xref: /llvm-project/llvm/test/CodeGen/RISCV/pr69586.ll (revision 9122c5235ec85ce0c0ad337e862b006e7b349d84) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
2; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+v,+xsfvcp \
3; RUN:   -riscv-use-rematerializable-movimm=false | FileCheck %s --check-prefix=NOREMAT
4; RUN: llc < %s -mtriple=riscv64 -mattr=+m,+v,+xsfvcp \
5; RUN:   --riscv-use-rematerializable-movimm=true | FileCheck %s --check-prefix=REMAT
6
7define void @test(ptr %0, ptr %1, i64 %2) {
8; NOREMAT-LABEL: test:
9; NOREMAT:       # %bb.0:
10; NOREMAT-NEXT:    addi sp, sp, -752
11; NOREMAT-NEXT:    .cfi_def_cfa_offset 752
12; NOREMAT-NEXT:    sd ra, 744(sp) # 8-byte Folded Spill
13; NOREMAT-NEXT:    sd s0, 736(sp) # 8-byte Folded Spill
14; NOREMAT-NEXT:    sd s1, 728(sp) # 8-byte Folded Spill
15; NOREMAT-NEXT:    sd s2, 720(sp) # 8-byte Folded Spill
16; NOREMAT-NEXT:    sd s3, 712(sp) # 8-byte Folded Spill
17; NOREMAT-NEXT:    sd s4, 704(sp) # 8-byte Folded Spill
18; NOREMAT-NEXT:    sd s5, 696(sp) # 8-byte Folded Spill
19; NOREMAT-NEXT:    sd s6, 688(sp) # 8-byte Folded Spill
20; NOREMAT-NEXT:    sd s7, 680(sp) # 8-byte Folded Spill
21; NOREMAT-NEXT:    sd s8, 672(sp) # 8-byte Folded Spill
22; NOREMAT-NEXT:    sd s9, 664(sp) # 8-byte Folded Spill
23; NOREMAT-NEXT:    sd s10, 656(sp) # 8-byte Folded Spill
24; NOREMAT-NEXT:    sd s11, 648(sp) # 8-byte Folded Spill
25; NOREMAT-NEXT:    .cfi_offset ra, -8
26; NOREMAT-NEXT:    .cfi_offset s0, -16
27; NOREMAT-NEXT:    .cfi_offset s1, -24
28; NOREMAT-NEXT:    .cfi_offset s2, -32
29; NOREMAT-NEXT:    .cfi_offset s3, -40
30; NOREMAT-NEXT:    .cfi_offset s4, -48
31; NOREMAT-NEXT:    .cfi_offset s5, -56
32; NOREMAT-NEXT:    .cfi_offset s6, -64
33; NOREMAT-NEXT:    .cfi_offset s7, -72
34; NOREMAT-NEXT:    .cfi_offset s8, -80
35; NOREMAT-NEXT:    .cfi_offset s9, -88
36; NOREMAT-NEXT:    .cfi_offset s10, -96
37; NOREMAT-NEXT:    .cfi_offset s11, -104
38; NOREMAT-NEXT:    csrr a2, vlenb
39; NOREMAT-NEXT:    slli a2, a2, 1
40; NOREMAT-NEXT:    sub sp, sp, a2
41; NOREMAT-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0xf0, 0x05, 0x22, 0x11, 0x02, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 752 + 2 * vlenb
42; NOREMAT-NEXT:    mv a7, a0
43; NOREMAT-NEXT:    li a0, 32
44; NOREMAT-NEXT:    addi a5, a7, 512
45; NOREMAT-NEXT:    addi a4, a7, 1024
46; NOREMAT-NEXT:    addi a6, a7, 1536
47; NOREMAT-NEXT:    li t4, 1
48; NOREMAT-NEXT:    li a2, 5
49; NOREMAT-NEXT:    li t1, 3
50; NOREMAT-NEXT:    li t0, 7
51; NOREMAT-NEXT:    lui t5, 1
52; NOREMAT-NEXT:    li s4, 9
53; NOREMAT-NEXT:    li s6, 11
54; NOREMAT-NEXT:    li s9, 13
55; NOREMAT-NEXT:    li ra, 15
56; NOREMAT-NEXT:    lui t2, 2
57; NOREMAT-NEXT:    lui s1, 3
58; NOREMAT-NEXT:    lui t3, 4
59; NOREMAT-NEXT:    lui s0, 5
60; NOREMAT-NEXT:    lui s3, 6
61; NOREMAT-NEXT:    lui s7, 7
62; NOREMAT-NEXT:    vsetvli zero, a0, e32, m2, ta, ma
63; NOREMAT-NEXT:    slli t4, t4, 11
64; NOREMAT-NEXT:    sd t4, 512(sp) # 8-byte Folded Spill
65; NOREMAT-NEXT:    slli a3, a2, 9
66; NOREMAT-NEXT:    sd a3, 504(sp) # 8-byte Folded Spill
67; NOREMAT-NEXT:    slli t6, t1, 10
68; NOREMAT-NEXT:    slli s2, t0, 9
69; NOREMAT-NEXT:    add a0, a7, t5
70; NOREMAT-NEXT:    lui s11, 1
71; NOREMAT-NEXT:    slli s4, s4, 9
72; NOREMAT-NEXT:    slli s5, a2, 10
73; NOREMAT-NEXT:    slli s6, s6, 9
74; NOREMAT-NEXT:    slli s8, t1, 11
75; NOREMAT-NEXT:    vle32.v v8, (a5)
76; NOREMAT-NEXT:    slli s9, s9, 9
77; NOREMAT-NEXT:    li t5, 13
78; NOREMAT-NEXT:    vle32.v v10, (a4)
79; NOREMAT-NEXT:    vle32.v v2, (a4)
80; NOREMAT-NEXT:    slli s10, t0, 10
81; NOREMAT-NEXT:    vle32.v v0, (a6)
82; NOREMAT-NEXT:    vle32.v v12, (a6)
83; NOREMAT-NEXT:    slli ra, ra, 9
84; NOREMAT-NEXT:    vle32.v v4, (a0)
85; NOREMAT-NEXT:    vle32.v v20, (a0)
86; NOREMAT-NEXT:    add a4, a7, t2
87; NOREMAT-NEXT:    vle32.v v6, (a4)
88; NOREMAT-NEXT:    vle32.v v30, (a4)
89; NOREMAT-NEXT:    add a4, a7, s1
90; NOREMAT-NEXT:    vle32.v v28, (a4)
91; NOREMAT-NEXT:    vle32.v v26, (a4)
92; NOREMAT-NEXT:    add a4, a7, t3
93; NOREMAT-NEXT:    vle32.v v24, (a4)
94; NOREMAT-NEXT:    vle32.v v22, (a4)
95; NOREMAT-NEXT:    add a4, a7, s0
96; NOREMAT-NEXT:    vle32.v v14, (a7)
97; NOREMAT-NEXT:    vle32.v v18, (a4)
98; NOREMAT-NEXT:    vle32.v v16, (a4)
99; NOREMAT-NEXT:    add a4, a7, s3
100; NOREMAT-NEXT:    sf.vc.vv 3, 0, v14, v8
101; NOREMAT-NEXT:    vle32.v v14, (a4)
102; NOREMAT-NEXT:    sf.vc.vv 3, 0, v8, v10
103; NOREMAT-NEXT:    vle32.v v8, (a4)
104; NOREMAT-NEXT:    addi a0, sp, 640
105; NOREMAT-NEXT:    vs2r.v v8, (a0) # Unknown-size Folded Spill
106; NOREMAT-NEXT:    add a4, a7, t4
107; NOREMAT-NEXT:    vle32.v v10, (a4)
108; NOREMAT-NEXT:    sf.vc.vv 3, 0, v2, v0
109; NOREMAT-NEXT:    vle32.v v2, (a4)
110; NOREMAT-NEXT:    add a4, a7, a3
111; NOREMAT-NEXT:    vle32.v v0, (a4)
112; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v10
113; NOREMAT-NEXT:    vle32.v v10, (a4)
114; NOREMAT-NEXT:    add a4, a7, t6
115; NOREMAT-NEXT:    vle32.v v12, (a4)
116; NOREMAT-NEXT:    sf.vc.vv 3, 0, v2, v0
117; NOREMAT-NEXT:    vle32.v v2, (a4)
118; NOREMAT-NEXT:    add a4, a7, s2
119; NOREMAT-NEXT:    vle32.v v8, (a4)
120; NOREMAT-NEXT:    sf.vc.vv 3, 0, v10, v12
121; NOREMAT-NEXT:    vle32.v v12, (a4)
122; NOREMAT-NEXT:    add a4, a7, s7
123; NOREMAT-NEXT:    vle32.v v0, (a4)
124; NOREMAT-NEXT:    sf.vc.vv 3, 0, v2, v8
125; NOREMAT-NEXT:    vle32.v v10, (a4)
126; NOREMAT-NEXT:    add a4, a7, s4
127; NOREMAT-NEXT:    vle32.v v8, (a4)
128; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v4
129; NOREMAT-NEXT:    vle32.v v12, (a4)
130; NOREMAT-NEXT:    add a4, a7, s5
131; NOREMAT-NEXT:    vle32.v v4, (a4)
132; NOREMAT-NEXT:    sf.vc.vv 3, 0, v20, v8
133; NOREMAT-NEXT:    vle32.v v8, (a4)
134; NOREMAT-NEXT:    add a4, a7, s6
135; NOREMAT-NEXT:    vle32.v v20, (a4)
136; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v4
137; NOREMAT-NEXT:    vle32.v v12, (a4)
138; NOREMAT-NEXT:    add a4, a7, s8
139; NOREMAT-NEXT:    vle32.v v4, (a4)
140; NOREMAT-NEXT:    sf.vc.vv 3, 0, v8, v20
141; NOREMAT-NEXT:    vle32.v v8, (a4)
142; NOREMAT-NEXT:    add a4, a7, s9
143; NOREMAT-NEXT:    vle32.v v20, (a4)
144; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v4
145; NOREMAT-NEXT:    vle32.v v12, (a4)
146; NOREMAT-NEXT:    add a4, a7, s10
147; NOREMAT-NEXT:    vle32.v v4, (a4)
148; NOREMAT-NEXT:    sf.vc.vv 3, 0, v8, v20
149; NOREMAT-NEXT:    vle32.v v8, (a4)
150; NOREMAT-NEXT:    add a4, a7, ra
151; NOREMAT-NEXT:    vle32.v v2, (a4)
152; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v4
153; NOREMAT-NEXT:    lui t4, 8
154; NOREMAT-NEXT:    add a5, a7, t4
155; NOREMAT-NEXT:    vle32.v v20, (a5)
156; NOREMAT-NEXT:    vle32.v v12, (a4)
157; NOREMAT-NEXT:    sf.vc.vv 3, 0, v8, v2
158; NOREMAT-NEXT:    li a4, 17
159; NOREMAT-NEXT:    slli a4, a4, 9
160; NOREMAT-NEXT:    li s1, 17
161; NOREMAT-NEXT:    sd a4, 624(sp) # 8-byte Folded Spill
162; NOREMAT-NEXT:    add a4, a7, a4
163; NOREMAT-NEXT:    vle32.v v8, (a4)
164; NOREMAT-NEXT:    vle32.v v4, (a4)
165; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v6
166; NOREMAT-NEXT:    li a5, 9
167; NOREMAT-NEXT:    slli a4, a5, 10
168; NOREMAT-NEXT:    sd a4, 616(sp) # 8-byte Folded Spill
169; NOREMAT-NEXT:    add a4, a7, a4
170; NOREMAT-NEXT:    vle32.v v12, (a4)
171; NOREMAT-NEXT:    vle32.v v6, (a4)
172; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v8
173; NOREMAT-NEXT:    li a4, 19
174; NOREMAT-NEXT:    slli a4, a4, 9
175; NOREMAT-NEXT:    li t2, 19
176; NOREMAT-NEXT:    sd a4, 608(sp) # 8-byte Folded Spill
177; NOREMAT-NEXT:    add a4, a7, a4
178; NOREMAT-NEXT:    vle32.v v8, (a4)
179; NOREMAT-NEXT:    vle32.v v30, (a4)
180; NOREMAT-NEXT:    slli a3, a2, 11
181; NOREMAT-NEXT:    sd a3, 600(sp) # 8-byte Folded Spill
182; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v12
183; NOREMAT-NEXT:    add a3, a7, a3
184; NOREMAT-NEXT:    vle32.v v12, (a3)
185; NOREMAT-NEXT:    vle32.v v4, (a3)
186; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v8
187; NOREMAT-NEXT:    li s7, 21
188; NOREMAT-NEXT:    slli a3, s7, 9
189; NOREMAT-NEXT:    sd a3, 592(sp) # 8-byte Folded Spill
190; NOREMAT-NEXT:    add a3, a7, a3
191; NOREMAT-NEXT:    vle32.v v8, (a3)
192; NOREMAT-NEXT:    vle32.v v6, (a3)
193; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v12
194; NOREMAT-NEXT:    li a6, 11
195; NOREMAT-NEXT:    slli a3, a6, 10
196; NOREMAT-NEXT:    sd a3, 584(sp) # 8-byte Folded Spill
197; NOREMAT-NEXT:    add a3, a7, a3
198; NOREMAT-NEXT:    vle32.v v12, (a3)
199; NOREMAT-NEXT:    vle32.v v30, (a3)
200; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v8
201; NOREMAT-NEXT:    li s3, 23
202; NOREMAT-NEXT:    slli a3, s3, 9
203; NOREMAT-NEXT:    sd a3, 576(sp) # 8-byte Folded Spill
204; NOREMAT-NEXT:    add a3, a7, a3
205; NOREMAT-NEXT:    vle32.v v8, (a3)
206; NOREMAT-NEXT:    vle32.v v4, (a3)
207; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v12
208; NOREMAT-NEXT:    li s0, 25
209; NOREMAT-NEXT:    slli a3, s0, 9
210; NOREMAT-NEXT:    sd a3, 568(sp) # 8-byte Folded Spill
211; NOREMAT-NEXT:    add a3, a7, a3
212; NOREMAT-NEXT:    vle32.v v12, (a3)
213; NOREMAT-NEXT:    vle32.v v6, (a3)
214; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v8
215; NOREMAT-NEXT:    slli a3, t5, 10
216; NOREMAT-NEXT:    sd a3, 560(sp) # 8-byte Folded Spill
217; NOREMAT-NEXT:    add a3, a7, a3
218; NOREMAT-NEXT:    vle32.v v8, (a3)
219; NOREMAT-NEXT:    vle32.v v30, (a3)
220; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v28
221; NOREMAT-NEXT:    li t3, 27
222; NOREMAT-NEXT:    slli a3, t3, 9
223; NOREMAT-NEXT:    sd a3, 552(sp) # 8-byte Folded Spill
224; NOREMAT-NEXT:    add a3, a7, a3
225; NOREMAT-NEXT:    vle32.v v28, (a3)
226; NOREMAT-NEXT:    vle32.v v4, (a3)
227; NOREMAT-NEXT:    slli a2, t0, 11
228; NOREMAT-NEXT:    sd a2, 544(sp) # 8-byte Folded Spill
229; NOREMAT-NEXT:    sf.vc.vv 3, 0, v26, v12
230; NOREMAT-NEXT:    add a2, a7, a2
231; NOREMAT-NEXT:    vle32.v v12, (a2)
232; NOREMAT-NEXT:    vle32.v v26, (a2)
233; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v8
234; NOREMAT-NEXT:    li t0, 29
235; NOREMAT-NEXT:    slli a2, t0, 9
236; NOREMAT-NEXT:    sd a2, 536(sp) # 8-byte Folded Spill
237; NOREMAT-NEXT:    add a2, a7, a2
238; NOREMAT-NEXT:    vle32.v v8, (a2)
239; NOREMAT-NEXT:    vle32.v v6, (a2)
240; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v28
241; NOREMAT-NEXT:    li a3, 15
242; NOREMAT-NEXT:    slli a2, a3, 10
243; NOREMAT-NEXT:    sd a2, 528(sp) # 8-byte Folded Spill
244; NOREMAT-NEXT:    add a2, a7, a2
245; NOREMAT-NEXT:    vle32.v v28, (a2)
246; NOREMAT-NEXT:    vle32.v v30, (a2)
247; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v12
248; NOREMAT-NEXT:    li t1, 31
249; NOREMAT-NEXT:    slli a2, t1, 9
250; NOREMAT-NEXT:    sd a2, 520(sp) # 8-byte Folded Spill
251; NOREMAT-NEXT:    add a2, a7, a2
252; NOREMAT-NEXT:    vle32.v v12, (a2)
253; NOREMAT-NEXT:    vle32.v v4, (a2)
254; NOREMAT-NEXT:    sf.vc.vv 3, 0, v26, v8
255; NOREMAT-NEXT:    lui a4, 4
256; NOREMAT-NEXT:    addiw a0, a4, 512
257; NOREMAT-NEXT:    sd a0, 496(sp) # 8-byte Folded Spill
258; NOREMAT-NEXT:    add a0, a7, a0
259; NOREMAT-NEXT:    vle32.v v8, (a0)
260; NOREMAT-NEXT:    vle32.v v26, (a0)
261; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v28
262; NOREMAT-NEXT:    slli a2, s1, 10
263; NOREMAT-NEXT:    sd a2, 488(sp) # 8-byte Folded Spill
264; NOREMAT-NEXT:    add a2, a7, a2
265; NOREMAT-NEXT:    vle32.v v28, (a2)
266; NOREMAT-NEXT:    vle32.v v6, (a2)
267; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v12
268; NOREMAT-NEXT:    addiw a2, a4, 1536
269; NOREMAT-NEXT:    sd a2, 480(sp) # 8-byte Folded Spill
270; NOREMAT-NEXT:    add a2, a7, a2
271; NOREMAT-NEXT:    vle32.v v12, (a2)
272; NOREMAT-NEXT:    vle32.v v30, (a2)
273; NOREMAT-NEXT:    slli a2, a5, 11
274; NOREMAT-NEXT:    sd a2, 472(sp) # 8-byte Folded Spill
275; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v24
276; NOREMAT-NEXT:    add a2, a7, a2
277; NOREMAT-NEXT:    vle32.v v24, (a2)
278; NOREMAT-NEXT:    vle32.v v4, (a2)
279; NOREMAT-NEXT:    sf.vc.vv 3, 0, v22, v8
280; NOREMAT-NEXT:    lui a5, 5
281; NOREMAT-NEXT:    addiw a2, a5, -1536
282; NOREMAT-NEXT:    sd a2, 464(sp) # 8-byte Folded Spill
283; NOREMAT-NEXT:    add a2, a7, a2
284; NOREMAT-NEXT:    vle32.v v8, (a2)
285; NOREMAT-NEXT:    vle32.v v22, (a2)
286; NOREMAT-NEXT:    sf.vc.vv 3, 0, v26, v28
287; NOREMAT-NEXT:    slli a2, t2, 10
288; NOREMAT-NEXT:    sd a2, 456(sp) # 8-byte Folded Spill
289; NOREMAT-NEXT:    li t2, 19
290; NOREMAT-NEXT:    add a2, a7, a2
291; NOREMAT-NEXT:    vle32.v v26, (a2)
292; NOREMAT-NEXT:    vle32.v v28, (a2)
293; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v12
294; NOREMAT-NEXT:    addiw a2, a5, -512
295; NOREMAT-NEXT:    sd a2, 448(sp) # 8-byte Folded Spill
296; NOREMAT-NEXT:    add a2, a7, a2
297; NOREMAT-NEXT:    vle32.v v12, (a2)
298; NOREMAT-NEXT:    vle32.v v6, (a2)
299; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v24
300; NOREMAT-NEXT:    addiw a2, a5, 512
301; NOREMAT-NEXT:    sd a2, 440(sp) # 8-byte Folded Spill
302; NOREMAT-NEXT:    add a2, a7, a2
303; NOREMAT-NEXT:    vle32.v v24, (a2)
304; NOREMAT-NEXT:    vle32.v v30, (a2)
305; NOREMAT-NEXT:    slli a2, s7, 10
306; NOREMAT-NEXT:    sd a2, 432(sp) # 8-byte Folded Spill
307; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v8
308; NOREMAT-NEXT:    add a2, a7, a2
309; NOREMAT-NEXT:    vle32.v v8, (a2)
310; NOREMAT-NEXT:    vle32.v v4, (a2)
311; NOREMAT-NEXT:    sf.vc.vv 3, 0, v22, v26
312; NOREMAT-NEXT:    addiw a2, a5, 1536
313; NOREMAT-NEXT:    sd a2, 424(sp) # 8-byte Folded Spill
314; NOREMAT-NEXT:    add a2, a7, a2
315; NOREMAT-NEXT:    vle32.v v22, (a2)
316; NOREMAT-NEXT:    vle32.v v26, (a2)
317; NOREMAT-NEXT:    slli a2, a6, 11
318; NOREMAT-NEXT:    sd a2, 416(sp) # 8-byte Folded Spill
319; NOREMAT-NEXT:    sf.vc.vv 3, 0, v28, v12
320; NOREMAT-NEXT:    add a2, a7, a2
321; NOREMAT-NEXT:    vle32.v v12, (a2)
322; NOREMAT-NEXT:    vle32.v v28, (a2)
323; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v18
324; NOREMAT-NEXT:    lui a6, 6
325; NOREMAT-NEXT:    addiw a2, a6, -1536
326; NOREMAT-NEXT:    sd a2, 408(sp) # 8-byte Folded Spill
327; NOREMAT-NEXT:    add a2, a7, a2
328; NOREMAT-NEXT:    vle32.v v18, (a2)
329; NOREMAT-NEXT:    vle32.v v6, (a2)
330; NOREMAT-NEXT:    slli a2, s3, 10
331; NOREMAT-NEXT:    sd a2, 400(sp) # 8-byte Folded Spill
332; NOREMAT-NEXT:    sf.vc.vv 3, 0, v16, v24
333; NOREMAT-NEXT:    add a2, a7, a2
334; NOREMAT-NEXT:    vle32.v v16, (a2)
335; NOREMAT-NEXT:    vle32.v v24, (a2)
336; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v8
337; NOREMAT-NEXT:    addiw a2, a6, -512
338; NOREMAT-NEXT:    sd a2, 392(sp) # 8-byte Folded Spill
339; NOREMAT-NEXT:    add a2, a7, a2
340; NOREMAT-NEXT:    vle32.v v8, (a2)
341; NOREMAT-NEXT:    vle32.v v30, (a2)
342; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v22
343; NOREMAT-NEXT:    addiw a2, a6, 512
344; NOREMAT-NEXT:    sd a2, 384(sp) # 8-byte Folded Spill
345; NOREMAT-NEXT:    add a2, a7, a2
346; NOREMAT-NEXT:    vle32.v v22, (a2)
347; NOREMAT-NEXT:    vle32.v v4, (a2)
348; NOREMAT-NEXT:    slli a2, s0, 10
349; NOREMAT-NEXT:    sd a2, 376(sp) # 8-byte Folded Spill
350; NOREMAT-NEXT:    sf.vc.vv 3, 0, v26, v12
351; NOREMAT-NEXT:    add a2, a7, a2
