xref: /llvm-project/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-ext-loads.ll (revision 61510b51c33464a6bc15e4cf5b1ee07e2e0ec1c9)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mattr=+sve -force-streaming-compatible  < %s | FileCheck %s
3; RUN: llc -mattr=+sme -force-streaming  < %s | FileCheck %s
4; RUN: llc -force-streaming-compatible < %s | FileCheck %s --check-prefix=NONEON-NOSVE
5
6target triple = "aarch64-unknown-linux-gnu"
7
8define <8 x i16> @load_zext_v8i8i16(ptr %ap)  {
9; CHECK-LABEL: load_zext_v8i8i16:
10; CHECK:       // %bb.0:
11; CHECK-NEXT:    ptrue p0.h, vl8
12; CHECK-NEXT:    ld1b { z0.h }, p0/z, [x0]
13; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
14; CHECK-NEXT:    ret
15;
16; NONEON-NOSVE-LABEL: load_zext_v8i8i16:
17; NONEON-NOSVE:       // %bb.0:
18; NONEON-NOSVE-NEXT:    sub sp, sp, #32
19; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
20; NONEON-NOSVE-NEXT:    ldr d0, [x0]
21; NONEON-NOSVE-NEXT:    str d0, [sp, #8]
22; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
23; NONEON-NOSVE-NEXT:    strh w8, [sp, #30]
24; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
25; NONEON-NOSVE-NEXT:    strh w8, [sp, #28]
26; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
27; NONEON-NOSVE-NEXT:    strh w8, [sp, #26]
28; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
29; NONEON-NOSVE-NEXT:    strh w8, [sp, #24]
30; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
31; NONEON-NOSVE-NEXT:    strh w8, [sp, #22]
32; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
33; NONEON-NOSVE-NEXT:    strh w8, [sp, #20]
34; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
35; NONEON-NOSVE-NEXT:    strh w8, [sp, #18]
36; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
37; NONEON-NOSVE-NEXT:    strh w8, [sp, #16]
38; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
39; NONEON-NOSVE-NEXT:    add sp, sp, #32
40; NONEON-NOSVE-NEXT:    ret
41  %a = load <8 x i8>, ptr %ap
42  %val = zext <8 x i8> %a to <8 x i16>
43  ret <8 x i16> %val
44}
45
46define <4 x i32> @load_zext_v4i16i32(ptr %ap)  {
47; CHECK-LABEL: load_zext_v4i16i32:
48; CHECK:       // %bb.0:
49; CHECK-NEXT:    ptrue p0.s, vl4
50; CHECK-NEXT:    ld1h { z0.s }, p0/z, [x0]
51; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
52; CHECK-NEXT:    ret
53;
54; NONEON-NOSVE-LABEL: load_zext_v4i16i32:
55; NONEON-NOSVE:       // %bb.0:
56; NONEON-NOSVE-NEXT:    sub sp, sp, #32
57; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
58; NONEON-NOSVE-NEXT:    ldr d0, [x0]
59; NONEON-NOSVE-NEXT:    str d0, [sp, #8]
60; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #14]
61; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #12]
62; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #24]
63; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #10]
64; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #8]
65; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #16]
66; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
67; NONEON-NOSVE-NEXT:    add sp, sp, #32
68; NONEON-NOSVE-NEXT:    ret
69  %a = load <4 x i16>, ptr %ap
70  %val = zext <4 x i16> %a to <4 x i32>
71  ret <4 x i32> %val
72}
73
74define <2 x i64> @load_zext_v2i32i64(ptr %ap) {
75; CHECK-LABEL: load_zext_v2i32i64:
76; CHECK:       // %bb.0:
77; CHECK-NEXT:    ptrue p0.d, vl2
78; CHECK-NEXT:    ld1w { z0.d }, p0/z, [x0]
79; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
80; CHECK-NEXT:    ret
81;
82; NONEON-NOSVE-LABEL: load_zext_v2i32i64:
83; NONEON-NOSVE:       // %bb.0:
84; NONEON-NOSVE-NEXT:    sub sp, sp, #32
85; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
86; NONEON-NOSVE-NEXT:    ldr d0, [x0]
87; NONEON-NOSVE-NEXT:    str d0, [sp, #8]
88; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #8]
89; NONEON-NOSVE-NEXT:    stp w9, wzr, [sp, #24]
90; NONEON-NOSVE-NEXT:    stp w8, wzr, [sp, #16]
91; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
92; NONEON-NOSVE-NEXT:    add sp, sp, #32
93; NONEON-NOSVE-NEXT:    ret
94  %a = load <2 x i32>, ptr %ap
95  %val = zext <2 x i32> %a to <2 x i64>
96  ret <2 x i64> %val
