 16-byte Folded Spill
7667; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2016
7668; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7669; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1764 ; 16-byte Folded Spill
7670; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v[20:21], off offset:2032
7671; GFX10-FLATSCR-NEXT:    v_add_co_u32 v20, vcc_lo, 0x780, v4
7672; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v21, vcc_lo, 0, v22, vcc_lo
7673; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v0
7674; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
7675; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7676; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[23:26], off offset:1780 ; 16-byte Folded Spill
7677; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:1920
7678; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7679; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1796 ; 16-byte Folded Spill
7680; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:1936
7681; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7682; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1812 ; 16-byte Folded Spill
7683; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:1952
7684; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7685; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1828 ; 16-byte Folded Spill
7686; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:1968
7687; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7688; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1844 ; 16-byte Folded Spill
7689; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:1984
7690; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7691; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1860 ; 16-byte Folded Spill
7692; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:2000
7693; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7694; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1876 ; 16-byte Folded Spill
7695; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[22:25], v[20:21], off offset:2016
7696; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7697; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[22:25], off offset:1892 ; 16-byte Folded Spill
7698; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[20:21], off offset:2032
7699; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7700; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1908 ; 16-byte Folded Spill
7701; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1920
7702; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7703; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1924 ; 16-byte Folded Spill
7704; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1936
7705; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7706; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1940 ; 16-byte Folded Spill
7707; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1952
7708; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7709; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1956 ; 16-byte Folded Spill
7710; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1968
7711; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7712; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1972 ; 16-byte Folded Spill
7713; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1984
7714; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7715; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:1988 ; 16-byte Folded Spill
7716; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2000
7717; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7718; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:2004 ; 16-byte Folded Spill
7719; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2016
7720; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7721; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:2020 ; 16-byte Folded Spill
7722; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2032
7723; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v2
7724; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v3, vcc_lo
7725; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7726; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], off offset:2036 ; 16-byte Folded Spill
7727; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1920
7728; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7729; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7730; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1936
7731; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x814
7732; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7733; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7734; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1952
7735; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x824
7736; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7737; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7738; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1968
7739; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x834
7740; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7741; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7742; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1984
7743; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x844
7744; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7745; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7746; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2000
7747; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x854
7748; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7749; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7750; GFX10-FLATSCR-NEXT:    s_clause 0x1
7751; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2016
7752; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7753; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x864
7754; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7755; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7756; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x874
7757; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7758; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7759; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v6
7760; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v7, vcc_lo
7761; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x884
7762; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1920
7763; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7764; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7765; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1936
7766; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x894
7767; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7768; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7769; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1952
7770; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8a4
7771; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7772; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7773; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1968
7774; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8b4
7775; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7776; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7777; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:1984
7778; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8c4
7779; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7780; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7781; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2000
7782; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8d4
7783; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7784; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7785; GFX10-FLATSCR-NEXT:    s_clause 0x1
7786; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[20:23], v[0:1], off offset:2016
7787; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7788; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8e4
7789; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7790; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[20:23], s4 ; 16-byte Folded Spill
7791; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x8f4
7792; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7793; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7794; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v8
7795; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v9, vcc_lo
7796; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x904
7797; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7798; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7799; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7800; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7801; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x914
7802; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7803; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7804; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7805; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x924
7806; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7807; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7808; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7809; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x934
7810; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7811; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7812; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7813; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x944
7814; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7815; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7816; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7817; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x954
7818; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7819; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7820; GFX10-FLATSCR-NEXT:    s_clause 0x1
7821; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7822; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7823; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x964
7824; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7825; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7826; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x974
7827; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7828; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7829; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v10
7830; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v11, vcc_lo
7831; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x984
7832; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7833; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7834; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7835; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7836; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x994
7837; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7838; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7839; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7840; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9a4
7841; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7842; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7843; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7844; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9b4
7845; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7846; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7847; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7848; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9c4
7849; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7850; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7851; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7852; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9d4
7853; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7854; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7855; GFX10-FLATSCR-NEXT:    s_clause 0x1
7856; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7857; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7858; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9e4
7859; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7860; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7861; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x9f4
7862; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7863; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7864; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v12
7865; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v13, vcc_lo
7866; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa04
7867; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7868; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7869; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7870; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7871; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa14
7872; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7873; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7874; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7875; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa24
7876; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7877; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7878; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7879; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa34
7880; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7881; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7882; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7883; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa44
7884; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7885; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7886; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7887; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa54
7888; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7889; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7890; GFX10-FLATSCR-NEXT:    s_clause 0x1
7891; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7892; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7893; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa64
7894; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7895; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7896; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa74
7897; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7898; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7899; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v14
7900; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v15, vcc_lo
7901; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa84
7902; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7903; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7904; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7905; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7906; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xa94
7907; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7908; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7909; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7910; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xaa4
7911; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7912; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7913; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7914; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xab4
7915; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7916; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7917; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7918; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xac4
7919; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7920; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7921; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7922; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xad4
7923; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7924; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7925; GFX10-FLATSCR-NEXT:    s_clause 0x1
7926; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7927; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7928; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xae4
7929; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7930; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7931; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xaf4
7932; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7933; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7934; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v16
7935; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v17, vcc_lo
7936; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb04
7937; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7938; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7939; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7940; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7941; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb14
7942; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7943; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7944; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7945; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb24
7946; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7947; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7948; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7949; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb34
7950; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7951; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7952; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7953; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb44
7954; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7955; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7956; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7957; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb54
7958; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7959; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7960; GFX10-FLATSCR-NEXT:    s_clause 0x1
7961; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7962; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7963; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb64
7964; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
7965; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7966; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb74
7967; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7968; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
7969; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x780, v18
7970; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v19, vcc_lo
7971; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb84
7972; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1920
7973; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7974; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7975; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1936
7976; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xb94
7977; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7978; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7979; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1952
7980; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xba4
7981; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7982; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7983; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1968
7984; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xbb4
7985; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7986; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7987; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:1984
7988; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xbc4
7989; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7990; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7991; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2000
7992; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xbd4
7993; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
7994; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
7995; GFX10-FLATSCR-NEXT:    s_clause 0x1
7996; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v[0:1], off offset:2016
7997; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v[0:1], off offset:2032
7998; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xbe4
7999; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
8000; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s4 ; 16-byte Folded Spill
8001; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xbf4
8002; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8003; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8004; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3]
8005; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc04
8006; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8007; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8008; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:16
8009; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc14
8010; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8011; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8012; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:32
8013; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc24
8014; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8015; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8016; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:48
8017; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc34
8018; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8019; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8020; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:64
8021; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc44
8022; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8023; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8024; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:80
8025; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc54
8026; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8027; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8028; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:96
8029; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc64
8030; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8031; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8032; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:112
8033; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc74
8034; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8035; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8036; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:128
8037; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc84
8038; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8039; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8040; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:144
8041; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xc94
8042; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8043; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8044; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:160
8045; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xca4
8046; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8047; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8048; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:176