352; NOREMAT-NEXT:    vle32.v v26, (a2)
353; NOREMAT-NEXT:    vle32.v v2, (a2)
354; NOREMAT-NEXT:    sf.vc.vv 3, 0, v28, v18
355; NOREMAT-NEXT:    addiw a2, a6, 1536
356; NOREMAT-NEXT:    sd a2, 368(sp) # 8-byte Folded Spill
357; NOREMAT-NEXT:    add a2, a7, a2
358; NOREMAT-NEXT:    vle32.v v18, (a2)
359; NOREMAT-NEXT:    vle32.v v28, (a2)
360; NOREMAT-NEXT:    slli a2, t5, 11
361; NOREMAT-NEXT:    sd a2, 360(sp) # 8-byte Folded Spill
362; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v16
363; NOREMAT-NEXT:    add a2, a7, a2
364; NOREMAT-NEXT:    vle32.v v16, (a2)
365; NOREMAT-NEXT:    vle32.v v6, (a2)
366; NOREMAT-NEXT:    sf.vc.vv 3, 0, v24, v8
367; NOREMAT-NEXT:    lui s0, 7
368; NOREMAT-NEXT:    addiw a2, s0, -1536
369; NOREMAT-NEXT:    sd a2, 352(sp) # 8-byte Folded Spill
370; NOREMAT-NEXT:    add a2, a7, a2
371; NOREMAT-NEXT:    vle32.v v8, (a2)
372; NOREMAT-NEXT:    vle32.v v24, (a2)
373; NOREMAT-NEXT:    slli a2, t3, 10
374; NOREMAT-NEXT:    sd a2, 344(sp) # 8-byte Folded Spill
375; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v14
376; NOREMAT-NEXT:    add a2, a7, a2
377; NOREMAT-NEXT:    vle32.v v14, (a2)
378; NOREMAT-NEXT:    vle32.v v30, (a2)
379; NOREMAT-NEXT:    addi a0, sp, 640
380; NOREMAT-NEXT:    vl2r.v v12, (a0) # Unknown-size Folded Reload
381; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v22
382; NOREMAT-NEXT:    addiw a2, s0, -512
383; NOREMAT-NEXT:    sd a2, 336(sp) # 8-byte Folded Spill
384; NOREMAT-NEXT:    add a2, a7, a2
385; NOREMAT-NEXT:    vle32.v v22, (a2)
386; NOREMAT-NEXT:    vle32.v v12, (a2)
387; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v26
388; NOREMAT-NEXT:    addiw a2, s0, 512
389; NOREMAT-NEXT:    sd a2, 328(sp) # 8-byte Folded Spill
390; NOREMAT-NEXT:    lui t3, 7
391; NOREMAT-NEXT:    add a2, a7, a2
392; NOREMAT-NEXT:    vle32.v v26, (a2)
393; NOREMAT-NEXT:    vle32.v v4, (a2)
394; NOREMAT-NEXT:    slli a2, t0, 10
395; NOREMAT-NEXT:    sd a2, 320(sp) # 8-byte Folded Spill
396; NOREMAT-NEXT:    sf.vc.vv 3, 0, v2, v18
397; NOREMAT-NEXT:    add a2, a7, a2
398; NOREMAT-NEXT:    vle32.v v18, (a2)
399; NOREMAT-NEXT:    vle32.v v2, (a2)
400; NOREMAT-NEXT:    sf.vc.vv 3, 0, v28, v16
401; NOREMAT-NEXT:    addiw a2, t3, 1536
402; NOREMAT-NEXT:    sd a2, 312(sp) # 8-byte Folded Spill
403; NOREMAT-NEXT:    add a2, a7, a2
404; NOREMAT-NEXT:    vle32.v v16, (a2)
405; NOREMAT-NEXT:    vle32.v v28, (a2)
406; NOREMAT-NEXT:    slli a2, a3, 11
407; NOREMAT-NEXT:    sd a2, 304(sp) # 8-byte Folded Spill
408; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v8
409; NOREMAT-NEXT:    add a2, a7, a2
410; NOREMAT-NEXT:    vle32.v v8, (a2)
411; NOREMAT-NEXT:    vle32.v v6, (a2)
412; NOREMAT-NEXT:    sf.vc.vv 3, 0, v24, v14
413; NOREMAT-NEXT:    addiw a2, t4, -1536
414; NOREMAT-NEXT:    sd a2, 296(sp) # 8-byte Folded Spill
415; NOREMAT-NEXT:    add a2, a7, a2
416; NOREMAT-NEXT:    vle32.v v14, (a2)
417; NOREMAT-NEXT:    vle32.v v24, (a2)
418; NOREMAT-NEXT:    slli a2, t1, 10
419; NOREMAT-NEXT:    sd a2, 288(sp) # 8-byte Folded Spill
420; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v22
421; NOREMAT-NEXT:    add a2, a7, a2
422; NOREMAT-NEXT:    vle32.v v22, (a2)
423; NOREMAT-NEXT:    vle32.v v30, (a2)
424; NOREMAT-NEXT:    addiw a0, t4, -512
425; NOREMAT-NEXT:    sd a0, 280(sp) # 8-byte Folded Spill
426; NOREMAT-NEXT:    add a0, a7, a0
427; NOREMAT-NEXT:    sf.vc.vv 3, 0, v12, v0
428; NOREMAT-NEXT:    vle32.v v12, (a0)
429; NOREMAT-NEXT:    vle32.v v0, (a0)
430; NOREMAT-NEXT:    sf.vc.vv 3, 0, v10, v26
431; NOREMAT-NEXT:    sf.vc.vv 3, 0, v4, v18
432; NOREMAT-NEXT:    sf.vc.vv 3, 0, v2, v16
433; NOREMAT-NEXT:    sf.vc.vv 3, 0, v28, v8
434; NOREMAT-NEXT:    sf.vc.vv 3, 0, v6, v14
435; NOREMAT-NEXT:    sf.vc.vv 3, 0, v24, v22
436; NOREMAT-NEXT:    sf.vc.vv 3, 0, v30, v12
437; NOREMAT-NEXT:    sf.vc.vv 3, 0, v0, v20
438; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
439; NOREMAT-NEXT:    addi a0, a1, 1024
440; NOREMAT-NEXT:    vse32.v v8, (a0)
441; NOREMAT-NEXT:    add s11, a1, s11
442; NOREMAT-NEXT:    sd s11, 272(sp) # 8-byte Folded Spill
443; NOREMAT-NEXT:    lui a0, 2
444; NOREMAT-NEXT:    add a0, a1, a0
445; NOREMAT-NEXT:    sd a0, 264(sp) # 8-byte Folded Spill
446; NOREMAT-NEXT:    lui a0, 3
447; NOREMAT-NEXT:    add a0, a1, a0
448; NOREMAT-NEXT:    sd a0, 256(sp) # 8-byte Folded Spill
449; NOREMAT-NEXT:    add a4, a1, a4
450; NOREMAT-NEXT:    sd a4, 248(sp) # 8-byte Folded Spill
451; NOREMAT-NEXT:    add a5, a1, a5
452; NOREMAT-NEXT:    sd a5, 240(sp) # 8-byte Folded Spill
453; NOREMAT-NEXT:    add a6, a1, a6
454; NOREMAT-NEXT:    sd a6, 232(sp) # 8-byte Folded Spill
455; NOREMAT-NEXT:    add t3, a1, t3
456; NOREMAT-NEXT:    sd t3, 224(sp) # 8-byte Folded Spill
457; NOREMAT-NEXT:    add a0, a1, t4
458; NOREMAT-NEXT:    sd a0, 216(sp) # 8-byte Folded Spill
459; NOREMAT-NEXT:    addiw a0, t4, 512
460; NOREMAT-NEXT:    sd a0, 192(sp) # 8-byte Folded Spill
461; NOREMAT-NEXT:    addiw a0, t4, 1024
462; NOREMAT-NEXT:    sd a0, 176(sp) # 8-byte Folded Spill
463; NOREMAT-NEXT:    addiw a0, t4, 1536
464; NOREMAT-NEXT:    sd a0, 160(sp) # 8-byte Folded Spill
465; NOREMAT-NEXT:    slli s1, s1, 11
466; NOREMAT-NEXT:    sd s1, 128(sp) # 8-byte Folded Spill
467; NOREMAT-NEXT:    lui a0, 9
468; NOREMAT-NEXT:    addiw a2, a0, -1536
469; NOREMAT-NEXT:    sd a2, 88(sp) # 8-byte Folded Spill
470; NOREMAT-NEXT:    addiw a2, a0, -1024
471; NOREMAT-NEXT:    sd a2, 72(sp) # 8-byte Folded Spill
472; NOREMAT-NEXT:    addiw a2, a0, -512
473; NOREMAT-NEXT:    sd a2, 40(sp) # 8-byte Folded Spill
474; NOREMAT-NEXT:    add a2, a1, a0
475; NOREMAT-NEXT:    sd a2, 208(sp) # 8-byte Folded Spill
476; NOREMAT-NEXT:    addiw s11, a0, 512
477; NOREMAT-NEXT:    addiw s7, a0, 1024
478; NOREMAT-NEXT:    addiw s3, a0, 1536
479; NOREMAT-NEXT:    slli s1, t2, 11
480; NOREMAT-NEXT:    lui a0, 10
481; NOREMAT-NEXT:    addiw t2, a0, -1536
482; NOREMAT-NEXT:    addiw a7, a0, -1024
483; NOREMAT-NEXT:    addiw a4, a0, -512
484; NOREMAT-NEXT:    add a2, a1, a0
485; NOREMAT-NEXT:    sd a2, 200(sp) # 8-byte Folded Spill
486; NOREMAT-NEXT:    addiw a0, a0, 512
487; NOREMAT-NEXT:    ld a2, 512(sp) # 8-byte Folded Reload
488; NOREMAT-NEXT:    add a2, a1, a2
489; NOREMAT-NEXT:    ld a3, 504(sp) # 8-byte Folded Reload
490; NOREMAT-NEXT:    add a3, a1, a3
491; NOREMAT-NEXT:    add a5, a1, t6
492; NOREMAT-NEXT:    add a6, a1, s2
493; NOREMAT-NEXT:    add t0, a1, s4
494; NOREMAT-NEXT:    add t1, a1, s5
495; NOREMAT-NEXT:    add t3, a1, s6
496; NOREMAT-NEXT:    add t4, a1, s8
497; NOREMAT-NEXT:    add t5, a1, s9
498; NOREMAT-NEXT:    add t6, a1, s10
499; NOREMAT-NEXT:    add s0, a1, ra
500; NOREMAT-NEXT:    ld s2, 624(sp) # 8-byte Folded Reload
501; NOREMAT-NEXT:    add s2, a1, s2
502; NOREMAT-NEXT:    ld s4, 616(sp) # 8-byte Folded Reload
503; NOREMAT-NEXT:    add s4, a1, s4
504; NOREMAT-NEXT:    ld s5, 608(sp) # 8-byte Folded Reload
505; NOREMAT-NEXT:    add s5, a1, s5
506; NOREMAT-NEXT:    ld s6, 600(sp) # 8-byte Folded Reload
507; NOREMAT-NEXT:    add s6, a1, s6
508; NOREMAT-NEXT:    ld s8, 592(sp) # 8-byte Folded Reload
509; NOREMAT-NEXT:    add s8, a1, s8
510; NOREMAT-NEXT:    ld s9, 584(sp) # 8-byte Folded Reload
511; NOREMAT-NEXT:    add s9, a1, s9
512; NOREMAT-NEXT:    ld s10, 576(sp) # 8-byte Folded Reload
513; NOREMAT-NEXT:    add s10, a1, s10
514; NOREMAT-NEXT:    ld ra, 568(sp) # 8-byte Folded Reload
515; NOREMAT-NEXT:    add ra, a1, ra
516; NOREMAT-NEXT:    sd ra, 16(sp) # 8-byte Folded Spill
517; NOREMAT-NEXT:    ld ra, 560(sp) # 8-byte Folded Reload
518; NOREMAT-NEXT:    add ra, a1, ra
519; NOREMAT-NEXT:    sd ra, 24(sp) # 8-byte Folded Spill
520; NOREMAT-NEXT:    ld ra, 552(sp) # 8-byte Folded Reload
521; NOREMAT-NEXT:    add ra, a1, ra
522; NOREMAT-NEXT:    sd ra, 32(sp) # 8-byte Folded Spill
523; NOREMAT-NEXT:    ld ra, 544(sp) # 8-byte Folded Reload
524; NOREMAT-NEXT:    add ra, a1, ra
525; NOREMAT-NEXT:    sd ra, 48(sp) # 8-byte Folded Spill
526; NOREMAT-NEXT:    ld ra, 536(sp) # 8-byte Folded Reload
527; NOREMAT-NEXT:    add ra, a1, ra
528; NOREMAT-NEXT:    sd ra, 56(sp) # 8-byte Folded Spill
529; NOREMAT-NEXT:    ld ra, 528(sp) # 8-byte Folded Reload
530; NOREMAT-NEXT:    add ra, a1, ra
531; NOREMAT-NEXT:    sd ra, 64(sp) # 8-byte Folded Spill
532; NOREMAT-NEXT:    ld ra, 520(sp) # 8-byte Folded Reload
533; NOREMAT-NEXT:    add ra, a1, ra
534; NOREMAT-NEXT:    sd ra, 80(sp) # 8-byte Folded Spill
535; NOREMAT-NEXT:    ld ra, 496(sp) # 8-byte Folded Reload
536; NOREMAT-NEXT:    add ra, a1, ra
537; NOREMAT-NEXT:    sd ra, 96(sp) # 8-byte Folded Spill
538; NOREMAT-NEXT:    ld ra, 488(sp) # 8-byte Folded Reload
539; NOREMAT-NEXT:    add ra, a1, ra
540; NOREMAT-NEXT:    sd ra, 104(sp) # 8-byte Folded Spill
541; NOREMAT-NEXT:    ld ra, 480(sp) # 8-byte Folded Reload
542; NOREMAT-NEXT:    add ra, a1, ra
543; NOREMAT-NEXT:    sd ra, 112(sp) # 8-byte Folded Spill
544; NOREMAT-NEXT:    ld ra, 472(sp) # 8-byte Folded Reload
545; NOREMAT-NEXT:    add ra, a1, ra
546; NOREMAT-NEXT:    sd ra, 120(sp) # 8-byte Folded Spill
547; NOREMAT-NEXT:    ld ra, 464(sp) # 8-byte Folded Reload
548; NOREMAT-NEXT:    add ra, a1, ra
549; NOREMAT-NEXT:    sd ra, 136(sp) # 8-byte Folded Spill
550; NOREMAT-NEXT:    ld ra, 456(sp) # 8-byte Folded Reload
551; NOREMAT-NEXT:    add ra, a1, ra
552; NOREMAT-NEXT:    sd ra, 144(sp) # 8-byte Folded Spill
553; NOREMAT-NEXT:    ld ra, 448(sp) # 8-byte Folded Reload
554; NOREMAT-NEXT:    add ra, a1, ra
555; NOREMAT-NEXT:    sd ra, 152(sp) # 8-byte Folded Spill
556; NOREMAT-NEXT:    ld ra, 440(sp) # 8-byte Folded Reload
557; NOREMAT-NEXT:    add ra, a1, ra
558; NOREMAT-NEXT:    sd ra, 168(sp) # 8-byte Folded Spill
559; NOREMAT-NEXT:    ld ra, 432(sp) # 8-byte Folded Reload
560; NOREMAT-NEXT:    add ra, a1, ra
561; NOREMAT-NEXT:    sd ra, 184(sp) # 8-byte Folded Spill
562; NOREMAT-NEXT:    ld ra, 424(sp) # 8-byte Folded Reload
563; NOREMAT-NEXT:    add ra, a1, ra
564; NOREMAT-NEXT:    sd ra, 424(sp) # 8-byte Folded Spill
565; NOREMAT-NEXT:    ld ra, 416(sp) # 8-byte Folded Reload
566; NOREMAT-NEXT:    add ra, a1, ra
567; NOREMAT-NEXT:    sd ra, 432(sp) # 8-byte Folded Spill
568; NOREMAT-NEXT:    ld ra, 408(sp) # 8-byte Folded Reload
569; NOREMAT-NEXT:    add ra, a1, ra
570; NOREMAT-NEXT:    sd ra, 440(sp) # 8-byte Folded Spill
571; NOREMAT-NEXT:    ld ra, 400(sp) # 8-byte Folded Reload
572; NOREMAT-NEXT:    add ra, a1, ra
573; NOREMAT-NEXT:    sd ra, 448(sp) # 8-byte Folded Spill
574; NOREMAT-NEXT:    ld ra, 392(sp) # 8-byte Folded Reload
575; NOREMAT-NEXT:    add ra, a1, ra
576; NOREMAT-NEXT:    sd ra, 456(sp) # 8-byte Folded Spill
577; NOREMAT-NEXT:    ld ra, 384(sp) # 8-byte Folded Reload
578; NOREMAT-NEXT:    add ra, a1, ra
579; NOREMAT-NEXT:    sd ra, 464(sp) # 8-byte Folded Spill
580; NOREMAT-NEXT:    ld ra, 376(sp) # 8-byte Folded Reload
581; NOREMAT-NEXT:    add ra, a1, ra
582; NOREMAT-NEXT:    sd ra, 472(sp) # 8-byte Folded Spill
583; NOREMAT-NEXT:    ld ra, 368(sp) # 8-byte Folded Reload
584; NOREMAT-NEXT:    add ra, a1, ra
585; NOREMAT-NEXT:    sd ra, 480(sp) # 8-byte Folded Spill
586; NOREMAT-NEXT:    ld ra, 360(sp) # 8-byte Folded Reload
587; NOREMAT-NEXT:    add ra, a1, ra
588; NOREMAT-NEXT:    sd ra, 488(sp) # 8-byte Folded Spill
589; NOREMAT-NEXT:    ld ra, 352(sp) # 8-byte Folded Reload
590; NOREMAT-NEXT:    add ra, a1, ra
591; NOREMAT-NEXT:    sd ra, 496(sp) # 8-byte Folded Spill
592; NOREMAT-NEXT:    ld ra, 344(sp) # 8-byte Folded Reload
593; NOREMAT-NEXT:    add ra, a1, ra
594; NOREMAT-NEXT:    sd ra, 504(sp) # 8-byte Folded Spill
595; NOREMAT-NEXT:    ld ra, 336(sp) # 8-byte Folded Reload
596; NOREMAT-NEXT:    add ra, a1, ra
597; NOREMAT-NEXT:    sd ra, 512(sp) # 8-byte Folded Spill
598; NOREMAT-NEXT:    ld ra, 328(sp) # 8-byte Folded Reload
599; NOREMAT-NEXT:    add ra, a1, ra
600; NOREMAT-NEXT:    sd ra, 520(sp) # 8-byte Folded Spill
601; NOREMAT-NEXT:    ld ra, 320(sp) # 8-byte Folded Reload
602; NOREMAT-NEXT:    add ra, a1, ra
603; NOREMAT-NEXT:    sd ra, 528(sp) # 8-byte Folded Spill
604; NOREMAT-NEXT:    ld ra, 312(sp) # 8-byte Folded Reload
605; NOREMAT-NEXT:    add ra, a1, ra
606; NOREMAT-NEXT:    sd ra, 536(sp) # 8-byte Folded Spill
607; NOREMAT-NEXT:    ld ra, 304(sp) # 8-byte Folded Reload
608; NOREMAT-NEXT:    add ra, a1, ra
609; NOREMAT-NEXT:    sd ra, 544(sp) # 8-byte Folded Spill
610; NOREMAT-NEXT:    ld ra, 296(sp) # 8-byte Folded Reload
611; NOREMAT-NEXT:    add ra, a1, ra
612; NOREMAT-NEXT:    sd ra, 552(sp) # 8-byte Folded Spill
613; NOREMAT-NEXT:    ld ra, 288(sp) # 8-byte Folded Reload
614; NOREMAT-NEXT:    add ra, a1, ra
615; NOREMAT-NEXT:    sd ra, 560(sp) # 8-byte Folded Spill
616; NOREMAT-NEXT:    ld ra, 280(sp) # 8-byte Folded Reload
617; NOREMAT-NEXT:    add ra, a1, ra
618; NOREMAT-NEXT:    sd ra, 568(sp) # 8-byte Folded Spill
619; NOREMAT-NEXT:    ld ra, 192(sp) # 8-byte Folded Reload
620; NOREMAT-NEXT:    add ra, a1, ra
621; NOREMAT-NEXT:    sd ra, 576(sp) # 8-byte Folded Spill
622; NOREMAT-NEXT:    ld ra, 176(sp) # 8-byte Folded Reload
623; NOREMAT-NEXT:    add ra, a1, ra
624; NOREMAT-NEXT:    sd ra, 584(sp) # 8-byte Folded Spill
625; NOREMAT-NEXT:    ld ra, 160(sp) # 8-byte Folded Reload
626; NOREMAT-NEXT:    add ra, a1, ra
627; NOREMAT-NEXT:    sd ra, 592(sp) # 8-byte Folded Spill
628; NOREMAT-NEXT:    ld ra, 128(sp) # 8-byte Folded Reload
629; NOREMAT-NEXT:    add ra, a1, ra
630; NOREMAT-NEXT:    sd ra, 600(sp) # 8-byte Folded Spill
631; NOREMAT-NEXT:    ld ra, 88(sp) # 8-byte Folded Reload
632; NOREMAT-NEXT:    add ra, a1, ra
633; NOREMAT-NEXT:    sd ra, 608(sp) # 8-byte Folded Spill
634; NOREMAT-NEXT:    ld ra, 72(sp) # 8-byte Folded Reload
635; NOREMAT-NEXT:    add ra, a1, ra
636; NOREMAT-NEXT:    sd ra, 616(sp) # 8-byte Folded Spill
637; NOREMAT-NEXT:    ld ra, 40(sp) # 8-byte Folded Reload
638; NOREMAT-NEXT:    add ra, a1, ra
639; NOREMAT-NEXT:    sd ra, 624(sp) # 8-byte Folded Spill
640; NOREMAT-NEXT:    add ra, a1, s11
641; NOREMAT-NEXT:    add s11, a1, s7
642; NOREMAT-NEXT:    add s7, a1, s3
643; NOREMAT-NEXT:    add s3, a1, s1
644; NOREMAT-NEXT:    add s1, a1, t2
645; NOREMAT-NEXT:    add t2, a1, a7
646; NOREMAT-NEXT:    add a7, a1, a4
647; NOREMAT-NEXT:    add a4, a1, a0
648; NOREMAT-NEXT:    addi a0, a1, 1536
649; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
650; NOREMAT-NEXT:    vse32.v v8, (a0)
651; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
652; NOREMAT-NEXT:    vse32.v v8, (a2)
653; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
654; NOREMAT-NEXT:    vse32.v v8, (a3)
655; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
656; NOREMAT-NEXT:    vse32.v v8, (a5)
657; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
658; NOREMAT-NEXT:    vse32.v v8, (a6)
659; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
660; NOREMAT-NEXT:    ld a0, 272(sp) # 8-byte Folded Reload
661; NOREMAT-NEXT:    vse32.v v8, (a0)
662; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
663; NOREMAT-NEXT:    vse32.v v8, (t0)
664; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
665; NOREMAT-NEXT:    vse32.v v8, (t1)
666; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
667; NOREMAT-NEXT:    vse32.v v8, (t3)
668; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
669; NOREMAT-NEXT:    vse32.v v8, (t4)
670; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