97}
98
99define <2 x i256> @load_zext_v2i64i256(ptr %ap) {
100; CHECK-LABEL: load_zext_v2i64i256:
101; CHECK:       // %bb.0:
102; CHECK-NEXT:    ldr q0, [x0]
103; CHECK-NEXT:    mov x1, xzr
104; CHECK-NEXT:    mov x2, xzr
105; CHECK-NEXT:    mov x3, xzr
106; CHECK-NEXT:    mov x5, xzr
107; CHECK-NEXT:    mov x6, xzr
108; CHECK-NEXT:    mov z1.d, z0.d[1]
109; CHECK-NEXT:    fmov x0, d0
110; CHECK-NEXT:    mov x7, xzr
111; CHECK-NEXT:    fmov x4, d1
112; CHECK-NEXT:    ret
113;
114; NONEON-NOSVE-LABEL: load_zext_v2i64i256:
115; NONEON-NOSVE:       // %bb.0:
116; NONEON-NOSVE-NEXT:    ldr q0, [x0]
117; NONEON-NOSVE-NEXT:    str q0, [sp, #-16]!
118; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
119; NONEON-NOSVE-NEXT:    ldp x0, x4, [sp], #16
120; NONEON-NOSVE-NEXT:    mov x1, xzr
121; NONEON-NOSVE-NEXT:    mov x2, xzr
122; NONEON-NOSVE-NEXT:    mov x3, xzr
123; NONEON-NOSVE-NEXT:    mov x5, xzr
124; NONEON-NOSVE-NEXT:    mov x6, xzr
125; NONEON-NOSVE-NEXT:    mov x7, xzr
126; NONEON-NOSVE-NEXT:    ret
127  %a = load <2 x i64>, ptr %ap
128  %val = zext <2 x i64> %a to <2 x i256>
129  ret <2 x i256> %val
130}
131
132define <16 x i32> @load_sext_v16i8i32(ptr %ap)  {
133; CHECK-LABEL: load_sext_v16i8i32:
134; CHECK:       // %bb.0:
135; CHECK-NEXT:    ptrue p0.s, vl4
136; CHECK-NEXT:    mov w8, #4 // =0x4
137; CHECK-NEXT:    mov w9, #8 // =0x8
138; CHECK-NEXT:    mov w10, #12 // =0xc
139; CHECK-NEXT:    ld1sb { z0.s }, p0/z, [x0]
140; CHECK-NEXT:    ld1sb { z1.s }, p0/z, [x0, x8]
141; CHECK-NEXT:    ld1sb { z2.s }, p0/z, [x0, x9]
142; CHECK-NEXT:    ld1sb { z3.s }, p0/z, [x0, x10]
143; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
144; CHECK-NEXT:    // kill: def $q1 killed $q1 killed $z1
145; CHECK-NEXT:    // kill: def $q2 killed $q2 killed $z2
146; CHECK-NEXT:    // kill: def $q3 killed $q3 killed $z3
147; CHECK-NEXT:    ret
148;
149; NONEON-NOSVE-LABEL: load_sext_v16i8i32:
150; NONEON-NOSVE:       // %bb.0:
151; NONEON-NOSVE-NEXT:    sub sp, sp, #160
152; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 160
153; NONEON-NOSVE-NEXT:    ldr q0, [x0]
154; NONEON-NOSVE-NEXT:    str q0, [sp]
155; NONEON-NOSVE-NEXT:    ldp d1, d0, [sp]
156; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #16]
157; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #27]
158; NONEON-NOSVE-NEXT:    strh w8, [sp, #62]
159; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #26]
160; NONEON-NOSVE-NEXT:    strh w8, [sp, #60]
161; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #25]
162; NONEON-NOSVE-NEXT:    strh w8, [sp, #58]
163; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #24]
164; NONEON-NOSVE-NEXT:    strh w8, [sp, #56]
165; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #31]
166; NONEON-NOSVE-NEXT:    strh w8, [sp, #54]
167; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #30]
168; NONEON-NOSVE-NEXT:    strh w8, [sp, #52]
169; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #29]
170; NONEON-NOSVE-NEXT:    strh w8, [sp, #50]
171; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #28]
172; NONEON-NOSVE-NEXT:    strh w8, [sp, #48]
173; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #19]
174; NONEON-NOSVE-NEXT:    ldp d0, d1, [sp, #48]
175; NONEON-NOSVE-NEXT:    strh w8, [sp, #46]
176; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #18]
177; NONEON-NOSVE-NEXT:    strh w8, [sp, #44]
178; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #17]
179; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #80]
180; NONEON-NOSVE-NEXT:    strh w8, [sp, #42]
181; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #16]
182; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #94]
183; NONEON-NOSVE-NEXT:    strh w8, [sp, #40]
184; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #23]
185; NONEON-NOSVE-NEXT:    strh w8, [sp, #38]
186; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #22]
187; NONEON-NOSVE-NEXT:    strh w8, [sp, #36]
188; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #21]
189; NONEON-NOSVE-NEXT:    strh w8, [sp, #34]
190; NONEON-NOSVE-NEXT:    ldrsb w8, [sp, #20]