8049; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xcb4
8050; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8051; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8052; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:192
8053; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xcc4
8054; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8055; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8056; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:208
8057; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xcd4
8058; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8059; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8060; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:224
8061; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xce4
8062; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8063; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8064; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:240
8065; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xcf4
8066; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8067; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8068; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:256
8069; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd04
8070; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8071; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8072; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:272
8073; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd14
8074; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8075; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8076; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:288
8077; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd24
8078; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8079; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8080; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:304
8081; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd34
8082; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8083; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8084; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:320
8085; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd44
8086; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8087; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8088; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:336
8089; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd54
8090; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8091; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8092; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:352
8093; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd64
8094; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8095; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8096; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:368
8097; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd74
8098; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8099; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8100; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:384
8101; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd84
8102; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8103; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8104; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:400
8105; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xd94
8106; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8107; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8108; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:416
8109; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xda4
8110; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8111; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8112; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:432
8113; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xdb4
8114; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8115; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8116; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:448
8117; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xdc4
8118; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8119; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8120; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:464
8121; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xdd4
8122; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8123; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8124; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:480
8125; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xde4
8126; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8127; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8128; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:496
8129; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xdf4
8130; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8131; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8132; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:512
8133; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe04
8134; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8135; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8136; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:528
8137; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe14
8138; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8139; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8140; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:544
8141; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe24
8142; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8143; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8144; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:560
8145; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe34
8146; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8147; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8148; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:576
8149; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe44
8150; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8151; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8152; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:592
8153; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe54
8154; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8155; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8156; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:608
8157; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe64
8158; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8159; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8160; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:624
8161; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe74
8162; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8163; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8164; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:640
8165; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe84
8166; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8167; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8168; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:656
8169; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xe94
8170; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8171; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8172; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:672
8173; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xea4
8174; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8175; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8176; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:688
8177; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xeb4
8178; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8179; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8180; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:704
8181; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xec4
8182; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8183; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8184; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:720
8185; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xed4
8186; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8187; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8188; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:736
8189; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xee4
8190; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8191; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8192; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:752
8193; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xef4
8194; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8195; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8196; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:768
8197; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf04
8198; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8199; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8200; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:784
8201; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf14
8202; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8203; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8204; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:800
8205; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf24
8206; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8207; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8208; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:816
8209; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf34
8210; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8211; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8212; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:832
8213; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf44
8214; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8215; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8216; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:848
8217; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf54
8218; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8219; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8220; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:864
8221; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf64
8222; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8223; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8224; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:880
8225; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf74
8226; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8227; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8228; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:896
8229; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf84
8230; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8231; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8232; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:912
8233; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xf94
8234; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8235; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8236; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:928
8237; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xfa4
8238; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8239; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8240; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:944
8241; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xfb4
8242; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8243; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8244; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:960
8245; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xfc4
8246; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8247; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8248; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:976
8249; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xfd4
8250; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8251; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8252; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:992
8253; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xfe4
8254; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8255; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8256; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1008
8257; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0xff4
8258; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8259; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8260; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1024
8261; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1004
8262; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8263; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8264; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1040
8265; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1014
8266; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8267; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8268; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1056
8269; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1024
8270; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8271; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8272; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1072
8273; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1034
8274; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8275; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8276; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1088
8277; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1044
8278; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8279; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8280; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1104
8281; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1054
8282; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8283; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8284; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1120
8285; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1064
8286; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8287; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8288; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1136
8289; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1074
8290; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8291; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8292; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1152
8293; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1084
8294; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8295; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8296; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1168
8297; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1094
8298; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8299; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8300; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1184
8301; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10a4
8302; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8303; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8304; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1200
8305; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10b4
8306; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8307; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8308; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1216
8309; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10c4
8310; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8311; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8312; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1232
8313; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10d4
8314; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8315; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8316; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1248
8317; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10e4
8318; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8319; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8320; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1264
8321; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x10f4
8322; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8323; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8324; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1280
8325; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1104
8326; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8327; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8328; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1296
8329; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1114
8330; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8331; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8332; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1312
8333; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1124
8334; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8335; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8336; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1328
8337; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1134
8338; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8339; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8340; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1344
8341; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1144
8342; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8343; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8344; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1360
8345; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1154
8346; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8347; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8348; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1376
8349; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1164
8350; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8351; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8352; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1392
8353; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1174
8354; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8355; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8356; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1408
8357; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1184
8358; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8359; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8360; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1424
8361; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1194
8362; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8363; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8364; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1440
8365; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11a4
8366; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8367; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8368; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1456
8369; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11b4
8370; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8371; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8372; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1472
8373; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11c4
8374; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8375; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8376; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1488
8377; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11d4
8378; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8379; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8380; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1504
8381; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11e4
8382; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8383; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8384; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1520
8385; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x11f4
8386; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8387; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8388; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1536
8389; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1204
8390; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8391; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8392; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1552
8393; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1214
8394; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8395; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8396; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1568
8397; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1224
8398; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8399; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8400; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1584
8401; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1234
8402; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8403; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8404; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1600
8405; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1244
8406; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8407; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8408; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1616
8409; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1254
8410; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8411; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8412; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1632
8413; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1264
8414; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8415; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8416; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1648
8417; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1274
8418; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8419; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8420; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1664
8421; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1284
8422; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8423; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8424; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1680
8425; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1294
8426; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8427; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8428; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1696
8429; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12a4
8430; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8431; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8432; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1712
8433; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12b4
8434; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8435; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8436; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1728