671; NOREMAT-NEXT:    vse32.v v8, (t5)
672; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
673; NOREMAT-NEXT:    vse32.v v8, (t6)
674; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
675; NOREMAT-NEXT:    vse32.v v8, (s0)
676; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
677; NOREMAT-NEXT:    ld a0, 264(sp) # 8-byte Folded Reload
678; NOREMAT-NEXT:    vse32.v v8, (a0)
679; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
680; NOREMAT-NEXT:    vse32.v v8, (s2)
681; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
682; NOREMAT-NEXT:    vse32.v v8, (s4)
683; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
684; NOREMAT-NEXT:    vse32.v v8, (s5)
685; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
686; NOREMAT-NEXT:    vse32.v v8, (s6)
687; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
688; NOREMAT-NEXT:    vse32.v v8, (s8)
689; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
690; NOREMAT-NEXT:    vse32.v v8, (s9)
691; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
692; NOREMAT-NEXT:    vse32.v v8, (s10)
693; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
694; NOREMAT-NEXT:    ld a0, 256(sp) # 8-byte Folded Reload
695; NOREMAT-NEXT:    vse32.v v8, (a0)
696; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
697; NOREMAT-NEXT:    ld a0, 16(sp) # 8-byte Folded Reload
698; NOREMAT-NEXT:    vse32.v v8, (a0)
699; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
700; NOREMAT-NEXT:    ld a0, 24(sp) # 8-byte Folded Reload
701; NOREMAT-NEXT:    vse32.v v8, (a0)
702; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
703; NOREMAT-NEXT:    ld a0, 32(sp) # 8-byte Folded Reload
704; NOREMAT-NEXT:    vse32.v v8, (a0)
705; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
706; NOREMAT-NEXT:    ld a0, 48(sp) # 8-byte Folded Reload
707; NOREMAT-NEXT:    vse32.v v8, (a0)
708; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
709; NOREMAT-NEXT:    ld a0, 56(sp) # 8-byte Folded Reload
710; NOREMAT-NEXT:    vse32.v v8, (a0)
711; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
712; NOREMAT-NEXT:    ld a0, 64(sp) # 8-byte Folded Reload
713; NOREMAT-NEXT:    vse32.v v8, (a0)
714; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
715; NOREMAT-NEXT:    ld a0, 80(sp) # 8-byte Folded Reload
716; NOREMAT-NEXT:    vse32.v v8, (a0)
717; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
718; NOREMAT-NEXT:    ld a0, 248(sp) # 8-byte Folded Reload
719; NOREMAT-NEXT:    vse32.v v8, (a0)
720; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
721; NOREMAT-NEXT:    ld a0, 96(sp) # 8-byte Folded Reload
722; NOREMAT-NEXT:    vse32.v v8, (a0)
723; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
724; NOREMAT-NEXT:    ld a0, 104(sp) # 8-byte Folded Reload
725; NOREMAT-NEXT:    vse32.v v8, (a0)
726; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
727; NOREMAT-NEXT:    ld a0, 112(sp) # 8-byte Folded Reload
728; NOREMAT-NEXT:    vse32.v v8, (a0)
729; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
730; NOREMAT-NEXT:    ld a0, 120(sp) # 8-byte Folded Reload
731; NOREMAT-NEXT:    vse32.v v8, (a0)
732; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
733; NOREMAT-NEXT:    ld a0, 136(sp) # 8-byte Folded Reload
734; NOREMAT-NEXT:    vse32.v v8, (a0)
735; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
736; NOREMAT-NEXT:    ld a0, 144(sp) # 8-byte Folded Reload
737; NOREMAT-NEXT:    vse32.v v8, (a0)
738; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
739; NOREMAT-NEXT:    ld a0, 152(sp) # 8-byte Folded Reload
740; NOREMAT-NEXT:    vse32.v v8, (a0)
741; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
742; NOREMAT-NEXT:    ld a0, 240(sp) # 8-byte Folded Reload
743; NOREMAT-NEXT:    vse32.v v8, (a0)
744; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
745; NOREMAT-NEXT:    ld a0, 168(sp) # 8-byte Folded Reload
746; NOREMAT-NEXT:    vse32.v v8, (a0)
747; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
748; NOREMAT-NEXT:    ld a0, 184(sp) # 8-byte Folded Reload
749; NOREMAT-NEXT:    vse32.v v8, (a0)
750; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
751; NOREMAT-NEXT:    ld a0, 424(sp) # 8-byte Folded Reload
752; NOREMAT-NEXT:    vse32.v v8, (a0)
753; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
754; NOREMAT-NEXT:    ld a0, 432(sp) # 8-byte Folded Reload
755; NOREMAT-NEXT:    vse32.v v8, (a0)
756; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
757; NOREMAT-NEXT:    ld a0, 440(sp) # 8-byte Folded Reload
758; NOREMAT-NEXT:    vse32.v v8, (a0)
759; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
760; NOREMAT-NEXT:    ld a0, 448(sp) # 8-byte Folded Reload
761; NOREMAT-NEXT:    vse32.v v8, (a0)
762; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
763; NOREMAT-NEXT:    ld a0, 456(sp) # 8-byte Folded Reload
764; NOREMAT-NEXT:    vse32.v v8, (a0)
765; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
766; NOREMAT-NEXT:    ld a0, 232(sp) # 8-byte Folded Reload
767; NOREMAT-NEXT:    vse32.v v8, (a0)
768; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
769; NOREMAT-NEXT:    ld a0, 464(sp) # 8-byte Folded Reload
770; NOREMAT-NEXT:    vse32.v v8, (a0)
771; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
772; NOREMAT-NEXT:    ld a0, 472(sp) # 8-byte Folded Reload
773; NOREMAT-NEXT:    vse32.v v8, (a0)
774; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
775; NOREMAT-NEXT:    ld a0, 480(sp) # 8-byte Folded Reload
776; NOREMAT-NEXT:    vse32.v v8, (a0)
777; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
778; NOREMAT-NEXT:    ld a0, 488(sp) # 8-byte Folded Reload
779; NOREMAT-NEXT:    vse32.v v8, (a0)
780; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
781; NOREMAT-NEXT:    ld a0, 496(sp) # 8-byte Folded Reload
782; NOREMAT-NEXT:    vse32.v v8, (a0)
783; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
784; NOREMAT-NEXT:    ld a0, 504(sp) # 8-byte Folded Reload
785; NOREMAT-NEXT:    vse32.v v8, (a0)
786; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
787; NOREMAT-NEXT:    ld a0, 512(sp) # 8-byte Folded Reload
788; NOREMAT-NEXT:    vse32.v v8, (a0)
789; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
790; NOREMAT-NEXT:    ld a0, 224(sp) # 8-byte Folded Reload
791; NOREMAT-NEXT:    vse32.v v8, (a0)
792; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
793; NOREMAT-NEXT:    ld a0, 520(sp) # 8-byte Folded Reload
794; NOREMAT-NEXT:    vse32.v v8, (a0)
795; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
796; NOREMAT-NEXT:    ld a0, 528(sp) # 8-byte Folded Reload
797; NOREMAT-NEXT:    vse32.v v8, (a0)
798; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
799; NOREMAT-NEXT:    ld a0, 536(sp) # 8-byte Folded Reload
800; NOREMAT-NEXT:    vse32.v v8, (a0)
801; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
802; NOREMAT-NEXT:    ld a0, 544(sp) # 8-byte Folded Reload
803; NOREMAT-NEXT:    vse32.v v8, (a0)
804; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
805; NOREMAT-NEXT:    ld a0, 552(sp) # 8-byte Folded Reload
806; NOREMAT-NEXT:    vse32.v v8, (a0)
807; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
808; NOREMAT-NEXT:    ld a0, 560(sp) # 8-byte Folded Reload
809; NOREMAT-NEXT:    vse32.v v8, (a0)
810; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
811; NOREMAT-NEXT:    ld a0, 568(sp) # 8-byte Folded Reload
812; NOREMAT-NEXT:    vse32.v v8, (a0)
813; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
814; NOREMAT-NEXT:    ld a0, 216(sp) # 8-byte Folded Reload
815; NOREMAT-NEXT:    vse32.v v8, (a0)
816; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
817; NOREMAT-NEXT:    ld a0, 576(sp) # 8-byte Folded Reload
818; NOREMAT-NEXT:    vse32.v v8, (a0)
819; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
820; NOREMAT-NEXT:    ld a0, 584(sp) # 8-byte Folded Reload
821; NOREMAT-NEXT:    vse32.v v8, (a0)
822; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
823; NOREMAT-NEXT:    ld a0, 592(sp) # 8-byte Folded Reload
824; NOREMAT-NEXT:    vse32.v v8, (a0)
825; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
826; NOREMAT-NEXT:    ld a0, 600(sp) # 8-byte Folded Reload
827; NOREMAT-NEXT:    vse32.v v8, (a0)
828; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
829; NOREMAT-NEXT:    ld a0, 608(sp) # 8-byte Folded Reload
830; NOREMAT-NEXT:    vse32.v v8, (a0)
831; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
832; NOREMAT-NEXT:    ld a0, 616(sp) # 8-byte Folded Reload
833; NOREMAT-NEXT:    vse32.v v8, (a0)
834; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
835; NOREMAT-NEXT:    ld a0, 624(sp) # 8-byte Folded Reload
836; NOREMAT-NEXT:    vse32.v v8, (a0)
837; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
838; NOREMAT-NEXT:    ld a0, 208(sp) # 8-byte Folded Reload
839; NOREMAT-NEXT:    vse32.v v8, (a0)
840; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
841; NOREMAT-NEXT:    vse32.v v8, (ra)
842; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
843; NOREMAT-NEXT:    vse32.v v8, (s11)
844; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
845; NOREMAT-NEXT:    vse32.v v8, (s7)
846; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
847; NOREMAT-NEXT:    vse32.v v8, (s3)
848; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
849; NOREMAT-NEXT:    vse32.v v8, (s1)
850; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
851; NOREMAT-NEXT:    vse32.v v8, (t2)
852; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
853; NOREMAT-NEXT:    vse32.v v8, (a7)
854; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
855; NOREMAT-NEXT:    ld a0, 200(sp) # 8-byte Folded Reload
856; NOREMAT-NEXT:    vse32.v v8, (a0)
857; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
858; NOREMAT-NEXT:    vse32.v v8, (a4)
859; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
860; NOREMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
861; NOREMAT-NEXT:    csrr a0, vlenb
862; NOREMAT-NEXT:    slli a0, a0, 1
863; NOREMAT-NEXT:    add sp, sp, a0
864; NOREMAT-NEXT:    .cfi_def_cfa sp, 752
865; NOREMAT-NEXT:    ld ra, 744(sp) # 8-byte Folded Reload
866; NOREMAT-NEXT:    ld s0, 736(sp) # 8-byte Folded Reload
867; NOREMAT-NEXT:    ld s1, 728(sp) # 8-byte Folded Reload
868; NOREMAT-NEXT:    ld s2, 720(sp) # 8-byte Folded Reload
869; NOREMAT-NEXT:    ld s3, 712(sp) # 8-byte Folded Reload
870; NOREMAT-NEXT:    ld s4, 704(sp) # 8-byte Folded Reload
871; NOREMAT-NEXT:    ld s5, 696(sp) # 8-byte Folded Reload
872; NOREMAT-NEXT:    ld s6, 688(sp) # 8-byte Folded Reload
873; NOREMAT-NEXT:    ld s7, 680(sp) # 8-byte Folded Reload
874; NOREMAT-NEXT:    ld s8, 672(sp) # 8-byte Folded Reload
875; NOREMAT-NEXT:    ld s9, 664(sp) # 8-byte Folded Reload
876; NOREMAT-NEXT:    ld s10, 656(sp) # 8-byte Folded Reload
877; NOREMAT-NEXT:    ld s11, 648(sp) # 8-byte Folded Reload
878; NOREMAT-NEXT:    .cfi_restore ra
879; NOREMAT-NEXT:    .cfi_restore s0
880; NOREMAT-NEXT:    .cfi_restore s1
881; NOREMAT-NEXT:    .cfi_restore s2
882; NOREMAT-NEXT:    .cfi_restore s3
883; NOREMAT-NEXT:    .cfi_restore s4
884; NOREMAT-NEXT:    .cfi_restore s5
885; NOREMAT-NEXT:    .cfi_restore s6
886; NOREMAT-NEXT:    .cfi_restore s7
887; NOREMAT-NEXT:    .cfi_restore s8
888; NOREMAT-NEXT:    .cfi_restore s9
889; NOREMAT-NEXT:    .cfi_restore s10
890; NOREMAT-NEXT:    .cfi_restore s11
891; NOREMAT-NEXT:    addi sp, sp, 752
892; NOREMAT-NEXT:    .cfi_def_cfa_offset 0
893; NOREMAT-NEXT:    ret
894;
895; REMAT-LABEL: test:
896; REMAT:       # %bb.0:
897; REMAT-NEXT:    addi sp, sp, -544
898; REMAT-NEXT:    .cfi_def_cfa_offset 544
899; REMAT-NEXT:    sd ra, 536(sp) # 8-byte Folded Spill
900; REMAT-NEXT:    sd s0, 528(sp) # 8-byte Folded Spill
901; REMAT-NEXT:    sd s1, 520(sp) # 8-byte Folded Spill
902; REMAT-NEXT:    sd s2, 512(sp) # 8-byte Folded Spill
903; REMAT-NEXT:    sd s3, 504(sp) # 8-byte Folded Spill
904; REMAT-NEXT:    sd s4, 496(sp) # 8-byte Folded Spill
905; REMAT-NEXT:    sd s5, 488(sp) # 8-byte Folded Spill
906; REMAT-NEXT:    sd s6, 480(sp) # 8-byte Folded Spill
907; REMAT-NEXT:    sd s7, 472(sp) # 8-byte Folded Spill
908; REMAT-NEXT:    sd s8, 464(sp) # 8-byte Folded Spill
909; REMAT-NEXT:    sd s9, 456(sp) # 8-byte Folded Spill
910; REMAT-NEXT:    sd s10, 448(sp) # 8-byte Folded Spill
911; REMAT-NEXT:    sd s11, 440(sp) # 8-byte Folded Spill
912; REMAT-NEXT:    .cfi_offset ra, -8
913; REMAT-NEXT:    .cfi_offset s0, -16
914; REMAT-NEXT:    .cfi_offset s1, -24
915; REMAT-NEXT:    .cfi_offset s2, -32
916; REMAT-NEXT:    .cfi_offset s3, -40
917; REMAT-NEXT:    .cfi_offset s4, -48
918; REMAT-NEXT:    .cfi_offset s5, -56
919; REMAT-NEXT:    .cfi_offset s6, -64
920; REMAT-NEXT:    .cfi_offset s7, -72
921; REMAT-NEXT:    .cfi_offset s8, -80
922; REMAT-NEXT:    .cfi_offset s9, -88
923; REMAT-NEXT:    .cfi_offset s10, -96
924; REMAT-NEXT:    .cfi_offset s11, -104
925; REMAT-NEXT:    csrr a2, vlenb
926; REMAT-NEXT:    li a3, 18
927; REMAT-NEXT:    mul a2, a2, a3
928; REMAT-NEXT:    sub sp, sp, a2
929; REMAT-NEXT:    .cfi_escape 0x0f, 0x0e, 0x72, 0x00, 0x11, 0xa0, 0x04, 0x22, 0x11, 0x12, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 544 + 18 * vlenb
930; REMAT-NEXT:    li a4, 32
931; REMAT-NEXT:    addi a5, a0, 512
932; REMAT-NEXT:    addi a3, a0, 1024
933; REMAT-NEXT:    addi a2, a0, 1536
934; REMAT-NEXT:    li a6, 1
935; REMAT-NEXT:    slli a6, a6, 11
936; REMAT-NEXT:    li a7, 5
937; REMAT-NEXT:    slli a7, a7, 9
938; REMAT-NEXT:    li t0, 3
939; REMAT-NEXT:    slli t0, t0, 10
940; REMAT-NEXT:    li t1, 7
941; REMAT-NEXT:    slli t1, t1, 9
942; REMAT-NEXT:    lui t2, 1
943; REMAT-NEXT:    li t3, 9
944; REMAT-NEXT:    slli t3, t3, 9
945; REMAT-NEXT:    li t4, 5
946; REMAT-NEXT:    slli t4, t4, 10
947; REMAT-NEXT:    li t5, 11
948; REMAT-NEXT:    slli t5, t5, 9
949; REMAT-NEXT:    li t6, 3
950; REMAT-NEXT:    slli t6, t6, 11
951; REMAT-NEXT:    li s0, 13
952; REMAT-NEXT:    slli s0, s0, 9
953; REMAT-NEXT:    li s1, 7
954; REMAT-NEXT:    slli s1, s1, 10
955; REMAT-NEXT:    li s2, 15
956; REMAT-NEXT:    slli s2, s2, 9
957; REMAT-NEXT:    lui s3, 2
958; REMAT-NEXT:    li s4, 17
959; REMAT-NEXT:    slli s4, s4, 9
960; REMAT-NEXT:    li s5, 9
961; REMAT-NEXT:    slli s5, s5, 10
962; REMAT-NEXT:    li s6, 19
963; REMAT-NEXT:    slli s6, s6, 9
964; REMAT-NEXT:    li s7, 5
965; REMAT-NEXT:    slli s7, s7, 11
966; REMAT-NEXT:    li s8, 21
967; REMAT-NEXT:    slli s8, s8, 9
968; REMAT-NEXT:    li s9, 11
969; REMAT-NEXT:    slli s9, s9, 10
970; REMAT-NEXT:    li s10, 23
971; REMAT-NEXT:    slli s10, s10, 9
972; REMAT-NEXT:    lui s11, 3
973; REMAT-NEXT:    li ra, 25
974; REMAT-NEXT:    slli ra, ra, 9
975; REMAT-NEXT:    vsetvli zero, a4, e32, m2, ta, ma
976; REMAT-NEXT:    vle32.v v8, (a5)
977; REMAT-NEXT:    li a4, 13
978; REMAT-NEXT:    slli a4, a4, 10
979; REMAT-NEXT:    vle32.v v10, (a3)
980; REMAT-NEXT:    vle32.v v12, (a3)
981; REMAT-NEXT:    li a3, 27
982; REMAT-NEXT:    slli a3, a3, 9
983; REMAT-NEXT:    vle32.v v14, (a2)
984; REMAT-NEXT:    vle32.v v16, (a2)
985; REMAT-NEXT:    add a2, a0, a6
986; REMAT-NEXT:    vle32.v v18, (a2)
987; REMAT-NEXT:    vle32.v v20, (a2)
988; REMAT-NEXT:    add a2, a0, a7
989; REMAT-NEXT:    vle32.v v22, (a2)
990; REMAT-NEXT:    vle32.v v24, (a2)
991; REMAT-NEXT:    add a2, a0, t0
992; REMAT-NEXT:    vle32.v v26, (a2)
993; REMAT-NEXT:    vle32.v v28, (a2)
994; REMAT-NEXT:    add a2, a0, t1