8437; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12c4
8438; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8439; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8440; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1744
8441; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12d4
8442; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8443; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8444; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1760
8445; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12e4
8446; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8447; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8448; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1776
8449; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x12f4
8450; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8451; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8452; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1792
8453; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1304
8454; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8455; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8456; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1808
8457; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1314
8458; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8459; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8460; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1824
8461; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1324
8462; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8463; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8464; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1840
8465; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1334
8466; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8467; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8468; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1856
8469; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1344
8470; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8471; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8472; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1872
8473; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1354
8474; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8475; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8476; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1888
8477; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1364
8478; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8479; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8480; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1904
8481; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1374
8482; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8483; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8484; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1920
8485; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1384
8486; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8487; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8488; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1936
8489; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x1394
8490; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8491; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8492; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1952
8493; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x13a4
8494; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8495; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8496; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1968
8497; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x13b4
8498; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8499; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8500; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:1984
8501; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x13c4
8502; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8503; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8504; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2000
8505; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x13d4
8506; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8507; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8508; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2016
8509; GFX10-FLATSCR-NEXT:    s_movk_i32 s4, 0x13e4
8510; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8511; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s4 ; 16-byte Folded Spill
8512; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[2:3] offset:2032
8513; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8514; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8515; GFX10-FLATSCR-NEXT:    v_add_co_u32 v4, s2, s0, v5
8516; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e64 v6, null, s1, 0, s2
8517; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x13e4
8518; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8519; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8520; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8521; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8522; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8523; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8524; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8525; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8526; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8527; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8528; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
8529; GFX10-FLATSCR-NEXT:    ;;#ASMEND
8530; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8531; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2032
8532; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8533; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x13d4
8534; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8535; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2016
8536; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8537; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x13c4
8538; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8539; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:2000
8540; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8541; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x13b4
8542; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8543; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1984
8544; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8545; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x13a4
8546; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8547; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1968
8548; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8549; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1394
8550; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8551; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1952
8552; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8553; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1384
8554; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8555; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1936
8556; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8557; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1374
8558; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8559; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1920
8560; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8561; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1364
8562; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8563; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1904
8564; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8565; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1354
8566; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8567; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1888
8568; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8569; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1344
8570; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8571; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1872
8572; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8573; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1334
8574; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8575; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1856
8576; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8577; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1324
8578; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8579; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1840
8580; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8581; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1314
8582; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8583; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1824
8584; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8585; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1304
8586; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8587; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1808
8588; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8589; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12f4
8590; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8591; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1792
8592; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8593; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12e4
8594; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8595; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1776
8596; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8597; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12d4
8598; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8599; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1760
8600; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8601; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12c4
8602; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8603; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1744
8604; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8605; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12b4
8606; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8607; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1728
8608; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8609; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x12a4
8610; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8611; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1712
8612; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8613; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1294
8614; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8615; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1696
8616; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8617; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1284
8618; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8619; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1680
8620; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8621; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1274
8622; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8623; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1664
8624; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8625; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1264
8626; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8627; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1648
8628; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8629; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1254
8630; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8631; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1632
8632; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8633; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1244
8634; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8635; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1616
8636; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8637; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1234
8638; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8639; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1600
8640; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8641; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1224
8642; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8643; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1584
8644; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8645; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1214
8646; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8647; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1568
8648; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8649; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1204
8650; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8651; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1552
8652; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8653; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11f4
8654; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8655; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1536
8656; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8657; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11e4
8658; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8659; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1520
8660; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8661; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11d4
8662; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8663; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1504
8664; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8665; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11c4
8666; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8667; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1488
8668; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8669; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11b4
8670; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8671; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1472
8672; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8673; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x11a4
8674; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8675; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1456
8676; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8677; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1194
8678; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8679; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1440
8680; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8681; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1184
8682; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8683; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1424
8684; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8685; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1174
8686; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8687; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1408
8688; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8689; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1164
8690; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8691; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1392
8692; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8693; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1154
8694; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8695; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1376
8696; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8697; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1144
8698; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8699; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1360
8700; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8701; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1134
8702; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8703; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1344
8704; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8705; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1124
8706; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8707; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1328
8708; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8709; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1114
8710; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8711; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1312
8712; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8713; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1104
8714; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8715; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1296
8716; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8717; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10f4
8718; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8719; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1280
8720; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8721; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10e4
8722; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8723; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1264
8724; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8725; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10d4
8726; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8727; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1248
8728; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8729; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10c4
8730; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8731; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1232
8732; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8733; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10b4
8734; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8735; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1216
8736; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8737; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x10a4
8738; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8739; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1200
8740; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8741; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1094
8742; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8743; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1184
8744; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8745; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1084
8746; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8747; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1168
8748; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8749; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1074
8750; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8751; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1152
8752; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8753; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1064
8754; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8755; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1136
8756; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8757; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1054
8758; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8759; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1120
8760; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8761; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1044
8762; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8763; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1104
8764; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8765; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1034
8766; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8767; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1088
8768; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8769; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1024
8770; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8771; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1072
8772; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8773; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1014
8774; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8775; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1056
8776; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8777; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0x1004
8778; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8779; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1040
8780; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8781; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xff4
8782; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8783; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1024
8784; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8785; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xfe4
8786; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8787; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:1008
8788; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8789; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xfd4
8790; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8791; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:992
8792; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8793; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xfc4
8794; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8795; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:976
8796; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8797; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xfb4
8798; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8799; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:960
8800; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8801; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xfa4
8802; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8803; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:944
8804; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8805; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf94
8806; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8807; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:928
8808; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8809; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf84
8810; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8811; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:912
8812; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8813; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf74
8814; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8815; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:896
8816; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8817; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf64
8818; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8819; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:880
8820; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8821; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf54
8822; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8823; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:864
8824; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8825; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf44
8826; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8827; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:848
8828; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8829; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf34
8830; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8831; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:832
8832; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8833; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf24
8834; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8835; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:816
8836; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8837; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf14
8838; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8839; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:800
8840; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8841; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xf04
8842; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8843; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:784
8844; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8845; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xef4
8846; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8847; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:768
8848; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8849; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xee4
8850; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8851; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:752
8852; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8853; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xed4
8854; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8855; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:736
8856; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8857; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xec4
8858; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8859; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:720
8860; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8861; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xeb4
8862; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8863; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:704
8864; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8865; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xea4
8866; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8867; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:688
8868; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8869; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe94
8870; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8871; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:672
8872; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8873; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe84
8874; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8875; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:656
8876; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8877; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe74
8878; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8879; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:640
8880; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8881; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe64
8882; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8883; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:624
8884; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8885; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe54
8886; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8887; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:608
8888; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8889; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe44
8890; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8891; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:592
8892; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8893; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe34
8894; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8895; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:576
8896; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8897; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe24
8898; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8899; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:560
8900; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8901; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe14
8902; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8903; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:544
8904; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8905; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xe04
8906; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8907; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:528
8908; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8909; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xdf4
8910; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8911; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:512
8912; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8913; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xde4
8914; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8915; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:496
8916; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8917; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xdd4
8918; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8919; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:480
8920; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8921; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xdc4
8922; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8923; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:464
8924; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8925; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xdb4
8926; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8927; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:448
8928; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8929; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xda4
8930; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8931; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:432
8932; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8933; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd94
8934; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8935; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:416
8936; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8937; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd84
8938; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8939; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:400
8940; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8941; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd74
8942; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8943; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:384
8944; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8945; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd64
8946; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8947; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:368
8948; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8949; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd54
8950; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8951; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:352
8952; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8953; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd44
8954; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8955; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:336
8956; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8957; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd34
8958; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8959; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:320
8960; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8961; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd24
8962; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8963; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:304
8964; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8965; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd14
8966; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8967; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:288
8968; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8969; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xd04
8970; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8971; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:272
8972; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8973; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xcf4
8974; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8975; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:256
8976; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8977; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xce4
8978; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8979; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:240
8980; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8981; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xcd4
8982; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8983; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:224
8984; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8985; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xcc4
8986; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8987; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:208
8988; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8989; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xcb4
8990; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8991; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:192
8992; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8993; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xca4
8994; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8995; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:176
8996; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
8997; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc94
8998; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
8999; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:160
9000; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9001; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc84
9002; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9003; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:144
9004; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9005; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc74
9006; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9007; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:128
9008; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9009; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc64
9010; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9011; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:112
9012; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9013; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc54
9014; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9015; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:96
9016; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9017; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc44
9018; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9019; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:80
9020; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9021; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc34
9022; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9023; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:64
9024; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9025; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc24
9026; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9027; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:48
9028; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9029; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc14
9030; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9031; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:32
9032; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9033; GFX10-FLATSCR-NEXT:    s_movk_i32 s2, 0xc04
9034; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9035; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1] offset:16
9036; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s2 ; 16-byte Folded Reload
9037; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9038; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1]
9039; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xbf4
9040; GFX10-FLATSCR-NEXT:    v_add_co_u32 v0, vcc_lo, 0x480, v4
9041; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9042; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v6, vcc_lo
9043; GFX10-FLATSCR-NEXT:    v_add_co_u32 v2, vcc_lo, 0x780, v0
9044; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xbe4
9045; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v1, vcc_lo
9046; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9047; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:2032
9048; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9049; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xbd4
9050; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9051; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:2016
9052; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9053; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xbc4
9054; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9055; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:2000
9056; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9057; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xbb4
9058; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9059; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:1984
9060; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9061; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xba4
9062; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9063; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:1968
9064; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9065; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb94
9066; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9067; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:1952
9068; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9069; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb84
9070; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9071; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:1936
9072; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
9073; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb74
9074; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9075; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[7:10], off offset:1920
9076; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9077; GFX10-FLATSCR-NEXT:    v_add_co_u32 v2, vcc_lo, 0x400, v4
9078; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v6, vcc_lo
9079; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb64
9080; GFX10-FLATSCR-NEXT:    v_add_co_u32 v7, vcc_lo, 0x780, v2
9081; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v8, vcc_lo, 0, v3, vcc_lo
9082; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9083; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:2032
9084; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9085; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb54
9086; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9087; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:2016
9088; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9089; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb44
9090; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9091; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:2000
9092; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9093; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb34
9094; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9095; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:1984
9096; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9097; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb24
9098; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9099; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:1968
9100; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9101; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb14
9102; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9103; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:1952
9104; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9105; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xb04
9106; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9107; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:1936
9108; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[9:12], off, s0 ; 16-byte Folded Reload
9109; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xaf4
9110; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9111; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[9:12], off offset:1920
9112; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9113; GFX10-FLATSCR-NEXT:    v_add_co_u32 v7, vcc_lo, 0x380, v4
9114; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v8, vcc_lo, 0, v6, vcc_lo
9115; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xae4
9116; GFX10-FLATSCR-NEXT:    v_add_co_u32 v9, vcc_lo, 0x780, v7
9117; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, 0, v8, vcc_lo
9118; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9119; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:2032
9120; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9121; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xad4
9122; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9123; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:2016
9124; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9125; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xac4
9126; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9127; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:2000
9128; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9129; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xab4
9130; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9131; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:1984
9132; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9133; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xaa4
9134; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9135; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:1968
9136; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9137; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa94
9138; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9139; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:1952
9140; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9141; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa84
9142; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9143; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:1936
9144; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[11:14], off, s0 ; 16-byte Folded Reload
9145; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa74
9146; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9147; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[11:14], off offset:1920
9148; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9149; GFX10-FLATSCR-NEXT:    v_add_co_u32 v9, vcc_lo, 0x300, v4
9150; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v10, vcc_lo, 0, v6, vcc_lo
9151; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa64
9152; GFX10-FLATSCR-NEXT:    v_add_co_u32 v11, vcc_lo, 0x780, v9
9153; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v12, vcc_lo, 0, v10, vcc_lo
9154; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9155; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:2032
9156; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9157; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa54
9158; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9159; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:2016
9160; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9161; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa44
9162; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9163; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:2000
9164; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9165; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa34
9166; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9167; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:1984
9168; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9169; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa24
9170; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9171; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:1968
9172; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9173; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa14
9174; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9175; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:1952
9176; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9177; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0xa04
9178; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9179; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:1936
9180; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[13:16], off, s0 ; 16-byte Folded Reload
9181; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9f4
9182; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9183; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[13:16], off offset:1920
9184; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9185; GFX10-FLATSCR-NEXT:    v_add_co_u32 v11, vcc_lo, 0x280, v4
9186; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v12, vcc_lo, 0, v6, vcc_lo
9187; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9e4
9188; GFX10-FLATSCR-NEXT:    v_add_co_u32 v13, vcc_lo, 0x780, v11
9189; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v14, vcc_lo, 0, v12, vcc_lo
9190; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9191; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:2032
9192; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9193; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9d4
9194; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9195; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:2016
9196; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9197; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9c4
9198; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9199; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:2000
9200; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9201; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9b4
9202; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9203; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:1984
9204; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9205; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x9a4
9206; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9207; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:1968