995; REMAT-NEXT:    vle32.v v30, (a2)
996; REMAT-NEXT:    vle32.v v6, (a2)
997; REMAT-NEXT:    csrr a2, vlenb
998; REMAT-NEXT:    slli a2, a2, 4
999; REMAT-NEXT:    add a2, sp, a2
1000; REMAT-NEXT:    addi a2, a2, 432
1001; REMAT-NEXT:    vs2r.v v6, (a2) # Unknown-size Folded Spill
1002; REMAT-NEXT:    add a2, a0, t2
1003; REMAT-NEXT:    vle32.v v4, (a0)
1004; REMAT-NEXT:    vle32.v v2, (a2)
1005; REMAT-NEXT:    vle32.v v6, (a2)
1006; REMAT-NEXT:    csrr a2, vlenb
1007; REMAT-NEXT:    li a5, 14
1008; REMAT-NEXT:    mul a2, a2, a5
1009; REMAT-NEXT:    add a2, sp, a2
1010; REMAT-NEXT:    addi a2, a2, 432
1011; REMAT-NEXT:    vs2r.v v6, (a2) # Unknown-size Folded Spill
1012; REMAT-NEXT:    add a2, a0, t3
1013; REMAT-NEXT:    sf.vc.vv 3, 0, v4, v8
1014; REMAT-NEXT:    vle32.v v4, (a2)
1015; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v10
1016; REMAT-NEXT:    vle32.v v6, (a2)
1017; REMAT-NEXT:    add a2, a0, t4
1018; REMAT-NEXT:    vle32.v v10, (a2)
1019; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v14
1020; REMAT-NEXT:    vle32.v v0, (a2)
1021; REMAT-NEXT:    add a2, a0, t5
1022; REMAT-NEXT:    vle32.v v14, (a2)
1023; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v18
1024; REMAT-NEXT:    vle32.v v8, (a2)
1025; REMAT-NEXT:    csrr a2, vlenb
1026; REMAT-NEXT:    li a5, 12
1027; REMAT-NEXT:    mul a2, a2, a5
1028; REMAT-NEXT:    add a2, sp, a2
1029; REMAT-NEXT:    addi a2, a2, 432
1030; REMAT-NEXT:    vs2r.v v8, (a2) # Unknown-size Folded Spill
1031; REMAT-NEXT:    add a2, a0, t6
1032; REMAT-NEXT:    vle32.v v18, (a2)
1033; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v22
1034; REMAT-NEXT:    vle32.v v20, (a2)
1035; REMAT-NEXT:    add a2, a0, s0
1036; REMAT-NEXT:    vle32.v v22, (a2)
1037; REMAT-NEXT:    sf.vc.vv 3, 0, v24, v26
1038; REMAT-NEXT:    vle32.v v24, (a2)
1039; REMAT-NEXT:    add a2, a0, s1
1040; REMAT-NEXT:    vle32.v v26, (a2)
1041; REMAT-NEXT:    sf.vc.vv 3, 0, v28, v30
1042; REMAT-NEXT:    vle32.v v28, (a2)
1043; REMAT-NEXT:    add a2, a0, s2
1044; REMAT-NEXT:    vle32.v v8, (a2)
1045; REMAT-NEXT:    csrr a5, vlenb
1046; REMAT-NEXT:    slli a5, a5, 4
1047; REMAT-NEXT:    add a5, sp, a5
1048; REMAT-NEXT:    addi a5, a5, 432
1049; REMAT-NEXT:    vl2r.v v12, (a5) # Unknown-size Folded Reload
1050; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v2
1051; REMAT-NEXT:    vle32.v v2, (a2)
1052; REMAT-NEXT:    add a2, a0, s3
1053; REMAT-NEXT:    vle32.v v12, (a2)
1054; REMAT-NEXT:    csrr a5, vlenb
1055; REMAT-NEXT:    li a6, 14
1056; REMAT-NEXT:    mul a5, a5, a6
1057; REMAT-NEXT:    add a5, sp, a5
1058; REMAT-NEXT:    addi a5, a5, 432
1059; REMAT-NEXT:    vl2r.v v16, (a5) # Unknown-size Folded Reload
1060; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1061; REMAT-NEXT:    vle32.v v30, (a2)
1062; REMAT-NEXT:    add a2, a0, s4
1063; REMAT-NEXT:    vle32.v v16, (a2)
1064; REMAT-NEXT:    sf.vc.vv 3, 0, v6, v10
1065; REMAT-NEXT:    vle32.v v6, (a2)
1066; REMAT-NEXT:    add a2, a0, s5
1067; REMAT-NEXT:    vle32.v v10, (a2)
1068; REMAT-NEXT:    sf.vc.vv 3, 0, v0, v14
1069; REMAT-NEXT:    vle32.v v4, (a2)
1070; REMAT-NEXT:    add a2, a0, s6
1071; REMAT-NEXT:    vle32.v v14, (a2)
1072; REMAT-NEXT:    csrr a5, vlenb
1073; REMAT-NEXT:    li a6, 12
1074; REMAT-NEXT:    mul a5, a5, a6
1075; REMAT-NEXT:    add a5, sp, a5
1076; REMAT-NEXT:    addi a5, a5, 432
1077; REMAT-NEXT:    vl2r.v v0, (a5) # Unknown-size Folded Reload
1078; REMAT-NEXT:    sf.vc.vv 3, 0, v0, v18
1079; REMAT-NEXT:    vle32.v v0, (a2)
1080; REMAT-NEXT:    add a2, a0, s7
1081; REMAT-NEXT:    vle32.v v18, (a2)
1082; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v22
1083; REMAT-NEXT:    vle32.v v22, (a2)
1084; REMAT-NEXT:    add a2, a0, s8
1085; REMAT-NEXT:    vle32.v v20, (a2)
1086; REMAT-NEXT:    sf.vc.vv 3, 0, v24, v26
1087; REMAT-NEXT:    vle32.v v26, (a2)
1088; REMAT-NEXT:    add a2, a0, s9
1089; REMAT-NEXT:    vle32.v v24, (a2)
1090; REMAT-NEXT:    sf.vc.vv 3, 0, v28, v8
1091; REMAT-NEXT:    vle32.v v28, (a2)
1092; REMAT-NEXT:    add a2, a0, s10
1093; REMAT-NEXT:    vle32.v v8, (a2)
1094; REMAT-NEXT:    sf.vc.vv 3, 0, v2, v12
1095; REMAT-NEXT:    vle32.v v12, (a2)
1096; REMAT-NEXT:    csrr a2, vlenb
1097; REMAT-NEXT:    slli a2, a2, 3
1098; REMAT-NEXT:    add a2, sp, a2
1099; REMAT-NEXT:    addi a2, a2, 432
1100; REMAT-NEXT:    vs2r.v v12, (a2) # Unknown-size Folded Spill
1101; REMAT-NEXT:    add a2, a0, s11
1102; REMAT-NEXT:    vle32.v v12, (a2)
1103; REMAT-NEXT:    sf.vc.vv 3, 0, v30, v16
1104; REMAT-NEXT:    vle32.v v16, (a2)
1105; REMAT-NEXT:    add a2, a0, ra
1106; REMAT-NEXT:    vle32.v v2, (a2)
1107; REMAT-NEXT:    sf.vc.vv 3, 0, v6, v10
1108; REMAT-NEXT:    vle32.v v10, (a2)
1109; REMAT-NEXT:    csrr a2, vlenb
1110; REMAT-NEXT:    slli a2, a2, 1
1111; REMAT-NEXT:    add a2, sp, a2
1112; REMAT-NEXT:    addi a2, a2, 432
1113; REMAT-NEXT:    vs2r.v v10, (a2) # Unknown-size Folded Spill
1114; REMAT-NEXT:    add a2, a0, a4
1115; REMAT-NEXT:    vle32.v v10, (a2)
1116; REMAT-NEXT:    sf.vc.vv 3, 0, v4, v14
1117; REMAT-NEXT:    vle32.v v14, (a2)
1118; REMAT-NEXT:    csrr a2, vlenb
1119; REMAT-NEXT:    slli a2, a2, 2
1120; REMAT-NEXT:    add a2, sp, a2
1121; REMAT-NEXT:    addi a2, a2, 432
1122; REMAT-NEXT:    vs2r.v v14, (a2) # Unknown-size Folded Spill
1123; REMAT-NEXT:    add a2, a0, a3
1124; REMAT-NEXT:    vle32.v v14, (a2)
1125; REMAT-NEXT:    sf.vc.vv 3, 0, v0, v18
1126; REMAT-NEXT:    vle32.v v18, (a2)
1127; REMAT-NEXT:    csrr a2, vlenb
1128; REMAT-NEXT:    slli a2, a2, 4
1129; REMAT-NEXT:    add a2, sp, a2
1130; REMAT-NEXT:    addi a2, a2, 432
1131; REMAT-NEXT:    vs2r.v v18, (a2) # Unknown-size Folded Spill
1132; REMAT-NEXT:    li a5, 7
1133; REMAT-NEXT:    slli a5, a5, 11
1134; REMAT-NEXT:    add a2, a0, a5
1135; REMAT-NEXT:    vle32.v v18, (a2)
1136; REMAT-NEXT:    addi a3, sp, 432
1137; REMAT-NEXT:    vs2r.v v18, (a3) # Unknown-size Folded Spill
1138; REMAT-NEXT:    sf.vc.vv 3, 0, v22, v20
1139; REMAT-NEXT:    vle32.v v18, (a2)
1140; REMAT-NEXT:    csrr a2, vlenb
1141; REMAT-NEXT:    li a3, 14
1142; REMAT-NEXT:    mul a2, a2, a3
1143; REMAT-NEXT:    add a2, sp, a2
1144; REMAT-NEXT:    addi a2, a2, 432
1145; REMAT-NEXT:    vs2r.v v18, (a2) # Unknown-size Folded Spill
1146; REMAT-NEXT:    li a2, 29
1147; REMAT-NEXT:    slli a2, a2, 9
1148; REMAT-NEXT:    add a2, a0, a2
1149; REMAT-NEXT:    vle32.v v18, (a2)
1150; REMAT-NEXT:    sf.vc.vv 3, 0, v26, v24
1151; REMAT-NEXT:    vle32.v v20, (a2)
1152; REMAT-NEXT:    csrr a2, vlenb
1153; REMAT-NEXT:    li a3, 12
1154; REMAT-NEXT:    mul a2, a2, a3
1155; REMAT-NEXT:    add a2, sp, a2
1156; REMAT-NEXT:    addi a2, a2, 432
1157; REMAT-NEXT:    vs2r.v v20, (a2) # Unknown-size Folded Spill
1158; REMAT-NEXT:    li a2, 15
1159; REMAT-NEXT:    slli a2, a2, 10
1160; REMAT-NEXT:    add a2, a0, a2
1161; REMAT-NEXT:    vle32.v v30, (a2)
1162; REMAT-NEXT:    sf.vc.vv 3, 0, v28, v8
1163; REMAT-NEXT:    vle32.v v8, (a2)
1164; REMAT-NEXT:    csrr a2, vlenb
1165; REMAT-NEXT:    li a3, 10
1166; REMAT-NEXT:    mul a2, a2, a3
1167; REMAT-NEXT:    add a2, sp, a2
1168; REMAT-NEXT:    addi a2, a2, 432
1169; REMAT-NEXT:    vs2r.v v8, (a2) # Unknown-size Folded Spill
1170; REMAT-NEXT:    li a2, 31
1171; REMAT-NEXT:    slli a2, a2, 9
1172; REMAT-NEXT:    add a2, a0, a2
1173; REMAT-NEXT:    vle32.v v6, (a2)
1174; REMAT-NEXT:    csrr a3, vlenb
1175; REMAT-NEXT:    slli a3, a3, 3
1176; REMAT-NEXT:    add a3, sp, a3
1177; REMAT-NEXT:    addi a3, a3, 432
1178; REMAT-NEXT:    vl2r.v v8, (a3) # Unknown-size Folded Reload
1179; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v12
1180; REMAT-NEXT:    vle32.v v8, (a2)
1181; REMAT-NEXT:    csrr a2, vlenb
1182; REMAT-NEXT:    slli a2, a2, 3
1183; REMAT-NEXT:    add a2, sp, a2
1184; REMAT-NEXT:    addi a2, a2, 432
1185; REMAT-NEXT:    vs2r.v v8, (a2) # Unknown-size Folded Spill
1186; REMAT-NEXT:    lui a2, 4
1187; REMAT-NEXT:    add a2, a0, a2
1188; REMAT-NEXT:    vle32.v v4, (a2)
1189; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v2
1190; REMAT-NEXT:    vle32.v v8, (a2)
1191; REMAT-NEXT:    csrr a2, vlenb
1192; REMAT-NEXT:    li a3, 6
1193; REMAT-NEXT:    mul a2, a2, a3
1194; REMAT-NEXT:    add a2, sp, a2
1195; REMAT-NEXT:    addi a2, a2, 432
1196; REMAT-NEXT:    vs2r.v v8, (a2) # Unknown-size Folded Spill
1197; REMAT-NEXT:    lui a2, 4
1198; REMAT-NEXT:    addiw a2, a2, 512
1199; REMAT-NEXT:    add a2, a0, a2
1200; REMAT-NEXT:    vle32.v v2, (a2)
1201; REMAT-NEXT:    csrr a3, vlenb
1202; REMAT-NEXT:    slli a3, a3, 1
1203; REMAT-NEXT:    add a3, sp, a3
1204; REMAT-NEXT:    addi a3, a3, 432
1205; REMAT-NEXT:    vl2r.v v8, (a3) # Unknown-size Folded Reload
1206; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v10
1207; REMAT-NEXT:    vle32.v v20, (a2)
1208; REMAT-NEXT:    li a2, 17
1209; REMAT-NEXT:    slli a2, a2, 10
1210; REMAT-NEXT:    add a2, a0, a2
1211; REMAT-NEXT:    vle32.v v0, (a2)
1212; REMAT-NEXT:    csrr a3, vlenb
1213; REMAT-NEXT:    slli a3, a3, 2
1214; REMAT-NEXT:    add a3, sp, a3
1215; REMAT-NEXT:    addi a3, a3, 432
1216; REMAT-NEXT:    vl2r.v v8, (a3) # Unknown-size Folded Reload
1217; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v14
1218; REMAT-NEXT:    vle32.v v22, (a2)
1219; REMAT-NEXT:    lui a2, 4
1220; REMAT-NEXT:    addiw a2, a2, 1536
1221; REMAT-NEXT:    add a2, a0, a2
1222; REMAT-NEXT:    vle32.v v24, (a2)
1223; REMAT-NEXT:    csrr a3, vlenb
1224; REMAT-NEXT:    slli a3, a3, 4
1225; REMAT-NEXT:    add a3, sp, a3
1226; REMAT-NEXT:    addi a3, a3, 432
1227; REMAT-NEXT:    vl2r.v v8, (a3) # Unknown-size Folded Reload
1228; REMAT-NEXT:    addi a3, sp, 432
1229; REMAT-NEXT:    vl2r.v v10, (a3) # Unknown-size Folded Reload
1230; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v10
1231; REMAT-NEXT:    vle32.v v8, (a2)
1232; REMAT-NEXT:    li a2, 9
1233; REMAT-NEXT:    slli a2, a2, 11
1234; REMAT-NEXT:    add a2, a0, a2
1235; REMAT-NEXT:    vle32.v v26, (a2)
1236; REMAT-NEXT:    csrr a3, vlenb
1237; REMAT-NEXT:    li a4, 14
1238; REMAT-NEXT:    mul a3, a3, a4
1239; REMAT-NEXT:    add a3, sp, a3
1240; REMAT-NEXT:    addi a3, a3, 432
1241; REMAT-NEXT:    vl2r.v v10, (a3) # Unknown-size Folded Reload
1242; REMAT-NEXT:    sf.vc.vv 3, 0, v10, v18
1243; REMAT-NEXT:    vle32.v v10, (a2)
1244; REMAT-NEXT:    lui a2, 5
1245; REMAT-NEXT:    addiw a2, a2, -1536
1246; REMAT-NEXT:    add a2, a0, a2
1247; REMAT-NEXT:    vle32.v v28, (a2)
1248; REMAT-NEXT:    csrr a3, vlenb
1249; REMAT-NEXT:    li a4, 12
1250; REMAT-NEXT:    mul a3, a3, a4
1251; REMAT-NEXT:    add a3, sp, a3
1252; REMAT-NEXT:    addi a3, a3, 432
1253; REMAT-NEXT:    vl2r.v v12, (a3) # Unknown-size Folded Reload
1254; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v30
1255; REMAT-NEXT:    vle32.v v12, (a2)
1256; REMAT-NEXT:    li a2, 19
1257; REMAT-NEXT:    slli a2, a2, 10
1258; REMAT-NEXT:    add a2, a0, a2
1259; REMAT-NEXT:    vle32.v v30, (a2)
1260; REMAT-NEXT:    csrr a3, vlenb
1261; REMAT-NEXT:    li a4, 10
1262; REMAT-NEXT:    mul a3, a3, a4
1263; REMAT-NEXT:    add a3, sp, a3
1264; REMAT-NEXT:    addi a3, a3, 432
1265; REMAT-NEXT:    vl2r.v v14, (a3) # Unknown-size Folded Reload
1266; REMAT-NEXT:    sf.vc.vv 3, 0, v14, v6
1267; REMAT-NEXT:    vle32.v v14, (a2)
1268; REMAT-NEXT:    lui a2, 5
1269; REMAT-NEXT:    addiw a2, a2, -512
1270; REMAT-NEXT:    add a2, a0, a2
1271; REMAT-NEXT:    vle32.v v6, (a2)
1272; REMAT-NEXT:    csrr a3, vlenb
1273; REMAT-NEXT:    slli a3, a3, 3
1274; REMAT-NEXT:    add a3, sp, a3
1275; REMAT-NEXT:    addi a3, a3, 432
1276; REMAT-NEXT:    vl2r.v v16, (a3) # Unknown-size Folded Reload
1277; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1278; REMAT-NEXT:    vle32.v v16, (a2)
1279; REMAT-NEXT:    lui a2, 5
1280; REMAT-NEXT:    add a2, a0, a2
1281; REMAT-NEXT:    vle32.v v4, (a2)
1282; REMAT-NEXT:    csrr a3, vlenb
1283; REMAT-NEXT:    li a4, 6
1284; REMAT-NEXT:    mul a3, a3, a4
1285; REMAT-NEXT:    add a3, sp, a3
1286; REMAT-NEXT:    addi a3, a3, 432
1287; REMAT-NEXT:    vl2r.v v18, (a3) # Unknown-size Folded Reload
1288; REMAT-NEXT:    sf.vc.vv 3, 0, v18, v2
1289; REMAT-NEXT:    vle32.v v18, (a2)
1290; REMAT-NEXT:    lui a2, 5
1291; REMAT-NEXT:    addiw a2, a2, 512
1292; REMAT-NEXT:    add a2, a0, a2
1293; REMAT-NEXT:    vle32.v v2, (a2)
1294; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v0
1295; REMAT-NEXT:    vle32.v v20, (a2)
1296; REMAT-NEXT:    li s7, 21
1297; REMAT-NEXT:    slli s7, s7, 10
1298; REMAT-NEXT:    add a2, a0, s7
1299; REMAT-NEXT:    vle32.v v0, (a2)
1300; REMAT-NEXT:    sf.vc.vv 3, 0, v22, v24
1301; REMAT-NEXT:    vle32.v v22, (a2)
1302; REMAT-NEXT:    lui s4, 5
1303; REMAT-NEXT:    addiw s4, s4, 1536
1304; REMAT-NEXT:    add a2, a0, s4
1305; REMAT-NEXT:    vle32.v v24, (a2)
1306; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v26
1307; REMAT-NEXT:    vle32.v v8, (a2)
1308; REMAT-NEXT:    li a2, 11
1309; REMAT-NEXT:    slli a2, a2, 11
1310; REMAT-NEXT:    add a2, a0, a2
1311; REMAT-NEXT:    vle32.v v26, (a2)
1312; REMAT-NEXT:    sf.vc.vv 3, 0, v10, v28
1313; REMAT-NEXT:    vle32.v v10, (a2)
1314; REMAT-NEXT:    lui s3, 6
1315; REMAT-NEXT:    addiw s3, s3, -1536
1316; REMAT-NEXT:    add a2, a0, s3
1317; REMAT-NEXT:    vle32.v v28, (a2)
1318; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v30
1319; REMAT-NEXT:    vle32.v v12, (a2)
1320; REMAT-NEXT:    li s2, 23
1321; REMAT-NEXT:    slli s2, s2, 10
1322; REMAT-NEXT:    add a2, a0, s2
1323; REMAT-NEXT:    vle32.v v30, (a2)
1324; REMAT-NEXT:    sf.vc.vv 3, 0, v14, v6
1325; REMAT-NEXT:    vle32.v v14, (a2)
1326; REMAT-NEXT:    lui a2, 6
1327; REMAT-NEXT:    addiw a2, a2, -512
1328; REMAT-NEXT:    add a2, a0, a2
1329; REMAT-NEXT:    vle32.v v6, (a2)
1330; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1331; REMAT-NEXT:    vle32.v v16, (a2)
1332; REMAT-NEXT:    lui a2, 6
1333; REMAT-NEXT:    add a2, a0, a2
1334; REMAT-NEXT:    lui s1, 6
1335; REMAT-NEXT:    vle32.v v4, (a2)
1336; REMAT-NEXT:    sf.vc.vv 3, 0, v18, v2
1337; REMAT-NEXT:    vle32.v v18, (a2)
1338; REMAT-NEXT:    lui s0, 6
1339; REMAT-NEXT:    addiw s0, s0, 512
1340; REMAT-NEXT:    add a2, a0, s0
1341; REMAT-NEXT:    vle32.v v2, (a2)
1342; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v0
1343; REMAT-NEXT:    vle32.v v20, (a2)
1344; REMAT-NEXT:    li a2, 25
1345; REMAT-NEXT:    slli a2, a2, 10
1346; REMAT-NEXT:    add a2, a0, a2
1347; REMAT-NEXT:    vle32.v v0, (a2)
1348; REMAT-NEXT:    sf.vc.vv 3, 0, v22, v24
1349; REMAT-NEXT:    vle32.v v22, (a2)
1350; REMAT-NEXT:    lui t6, 6
1351; REMAT-NEXT:    addiw t6, t6, 1536
1352; REMAT-NEXT:    add a2, a0, t6
1353; REMAT-NEXT:    vle32.v v24, (a2)
1354; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v26
1355; REMAT-NEXT:    vle32.v v8, (a2)
1356; REMAT-NEXT:    li t5, 13
1357; REMAT-NEXT:    slli t5, t5, 11
1358; REMAT-NEXT:    add a2, a0, t5
1359; REMAT-NEXT:    vle32.v v26, (a2)
1360; REMAT-NEXT:    sf.vc.vv 3, 0, v10, v28
1361; REMAT-NEXT:    vle32.v v10, (a2)
1362; REMAT-NEXT:    lui a2, 7
1363; REMAT-NEXT:    addiw a2, a2, -1536
1364; REMAT-NEXT:    add a2, a0, a2
1365; REMAT-NEXT:    vle32.v v28, (a2)
1366; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v30
1367; REMAT-NEXT:    vle32.v v12, (a2)
1368; REMAT-NEXT:    li t4, 27
1369; REMAT-NEXT:    slli t4, t4, 10
1370; REMAT-NEXT:    add a2, a0, t4
1371; REMAT-NEXT:    vle32.v v30, (a2)