9208; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9209; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x994
9210; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9211; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:1952
9212; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9213; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x984
9214; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9215; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:1936
9216; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
9217; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x974
9218; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9219; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[15:18], off offset:1920
9220; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9221; GFX10-FLATSCR-NEXT:    v_add_co_u32 v13, vcc_lo, 0x200, v4
9222; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v14, vcc_lo, 0, v6, vcc_lo
9223; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x964
9224; GFX10-FLATSCR-NEXT:    v_add_co_u32 v15, vcc_lo, 0x780, v13
9225; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v16, vcc_lo, 0, v14, vcc_lo
9226; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9227; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:2032
9228; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9229; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x954
9230; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9231; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:2016
9232; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9233; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x944
9234; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9235; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:2000
9236; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9237; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x934
9238; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9239; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:1984
9240; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9241; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x924
9242; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9243; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:1968
9244; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9245; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x914
9246; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9247; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:1952
9248; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9249; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x904
9250; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9251; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:1936
9252; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[17:20], off, s0 ; 16-byte Folded Reload
9253; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8f4
9254; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9255; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[17:20], off offset:1920
9256; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9257; GFX10-FLATSCR-NEXT:    v_add_co_u32 v15, vcc_lo, 0x180, v4
9258; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v16, vcc_lo, 0, v6, vcc_lo
9259; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8e4
9260; GFX10-FLATSCR-NEXT:    v_add_co_u32 v17, vcc_lo, 0x780, v15
9261; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v18, vcc_lo, 0, v16, vcc_lo
9262; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9263; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:2032
9264; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9265; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8d4
9266; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9267; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:2016
9268; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9269; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8c4
9270; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9271; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:2000
9272; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9273; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8b4
9274; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9275; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:1984
9276; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9277; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x8a4
9278; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9279; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:1968
9280; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9281; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x894
9282; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9283; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:1952
9284; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9285; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x884
9286; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9287; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:1936
9288; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
9289; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x874
9290; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9291; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[19:22], off offset:1920
9292; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9293; GFX10-FLATSCR-NEXT:    v_add_co_u32 v17, vcc_lo, 0x100, v4
9294; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v18, vcc_lo, 0, v6, vcc_lo
9295; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x864
9296; GFX10-FLATSCR-NEXT:    v_add_co_u32 v19, vcc_lo, 0x780, v17
9297; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v20, vcc_lo, 0, v18, vcc_lo
9298; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9299; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:2032
9300; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9301; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x854
9302; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9303; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:2016
9304; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9305; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x844
9306; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9307; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:2000
9308; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9309; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x834
9310; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9311; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:1984
9312; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9313; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x824
9314; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9315; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:1968
9316; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9317; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x814
9318; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9319; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:1952
9320; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9321; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x804
9322; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9323; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:1936
9324; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, s0 ; 16-byte Folded Reload
9325; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9326; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[21:24], off offset:1920
9327; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:2036 ; 16-byte Folded Reload
9328; GFX10-FLATSCR-NEXT:    v_add_co_u32 v19, vcc_lo, 0x80, v4
9329; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v20, vcc_lo, 0, v6, vcc_lo
9330; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x780, v19
9331; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v20, vcc_lo
9332; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9333; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9334; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:2020 ; 16-byte Folded Reload
9335; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9336; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9337; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:2004 ; 16-byte Folded Reload
9338; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9339; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9340; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1988 ; 16-byte Folded Reload
9341; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9342; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9343; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1972 ; 16-byte Folded Reload
9344; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9345; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9346; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1956 ; 16-byte Folded Reload
9347; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9348; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9349; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1940 ; 16-byte Folded Reload
9350; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9351; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9352; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1924 ; 16-byte Folded Reload
9353; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9354; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9355; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1908 ; 16-byte Folded Reload
9356; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x780, v4
9357; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v6, vcc_lo
9358; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9359; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9360; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1892 ; 16-byte Folded Reload
9361; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9362; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9363; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1876 ; 16-byte Folded Reload
9364; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9365; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9366; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1860 ; 16-byte Folded Reload
9367; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9368; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9369; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1844 ; 16-byte Folded Reload
9370; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9371; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9372; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1828 ; 16-byte Folded Reload
9373; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9374; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9375; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1812 ; 16-byte Folded Reload
9376; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9377; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9378; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1796 ; 16-byte Folded Reload
9379; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9380; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9381; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1780 ; 16-byte Folded Reload
9382; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x700, v4
9383; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v6, vcc_lo
9384; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9385; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9386; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1764 ; 16-byte Folded Reload
9387; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9388; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9389; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1748 ; 16-byte Folded Reload
9390; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9391; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9392; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1732 ; 16-byte Folded Reload
9393; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9394; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9395; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1716 ; 16-byte Folded Reload
9396; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9397; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9398; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1700 ; 16-byte Folded Reload
9399; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9400; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9401; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1684 ; 16-byte Folded Reload
9402; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9403; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9404; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1668 ; 16-byte Folded Reload
9405; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9406; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9407; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1652 ; 16-byte Folded Reload
9408; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x680, v4
9409; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v6, vcc_lo
9410; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9411; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9412; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1636 ; 16-byte Folded Reload
9413; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9414; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9415; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1620 ; 16-byte Folded Reload
9416; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9417; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9418; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1604 ; 16-byte Folded Reload
9419; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9420; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9421; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1588 ; 16-byte Folded Reload
9422; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9423; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9424; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1572 ; 16-byte Folded Reload
9425; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9426; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9427; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1556 ; 16-byte Folded Reload
9428; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9429; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9430; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1540 ; 16-byte Folded Reload
9431; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9432; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9433; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1524 ; 16-byte Folded Reload
9434; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x600, v4
9435; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v6, vcc_lo
9436; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9437; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9438; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1508 ; 16-byte Folded Reload
9439; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9440; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9441; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1492 ; 16-byte Folded Reload
9442; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9443; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9444; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1476 ; 16-byte Folded Reload
9445; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9446; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9447; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1460 ; 16-byte Folded Reload
9448; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9449; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9450; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1444 ; 16-byte Folded Reload
9451; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9452; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9453; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1428 ; 16-byte Folded Reload
9454; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9455; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9456; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1412 ; 16-byte Folded Reload
9457; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9458; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9459; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1396 ; 16-byte Folded Reload
9460; GFX10-FLATSCR-NEXT:    v_add_co_u32 v21, vcc_lo, 0x580, v4
9461; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v22, vcc_lo, 0, v6, vcc_lo
9462; GFX10-FLATSCR-NEXT:    v_add_co_u32 v4, vcc_lo, 0x500, v4
9463; GFX10-FLATSCR-NEXT:    v_add_co_ci_u32_e32 v5, vcc_lo, 0, v6, vcc_lo
9464; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9465; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2032
9466; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1380 ; 16-byte Folded Reload
9467; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9468; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2016
9469; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1364 ; 16-byte Folded Reload
9470; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9471; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:2000
9472; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1348 ; 16-byte Folded Reload
9473; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9474; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1984
9475; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1332 ; 16-byte Folded Reload
9476; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9477; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1968
9478; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1316 ; 16-byte Folded Reload
9479; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9480; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1952
9481; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1300 ; 16-byte Folded Reload
9482; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9483; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1936
9484; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[23:26], off, off offset:1284 ; 16-byte Folded Reload
9485; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9486; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[21:22], v[23:26], off offset:1920
9487; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1268 ; 16-byte Folded Reload
9488; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9489; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:2032
9490; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1252 ; 16-byte Folded Reload
9491; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9492; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:2016
9493; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1236 ; 16-byte Folded Reload
9494; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9495; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:2000
9496; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1220 ; 16-byte Folded Reload
9497; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9498; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:1984
9499; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1204 ; 16-byte Folded Reload
9500; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9501; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:1968
9502; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1188 ; 16-byte Folded Reload
9503; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9504; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:1952
9505; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1172 ; 16-byte Folded Reload
9506; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9507; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:1936
9508; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1156 ; 16-byte Folded Reload
9509; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9510; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[4:5], v[21:24], off offset:1920
9511; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1140 ; 16-byte Folded Reload
9512; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9513; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:2032
9514; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1124 ; 16-byte Folded Reload
9515; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9516; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:2016
9517; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1108 ; 16-byte Folded Reload
9518; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9519; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:2000
9520; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1092 ; 16-byte Folded Reload
9521; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9522; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:1984
9523; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1076 ; 16-byte Folded Reload
9524; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9525; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:1968
9526; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1060 ; 16-byte Folded Reload
9527; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9528; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:1952
9529; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1044 ; 16-byte Folded Reload
9530; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9531; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:1936
9532; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1028 ; 16-byte Folded Reload
9533; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9534; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[0:1], v[21:24], off offset:1920
9535; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:1012 ; 16-byte Folded Reload
9536; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9537; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:2032
9538; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:996 ; 16-byte Folded Reload
9539; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9540; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:2016
9541; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:980 ; 16-byte Folded Reload
9542; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9543; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:2000
9544; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:964 ; 16-byte Folded Reload
9545; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9546; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:1984
9547; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:948 ; 16-byte Folded Reload
9548; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9549; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:1968
9550; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:932 ; 16-byte Folded Reload
9551; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9552; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:1952
9553; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:916 ; 16-byte Folded Reload
9554; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9555; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:1936
9556; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[21:24], off, off offset:900 ; 16-byte Folded Reload