1372; REMAT-NEXT:    sf.vc.vv 3, 0, v14, v6
1373; REMAT-NEXT:    vle32.v v14, (a2)
1374; REMAT-NEXT:    lui a2, 7
1375; REMAT-NEXT:    addiw a2, a2, -512
1376; REMAT-NEXT:    add a2, a0, a2
1377; REMAT-NEXT:    vle32.v v6, (a2)
1378; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1379; REMAT-NEXT:    vle32.v v16, (a2)
1380; REMAT-NEXT:    lui a2, 7
1381; REMAT-NEXT:    add a2, a0, a2
1382; REMAT-NEXT:    lui t3, 7
1383; REMAT-NEXT:    vle32.v v4, (a2)
1384; REMAT-NEXT:    sf.vc.vv 3, 0, v18, v2
1385; REMAT-NEXT:    vle32.v v18, (a2)
1386; REMAT-NEXT:    lui t2, 7
1387; REMAT-NEXT:    addiw t2, t2, 512
1388; REMAT-NEXT:    add a2, a0, t2
1389; REMAT-NEXT:    vle32.v v2, (a2)
1390; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v0
1391; REMAT-NEXT:    vle32.v v20, (a2)
1392; REMAT-NEXT:    li t1, 29
1393; REMAT-NEXT:    slli t1, t1, 10
1394; REMAT-NEXT:    add a2, a0, t1
1395; REMAT-NEXT:    vle32.v v0, (a2)
1396; REMAT-NEXT:    sf.vc.vv 3, 0, v22, v24
1397; REMAT-NEXT:    vle32.v v22, (a2)
1398; REMAT-NEXT:    lui t0, 7
1399; REMAT-NEXT:    addiw t0, t0, 1536
1400; REMAT-NEXT:    add a2, a0, t0
1401; REMAT-NEXT:    vle32.v v24, (a2)
1402; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v26
1403; REMAT-NEXT:    vle32.v v8, (a2)
1404; REMAT-NEXT:    li a7, 15
1405; REMAT-NEXT:    slli a7, a7, 11
1406; REMAT-NEXT:    add a2, a0, a7
1407; REMAT-NEXT:    vle32.v v26, (a2)
1408; REMAT-NEXT:    sf.vc.vv 3, 0, v10, v28
1409; REMAT-NEXT:    vle32.v v10, (a2)
1410; REMAT-NEXT:    lui a6, 8
1411; REMAT-NEXT:    addiw a6, a6, -1536
1412; REMAT-NEXT:    add a2, a0, a6
1413; REMAT-NEXT:    vle32.v v28, (a2)
1414; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v30
1415; REMAT-NEXT:    vle32.v v12, (a2)
1416; REMAT-NEXT:    li a4, 31
1417; REMAT-NEXT:    slli a4, a4, 10
1418; REMAT-NEXT:    add a2, a0, a4
1419; REMAT-NEXT:    vle32.v v30, (a2)
1420; REMAT-NEXT:    sf.vc.vv 3, 0, v14, v6
1421; REMAT-NEXT:    vle32.v v14, (a2)
1422; REMAT-NEXT:    lui a3, 8
1423; REMAT-NEXT:    addiw a3, a3, -512
1424; REMAT-NEXT:    add a2, a0, a3
1425; REMAT-NEXT:    vle32.v v6, (a2)
1426; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1427; REMAT-NEXT:    vle32.v v16, (a2)
1428; REMAT-NEXT:    lui a2, 8
1429; REMAT-NEXT:    add a0, a0, a2
1430; REMAT-NEXT:    vle32.v v4, (a0)
1431; REMAT-NEXT:    sf.vc.vv 3, 0, v18, v2
1432; REMAT-NEXT:    sf.vc.vv 3, 0, v20, v0
1433; REMAT-NEXT:    sf.vc.vv 3, 0, v22, v24
1434; REMAT-NEXT:    sf.vc.vv 3, 0, v8, v26
1435; REMAT-NEXT:    sf.vc.vv 3, 0, v10, v28
1436; REMAT-NEXT:    sf.vc.vv 3, 0, v12, v30
1437; REMAT-NEXT:    sf.vc.vv 3, 0, v14, v6
1438; REMAT-NEXT:    sf.vc.vv 3, 0, v16, v4
1439; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1440; REMAT-NEXT:    addi a0, a1, 1024
1441; REMAT-NEXT:    vse32.v v8, (a0)
1442; REMAT-NEXT:    li a0, 1
1443; REMAT-NEXT:    slli a0, a0, 11
1444; REMAT-NEXT:    add a0, a1, a0
1445; REMAT-NEXT:    sd a0, 416(sp) # 8-byte Folded Spill
1446; REMAT-NEXT:    li a0, 5
1447; REMAT-NEXT:    slli a0, a0, 9
1448; REMAT-NEXT:    add a0, a1, a0
1449; REMAT-NEXT:    sd a0, 408(sp) # 8-byte Folded Spill
1450; REMAT-NEXT:    li a0, 3
1451; REMAT-NEXT:    slli a0, a0, 10
1452; REMAT-NEXT:    add a0, a1, a0
1453; REMAT-NEXT:    sd a0, 400(sp) # 8-byte Folded Spill
1454; REMAT-NEXT:    li a0, 7
1455; REMAT-NEXT:    slli a0, a0, 9
1456; REMAT-NEXT:    add a0, a1, a0
1457; REMAT-NEXT:    sd a0, 392(sp) # 8-byte Folded Spill
1458; REMAT-NEXT:    lui a0, 1
1459; REMAT-NEXT:    add a0, a1, a0
1460; REMAT-NEXT:    sd a0, 384(sp) # 8-byte Folded Spill
1461; REMAT-NEXT:    li a0, 9
1462; REMAT-NEXT:    slli a0, a0, 9
1463; REMAT-NEXT:    add a0, a1, a0
1464; REMAT-NEXT:    sd a0, 376(sp) # 8-byte Folded Spill
1465; REMAT-NEXT:    li a0, 5
1466; REMAT-NEXT:    slli a0, a0, 10
1467; REMAT-NEXT:    add a0, a1, a0
1468; REMAT-NEXT:    sd a0, 368(sp) # 8-byte Folded Spill
1469; REMAT-NEXT:    li a0, 11
1470; REMAT-NEXT:    slli a0, a0, 9
1471; REMAT-NEXT:    add a0, a1, a0
1472; REMAT-NEXT:    sd a0, 360(sp) # 8-byte Folded Spill
1473; REMAT-NEXT:    li a0, 3
1474; REMAT-NEXT:    slli a0, a0, 11
1475; REMAT-NEXT:    add a0, a1, a0
1476; REMAT-NEXT:    sd a0, 352(sp) # 8-byte Folded Spill
1477; REMAT-NEXT:    li a0, 13
1478; REMAT-NEXT:    slli a0, a0, 9
1479; REMAT-NEXT:    add a0, a1, a0
1480; REMAT-NEXT:    sd a0, 344(sp) # 8-byte Folded Spill
1481; REMAT-NEXT:    li a0, 7
1482; REMAT-NEXT:    slli a0, a0, 10
1483; REMAT-NEXT:    add a0, a1, a0
1484; REMAT-NEXT:    sd a0, 336(sp) # 8-byte Folded Spill
1485; REMAT-NEXT:    li a0, 15
1486; REMAT-NEXT:    slli a0, a0, 9
1487; REMAT-NEXT:    add a0, a1, a0
1488; REMAT-NEXT:    sd a0, 328(sp) # 8-byte Folded Spill
1489; REMAT-NEXT:    lui a0, 2
1490; REMAT-NEXT:    add a0, a1, a0
1491; REMAT-NEXT:    sd a0, 320(sp) # 8-byte Folded Spill
1492; REMAT-NEXT:    li a0, 17
1493; REMAT-NEXT:    slli a0, a0, 9
1494; REMAT-NEXT:    add a0, a1, a0
1495; REMAT-NEXT:    sd a0, 312(sp) # 8-byte Folded Spill
1496; REMAT-NEXT:    add s5, a1, s5
1497; REMAT-NEXT:    sd s5, 304(sp) # 8-byte Folded Spill
1498; REMAT-NEXT:    add s6, a1, s6
1499; REMAT-NEXT:    sd s6, 296(sp) # 8-byte Folded Spill
1500; REMAT-NEXT:    li a0, 5
1501; REMAT-NEXT:    slli a0, a0, 11
1502; REMAT-NEXT:    add a0, a1, a0
1503; REMAT-NEXT:    sd a0, 288(sp) # 8-byte Folded Spill
1504; REMAT-NEXT:    add s8, a1, s8
1505; REMAT-NEXT:    sd s8, 280(sp) # 8-byte Folded Spill
1506; REMAT-NEXT:    add s9, a1, s9
1507; REMAT-NEXT:    sd s9, 272(sp) # 8-byte Folded Spill
1508; REMAT-NEXT:    add s10, a1, s10
1509; REMAT-NEXT:    sd s10, 264(sp) # 8-byte Folded Spill
1510; REMAT-NEXT:    add s11, a1, s11
1511; REMAT-NEXT:    sd s11, 256(sp) # 8-byte Folded Spill
1512; REMAT-NEXT:    add ra, a1, ra
1513; REMAT-NEXT:    sd ra, 248(sp) # 8-byte Folded Spill
1514; REMAT-NEXT:    li a0, 13
1515; REMAT-NEXT:    slli a0, a0, 10
1516; REMAT-NEXT:    add a0, a1, a0
1517; REMAT-NEXT:    sd a0, 240(sp) # 8-byte Folded Spill
1518; REMAT-NEXT:    li a0, 27
1519; REMAT-NEXT:    slli a0, a0, 9
1520; REMAT-NEXT:    add a0, a1, a0
1521; REMAT-NEXT:    sd a0, 232(sp) # 8-byte Folded Spill
1522; REMAT-NEXT:    add a5, a1, a5
1523; REMAT-NEXT:    sd a5, 224(sp) # 8-byte Folded Spill
1524; REMAT-NEXT:    li a0, 29
1525; REMAT-NEXT:    slli a0, a0, 9
1526; REMAT-NEXT:    add a0, a1, a0
1527; REMAT-NEXT:    sd a0, 216(sp) # 8-byte Folded Spill
1528; REMAT-NEXT:    li a0, 15
1529; REMAT-NEXT:    slli a0, a0, 10
1530; REMAT-NEXT:    add a0, a1, a0
1531; REMAT-NEXT:    sd a0, 208(sp) # 8-byte Folded Spill
1532; REMAT-NEXT:    li a0, 31
1533; REMAT-NEXT:    slli a0, a0, 9
1534; REMAT-NEXT:    add a0, a1, a0
1535; REMAT-NEXT:    sd a0, 200(sp) # 8-byte Folded Spill
1536; REMAT-NEXT:    lui a0, 4
1537; REMAT-NEXT:    add a0, a1, a0
1538; REMAT-NEXT:    sd a0, 192(sp) # 8-byte Folded Spill
1539; REMAT-NEXT:    lui a0, 4
1540; REMAT-NEXT:    addiw a0, a0, 512
1541; REMAT-NEXT:    add a0, a1, a0
1542; REMAT-NEXT:    sd a0, 184(sp) # 8-byte Folded Spill
1543; REMAT-NEXT:    li a0, 17
1544; REMAT-NEXT:    slli a0, a0, 10
1545; REMAT-NEXT:    add a0, a1, a0
1546; REMAT-NEXT:    sd a0, 176(sp) # 8-byte Folded Spill
1547; REMAT-NEXT:    lui a0, 4
1548; REMAT-NEXT:    addiw a0, a0, 1536
1549; REMAT-NEXT:    add a0, a1, a0
1550; REMAT-NEXT:    sd a0, 168(sp) # 8-byte Folded Spill
1551; REMAT-NEXT:    li a0, 9
1552; REMAT-NEXT:    slli a0, a0, 11
1553; REMAT-NEXT:    add a0, a1, a0
1554; REMAT-NEXT:    sd a0, 160(sp) # 8-byte Folded Spill
1555; REMAT-NEXT:    lui a0, 5
1556; REMAT-NEXT:    addiw a0, a0, -1536
1557; REMAT-NEXT:    add a0, a1, a0
1558; REMAT-NEXT:    sd a0, 152(sp) # 8-byte Folded Spill
1559; REMAT-NEXT:    li a0, 19
1560; REMAT-NEXT:    slli a0, a0, 10
1561; REMAT-NEXT:    add a0, a1, a0
1562; REMAT-NEXT:    sd a0, 144(sp) # 8-byte Folded Spill
1563; REMAT-NEXT:    lui a0, 5
1564; REMAT-NEXT:    addiw a0, a0, -512
1565; REMAT-NEXT:    add a0, a1, a0
1566; REMAT-NEXT:    sd a0, 136(sp) # 8-byte Folded Spill
1567; REMAT-NEXT:    lui a0, 5
1568; REMAT-NEXT:    add a0, a1, a0
1569; REMAT-NEXT:    sd a0, 128(sp) # 8-byte Folded Spill
1570; REMAT-NEXT:    lui a0, 5
1571; REMAT-NEXT:    addiw a0, a0, 512
1572; REMAT-NEXT:    add a0, a1, a0
1573; REMAT-NEXT:    sd a0, 120(sp) # 8-byte Folded Spill
1574; REMAT-NEXT:    add s7, a1, s7
1575; REMAT-NEXT:    sd s7, 112(sp) # 8-byte Folded Spill
1576; REMAT-NEXT:    add s4, a1, s4
1577; REMAT-NEXT:    sd s4, 104(sp) # 8-byte Folded Spill
1578; REMAT-NEXT:    li a0, 11
1579; REMAT-NEXT:    slli a0, a0, 11
1580; REMAT-NEXT:    add a0, a1, a0
1581; REMAT-NEXT:    sd a0, 96(sp) # 8-byte Folded Spill
1582; REMAT-NEXT:    add s3, a1, s3
1583; REMAT-NEXT:    sd s3, 88(sp) # 8-byte Folded Spill
1584; REMAT-NEXT:    add s2, a1, s2
1585; REMAT-NEXT:    sd s2, 80(sp) # 8-byte Folded Spill
1586; REMAT-NEXT:    lui a0, 6
1587; REMAT-NEXT:    addiw a0, a0, -512
1588; REMAT-NEXT:    add a0, a1, a0
1589; REMAT-NEXT:    sd a0, 72(sp) # 8-byte Folded Spill
1590; REMAT-NEXT:    add s1, a1, s1
1591; REMAT-NEXT:    sd s1, 64(sp) # 8-byte Folded Spill
1592; REMAT-NEXT:    add s0, a1, s0
1593; REMAT-NEXT:    sd s0, 56(sp) # 8-byte Folded Spill
1594; REMAT-NEXT:    li a0, 25
1595; REMAT-NEXT:    slli a0, a0, 10
1596; REMAT-NEXT:    add a0, a1, a0
1597; REMAT-NEXT:    sd a0, 48(sp) # 8-byte Folded Spill
1598; REMAT-NEXT:    add t6, a1, t6
1599; REMAT-NEXT:    sd t6, 40(sp) # 8-byte Folded Spill
1600; REMAT-NEXT:    add t5, a1, t5
1601; REMAT-NEXT:    sd t5, 32(sp) # 8-byte Folded Spill
1602; REMAT-NEXT:    lui a0, 7
1603; REMAT-NEXT:    addiw a0, a0, -1536
1604; REMAT-NEXT:    add a0, a1, a0
1605; REMAT-NEXT:    sd a0, 24(sp) # 8-byte Folded Spill
1606; REMAT-NEXT:    add t4, a1, t4
1607; REMAT-NEXT:    sd t4, 16(sp) # 8-byte Folded Spill
1608; REMAT-NEXT:    lui ra, 7
1609; REMAT-NEXT:    addiw ra, ra, -512
1610; REMAT-NEXT:    add ra, a1, ra
1611; REMAT-NEXT:    add s11, a1, t3
1612; REMAT-NEXT:    add s10, a1, t2
1613; REMAT-NEXT:    add s9, a1, t1
1614; REMAT-NEXT:    add s8, a1, t0
1615; REMAT-NEXT:    add s7, a1, a7
1616; REMAT-NEXT:    add s6, a1, a6
1617; REMAT-NEXT:    add s5, a1, a4
1618; REMAT-NEXT:    add s4, a1, a3
1619; REMAT-NEXT:    add s3, a1, a2
1620; REMAT-NEXT:    lui s2, 8
1621; REMAT-NEXT:    addiw s2, s2, 512
1622; REMAT-NEXT:    add s2, a1, s2
1623; REMAT-NEXT:    lui s1, 8
1624; REMAT-NEXT:    addiw s1, s1, 1024
1625; REMAT-NEXT:    add s1, a1, s1
1626; REMAT-NEXT:    lui s0, 8
1627; REMAT-NEXT:    addiw s0, s0, 1536
1628; REMAT-NEXT:    add s0, a1, s0
1629; REMAT-NEXT:    li t6, 17
1630; REMAT-NEXT:    slli t6, t6, 11
1631; REMAT-NEXT:    add t6, a1, t6
1632; REMAT-NEXT:    lui t5, 9
1633; REMAT-NEXT:    addiw t5, t5, -1536
1634; REMAT-NEXT:    add t5, a1, t5
1635; REMAT-NEXT:    lui t4, 9
1636; REMAT-NEXT:    addiw t4, t4, -1024
1637; REMAT-NEXT:    add t4, a1, t4
1638; REMAT-NEXT:    lui t3, 9
1639; REMAT-NEXT:    addiw t3, t3, -512
1640; REMAT-NEXT:    add t3, a1, t3
1641; REMAT-NEXT:    lui t2, 9
1642; REMAT-NEXT:    add t2, a1, t2
1643; REMAT-NEXT:    lui t1, 9
1644; REMAT-NEXT:    addiw t1, t1, 512
1645; REMAT-NEXT:    add t1, a1, t1
1646; REMAT-NEXT:    lui t0, 9
1647; REMAT-NEXT:    addiw t0, t0, 1024
1648; REMAT-NEXT:    add t0, a1, t0
1649; REMAT-NEXT:    lui a7, 9
1650; REMAT-NEXT:    addiw a7, a7, 1536
1651; REMAT-NEXT:    add a7, a1, a7
1652; REMAT-NEXT:    li a6, 19
1653; REMAT-NEXT:    slli a6, a6, 11
1654; REMAT-NEXT:    add a6, a1, a6
1655; REMAT-NEXT:    lui a5, 10
1656; REMAT-NEXT:    addiw a5, a5, -1536
1657; REMAT-NEXT:    add a5, a1, a5
1658; REMAT-NEXT:    lui a4, 10
1659; REMAT-NEXT:    addiw a4, a4, -1024
1660; REMAT-NEXT:    add a4, a1, a4
1661; REMAT-NEXT:    lui a3, 10
1662; REMAT-NEXT:    addiw a3, a3, -512
1663; REMAT-NEXT:    add a3, a1, a3
1664; REMAT-NEXT:    lui a2, 10
1665; REMAT-NEXT:    add a2, a1, a2
1666; REMAT-NEXT:    lui a0, 10
1667; REMAT-NEXT:    addiw a0, a0, 512
1668; REMAT-NEXT:    add a0, a1, a0
1669; REMAT-NEXT:    addi a1, a1, 1536
1670; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1671; REMAT-NEXT:    vse32.v v8, (a1)
1672; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1673; REMAT-NEXT:    ld a1, 416(sp) # 8-byte Folded Reload
1674; REMAT-NEXT:    vse32.v v8, (a1)
1675; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1676; REMAT-NEXT:    ld a1, 408(sp) # 8-byte Folded Reload
1677; REMAT-NEXT:    vse32.v v8, (a1)
1678; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1679; REMAT-NEXT:    ld a1, 400(sp) # 8-byte Folded Reload
1680; REMAT-NEXT:    vse32.v v8, (a1)
1681; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1682; REMAT-NEXT:    ld a1, 392(sp) # 8-byte Folded Reload
1683; REMAT-NEXT:    vse32.v v8, (a1)
1684; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1685; REMAT-NEXT:    ld a1, 384(sp) # 8-byte Folded Reload
1686; REMAT-NEXT:    vse32.v v8, (a1)
1687; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1688; REMAT-NEXT:    ld a1, 376(sp) # 8-byte Folded Reload
1689; REMAT-NEXT:    vse32.v v8, (a1)
1690; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1691; REMAT-NEXT:    ld a1, 368(sp) # 8-byte Folded Reload
1692; REMAT-NEXT:    vse32.v v8, (a1)
1693; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1694; REMAT-NEXT:    ld a1, 360(sp) # 8-byte Folded Reload
1695; REMAT-NEXT:    vse32.v v8, (a1)
1696; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1697; REMAT-NEXT:    ld a1, 352(sp) # 8-byte Folded Reload
1698; REMAT-NEXT:    vse32.v v8, (a1)
1699; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1700; REMAT-NEXT:    ld a1, 344(sp) # 8-byte Folded Reload
1701; REMAT-NEXT:    vse32.v v8, (a1)
1702; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1703; REMAT-NEXT:    ld a1, 336(sp) # 8-byte Folded Reload
1704; REMAT-NEXT:    vse32.v v8, (a1)
1705; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1706; REMAT-NEXT:    ld a1, 328(sp) # 8-byte Folded Reload
1707; REMAT-NEXT:    vse32.v v8, (a1)
1708; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1709; REMAT-NEXT:    ld a1, 320(sp) # 8-byte Folded Reload
1710; REMAT-NEXT:    vse32.v v8, (a1)
1711; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1712; REMAT-NEXT:    ld a1, 312(sp) # 8-byte Folded Reload
1713; REMAT-NEXT:    vse32.v v8, (a1)
1714; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1715; REMAT-NEXT:    ld a1, 304(sp) # 8-byte Folded Reload
1716; REMAT-NEXT:    vse32.v v8, (a1)
1717; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1718; REMAT-NEXT:    ld a1, 296(sp) # 8-byte Folded Reload
1719; REMAT-NEXT:    vse32.v v8, (a1)
1720; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1721; REMAT-NEXT:    ld a1, 288(sp) # 8-byte Folded Reload
1722; REMAT-NEXT:    vse32.v v8, (a1)
1723; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1724; REMAT-NEXT:    ld a1, 280(sp) # 8-byte Folded Reload
1725; REMAT-NEXT:    vse32.v v8, (a1)
1726; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1727; REMAT-NEXT:    ld a1, 272(sp) # 8-byte Folded Reload
1728; REMAT-NEXT:    vse32.v v8, (a1)
1729; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1730; REMAT-NEXT:    ld a1, 264(sp) # 8-byte Folded Reload
1731; REMAT-NEXT:    vse32.v v8, (a1)
1732; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1733; REMAT-NEXT:    ld a1, 256(sp) # 8-byte Folded Reload
1734; REMAT-NEXT:    vse32.v v8, (a1)
1735; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1736; REMAT-NEXT:    ld a1, 248(sp) # 8-byte Folded Reload
1737; REMAT-NEXT:    vse32.v v8, (a1)
1738; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1739; REMAT-NEXT:    ld a1, 240(sp) # 8-byte Folded Reload
1740; REMAT-NEXT:    vse32.v v8, (a1)
1741; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1742; REMAT-NEXT:    ld a1, 232(sp) # 8-byte Folded Reload
1743; REMAT-NEXT:    vse32.v v8, (a1)
1744; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1745; REMAT-NEXT:    ld a1, 224(sp) # 8-byte Folded Reload
1746; REMAT-NEXT:    vse32.v v8, (a1)
1747; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1748; REMAT-NEXT:    ld a1, 216(sp) # 8-byte Folded Reload
1749; REMAT-NEXT:    vse32.v v8, (a1)
1750; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1751; REMAT-NEXT:    ld a1, 208(sp) # 8-byte Folded Reload
1752; REMAT-NEXT:    vse32.v v8, (a1)
1753; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1754; REMAT-NEXT:    ld a1, 200(sp) # 8-byte Folded Reload
1755; REMAT-NEXT:    vse32.v v8, (a1)
1756; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1757; REMAT-NEXT:    ld a1, 192(sp) # 8-byte Folded Reload
1758; REMAT-NEXT:    vse32.v v8, (a1)
1759; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1760; REMAT-NEXT:    ld a1, 184(sp) # 8-byte Folded Reload
1761; REMAT-NEXT:    vse32.v v8, (a1)
1762; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1763; REMAT-NEXT:    ld a1, 176(sp) # 8-byte Folded Reload
1764; REMAT-NEXT:    vse32.v v8, (a1)
1765; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1766; REMAT-NEXT:    ld a1, 168(sp) # 8-byte Folded Reload
1767; REMAT-NEXT:    vse32.v v8, (a1)
1768; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1769; REMAT-NEXT:    ld a1, 160(sp) # 8-byte Folded Reload
1770; REMAT-NEXT:    vse32.v v8, (a1)
1771; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1772; REMAT-NEXT:    ld a1, 152(sp) # 8-byte Folded Reload
1773; REMAT-NEXT:    vse32.v v8, (a1)
1774; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1775; REMAT-NEXT:    ld a1, 144(sp) # 8-byte Folded Reload
1776; REMAT-NEXT:    vse32.v v8, (a1)
1777; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1778; REMAT-NEXT:    ld a1, 136(sp) # 8-byte Folded Reload
1779; REMAT-NEXT:    vse32.v v8, (a1)
1780; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1781; REMAT-NEXT:    ld a1, 128(sp) # 8-byte Folded Reload
1782; REMAT-NEXT:    vse32.v v8, (a1)
1783; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1784; REMAT-NEXT:    ld a1, 120(sp) # 8-byte Folded Reload
1785; REMAT-NEXT:    vse32.v v8, (a1)
1786; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1787; REMAT-NEXT:    ld a1, 112(sp) # 8-byte Folded Reload
1788; REMAT-NEXT:    vse32.v v8, (a1)
1789; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1790; REMAT-NEXT:    ld a1, 104(sp) # 8-byte Folded Reload
1791; REMAT-NEXT:    vse32.v v8, (a1)
1792; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1793; REMAT-NEXT:    ld a1, 96(sp) # 8-byte Folded Reload
1794; REMAT-NEXT:    vse32.v v8, (a1)
1795; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1796; REMAT-NEXT:    ld a1, 88(sp) # 8-byte Folded Reload
1797; REMAT-NEXT:    vse32.v v8, (a1)
1798; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1799; REMAT-NEXT:    ld a1, 80(sp) # 8-byte Folded Reload
1800; REMAT-NEXT:    vse32.v v8, (a1)
1801; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1802; REMAT-NEXT:    ld a1, 72(sp) # 8-byte Folded Reload
1803; REMAT-NEXT:    vse32.v v8, (a1)
1804; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1805; REMAT-NEXT:    ld a1, 64(sp) # 8-byte Folded Reload
1806; REMAT-NEXT:    vse32.v v8, (a1)
1807; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1808; REMAT-NEXT:    ld a1, 56(sp) # 8-byte Folded Reload
1809; REMAT-NEXT:    vse32.v v8, (a1)
1810; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1811; REMAT-NEXT:    ld a1, 48(sp) # 8-byte Folded Reload
1812; REMAT-NEXT:    vse32.v v8, (a1)
1813; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1814; REMAT-NEXT:    ld a1, 40(sp) # 8-byte Folded Reload
1815; REMAT-NEXT:    vse32.v v8, (a1)
1816; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1817; REMAT-NEXT:    ld a1, 32(sp) # 8-byte Folded Reload
1818; REMAT-NEXT:    vse32.v v8, (a1)
1819; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1820; REMAT-NEXT:    ld a1, 24(sp) # 8-byte Folded Reload
1821; REMAT-NEXT:    vse32.v v8, (a1)
1822; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1823; REMAT-NEXT:    ld a1, 16(sp) # 8-byte Folded Reload
1824; REMAT-NEXT:    vse32.v v8, (a1)
1825; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1826; REMAT-NEXT:    vse32.v v8, (ra)
1827; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1828; REMAT-NEXT:    vse32.v v8, (s11)
1829; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1830; REMAT-NEXT:    vse32.v v8, (s10)
1831; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1832; REMAT-NEXT:    vse32.v v8, (s9)
1833; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1834; REMAT-NEXT:    vse32.v v8, (s8)
1835; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1836; REMAT-NEXT:    vse32.v v8, (s7)
1837; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1838; REMAT-NEXT:    vse32.v v8, (s6)
1839; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1840; REMAT-NEXT:    vse32.v v8, (s5)
1841; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1842; REMAT-NEXT:    vse32.v v8, (s4)
1843; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1844; REMAT-NEXT:    vse32.v v8, (s3)
1845; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1846; REMAT-NEXT:    vse32.v v8, (s2)
1847; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1848; REMAT-NEXT:    vse32.v v8, (s1)
1849; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1850; REMAT-NEXT:    vse32.v v8, (s0)
1851; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1852; REMAT-NEXT:    vse32.v v8, (t6)
1853; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1854; REMAT-NEXT:    vse32.v v8, (t5)
1855; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1856; REMAT-NEXT:    vse32.v v8, (t4)
1857; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1858; REMAT-NEXT:    vse32.v v8, (t3)
1859; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1860; REMAT-NEXT:    vse32.v v8, (t2)
1861; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1862; REMAT-NEXT:    vse32.v v8, (t1)
1863; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1864; REMAT-NEXT:    vse32.v v8, (t0)
1865; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1866; REMAT-NEXT:    vse32.v v8, (a7)
1867; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1868; REMAT-NEXT:    vse32.v v8, (a6)
1869; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1870; REMAT-NEXT:    vse32.v v8, (a5)
1871; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1872; REMAT-NEXT:    vse32.v v8, (a4)
1873; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1874; REMAT-NEXT:    vse32.v v8, (a3)
1875; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1876; REMAT-NEXT:    vse32.v v8, (a2)
1877; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1878; REMAT-NEXT:    vse32.v v8, (a0)
1879; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1880; REMAT-NEXT:    sf.vc.v.i 2, 0, v8, 0
1881; REMAT-NEXT:    csrr a0, vlenb
1882; REMAT-NEXT:    li a1, 18
1883; REMAT-NEXT:    mul a0, a0, a1
1884; REMAT-NEXT:    add sp, sp, a0
1885; REMAT-NEXT:    .cfi_def_cfa sp, 544
1886; REMAT-NEXT:    ld ra, 536(sp) # 8-byte Folded Reload
1887; REMAT-NEXT:    ld s0, 528(sp) # 8-byte Folded Reload
1888; REMAT-NEXT:    ld s1, 520(sp) # 8-byte Folded Reload
1889; REMAT-NEXT:    ld s2, 512(sp) # 8-byte Folded Reload
1890; REMAT-NEXT:    ld s3, 504(sp) # 8-byte Folded Reload
1891; REMAT-NEXT:    ld s4, 496(sp) # 8-byte Folded Reload
1892; REMAT-NEXT:    ld s5, 488(sp) # 8-byte Folded Reload
1893; REMAT-NEXT:    ld s6, 480(sp) # 8-byte Folded Reload
1894; REMAT-NEXT:    ld s7, 472(sp) # 8-byte Folded Reload
1895; REMAT-NEXT:    ld s8, 464(sp) # 8-byte Folded Reload
1896; REMAT-NEXT:    ld s9, 456(sp) # 8-byte Folded Reload
1897; REMAT-NEXT:    ld s10, 448(sp) # 8-byte Folded Reload
1898; REMAT-NEXT:    ld s11, 440(sp) # 8-byte Folded Reload
1899; REMAT-NEXT:    .cfi_restore ra
1900; REMAT-NEXT:    .cfi_restore s0
1901; REMAT-NEXT:    .cfi_restore s1
1902; REMAT-NEXT:    .cfi_restore s2
1903; REMAT-NEXT:    .cfi_restore s3
1904; REMAT-NEXT:    .cfi_restore s4
1905; REMAT-NEXT:    .cfi_restore s5
1906; REMAT-NEXT:    .cfi_restore s6
1907; REMAT-NEXT:    .cfi_restore s7
1908; REMAT-NEXT:    .cfi_restore s8
1909; REMAT-NEXT:    .cfi_restore s9
1910; REMAT-NEXT:    .cfi_restore s10
1911; REMAT-NEXT:    .cfi_restore s11
1912; REMAT-NEXT:    addi sp, sp, 544
1913; REMAT-NEXT:    .cfi_def_cfa_offset 0
1914; REMAT-NEXT:    ret
1915  %4 = tail call i64 @llvm.riscv.vsetvli.i64(i64 32, i64 2, i64 1)
1916  %5 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %0, i64 %4)
1917  %6 = getelementptr inbounds i32, ptr %0, i64 128
1918  %7 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %6, i64 %4)
1919  %8 = getelementptr inbounds i32, ptr %0, i64 256
1920  %9 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %8, i64 %4)
1921  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %5, <vscale x 4 x i32> %7, i64 %4)
1922  %10 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %8, i64 %4)
1923  %11 = getelementptr inbounds i32, ptr %0, i64 384
1924  %12 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %11, i64 %4)
1925  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %7, <vscale x 4 x i32> %9, i64 %4)
1926  %13 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %11, i64 %4)
1927  %14 = getelementptr inbounds i32, ptr %0, i64 512
1928  %15 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %14, i64 %4)
1929  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %10, <vscale x 4 x i32> %12, i64 %4)
1930  %16 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %14, i64 %4)
1931  %17 = getelementptr inbounds i32, ptr %0, i64 640
1932  %18 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %17, i64 %4)
1933  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %13, <vscale x 4 x i32> %15, i64 %4)
1934  %19 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %17, i64 %4)
1935  %20 = getelementptr inbounds i32, ptr %0, i64 768
1936  %21 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %20, i64 %4)
1937  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %16, <vscale x 4 x i32> %18, i64 %4)
1938  %22 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %20, i64 %4)
1939  %23 = getelementptr inbounds i32, ptr %0, i64 896
1940  %24 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %23, i64 %4)
1941  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %19, <vscale x 4 x i32> %21, i64 %4)
1942  %25 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %23, i64 %4)
1943  %26 = getelementptr inbounds i32, ptr %0, i64 1024
1944  %27 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %26, i64 %4)
1945  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %22, <vscale x 4 x i32> %24, i64 %4)
1946  %28 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %26, i64 %4)
1947  %29 = getelementptr inbounds i32, ptr %0, i64 1152
1948  %30 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %29, i64 %4)
1949  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %25, <vscale x 4 x i32> %27, i64 %4)
1950  %31 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %29, i64 %4)
1951  %32 = getelementptr inbounds i32, ptr %0, i64 1280
1952  %33 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %32, i64 %4)
1953  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %28, <vscale x 4 x i32> %30, i64 %4)
1954  %34 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %32, i64 %4)
1955  %35 = getelementptr inbounds i32, ptr %0, i64 1408
1956  %36 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %35, i64 %4)
1957  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %31, <vscale x 4 x i32> %33, i64 %4)
1958  %37 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %35, i64 %4)
1959  %38 = getelementptr inbounds i32, ptr %0, i64 1536
1960  %39 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %38, i64 %4)
1961  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %34, <vscale x 4 x i32> %36, i64 %4)
1962  %40 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %38, i64 %4)
1963  %41 = getelementptr inbounds i32, ptr %0, i64 1664
1964  %42 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %41, i64 %4)
1965  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %37, <vscale x 4 x i32> %39, i64 %4)
1966  %43 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %41, i64 %4)
1967  %44 = getelementptr inbounds i32, ptr %0, i64 1792
1968  %45 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %44, i64 %4)
1969  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %40, <vscale x 4 x i32> %42, i64 %4)
1970  %46 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %44, i64 %4)
1971  %47 = getelementptr inbounds i32, ptr %0, i64 1920
1972  %48 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %47, i64 %4)
1973  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %43, <vscale x 4 x i32> %45, i64 %4)
1974  %49 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %47, i64 %4)
1975  %50 = getelementptr inbounds i32, ptr %0, i64 2048
1976  %51 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %50, i64 %4)
1977  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %46, <vscale x 4 x i32> %48, i64 %4)
1978  %52 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %50, i64 %4)
1979  %53 = getelementptr inbounds i32, ptr %0, i64 2176
1980  %54 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %53, i64 %4)
1981  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %49, <vscale x 4 x i32> %51, i64 %4)
1982  %55 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %53, i64 %4)
1983  %56 = getelementptr inbounds i32, ptr %0, i64 2304
1984  %57 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %56, i64 %4)
1985  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %52, <vscale x 4 x i32> %54, i64 %4)
1986  %58 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %56, i64 %4)
1987  %59 = getelementptr inbounds i32, ptr %0, i64 2432
1988  %60 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %59, i64 %4)
1989  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %55, <vscale x 4 x i32> %57, i64 %4)
1990  %61 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %59, i64 %4)
1991  %62 = getelementptr inbounds i32, ptr %0, i64 2560
1992  %63 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %62, i64 %4)
1993  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %58, <vscale x 4 x i32> %60, i64 %4)
1994  %64 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %62, i64 %4)
1995  %65 = getelementptr inbounds i32, ptr %0, i64 2688
1996  %66 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %65, i64 %4)
1997  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %61, <vscale x 4 x i32> %63, i64 %4)
1998  %67 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %65, i64 %4)