9557; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9558; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[2:3], v[21:24], off offset:1920
9559; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:884 ; 16-byte Folded Reload
9560; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9561; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:2032
9562; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:868 ; 16-byte Folded Reload
9563; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9564; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:2016
9565; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:852 ; 16-byte Folded Reload
9566; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9567; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:2000
9568; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:836 ; 16-byte Folded Reload
9569; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9570; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:1984
9571; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:820 ; 16-byte Folded Reload
9572; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9573; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:1968
9574; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:804 ; 16-byte Folded Reload
9575; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9576; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:1952
9577; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:788 ; 16-byte Folded Reload
9578; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9579; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:1936
9580; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:772 ; 16-byte Folded Reload
9581; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9582; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[7:8], v[0:3], off offset:1920
9583; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:756 ; 16-byte Folded Reload
9584; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9585; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:2032
9586; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:740 ; 16-byte Folded Reload
9587; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9588; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:2016
9589; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:724 ; 16-byte Folded Reload
9590; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9591; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:2000
9592; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:708 ; 16-byte Folded Reload
9593; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9594; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:1984
9595; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:692 ; 16-byte Folded Reload
9596; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9597; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:1968
9598; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:676 ; 16-byte Folded Reload
9599; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9600; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:1952
9601; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:660 ; 16-byte Folded Reload
9602; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9603; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:1936
9604; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:644 ; 16-byte Folded Reload
9605; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9606; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[9:10], v[0:3], off offset:1920
9607; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:628 ; 16-byte Folded Reload
9608; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9609; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:2032
9610; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:612 ; 16-byte Folded Reload
9611; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9612; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:2016
9613; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:596 ; 16-byte Folded Reload
9614; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9615; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:2000
9616; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:580 ; 16-byte Folded Reload
9617; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9618; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:1984
9619; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:564 ; 16-byte Folded Reload
9620; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9621; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:1968
9622; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:548 ; 16-byte Folded Reload
9623; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9624; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:1952
9625; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:532 ; 16-byte Folded Reload
9626; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9627; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:1936
9628; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:516 ; 16-byte Folded Reload
9629; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9630; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[11:12], v[0:3], off offset:1920
9631; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:500 ; 16-byte Folded Reload
9632; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9633; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:2032
9634; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:484 ; 16-byte Folded Reload
9635; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9636; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:2016
9637; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:468 ; 16-byte Folded Reload
9638; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9639; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:2000
9640; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:452 ; 16-byte Folded Reload
9641; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9642; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:1984
9643; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:436 ; 16-byte Folded Reload
9644; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9645; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:1968
9646; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:420 ; 16-byte Folded Reload
9647; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9648; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:1952
9649; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:404 ; 16-byte Folded Reload
9650; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9651; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:1936
9652; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:388 ; 16-byte Folded Reload
9653; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9654; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[13:14], v[0:3], off offset:1920
9655; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:372 ; 16-byte Folded Reload
9656; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9657; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:2032
9658; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:356 ; 16-byte Folded Reload
9659; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9660; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:2016
9661; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:340 ; 16-byte Folded Reload
9662; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9663; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:2000
9664; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:324 ; 16-byte Folded Reload
9665; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9666; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:1984
9667; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:308 ; 16-byte Folded Reload
9668; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9669; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:1968
9670; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:292 ; 16-byte Folded Reload
9671; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9672; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:1952
9673; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:276 ; 16-byte Folded Reload
9674; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9675; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:1936
9676; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:260 ; 16-byte Folded Reload
9677; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9678; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[15:16], v[0:3], off offset:1920
9679; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:244 ; 16-byte Folded Reload
9680; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9681; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:2032
9682; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:228 ; 16-byte Folded Reload
9683; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9684; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:2016
9685; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:212 ; 16-byte Folded Reload
9686; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9687; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:2000
9688; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:196 ; 16-byte Folded Reload
9689; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9690; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:1984
9691; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:180 ; 16-byte Folded Reload
9692; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9693; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:1968
9694; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:164 ; 16-byte Folded Reload
9695; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9696; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:1952
9697; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:148 ; 16-byte Folded Reload
9698; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9699; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:1936
9700; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:132 ; 16-byte Folded Reload
9701; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9702; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[17:18], v[0:3], off offset:1920
9703; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:116 ; 16-byte Folded Reload
9704; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9705; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:2032
9706; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:100 ; 16-byte Folded Reload
9707; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9708; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:2016
9709; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:84 ; 16-byte Folded Reload
9710; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9711; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:2000
9712; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:68 ; 16-byte Folded Reload
9713; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9714; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:1984
9715; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:52 ; 16-byte Folded Reload
9716; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9717; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:1968
9718; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:36 ; 16-byte Folded Reload
9719; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9720; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:1952
9721; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:20 ; 16-byte Folded Reload
9722; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9723; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:1936
9724; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, off offset:4 ; 16-byte Folded Reload
9725; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
9726; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v[19:20], v[0:3], off offset:1920
9727; GFX10-FLATSCR-NEXT:    s_endpgm
9728entry:
9729  %lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
9730  %tid = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %lo)
9731
9732  %aptr = getelementptr <1280 x i32>, ptr addrspace(1) %in, i32 %tid
9733  %a = load <1280 x i32>, ptr addrspace(1) %aptr
9734
9735; mark most VGPR registers as used to increase register pressure
9736  call void asm sideeffect "", "~{v4},~{v8},~{v12},~{v16},~{v20},~{v24},~{v28},~{v32}" ()
9737  call void asm sideeffect "", "~{v36},~{v40},~{v44},~{v48},~{v52},~{v56},~{v60},~{v64}" ()
9738  call void asm sideeffect "", "~{v68},~{v72},~{v76},~{v80},~{v84},~{v88},~{v92},~{v96}" ()
9739  call void asm sideeffect "", "~{v100},~{v104},~{v108},~{v112},~{v116},~{v120},~{v124},~{v128}" ()
9740  call void asm sideeffect "", "~{v132},~{v136},~{v140},~{v144},~{v148},~{v152},~{v156},~{v160}" ()
9741  call void asm sideeffect "", "~{v164},~{v168},~{v172},~{v176},~{v180},~{v184},~{v188},~{v192}" ()
9742  call void asm sideeffect "", "~{v196},~{v200},~{v204},~{v208},~{v212},~{v216},~{v220},~{v224}" ()
9743
9744  %outptr = getelementptr <1280 x i32>, ptr addrspace(1) %out, i32 %tid
9745  store <1280 x i32> %a, ptr addrspace(1) %outptr
9746
9747  ret void
9748}
9749
9750define amdgpu_kernel void @test_limited_sgpr(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
9751; GFX6-LABEL: test_limited_sgpr:
9752; GFX6:       ; %bb.0: ; %entry
9753; GFX6-NEXT:    s_mov_b32 s40, SCRATCH_RSRC_DWORD0
9754; GFX6-NEXT:    s_mov_b32 s41, SCRATCH_RSRC_DWORD1
9755; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
9756; GFX6-NEXT:    s_mov_b32 s42, -1
9757; GFX6-NEXT:    s_mov_b32 s43, 0xe8f000
9758; GFX6-NEXT:    s_add_u32 s40, s40, s11
9759; GFX6-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, -1, 0
9760; GFX6-NEXT:    s_addc_u32 s41, s41, 0
9761; GFX6-NEXT:    v_mbcnt_hi_u32_b32_e32 v0, -1, v0
9762; GFX6-NEXT:    v_mov_b32_e32 v6, 0
9763; GFX6-NEXT:    s_mov_b64 s[4:5], exec
9764; GFX6-NEXT:    s_mov_b32 s6, 0
9765; GFX6-NEXT:    s_mov_b32 s7, 0xf000
9766; GFX6-NEXT:    s_mov_b64 exec, 15
9767; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], 0
9768; GFX6-NEXT:    s_waitcnt expcnt(0) lgkmcnt(0)
9769; GFX6-NEXT:    v_writelane_b32 v1, s0, 0
9770; GFX6-NEXT:    v_writelane_b32 v1, s1, 1
9771; GFX6-NEXT:    v_writelane_b32 v1, s2, 2
9772; GFX6-NEXT:    v_writelane_b32 v1, s3, 3
9773; GFX6-NEXT:    s_mov_b32 s8, 0x80400
9774; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s8 ; 4-byte Folded Spill
9775; GFX6-NEXT:    s_waitcnt expcnt(0)
9776; GFX6-NEXT:    buffer_load_dword v1, off, s[40:43], 0
9777; GFX6-NEXT:    s_waitcnt vmcnt(0)
9778; GFX6-NEXT:    s_mov_b64 exec, s[4:5]
9779; GFX6-NEXT:    s_mov_b64 s[4:5], s[2:3]
9780; GFX6-NEXT:    v_lshlrev_b32_e32 v5, 8, v0
9781; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:240
9782; GFX6-NEXT:    s_mov_b32 s2, 0x86a00
9783; GFX6-NEXT:    s_mov_b64 s[8:9], exec
9784; GFX6-NEXT:    s_waitcnt vmcnt(0)
9785; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9786; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9787; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9788; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9789; GFX6-NEXT:    s_waitcnt expcnt(0)
9790; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:224
9791; GFX6-NEXT:    s_mov_b32 s2, 0x86600
9792; GFX6-NEXT:    s_waitcnt vmcnt(0)
9793; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9794; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9795; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9796; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9797; GFX6-NEXT:    s_waitcnt expcnt(0)
9798; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:208
9799; GFX6-NEXT:    s_mov_b32 s2, 0x86200
9800; GFX6-NEXT:    s_waitcnt vmcnt(0)
9801; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9802; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9803; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9804; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9805; GFX6-NEXT:    s_waitcnt expcnt(0)
9806; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:192
9807; GFX6-NEXT:    s_mov_b32 s2, 0x85e00
9808; GFX6-NEXT:    s_waitcnt vmcnt(0)
9809; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9810; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9811; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9812; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9813; GFX6-NEXT:    s_waitcnt expcnt(0)
9814; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:176
9815; GFX6-NEXT:    s_mov_b32 s2, 0x85a00
9816; GFX6-NEXT:    s_waitcnt vmcnt(0)
9817; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9818; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9819; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9820; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9821; GFX6-NEXT:    s_waitcnt expcnt(0)
9822; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:160
9823; GFX6-NEXT:    s_mov_b32 s2, 0x85600
9824; GFX6-NEXT:    s_waitcnt vmcnt(0)
9825; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9826; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9827; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9828; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9829; GFX6-NEXT:    s_waitcnt expcnt(0)
9830; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:144
9831; GFX6-NEXT:    s_mov_b32 s2, 0x85200
9832; GFX6-NEXT:    s_waitcnt vmcnt(0)
9833; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9834; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9835; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9836; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9837; GFX6-NEXT:    s_waitcnt expcnt(0)
9838; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:128
9839; GFX6-NEXT:    s_mov_b32 s2, 0x84e00
9840; GFX6-NEXT:    s_waitcnt vmcnt(0)
9841; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9842; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9843; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9844; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9845; GFX6-NEXT:    s_waitcnt expcnt(0)
9846; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:112
9847; GFX6-NEXT:    s_mov_b32 s2, 0x84a00
9848; GFX6-NEXT:    s_waitcnt vmcnt(0)
9849; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9850; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9851; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9852; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9853; GFX6-NEXT:    s_waitcnt expcnt(0)
9854; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:96
9855; GFX6-NEXT:    s_mov_b32 s2, 0x84600
9856; GFX6-NEXT:    s_waitcnt vmcnt(0)
9857; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9858; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9859; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9860; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9861; GFX6-NEXT:    s_waitcnt expcnt(0)
9862; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:80
9863; GFX6-NEXT:    s_mov_b32 s2, 0x84200
9864; GFX6-NEXT:    s_waitcnt vmcnt(0)
9865; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9866; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9867; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9868; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9869; GFX6-NEXT:    s_waitcnt expcnt(0)
9870; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64 offset:64
9871; GFX6-NEXT:    s_mov_b32 s2, 0x83a00
9872; GFX6-NEXT:    s_waitcnt vmcnt(0)
9873; GFX6-NEXT:    buffer_store_dword v0, off, s[40:43], s2 ; 4-byte Folded Spill
9874; GFX6-NEXT:    buffer_store_dword v1, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9875; GFX6-NEXT:    buffer_store_dword v2, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9876; GFX6-NEXT:    buffer_store_dword v3, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9877; GFX6-NEXT:    s_waitcnt expcnt(0)
9878; GFX6-NEXT:    buffer_load_dwordx4 v[0:3], v[5:6], s[4:7], 0 addr64
9879; GFX6-NEXT:    buffer_load_dwordx4 v[7:10], v[5:6], s[4:7], 0 addr64 offset:16
9880; GFX6-NEXT:    s_mov_b32 s2, 0x83200
9881; GFX6-NEXT:    s_waitcnt vmcnt(0)
9882; GFX6-NEXT:    buffer_store_dword v7, off, s[40:43], s2 ; 4-byte Folded Spill
9883; GFX6-NEXT:    buffer_store_dword v8, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9884; GFX6-NEXT:    buffer_store_dword v9, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9885; GFX6-NEXT:    buffer_store_dword v10, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9886; GFX6-NEXT:    s_waitcnt expcnt(0)
9887; GFX6-NEXT:    buffer_load_dwordx4 v[7:10], v[5:6], s[4:7], 0 addr64 offset:32
9888; GFX6-NEXT:    s_mov_b32 s2, 0x83600
9889; GFX6-NEXT:    s_waitcnt vmcnt(0)
9890; GFX6-NEXT:    buffer_store_dword v7, off, s[40:43], s2 ; 4-byte Folded Spill
9891; GFX6-NEXT:    buffer_store_dword v8, off, s[40:43], s2 offset:4 ; 4-byte Folded Spill
9892; GFX6-NEXT:    buffer_store_dword v9, off, s[40:43], s2 offset:8 ; 4-byte Folded Spill
9893; GFX6-NEXT:    buffer_store_dword v10, off, s[40:43], s2 offset:12 ; 4-byte Folded Spill
9894; GFX6-NEXT:    s_mov_b64 s[2:3], s[6:7]
9895; GFX6-NEXT:    s_mov_b64 exec, 15
9896; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
9897; GFX6-NEXT:    s_waitcnt expcnt(0)
9898; GFX6-NEXT:    v_writelane_b32 v4, s0, 0
9899; GFX6-NEXT:    v_writelane_b32 v4, s1, 1
9900; GFX6-NEXT:    v_writelane_b32 v4, s2, 2
9901; GFX6-NEXT:    v_writelane_b32 v4, s3, 3
9902; GFX6-NEXT:    s_mov_b32 s10, 0x80800
9903; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s10 ; 4-byte Folded Spill
9904; GFX6-NEXT:    s_waitcnt expcnt(0)
9905; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
9906; GFX6-NEXT:    s_waitcnt vmcnt(0)
9907; GFX6-NEXT:    s_mov_b64 exec, s[8:9]
9908; GFX6-NEXT:    buffer_load_dwordx4 v[7:10], v[5:6], s[4:7], 0 addr64 offset:48
9909; GFX6-NEXT:    s_mov_b32 s0, 0x83e00
9910; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 13, v0
9911; GFX6-NEXT:    v_add_i32_e32 v4, vcc, 16, v4
9912; GFX6-NEXT:    s_waitcnt vmcnt(0)
9913; GFX6-NEXT:    buffer_store_dword v7, off, s[40:43], s0 ; 4-byte Folded Spill
9914; GFX6-NEXT:    buffer_store_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Spill
9915; GFX6-NEXT:    buffer_store_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Spill
9916; GFX6-NEXT:    buffer_store_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Spill
9917; GFX6-NEXT:    s_waitcnt expcnt(3)
9918; GFX6-NEXT:    v_mov_b32_e32 v7, 1
9919; GFX6-NEXT:    s_mov_b64 s[0:1], exec
9920; GFX6-NEXT:    buffer_store_dword v7, v4, s[40:43], 0 offen
9921; GFX6-NEXT:    ;;#ASMSTART
9922; GFX6-NEXT:    ; def s[4:11]
9923; GFX6-NEXT:    ;;#ASMEND
9924; GFX6-NEXT:    s_mov_b64 exec, 0xff
9925; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
9926; GFX6-NEXT:    s_waitcnt expcnt(0)
9927; GFX6-NEXT:    v_writelane_b32 v4, s4, 0
9928; GFX6-NEXT:    v_writelane_b32 v4, s5, 1
9929; GFX6-NEXT:    v_writelane_b32 v4, s6, 2
9930; GFX6-NEXT:    v_writelane_b32 v4, s7, 3
9931; GFX6-NEXT:    v_writelane_b32 v4, s8, 4
9932; GFX6-NEXT:    v_writelane_b32 v4, s9, 5
9933; GFX6-NEXT:    v_writelane_b32 v4, s10, 6
9934; GFX6-NEXT:    v_writelane_b32 v4, s11, 7
9935; GFX6-NEXT:    s_mov_b32 s2, 0x80c00
9936; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s2 ; 4-byte Folded Spill
9937; GFX6-NEXT:    s_waitcnt expcnt(0)
9938; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