1999  %68 = getelementptr inbounds i32, ptr %0, i64 2816
2000  %69 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %68, i64 %4)
2001  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %64, <vscale x 4 x i32> %66, i64 %4)
2002  %70 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %68, i64 %4)
2003  %71 = getelementptr inbounds i32, ptr %0, i64 2944
2004  %72 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %71, i64 %4)
2005  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %67, <vscale x 4 x i32> %69, i64 %4)
2006  %73 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %71, i64 %4)
2007  %74 = getelementptr inbounds i32, ptr %0, i64 3072
2008  %75 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %74, i64 %4)
2009  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %70, <vscale x 4 x i32> %72, i64 %4)
2010  %76 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %74, i64 %4)
2011  %77 = getelementptr inbounds i32, ptr %0, i64 3200
2012  %78 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %77, i64 %4)
2013  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %73, <vscale x 4 x i32> %75, i64 %4)
2014  %79 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %77, i64 %4)
2015  %80 = getelementptr inbounds i32, ptr %0, i64 3328
2016  %81 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %80, i64 %4)
2017  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %76, <vscale x 4 x i32> %78, i64 %4)
2018  %82 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %80, i64 %4)
2019  %83 = getelementptr inbounds i32, ptr %0, i64 3456
2020  %84 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %83, i64 %4)
2021  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %79, <vscale x 4 x i32> %81, i64 %4)
2022  %85 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %83, i64 %4)
2023  %86 = getelementptr inbounds i32, ptr %0, i64 3584
2024  %87 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %86, i64 %4)
2025  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %82, <vscale x 4 x i32> %84, i64 %4)
2026  %88 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %86, i64 %4)
2027  %89 = getelementptr inbounds i32, ptr %0, i64 3712
2028  %90 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %89, i64 %4)
2029  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %85, <vscale x 4 x i32> %87, i64 %4)
2030  %91 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %89, i64 %4)
2031  %92 = getelementptr inbounds i32, ptr %0, i64 3840
2032  %93 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %92, i64 %4)
2033  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %88, <vscale x 4 x i32> %90, i64 %4)
2034  %94 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %92, i64 %4)
2035  %95 = getelementptr inbounds i32, ptr %0, i64 3968
2036  %96 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %95, i64 %4)
2037  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %91, <vscale x 4 x i32> %93, i64 %4)
2038  %97 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %95, i64 %4)
2039  %98 = getelementptr inbounds i32, ptr %0, i64 4096
2040  %99 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %98, i64 %4)
2041  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %94, <vscale x 4 x i32> %96, i64 %4)
2042  %100 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %98, i64 %4)
2043  %101 = getelementptr inbounds i32, ptr %0, i64 4224
2044  %102 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %101, i64 %4)
2045  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %97, <vscale x 4 x i32> %99, i64 %4)
2046  %103 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %101, i64 %4)
2047  %104 = getelementptr inbounds i32, ptr %0, i64 4352
2048  %105 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %104, i64 %4)
2049  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %100, <vscale x 4 x i32> %102, i64 %4)
2050  %106 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %104, i64 %4)
2051  %107 = getelementptr inbounds i32, ptr %0, i64 4480
2052  %108 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %107, i64 %4)
2053  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %103, <vscale x 4 x i32> %105, i64 %4)
2054  %109 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %107, i64 %4)
2055  %110 = getelementptr inbounds i32, ptr %0, i64 4608
2056  %111 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %110, i64 %4)
2057  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %106, <vscale x 4 x i32> %108, i64 %4)
2058  %112 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %110, i64 %4)
2059  %113 = getelementptr inbounds i32, ptr %0, i64 4736
2060  %114 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %113, i64 %4)
2061  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %109, <vscale x 4 x i32> %111, i64 %4)
2062  %115 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %113, i64 %4)
2063  %116 = getelementptr inbounds i32, ptr %0, i64 4864
2064  %117 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %116, i64 %4)
2065  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %112, <vscale x 4 x i32> %114, i64 %4)
2066  %118 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %116, i64 %4)
2067  %119 = getelementptr inbounds i32, ptr %0, i64 4992
2068  %120 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %119, i64 %4)
2069  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %115, <vscale x 4 x i32> %117, i64 %4)
2070  %121 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %119, i64 %4)
2071  %122 = getelementptr inbounds i32, ptr %0, i64 5120
2072  %123 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %122, i64 %4)
2073  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %118, <vscale x 4 x i32> %120, i64 %4)
2074  %124 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %122, i64 %4)
2075  %125 = getelementptr inbounds i32, ptr %0, i64 5248
2076  %126 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %125, i64 %4)
2077  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %121, <vscale x 4 x i32> %123, i64 %4)
2078  %127 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %125, i64 %4)
2079  %128 = getelementptr inbounds i32, ptr %0, i64 5376
2080  %129 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %128, i64 %4)
2081  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %124, <vscale x 4 x i32> %126, i64 %4)
2082  %130 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %128, i64 %4)
2083  %131 = getelementptr inbounds i32, ptr %0, i64 5504
2084  %132 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %131, i64 %4)
2085  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %127, <vscale x 4 x i32> %129, i64 %4)
2086  %133 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %131, i64 %4)
2087  %134 = getelementptr inbounds i32, ptr %0, i64 5632
2088  %135 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %134, i64 %4)
2089  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %130, <vscale x 4 x i32> %132, i64 %4)
2090  %136 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %134, i64 %4)
2091  %137 = getelementptr inbounds i32, ptr %0, i64 5760
2092  %138 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %137, i64 %4)
2093  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %133, <vscale x 4 x i32> %135, i64 %4)
2094  %139 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %137, i64 %4)
2095  %140 = getelementptr inbounds i32, ptr %0, i64 5888
2096  %141 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %140, i64 %4)
2097  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %136, <vscale x 4 x i32> %138, i64 %4)
2098  %142 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %140, i64 %4)
2099  %143 = getelementptr inbounds i32, ptr %0, i64 6016
2100  %144 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %143, i64 %4)
2101  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %139, <vscale x 4 x i32> %141, i64 %4)
2102  %145 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %143, i64 %4)
2103  %146 = getelementptr inbounds i32, ptr %0, i64 6144
2104  %147 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %146, i64 %4)
2105  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %142, <vscale x 4 x i32> %144, i64 %4)
2106  %148 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %146, i64 %4)
2107  %149 = getelementptr inbounds i32, ptr %0, i64 6272
2108  %150 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %149, i64 %4)
2109  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %145, <vscale x 4 x i32> %147, i64 %4)
2110  %151 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %149, i64 %4)
2111  %152 = getelementptr inbounds i32, ptr %0, i64 6400
2112  %153 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %152, i64 %4)
2113  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %148, <vscale x 4 x i32> %150, i64 %4)
2114  %154 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %152, i64 %4)
2115  %155 = getelementptr inbounds i32, ptr %0, i64 6528
2116  %156 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %155, i64 %4)
2117  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %151, <vscale x 4 x i32> %153, i64 %4)
2118  %157 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %155, i64 %4)
2119  %158 = getelementptr inbounds i32, ptr %0, i64 6656
2120  %159 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %158, i64 %4)
2121  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %154, <vscale x 4 x i32> %156, i64 %4)
2122  %160 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %158, i64 %4)
2123  %161 = getelementptr inbounds i32, ptr %0, i64 6784
2124  %162 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %161, i64 %4)
2125  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %157, <vscale x 4 x i32> %159, i64 %4)
2126  %163 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %161, i64 %4)
2127  %164 = getelementptr inbounds i32, ptr %0, i64 6912
2128  %165 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %164, i64 %4)
2129  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %160, <vscale x 4 x i32> %162, i64 %4)
2130  %166 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %164, i64 %4)
2131  %167 = getelementptr inbounds i32, ptr %0, i64 7040
2132  %168 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %167, i64 %4)
2133  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %163, <vscale x 4 x i32> %165, i64 %4)
2134  %169 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %167, i64 %4)
2135  %170 = getelementptr inbounds i32, ptr %0, i64 7168
2136  %171 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %170, i64 %4)
2137  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %166, <vscale x 4 x i32> %168, i64 %4)
2138  %172 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %170, i64 %4)
2139  %173 = getelementptr inbounds i32, ptr %0, i64 7296
2140  %174 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %173, i64 %4)
2141  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %169, <vscale x 4 x i32> %171, i64 %4)
2142  %175 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %173, i64 %4)
2143  %176 = getelementptr inbounds i32, ptr %0, i64 7424
2144  %177 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %176, i64 %4)
2145  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %172, <vscale x 4 x i32> %174, i64 %4)
2146  %178 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %176, i64 %4)
2147  %179 = getelementptr inbounds i32, ptr %0, i64 7552
2148  %180 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %179, i64 %4)
2149  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %175, <vscale x 4 x i32> %177, i64 %4)
2150  %181 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %179, i64 %4)
2151  %182 = getelementptr inbounds i32, ptr %0, i64 7680
2152  %183 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %182, i64 %4)
2153  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %178, <vscale x 4 x i32> %180, i64 %4)
2154  %184 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %182, i64 %4)
2155  %185 = getelementptr inbounds i32, ptr %0, i64 7808
2156  %186 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %185, i64 %4)
2157  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %181, <vscale x 4 x i32> %183, i64 %4)
2158  %187 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %185, i64 %4)
2159  %188 = getelementptr inbounds i32, ptr %0, i64 7936
2160  %189 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %188, i64 %4)
2161  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %184, <vscale x 4 x i32> %186, i64 %4)
2162  %190 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %188, i64 %4)
2163  %191 = getelementptr inbounds i32, ptr %0, i64 8064
2164  %192 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %191, i64 %4)
2165  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %187, <vscale x 4 x i32> %189, i64 %4)
2166  %193 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %191, i64 %4)
2167  %194 = getelementptr inbounds i32, ptr %0, i64 8192
2168  %195 = tail call <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32> poison, ptr %194, i64 %4)
2169  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %190, <vscale x 4 x i32> %192, i64 %4)
2170  tail call void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64 3, i64 0, <vscale x 4 x i32> %193, <vscale x 4 x i32> %195, i64 %4)
2171  %196 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2172  %197 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2173  %198 = getelementptr inbounds i32, ptr %1, i64 256
2174  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %196, ptr %198, i64 %4)
2175  %199 = getelementptr inbounds i32, ptr %1, i64 384
2176  %200 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2177  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %197, ptr %199, i64 %4)
2178  %201 = getelementptr inbounds i32, ptr %1, i64 512
2179  %202 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2180  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %200, ptr %201, i64 %4)
2181  %203 = getelementptr inbounds i32, ptr %1, i64 640
2182  %204 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2183  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %202, ptr %203, i64 %4)
2184  %205 = getelementptr inbounds i32, ptr %1, i64 768
2185  %206 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2186  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %204, ptr %205, i64 %4)
2187  %207 = getelementptr inbounds i32, ptr %1, i64 896
2188  %208 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2189  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %206, ptr %207, i64 %4)
2190  %209 = getelementptr inbounds i32, ptr %1, i64 1024
2191  %210 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2192  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %208, ptr %209, i64 %4)
2193  %211 = getelementptr inbounds i32, ptr %1, i64 1152
2194  %212 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2195  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %210, ptr %211, i64 %4)
2196  %213 = getelementptr inbounds i32, ptr %1, i64 1280