9939; GFX6-NEXT:    s_waitcnt vmcnt(0)
9940; GFX6-NEXT:    s_mov_b64 exec, s[0:1]
9941; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
9942; GFX6-NEXT:    ;;#ASMSTART
9943; GFX6-NEXT:    ; def s[8:15]
9944; GFX6-NEXT:    ;;#ASMEND
9945; GFX6-NEXT:    ;;#ASMSTART
9946; GFX6-NEXT:    ; def s[16:23]
9947; GFX6-NEXT:    ;;#ASMEND
9948; GFX6-NEXT:    ;;#ASMSTART
9949; GFX6-NEXT:    ; def s[24:31]
9950; GFX6-NEXT:    ;;#ASMEND
9951; GFX6-NEXT:    ;;#ASMSTART
9952; GFX6-NEXT:    ; def s[0:3]
9953; GFX6-NEXT:    ;;#ASMEND
9954; GFX6-NEXT:    ;;#ASMSTART
9955; GFX6-NEXT:    ; def s[4:5]
9956; GFX6-NEXT:    ;;#ASMEND
9957; GFX6-NEXT:    ;;#ASMSTART
9958; GFX6-NEXT:    ; def s33
9959; GFX6-NEXT:    ;;#ASMEND
9960; GFX6-NEXT:    s_and_saveexec_b64 s[6:7], vcc
9961; GFX6-NEXT:    s_mov_b64 vcc, s[6:7]
9962; GFX6-NEXT:    s_cbranch_execz .LBB1_2
9963; GFX6-NEXT:  ; %bb.1: ; %bb0
9964; GFX6-NEXT:    s_mov_b64 s[6:7], exec
9965; GFX6-NEXT:    s_mov_b64 exec, 0xff
9966; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
9967; GFX6-NEXT:    s_waitcnt expcnt(0)
9968; GFX6-NEXT:    v_writelane_b32 v4, s8, 0
9969; GFX6-NEXT:    v_writelane_b32 v4, s9, 1
9970; GFX6-NEXT:    v_writelane_b32 v4, s10, 2
9971; GFX6-NEXT:    v_writelane_b32 v4, s11, 3
9972; GFX6-NEXT:    v_writelane_b32 v4, s12, 4
9973; GFX6-NEXT:    v_writelane_b32 v4, s13, 5
9974; GFX6-NEXT:    v_writelane_b32 v4, s14, 6
9975; GFX6-NEXT:    v_writelane_b32 v4, s15, 7
9976; GFX6-NEXT:    s_mov_b32 s34, 0x81400
9977; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s34 ; 4-byte Folded Spill
9978; GFX6-NEXT:    s_waitcnt expcnt(0)
9979; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
9980; GFX6-NEXT:    s_waitcnt vmcnt(0)
9981; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
9982; GFX6-NEXT:    s_mov_b64 s[6:7], exec
9983; GFX6-NEXT:    s_mov_b64 exec, 0xff
9984; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
9985; GFX6-NEXT:    s_mov_b32 s34, 0x80c00
9986; GFX6-NEXT:    s_waitcnt expcnt(0)
9987; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s34 ; 4-byte Folded Reload
9988; GFX6-NEXT:    s_waitcnt vmcnt(0)
9989; GFX6-NEXT:    v_readlane_b32 s8, v4, 0
9990; GFX6-NEXT:    v_readlane_b32 s9, v4, 1
9991; GFX6-NEXT:    v_readlane_b32 s10, v4, 2
9992; GFX6-NEXT:    v_readlane_b32 s11, v4, 3
9993; GFX6-NEXT:    v_readlane_b32 s12, v4, 4
9994; GFX6-NEXT:    v_readlane_b32 s13, v4, 5
9995; GFX6-NEXT:    v_readlane_b32 s14, v4, 6
9996; GFX6-NEXT:    v_readlane_b32 s15, v4, 7
9997; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
9998; GFX6-NEXT:    s_waitcnt vmcnt(0)
9999; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10000; GFX6-NEXT:    s_mov_b64 s[6:7], exec
10001; GFX6-NEXT:    s_mov_b64 exec, 0xff
10002; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10003; GFX6-NEXT:    s_waitcnt expcnt(0)
10004; GFX6-NEXT:    v_writelane_b32 v4, s16, 0
10005; GFX6-NEXT:    v_writelane_b32 v4, s17, 1
10006; GFX6-NEXT:    v_writelane_b32 v4, s18, 2
10007; GFX6-NEXT:    v_writelane_b32 v4, s19, 3
10008; GFX6-NEXT:    v_writelane_b32 v4, s20, 4
10009; GFX6-NEXT:    v_writelane_b32 v4, s21, 5
10010; GFX6-NEXT:    v_writelane_b32 v4, s22, 6
10011; GFX6-NEXT:    v_writelane_b32 v4, s23, 7
10012; GFX6-NEXT:    s_mov_b32 s34, 0x81c00
10013; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s34 ; 4-byte Folded Spill
10014; GFX6-NEXT:    s_waitcnt expcnt(0)
10015; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10016; GFX6-NEXT:    s_waitcnt vmcnt(0)
10017; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10018; GFX6-NEXT:    s_mov_b64 s[6:7], exec
10019; GFX6-NEXT:    s_mov_b64 exec, 0xff
10020; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10021; GFX6-NEXT:    s_mov_b32 s34, 0x81400
10022; GFX6-NEXT:    s_waitcnt expcnt(0)
10023; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s34 ; 4-byte Folded Reload
10024; GFX6-NEXT:    s_waitcnt vmcnt(0)
10025; GFX6-NEXT:    v_readlane_b32 s16, v4, 0
10026; GFX6-NEXT:    v_readlane_b32 s17, v4, 1
10027; GFX6-NEXT:    v_readlane_b32 s18, v4, 2
10028; GFX6-NEXT:    v_readlane_b32 s19, v4, 3
10029; GFX6-NEXT:    v_readlane_b32 s20, v4, 4
10030; GFX6-NEXT:    v_readlane_b32 s21, v4, 5
10031; GFX6-NEXT:    v_readlane_b32 s22, v4, 6
10032; GFX6-NEXT:    v_readlane_b32 s23, v4, 7
10033; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10034; GFX6-NEXT:    s_waitcnt vmcnt(0)
10035; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10036; GFX6-NEXT:    s_mov_b64 s[6:7], exec
10037; GFX6-NEXT:    s_mov_b64 exec, 0xff
10038; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10039; GFX6-NEXT:    s_waitcnt expcnt(0)
10040; GFX6-NEXT:    v_writelane_b32 v4, s24, 0
10041; GFX6-NEXT:    v_writelane_b32 v4, s25, 1
10042; GFX6-NEXT:    v_writelane_b32 v4, s26, 2
10043; GFX6-NEXT:    v_writelane_b32 v4, s27, 3
10044; GFX6-NEXT:    v_writelane_b32 v4, s28, 4
10045; GFX6-NEXT:    v_writelane_b32 v4, s29, 5
10046; GFX6-NEXT:    v_writelane_b32 v4, s30, 6
10047; GFX6-NEXT:    v_writelane_b32 v4, s31, 7
10048; GFX6-NEXT:    s_mov_b32 s34, 0x82400
10049; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s34 ; 4-byte Folded Spill
10050; GFX6-NEXT:    s_waitcnt expcnt(0)
10051; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10052; GFX6-NEXT:    s_waitcnt vmcnt(0)
10053; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10054; GFX6-NEXT:    s_mov_b64 s[6:7], exec
10055; GFX6-NEXT:    s_mov_b64 exec, 0xff
10056; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10057; GFX6-NEXT:    s_mov_b32 s34, 0x81c00
10058; GFX6-NEXT:    s_waitcnt expcnt(0)
10059; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s34 ; 4-byte Folded Reload
10060; GFX6-NEXT:    s_waitcnt vmcnt(0)
10061; GFX6-NEXT:    v_readlane_b32 s24, v4, 0
10062; GFX6-NEXT:    v_readlane_b32 s25, v4, 1
10063; GFX6-NEXT:    v_readlane_b32 s26, v4, 2
10064; GFX6-NEXT:    v_readlane_b32 s27, v4, 3
10065; GFX6-NEXT:    v_readlane_b32 s28, v4, 4
10066; GFX6-NEXT:    v_readlane_b32 s29, v4, 5
10067; GFX6-NEXT:    v_readlane_b32 s30, v4, 6
10068; GFX6-NEXT:    v_readlane_b32 s31, v4, 7
10069; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10070; GFX6-NEXT:    s_waitcnt vmcnt(0)
10071; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10072; GFX6-NEXT:    s_mov_b64 s[6:7], exec
10073; GFX6-NEXT:    s_mov_b64 exec, 15
10074; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10075; GFX6-NEXT:    s_waitcnt expcnt(0)
10076; GFX6-NEXT:    v_writelane_b32 v4, s0, 0
10077; GFX6-NEXT:    v_writelane_b32 v4, s1, 1
10078; GFX6-NEXT:    v_writelane_b32 v4, s2, 2
10079; GFX6-NEXT:    v_writelane_b32 v4, s3, 3
10080; GFX6-NEXT:    s_mov_b32 s34, 0x82c00
10081; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s34 ; 4-byte Folded Spill
10082; GFX6-NEXT:    s_waitcnt expcnt(0)
10083; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10084; GFX6-NEXT:    s_waitcnt vmcnt(0)
10085; GFX6-NEXT:    s_mov_b64 exec, s[6:7]
10086; GFX6-NEXT:    s_mov_b64 s[0:1], exec
10087; GFX6-NEXT:    s_mov_b64 exec, 3
10088; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10089; GFX6-NEXT:    s_waitcnt expcnt(0)
10090; GFX6-NEXT:    v_writelane_b32 v4, s4, 0
10091; GFX6-NEXT:    v_writelane_b32 v4, s5, 1
10092; GFX6-NEXT:    s_mov_b32 s2, 0x83000
10093; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], s2 ; 4-byte Folded Spill
10094; GFX6-NEXT:    s_waitcnt expcnt(0)
10095; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10096; GFX6-NEXT:    s_waitcnt vmcnt(0)
10097; GFX6-NEXT:    s_mov_b64 exec, s[0:1]
10098; GFX6-NEXT:    s_mov_b64 s[34:35], exec
10099; GFX6-NEXT:    s_mov_b64 exec, 0xff
10100; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10101; GFX6-NEXT:    s_mov_b32 s36, 0x82400
10102; GFX6-NEXT:    s_waitcnt expcnt(0)
10103; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s36 ; 4-byte Folded Reload
10104; GFX6-NEXT:    s_waitcnt vmcnt(0)
10105; GFX6-NEXT:    v_readlane_b32 s0, v4, 0
10106; GFX6-NEXT:    v_readlane_b32 s1, v4, 1
10107; GFX6-NEXT:    v_readlane_b32 s2, v4, 2
10108; GFX6-NEXT:    v_readlane_b32 s3, v4, 3
10109; GFX6-NEXT:    v_readlane_b32 s4, v4, 4
10110; GFX6-NEXT:    v_readlane_b32 s5, v4, 5
10111; GFX6-NEXT:    v_readlane_b32 s6, v4, 6
10112; GFX6-NEXT:    v_readlane_b32 s7, v4, 7
10113; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10114; GFX6-NEXT:    s_waitcnt vmcnt(0)
10115; GFX6-NEXT:    s_mov_b64 exec, s[34:35]
10116; GFX6-NEXT:    s_mov_b64 s[34:35], exec
10117; GFX6-NEXT:    s_mov_b64 exec, 15
10118; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10119; GFX6-NEXT:    s_mov_b32 s44, 0x82c00
10120; GFX6-NEXT:    s_waitcnt expcnt(0)
10121; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s44 ; 4-byte Folded Reload
10122; GFX6-NEXT:    s_waitcnt vmcnt(0)
10123; GFX6-NEXT:    v_readlane_b32 s36, v4, 0
10124; GFX6-NEXT:    v_readlane_b32 s37, v4, 1
10125; GFX6-NEXT:    v_readlane_b32 s38, v4, 2
10126; GFX6-NEXT:    v_readlane_b32 s39, v4, 3
10127; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10128; GFX6-NEXT:    s_waitcnt vmcnt(0)
10129; GFX6-NEXT:    s_mov_b64 exec, s[34:35]
10130; GFX6-NEXT:    s_mov_b64 s[44:45], exec
10131; GFX6-NEXT:    s_mov_b64 exec, 3
10132; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10133; GFX6-NEXT:    v_mov_b32_e32 v7, 0x20c0
10134; GFX6-NEXT:    s_waitcnt expcnt(0)
10135; GFX6-NEXT:    buffer_load_dword v4, v7, s[40:43], 0 offen ; 4-byte Folded Reload
10136; GFX6-NEXT:    s_waitcnt vmcnt(0)
10137; GFX6-NEXT:    v_readlane_b32 s34, v4, 0
10138; GFX6-NEXT:    v_readlane_b32 s35, v4, 1
10139; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10140; GFX6-NEXT:    s_waitcnt vmcnt(0)
10141; GFX6-NEXT:    s_mov_b64 exec, s[44:45]
10142; GFX6-NEXT:    ;;#ASMSTART
10143; GFX6-NEXT:    ; use s[8:15],s[16:23],s[24:31],s[0:7],s[36:39],s[34:35]
10144; GFX6-NEXT:    ;;#ASMEND
10145; GFX6-NEXT:    ;;#ASMSTART
10146; GFX6-NEXT:    ;;#ASMEND
10147; GFX6-NEXT:    ;;#ASMSTART
10148; GFX6-NEXT:    ;;#ASMEND
10149; GFX6-NEXT:    ;;#ASMSTART
10150; GFX6-NEXT:    ;;#ASMEND
10151; GFX6-NEXT:    ;;#ASMSTART
10152; GFX6-NEXT:    ;;#ASMEND
10153; GFX6-NEXT:    ;;#ASMSTART
10154; GFX6-NEXT:    ;;#ASMEND
10155; GFX6-NEXT:    ;;#ASMSTART
10156; GFX6-NEXT:    ;;#ASMEND
10157; GFX6-NEXT:    ;;#ASMSTART
10158; GFX6-NEXT:    ;;#ASMEND
10159; GFX6-NEXT:  .LBB1_2: ; %ret
10160; GFX6-NEXT:    s_or_b64 exec, exec, vcc
10161; GFX6-NEXT:    s_mov_b64 s[4:5], exec
10162; GFX6-NEXT:    s_mov_b64 exec, 15
10163; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10164; GFX6-NEXT:    s_mov_b32 s6, 0x80400
10165; GFX6-NEXT:    s_waitcnt expcnt(0)
10166; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s6 ; 4-byte Folded Reload
10167; GFX6-NEXT:    s_waitcnt vmcnt(0)
10168; GFX6-NEXT:    v_readlane_b32 s0, v4, 0
10169; GFX6-NEXT:    v_readlane_b32 s1, v4, 1
10170; GFX6-NEXT:    v_readlane_b32 s2, v4, 2
10171; GFX6-NEXT:    v_readlane_b32 s3, v4, 3
10172; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10173; GFX6-NEXT:    s_waitcnt vmcnt(0)
10174; GFX6-NEXT:    s_mov_b64 exec, s[4:5]
10175; GFX6-NEXT:    s_mov_b64 s[36:37], s[0:1]
10176; GFX6-NEXT:    s_mov_b64 s[4:5], exec
10177; GFX6-NEXT:    s_mov_b64 exec, 15
10178; GFX6-NEXT:    buffer_store_dword v4, off, s[40:43], 0
10179; GFX6-NEXT:    s_mov_b32 s6, 0x80800
10180; GFX6-NEXT:    s_waitcnt expcnt(0)
10181; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], s6 ; 4-byte Folded Reload
10182; GFX6-NEXT:    s_waitcnt vmcnt(0)
10183; GFX6-NEXT:    v_readlane_b32 s0, v4, 0
10184; GFX6-NEXT:    v_readlane_b32 s1, v4, 1
10185; GFX6-NEXT:    v_readlane_b32 s2, v4, 2
10186; GFX6-NEXT:    v_readlane_b32 s3, v4, 3
10187; GFX6-NEXT:    buffer_load_dword v4, off, s[40:43], 0
10188; GFX6-NEXT:    s_waitcnt vmcnt(0)
10189; GFX6-NEXT:    s_mov_b64 exec, s[4:5]
10190; GFX6-NEXT:    s_mov_b32 s0, 0x86a00
10191; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10192; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10193; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10194; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10195; GFX6-NEXT:    s_mov_b64 s[38:39], s[2:3]
10196; GFX6-NEXT:    s_mov_b32 s0, 0x86600
10197; GFX6-NEXT:    s_waitcnt vmcnt(0)
10198; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:240
10199; GFX6-NEXT:    s_waitcnt expcnt(0)
10200; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10201; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10202; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10203; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10204; GFX6-NEXT:    s_mov_b32 s0, 0x86200
10205; GFX6-NEXT:    s_waitcnt vmcnt(0)
10206; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:224
10207; GFX6-NEXT:    s_waitcnt expcnt(0)
10208; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10209; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10210; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10211; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10212; GFX6-NEXT:    s_mov_b32 s0, 0x85e00
10213; GFX6-NEXT:    s_waitcnt vmcnt(0)
10214; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:208
10215; GFX6-NEXT:    s_waitcnt expcnt(0)
10216; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10217; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10218; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10219; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10220; GFX6-NEXT:    s_mov_b32 s0, 0x85a00
10221; GFX6-NEXT:    s_waitcnt vmcnt(0)
10222; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:192
10223; GFX6-NEXT:    s_waitcnt expcnt(0)
10224; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10225; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10226; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10227; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10228; GFX6-NEXT:    s_mov_b32 s0, 0x85600
10229; GFX6-NEXT:    s_waitcnt vmcnt(0)
10230; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:176
10231; GFX6-NEXT:    s_waitcnt expcnt(0)
10232; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10233; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10234; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10235; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10236; GFX6-NEXT:    s_mov_b32 s0, 0x85200
10237; GFX6-NEXT:    s_waitcnt vmcnt(0)
10238; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:160
10239; GFX6-NEXT:    s_waitcnt expcnt(0)
10240; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10241; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10242; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10243; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10244; GFX6-NEXT:    s_mov_b32 s0, 0x84e00
10245; GFX6-NEXT:    s_waitcnt vmcnt(0)
10246; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:144
10247; GFX6-NEXT:    s_waitcnt expcnt(0)
10248; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10249; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10250; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10251; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10252; GFX6-NEXT:    s_mov_b32 s0, 0x84a00
10253; GFX6-NEXT:    s_waitcnt vmcnt(0)
10254; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:128
10255; GFX6-NEXT:    s_waitcnt expcnt(0)
10256; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10257; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10258; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10259; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10260; GFX6-NEXT:    s_mov_b32 s0, 0x84600
10261; GFX6-NEXT:    s_waitcnt vmcnt(0)
10262; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:112
10263; GFX6-NEXT:    s_waitcnt expcnt(0)
10264; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10265; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10266; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10267; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10268; GFX6-NEXT:    s_mov_b32 s0, 0x84200
10269; GFX6-NEXT:    s_waitcnt vmcnt(0)
10270; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:96
10271; GFX6-NEXT:    s_waitcnt expcnt(0)
10272; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10273; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10274; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10275; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10276; GFX6-NEXT:    s_mov_b32 s0, 0x83a00
10277; GFX6-NEXT:    s_waitcnt vmcnt(0)
10278; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:80
10279; GFX6-NEXT:    s_waitcnt expcnt(0)
10280; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10281; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10282; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10283; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10284; GFX6-NEXT:    s_mov_b32 s0, 0x83e00
10285; GFX6-NEXT:    s_waitcnt vmcnt(0)
10286; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:64
10287; GFX6-NEXT:    s_waitcnt expcnt(0)
10288; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10289; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10290; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10291; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10292; GFX6-NEXT:    s_mov_b32 s0, 0x83600
10293; GFX6-NEXT:    s_waitcnt vmcnt(0)
10294; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:48
10295; GFX6-NEXT:    s_waitcnt expcnt(0)
10296; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10297; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10298; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10299; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10300; GFX6-NEXT:    s_mov_b32 s0, 0x83200
10301; GFX6-NEXT:    s_waitcnt vmcnt(0)
10302; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:32
10303; GFX6-NEXT:    s_waitcnt expcnt(0)
10304; GFX6-NEXT:    buffer_load_dword v7, off, s[40:43], s0 ; 4-byte Folded Reload
10305; GFX6-NEXT:    buffer_load_dword v8, off, s[40:43], s0 offset:4 ; 4-byte Folded Reload
10306; GFX6-NEXT:    buffer_load_dword v9, off, s[40:43], s0 offset:8 ; 4-byte Folded Reload
10307; GFX6-NEXT:    buffer_load_dword v10, off, s[40:43], s0 offset:12 ; 4-byte Folded Reload
10308; GFX6-NEXT:    s_waitcnt vmcnt(0)
10309; GFX6-NEXT:    buffer_store_dwordx4 v[7:10], v[5:6], s[36:39], 0 addr64 offset:16
10310; GFX6-NEXT:    buffer_store_dwordx4 v[0:3], v[5:6], s[36:39], 0 addr64
10311; GFX6-NEXT:    s_endpgm
10312;
10313; GFX9-FLATSCR-LABEL: test_limited_sgpr:
10314; GFX9-FLATSCR:       ; %bb.0: ; %entry
10315; GFX9-FLATSCR-NEXT:    s_load_dwordx4 s[36:39], s[4:5], 0x24
10316; GFX9-FLATSCR-NEXT:    v_mbcnt_lo_u32_b32 v0, -1, 0
10317; GFX9-FLATSCR-NEXT:    v_mbcnt_hi_u32_b32 v0, -1, v0
10318; GFX9-FLATSCR-NEXT:    v_lshlrev_b32_e32 v5, 8, v0
10319; GFX9-FLATSCR-NEXT:    s_add_u32 flat_scratch_lo, s8, s13
10320; GFX9-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
10321; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:240
10322; GFX9-FLATSCR-NEXT:    s_addc_u32 flat_scratch_hi, s9, 0
10323; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2050
10324; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v4, 16
10325; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10326; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10327; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:224
10328; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2040
10329; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10330; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10331; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:208
10332; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2030
10333; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10334; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10335; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[19:22], v5, s[38:39] offset:192
10336; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[15:18], v5, s[38:39] offset:176
10337; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[11:14], v5, s[38:39] offset:160
10338; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:144
10339; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2010
10340; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[6:9], v5, s[38:39] offset:128
10341; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
10342; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10343; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:112
10344; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2020
10345; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(2)
10346; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[6:9], s0 ; 16-byte Folded Spill
10347; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20c0
10348; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v6, 1
10349; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[7:10], v5, s[38:39]
10350; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(2)
10351; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10352; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:96
10353; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20b0
10354; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(2)
10355; GFX9-FLATSCR-NEXT:    v_lshl_add_u32 v4, v7, 13, v4
10356; GFX9-FLATSCR-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v7
10357; GFX9-FLATSCR-NEXT:    scratch_store_dword v4, v6, off
10358; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(1)
10359; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10360; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:80
10361; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20a0
10362; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10363; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10364; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:64
10365; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2090
10366; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10367; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10368; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:48
10369; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2080
10370; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10371; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10372; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:32
10373; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2070
10374; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10375; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10376; GFX9-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39] offset:16
10377; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2060
10378; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10379; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[0:3], s0 ; 16-byte Folded Spill
10380; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10381; GFX9-FLATSCR-NEXT:    ; def s[0:7]
10382; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10383; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10384; GFX9-FLATSCR-NEXT:    ; def s[8:15]
10385; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10386; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10387; GFX9-FLATSCR-NEXT:    ; def s[16:23]
10388; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10389; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10390; GFX9-FLATSCR-NEXT:    ; def s[24:31]
10391; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10392; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10393; GFX9-FLATSCR-NEXT:    ; def s[40:43]