2197  %214 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2198  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %212, ptr %213, i64 %4)
2199  %215 = getelementptr inbounds i32, ptr %1, i64 1408
2200  %216 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2201  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %214, ptr %215, i64 %4)
2202  %217 = getelementptr inbounds i32, ptr %1, i64 1536
2203  %218 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2204  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %216, ptr %217, i64 %4)
2205  %219 = getelementptr inbounds i32, ptr %1, i64 1664
2206  %220 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2207  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %218, ptr %219, i64 %4)
2208  %221 = getelementptr inbounds i32, ptr %1, i64 1792
2209  %222 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2210  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %220, ptr %221, i64 %4)
2211  %223 = getelementptr inbounds i32, ptr %1, i64 1920
2212  %224 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2213  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %222, ptr %223, i64 %4)
2214  %225 = getelementptr inbounds i32, ptr %1, i64 2048
2215  %226 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2216  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %224, ptr %225, i64 %4)
2217  %227 = getelementptr inbounds i32, ptr %1, i64 2176
2218  %228 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2219  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %226, ptr %227, i64 %4)
2220  %229 = getelementptr inbounds i32, ptr %1, i64 2304
2221  %230 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2222  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %228, ptr %229, i64 %4)
2223  %231 = getelementptr inbounds i32, ptr %1, i64 2432
2224  %232 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2225  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %230, ptr %231, i64 %4)
2226  %233 = getelementptr inbounds i32, ptr %1, i64 2560
2227  %234 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2228  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %232, ptr %233, i64 %4)
2229  %235 = getelementptr inbounds i32, ptr %1, i64 2688
2230  %236 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2231  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %234, ptr %235, i64 %4)
2232  %237 = getelementptr inbounds i32, ptr %1, i64 2816
2233  %238 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2234  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %236, ptr %237, i64 %4)
2235  %239 = getelementptr inbounds i32, ptr %1, i64 2944
2236  %240 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2237  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %238, ptr %239, i64 %4)
2238  %241 = getelementptr inbounds i32, ptr %1, i64 3072
2239  %242 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2240  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %240, ptr %241, i64 %4)
2241  %243 = getelementptr inbounds i32, ptr %1, i64 3200
2242  %244 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2243  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %242, ptr %243, i64 %4)
2244  %245 = getelementptr inbounds i32, ptr %1, i64 3328
2245  %246 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2246  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %244, ptr %245, i64 %4)
2247  %247 = getelementptr inbounds i32, ptr %1, i64 3456
2248  %248 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2249  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %246, ptr %247, i64 %4)
2250  %249 = getelementptr inbounds i32, ptr %1, i64 3584
2251  %250 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2252  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %248, ptr %249, i64 %4)
2253  %251 = getelementptr inbounds i32, ptr %1, i64 3712
2254  %252 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2255  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %250, ptr %251, i64 %4)
2256  %253 = getelementptr inbounds i32, ptr %1, i64 3840
2257  %254 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2258  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %252, ptr %253, i64 %4)
2259  %255 = getelementptr inbounds i32, ptr %1, i64 3968
2260  %256 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2261  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %254, ptr %255, i64 %4)
2262  %257 = getelementptr inbounds i32, ptr %1, i64 4096
2263  %258 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2264  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %256, ptr %257, i64 %4)
2265  %259 = getelementptr inbounds i32, ptr %1, i64 4224
2266  %260 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2267  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %258, ptr %259, i64 %4)
2268  %261 = getelementptr inbounds i32, ptr %1, i64 4352
2269  %262 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2270  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %260, ptr %261, i64 %4)
2271  %263 = getelementptr inbounds i32, ptr %1, i64 4480
2272  %264 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2273  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %262, ptr %263, i64 %4)
2274  %265 = getelementptr inbounds i32, ptr %1, i64 4608
2275  %266 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2276  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %264, ptr %265, i64 %4)
2277  %267 = getelementptr inbounds i32, ptr %1, i64 4736
2278  %268 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2279  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %266, ptr %267, i64 %4)
2280  %269 = getelementptr inbounds i32, ptr %1, i64 4864
2281  %270 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2282  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %268, ptr %269, i64 %4)
2283  %271 = getelementptr inbounds i32, ptr %1, i64 4992
2284  %272 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2285  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %270, ptr %271, i64 %4)
2286  %273 = getelementptr inbounds i32, ptr %1, i64 5120
2287  %274 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2288  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %272, ptr %273, i64 %4)
2289  %275 = getelementptr inbounds i32, ptr %1, i64 5248
2290  %276 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2291  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %274, ptr %275, i64 %4)
2292  %277 = getelementptr inbounds i32, ptr %1, i64 5376
2293  %278 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2294  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %276, ptr %277, i64 %4)
2295  %279 = getelementptr inbounds i32, ptr %1, i64 5504
2296  %280 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2297  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %278, ptr %279, i64 %4)
2298  %281 = getelementptr inbounds i32, ptr %1, i64 5632
2299  %282 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2300  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %280, ptr %281, i64 %4)
2301  %283 = getelementptr inbounds i32, ptr %1, i64 5760
2302  %284 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2303  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %282, ptr %283, i64 %4)
2304  %285 = getelementptr inbounds i32, ptr %1, i64 5888
2305  %286 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2306  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %284, ptr %285, i64 %4)
2307  %287 = getelementptr inbounds i32, ptr %1, i64 6016
2308  %288 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2309  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %286, ptr %287, i64 %4)
2310  %289 = getelementptr inbounds i32, ptr %1, i64 6144
2311  %290 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2312  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %288, ptr %289, i64 %4)
2313  %291 = getelementptr inbounds i32, ptr %1, i64 6272
2314  %292 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2315  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %290, ptr %291, i64 %4)
2316  %293 = getelementptr inbounds i32, ptr %1, i64 6400
2317  %294 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2318  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %292, ptr %293, i64 %4)
2319  %295 = getelementptr inbounds i32, ptr %1, i64 6528
2320  %296 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2321  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %294, ptr %295, i64 %4)
2322  %297 = getelementptr inbounds i32, ptr %1, i64 6656
2323  %298 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2324  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %296, ptr %297, i64 %4)
2325  %299 = getelementptr inbounds i32, ptr %1, i64 6784
2326  %300 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2327  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %298, ptr %299, i64 %4)
2328  %301 = getelementptr inbounds i32, ptr %1, i64 6912
2329  %302 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2330  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %300, ptr %301, i64 %4)
2331  %303 = getelementptr inbounds i32, ptr %1, i64 7040
2332  %304 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2333  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %302, ptr %303, i64 %4)
2334  %305 = getelementptr inbounds i32, ptr %1, i64 7168
2335  %306 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2336  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %304, ptr %305, i64 %4)
2337  %307 = getelementptr inbounds i32, ptr %1, i64 7296
2338  %308 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2339  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %306, ptr %307, i64 %4)
2340  %309 = getelementptr inbounds i32, ptr %1, i64 7424
2341  %310 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2342  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %308, ptr %309, i64 %4)
2343  %311 = getelementptr inbounds i32, ptr %1, i64 7552
2344  %312 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2345  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %310, ptr %311, i64 %4)
2346  %313 = getelementptr inbounds i32, ptr %1, i64 7680
2347  %314 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2348  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %312, ptr %313, i64 %4)
2349  %315 = getelementptr inbounds i32, ptr %1, i64 7808
2350  %316 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2351  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %314, ptr %315, i64 %4)
2352  %317 = getelementptr inbounds i32, ptr %1, i64 7936
2353  %318 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2354  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %316, ptr %317, i64 %4)
2355  %319 = getelementptr inbounds i32, ptr %1, i64 8064
2356  %320 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2357  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %318, ptr %319, i64 %4)
2358  %321 = getelementptr inbounds i32, ptr %1, i64 8192
2359  %322 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2360  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %320, ptr %321, i64 %4)
2361  %323 = getelementptr inbounds i32, ptr %1, i64 8320
2362  %324 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2363  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %322, ptr %323, i64 %4)
2364  %325 = getelementptr inbounds i32, ptr %1, i64 8448
2365  %326 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2366  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %324, ptr %325, i64 %4)
2367  %327 = getelementptr inbounds i32, ptr %1, i64 8576
2368  %328 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2369  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %326, ptr %327, i64 %4)
2370  %329 = getelementptr inbounds i32, ptr %1, i64 8704
2371  %330 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2372  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %328, ptr %329, i64 %4)
2373  %331 = getelementptr inbounds i32, ptr %1, i64 8832
2374  %332 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2375  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %330, ptr %331, i64 %4)
2376  %333 = getelementptr inbounds i32, ptr %1, i64 8960
2377  %334 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2378  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %332, ptr %333, i64 %4)
2379  %335 = getelementptr inbounds i32, ptr %1, i64 9088
2380  %336 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2381  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %334, ptr %335, i64 %4)
2382  %337 = getelementptr inbounds i32, ptr %1, i64 9216
2383  %338 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2384  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %336, ptr %337, i64 %4)
2385  %339 = getelementptr inbounds i32, ptr %1, i64 9344
2386  %340 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2387  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %338, ptr %339, i64 %4)
2388  %341 = getelementptr inbounds i32, ptr %1, i64 9472
2389  %342 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2390  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %340, ptr %341, i64 %4)
2391  %343 = getelementptr inbounds i32, ptr %1, i64 9600
2392  %344 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2393  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %342, ptr %343, i64 %4)
2394  %345 = getelementptr inbounds i32, ptr %1, i64 9728
2395  %346 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2396  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %344, ptr %345, i64 %4)
2397  %347 = getelementptr inbounds i32, ptr %1, i64 9856
2398  %348 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2399  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %346, ptr %347, i64 %4)
2400  %349 = getelementptr inbounds i32, ptr %1, i64 9984
2401  %350 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2402  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %348, ptr %349, i64 %4)
2403  %351 = getelementptr inbounds i32, ptr %1, i64 10112
2404  %352 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2405  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %350, ptr %351, i64 %4)
2406  %353 = getelementptr inbounds i32, ptr %1, i64 10240
2407  %354 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2408  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %352, ptr %353, i64 %4)
2409  %355 = getelementptr inbounds i32, ptr %1, i64 10368
2410  %356 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2411  tail call void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32> %354, ptr %355, i64 %4)
2412  %357 = tail call <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64 2, i64 0, i64 0, i64 %4)
2413  ret void
2414}
2415
2416declare i64 @llvm.riscv.vsetvli.i64(i64, i64, i64)
2417declare <vscale x 4 x i32> @llvm.riscv.vle.nxv4i32.i64(<vscale x 4 x i32>, ptr, i64)
2418declare void @llvm.riscv.sf.vc.vv.se.i64.nxv4i32.nxv4i32.i64(i64, i64, <vscale x 4 x i32>, <vscale x 4 x i32>, i64)
2419declare <vscale x 4 x i32> @llvm.riscv.sf.vc.v.i.se.nxv4i32.i64.i64.i64(i64, i64, i64, i64)
2420declare void @llvm.riscv.vse.nxv4i32.i64(<vscale x 4 x i32>, ptr, i64)
2421