10394; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10395; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10396; GFX9-FLATSCR-NEXT:    ; def s[38:39]
10397; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10398; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10399; GFX9-FLATSCR-NEXT:    ; def s33
10400; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10401; GFX9-FLATSCR-NEXT:    s_and_saveexec_b64 s[34:35], vcc
10402; GFX9-FLATSCR-NEXT:    s_cbranch_execz .LBB1_2
10403; GFX9-FLATSCR-NEXT:  ; %bb.1: ; %bb0
10404; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10405; GFX9-FLATSCR-NEXT:    ; use s[0:7],s[8:15],s[16:23],s[24:31],s[40:43],s[38:39]
10406; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10407; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20d0
10408; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[15:18], s0 ; 16-byte Folded Spill
10409; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20e0
10410; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v0, v11
10411; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[19:22], s0 ; 16-byte Folded Spill
10412; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20f0
10413; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v1, v12
10414; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v2, v13
10415; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v3, v14
10416; GFX9-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[7:10], s0 ; 16-byte Folded Spill
10417; GFX9-FLATSCR-NEXT:    s_nop 0
10418; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10419; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10420; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[7:10], off, s0 ; 16-byte Folded Reload
10421; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20e0
10422; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[19:22], off, s0 ; 16-byte Folded Reload
10423; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20d0
10424; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[15:18], off, s0 ; 16-byte Folded Reload
10425; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v14, v3
10426; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v13, v2
10427; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v12, v1
10428; GFX9-FLATSCR-NEXT:    v_mov_b32_e32 v11, v0
10429; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10430; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10431; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10432; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10433; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10434; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10435; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10436; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10437; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10438; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10439; GFX9-FLATSCR-NEXT:    ;;#ASMSTART
10440; GFX9-FLATSCR-NEXT:    ;;#ASMEND
10441; GFX9-FLATSCR-NEXT:  .LBB1_2: ; %ret
10442; GFX9-FLATSCR-NEXT:    s_or_b64 exec, exec, s[34:35]
10443; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20c0
10444; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10445; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20b0
10446; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10447; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:112
10448; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10449; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x20a0
10450; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10451; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:96
10452; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10453; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2090
10454; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10455; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:80
10456; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10457; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2080
10458; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10459; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:64
10460; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10461; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2070
10462; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10463; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:48
10464; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10465; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2060
10466; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10467; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:32
10468; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10469; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2050
10470; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10471; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:16
10472; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[7:10], s[36:37]
10473; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[6:9], off, s0 ; 16-byte Folded Reload
10474; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2040
10475; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10476; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[6:9], s[36:37] offset:240
10477; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[6:9], off, s0 ; 16-byte Folded Reload
10478; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2030
10479; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10480; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[6:9], s[36:37] offset:224
10481; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[6:9], off, s0 ; 16-byte Folded Reload
10482; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2010
10483; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10484; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[6:9], s[36:37] offset:208
10485; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[19:22], s[36:37] offset:192
10486; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[15:18], s[36:37] offset:176
10487; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[11:14], s[36:37] offset:160
10488; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10489; GFX9-FLATSCR-NEXT:    s_movk_i32 s0, 0x2020
10490; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10491; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:144
10492; GFX9-FLATSCR-NEXT:    scratch_load_dwordx4 v[0:3], off, s0 ; 16-byte Folded Reload
10493; GFX9-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10494; GFX9-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37] offset:128
10495; GFX9-FLATSCR-NEXT:    s_endpgm
10496;
10497; GFX10-FLATSCR-LABEL: test_limited_sgpr:
10498; GFX10-FLATSCR:       ; %bb.0: ; %entry
10499; GFX10-FLATSCR-NEXT:    s_add_u32 s8, s8, s13
10500; GFX10-FLATSCR-NEXT:    s_addc_u32 s9, s9, 0
10501; GFX10-FLATSCR-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s8
10502; GFX10-FLATSCR-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s9
10503; GFX10-FLATSCR-NEXT:    s_load_dwordx4 s[36:39], s[4:5], 0x24
10504; GFX10-FLATSCR-NEXT:    v_mbcnt_lo_u32_b32 v0, -1, 0
10505; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v6, 1
10506; GFX10-FLATSCR-NEXT:    s_mov_b32 s33, exec_lo
10507; GFX10-FLATSCR-NEXT:    v_mbcnt_hi_u32_b32 v0, -1, v0
10508; GFX10-FLATSCR-NEXT:    v_lshlrev_b32_e32 v5, 8, v0
10509; GFX10-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
10510; GFX10-FLATSCR-NEXT:    s_clause 0xf
10511; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[35:38], v5, s[38:39] offset:240
10512; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[31:34], v5, s[38:39] offset:224
10513; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[27:30], v5, s[38:39] offset:208
10514; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[23:26], v5, s[38:39] offset:192
10515; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[19:22], v5, s[38:39] offset:176
10516; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[15:18], v5, s[38:39] offset:160
10517; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[11:14], v5, s[38:39] offset:144
10518; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[7:10], v5, s[38:39] offset:128
10519; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[63:66], v5, s[38:39] offset:112
10520; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[59:62], v5, s[38:39] offset:96
10521; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[55:58], v5, s[38:39] offset:80
10522; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[51:54], v5, s[38:39] offset:64
10523; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[47:50], v5, s[38:39] offset:48
10524; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[43:46], v5, s[38:39] offset:32
10525; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[39:42], v5, s[38:39] offset:16
10526; GFX10-FLATSCR-NEXT:    global_load_dwordx4 v[0:3], v5, s[38:39]
10527; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10528; GFX10-FLATSCR-NEXT:    v_lshl_add_u32 v4, v0, 13, 16
10529; GFX10-FLATSCR-NEXT:    scratch_store_dword v4, v6, off
10530; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10531; GFX10-FLATSCR-NEXT:    ; def s[0:7]
10532; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10533; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10534; GFX10-FLATSCR-NEXT:    ; def s[8:15]
10535; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10536; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10537; GFX10-FLATSCR-NEXT:    ; def s[16:23]
10538; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10539; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10540; GFX10-FLATSCR-NEXT:    ; def s[24:31]
10541; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10542; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10543; GFX10-FLATSCR-NEXT:    ; def s[40:43]
10544; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10545; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10546; GFX10-FLATSCR-NEXT:    ; def s[34:35]
10547; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10548; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10549; GFX10-FLATSCR-NEXT:    ; def s38
10550; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10551; GFX10-FLATSCR-NEXT:    v_cmpx_eq_u32_e32 0, v0
10552; GFX10-FLATSCR-NEXT:    s_cbranch_execz .LBB1_2
10553; GFX10-FLATSCR-NEXT:  ; %bb.1: ; %bb0
10554; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10555; GFX10-FLATSCR-NEXT:    ; use s[0:7],s[8:15],s[16:23],s[24:31],s[40:43],s[34:35]
10556; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10557; GFX10-FLATSCR-NEXT:    s_movk_i32 s0, 0x2010
10558; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v88, v58
10559; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v92, v62
10560; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v87, v57
10561; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v86, v56
10562; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v85, v55
10563; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v91, v61
10564; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v90, v60
10565; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v89, v59
10566; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v60, v34
10567; GFX10-FLATSCR-NEXT:    scratch_store_dwordx4 off, v[63:66], s0 ; 16-byte Folded Spill
10568; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v68, v38
10569; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v59, v33
10570; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v58, v32
10571; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v57, v31
10572; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v67, v37
10573; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v66, v36
10574; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v65, v35
10575; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v36, v10
10576; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v72, v42
10577; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v76, v46
10578; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v80, v50
10579; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v84, v54
10580; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v33, v7
10581; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v71, v41
10582; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v70, v40
10583; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v69, v39
10584; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v40, v14
10585; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v75, v45
10586; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v74, v44
10587; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v73, v43
10588; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v44, v18
10589; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v79, v49
10590; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v78, v48
10591; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v77, v47
10592; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v48, v22
10593; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v83, v53
10594; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v82, v52
10595; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v81, v51
10596; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v52, v26
10597; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v56, v30
10598; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v35, v9
10599; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v34, v8
10600; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v37, v11
10601; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v41, v15
10602; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v45, v19
10603; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v49, v23
10604; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v53, v27
10605; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v39, v13
10606; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v38, v12
10607; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v43, v17
10608; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v42, v16
10609; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v47, v21
10610; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v46, v20
10611; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v51, v25
10612; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v50, v24
10613; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v55, v29
10614; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v54, v28
10615; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10616; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10617; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v7, v33
10618; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v27, v53
10619; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v23, v49
10620; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v19, v45
10621; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v15, v41
10622; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v11, v37
10623; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v8, v34
10624; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v9, v35
10625; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v10, v36
10626; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v31, v57
10627; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v28, v54
10628; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v29, v55
10629; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v30, v56
10630; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v24, v50
10631; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v25, v51
10632; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v26, v52
10633; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v20, v46
10634; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v21, v47
10635; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v22, v48
10636; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v16, v42
10637; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v17, v43
10638; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v18, v44
10639; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v12, v38
10640; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v13, v39
10641; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v14, v40
10642; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v32, v58
10643; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v33, v59
10644; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v34, v60
10645; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10646; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10647; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v35, v65
10648; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v36, v66
10649; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v37, v67
10650; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v38, v68
10651; GFX10-FLATSCR-NEXT:    scratch_load_dwordx4 v[63:66], off, s0 ; 16-byte Folded Reload
10652; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v59, v89
10653; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v55, v85
10654; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v51, v81
10655; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v47, v77
10656; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v43, v73
10657; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v39, v69
10658; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v60, v90
10659; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v61, v91
10660; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v62, v92
10661; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v56, v86
10662; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v57, v87
10663; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v58, v88
10664; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v52, v82
10665; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v53, v83
10666; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v54, v84
10667; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v48, v78
10668; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v49, v79
10669; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v50, v80
10670; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v44, v74
10671; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v45, v75
10672; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v46, v76
10673; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v40, v70
10674; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v41, v71
10675; GFX10-FLATSCR-NEXT:    v_mov_b32_e32 v42, v72
10676; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10677; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10678; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10679; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10680; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10681; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10682; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10683; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10684; GFX10-FLATSCR-NEXT:    ;;#ASMSTART
10685; GFX10-FLATSCR-NEXT:    ;;#ASMEND
10686; GFX10-FLATSCR-NEXT:  .LBB1_2: ; %ret
10687; GFX10-FLATSCR-NEXT:    s_or_b32 exec_lo, exec_lo, s33
10688; GFX10-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
10689; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[63:66], s[36:37] offset:112
10690; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[59:62], s[36:37] offset:96
10691; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[55:58], s[36:37] offset:80
10692; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[51:54], s[36:37] offset:64
10693; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[47:50], s[36:37] offset:48
10694; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[43:46], s[36:37] offset:32
10695; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[39:42], s[36:37] offset:16
10696; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[0:3], s[36:37]
10697; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[35:38], s[36:37] offset:240
10698; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[31:34], s[36:37] offset:224
10699; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[27:30], s[36:37] offset:208
10700; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[23:26], s[36:37] offset:192
10701; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[19:22], s[36:37] offset:176
10702; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[15:18], s[36:37] offset:160
10703; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[11:14], s[36:37] offset:144
10704; GFX10-FLATSCR-NEXT:    global_store_dwordx4 v5, v[7:10], s[36:37] offset:128
10705; GFX10-FLATSCR-NEXT:    s_endpgm
10706entry:
10707  %lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
10708  %tid = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %lo)
10709
10710; allocate enough scratch to go beyond 2^12 addressing
10711  %scratch = alloca <1280 x i32>, align 16, addrspace(5)
10712
10713; load VGPR data
10714  %aptr = getelementptr <64 x i32>, ptr addrspace(1) %in, i32 %tid
10715  %a = load <64 x i32>, ptr addrspace(1) %aptr
10716
10717; make sure scratch is used
10718  %x = extractelement <64 x i32> %a, i32 0
10719  %sptr0 = getelementptr <1280 x i32>, ptr addrspace(5) %scratch, i32 %x, i32 0
10720  store i32 1, ptr addrspace(5) %sptr0
10721
10722; fill up SGPRs
10723  %sgpr0 = call <8 x i32> asm sideeffect "; def $0", "=s" ()
10724  %sgpr1 = call <8 x i32> asm sideeffect "; def $0", "=s" ()
10725  %sgpr2 = call <8 x i32> asm sideeffect "; def $0", "=s" ()
10726  %sgpr3 = call <8 x i32> asm sideeffect "; def $0", "=s" ()
10727  %sgpr4 = call <4 x i32> asm sideeffect "; def $0", "=s" ()
10728  %sgpr5 = call <2 x i32> asm sideeffect "; def $0", "=s" ()
10729  %sgpr6 = call i32 asm sideeffect "; def $0", "=s" ()
10730
10731  %cmp = icmp eq i32 %x, 0
10732  br i1 %cmp, label %bb0, label %ret
10733
10734bb0:
10735; create SGPR pressure
10736  call void asm sideeffect "; use $0,$1,$2,$3,$4,$5", "s,s,s,s,s,s,s"(<8 x i32> %sgpr0, <8 x i32> %sgpr1, <8 x i32> %sgpr2, <8 x i32> %sgpr3, <4 x i32> %sgpr4, <2 x i32> %sgpr5, i32 %sgpr6)
10737
10738; mark most VGPR registers as used to increase register pressure
10739  call void asm sideeffect "", "~{v4},~{v8},~{v12},~{v16},~{v20},~{v24},~{v28},~{v32}" ()
10740  call void asm sideeffect "", "~{v36},~{v40},~{v44},~{v48},~{v52},~{v56},~{v60},~{v64}" ()
10741  call void asm sideeffect "", "~{v68},~{v72},~{v76},~{v80},~{v84},~{v88},~{v92},~{v96}" ()
10742  call void asm sideeffect "", "~{v100},~{v104},~{v108},~{v112},~{v116},~{v120},~{v124},~{v128}" ()
10743  call void asm sideeffect "", "~{v132},~{v136},~{v140},~{v144},~{v148},~{v152},~{v156},~{v160}" ()
10744  call void asm sideeffect "", "~{v164},~{v168},~{v172},~{v176},~{v180},~{v184},~{v188},~{v192}" ()
10745  call void asm sideeffect "", "~{v196},~{v200},~{v204},~{v208},~{v212},~{v216},~{v220},~{v224}" ()
10746  br label %ret
10747
10748ret:
10749  %outptr = getelementptr <64 x i32>, ptr addrspace(1) %out, i32 %tid
10750  store <64 x i32> %a, ptr addrspace(1) %outptr
10751
10752  ret void
10753}
10754
10755declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
10756declare i32 @llvm.amdgcn.mbcnt.hi(i32, i32) #1
10757
10758attributes #0 = { "amdgpu-waves-per-eu"="10,10" }
10759attributes #1 = { nounwind readnone }
10760;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
10761; CHECK: {{.*}}
10762; FLATSCR: {{.*}}
10763