191; NONEON-NOSVE-NEXT:    strh w8, [sp, #32]
192; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #92]
193; NONEON-NOSVE-NEXT:    ldp d0, d1, [sp, #32]
194; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #152]
195; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #90]
196; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #88]
197; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #144]
198; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #86]
199; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #84]
200; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #64]
201; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #136]
202; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #82]
203; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #80]
204; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #128]
205; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #78]
206; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #76]
207; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #128]
208; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #120]
209; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #74]
210; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #72]
211; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #112]
212; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #70]
213; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #68]
214; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #104]
215; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #66]
216; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #64]
217; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #96]
218; NONEON-NOSVE-NEXT:    ldp q3, q2, [sp, #96]
219; NONEON-NOSVE-NEXT:    add sp, sp, #160
220; NONEON-NOSVE-NEXT:    ret
221  %a = load <16 x i8>, ptr %ap
222  %val = sext <16 x i8> %a to <16 x i32>
223  ret <16 x i32> %val
224}
225
226define <8 x i32> @load_sext_v8i16i32(ptr %ap)  {
227; CHECK-LABEL: load_sext_v8i16i32:
228; CHECK:       // %bb.0:
229; CHECK-NEXT:    ptrue p0.s, vl4
230; CHECK-NEXT:    mov x8, #4 // =0x4
231; CHECK-NEXT:    ld1sh { z1.s }, p0/z, [x0, x8, lsl #1]
232; CHECK-NEXT:    ld1sh { z0.s }, p0/z, [x0]
233; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
234; CHECK-NEXT:    // kill: def $q1 killed $q1 killed $z1
235; CHECK-NEXT:    ret
236;
237; NONEON-NOSVE-LABEL: load_sext_v8i16i32:
238; NONEON-NOSVE:       // %bb.0:
239; NONEON-NOSVE-NEXT:    ldr q0, [x0]
240; NONEON-NOSVE-NEXT:    str q0, [sp, #-64]!
241; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
242; NONEON-NOSVE-NEXT:    ldp d1, d0, [sp]
243; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #16]
244; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #30]
245; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #28]
246; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #56]
247; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #26]
248; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #24]
249; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #48]
250; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #22]
251; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #20]
252; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #40]
253; NONEON-NOSVE-NEXT:    ldrsh w9, [sp, #18]
254; NONEON-NOSVE-NEXT:    ldrsh w8, [sp, #16]
255; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #32]
256; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #32]
257; NONEON-NOSVE-NEXT:    add sp, sp, #64
258; NONEON-NOSVE-NEXT:    ret
259  %a = load <8 x i16>, ptr %ap
260  %val = sext <8 x i16> %a to <8 x i32>
261  ret <8 x i32> %val
262}
263
264define <4 x i256> @load_sext_v4i32i256(ptr %ap) {
265; CHECK-LABEL: load_sext_v4i32i256:
266; CHECK:       // %bb.0:
267; CHECK-NEXT:    ldr q0, [x0]
268; CHECK-NEXT:    sunpklo z1.d, z0.s
269; CHECK-NEXT:    ext z0.b, z0.b, z0.b, #8
270; CHECK-NEXT:    sunpklo z0.d, z0.s
271; CHECK-NEXT:    fmov x9, d1
272; CHECK-NEXT:    mov z1.d, z1.d[1]
273; CHECK-NEXT:    fmov x11, d0
274; CHECK-NEXT:    mov z0.d, z0.d[1]
275; CHECK-NEXT:    asr x10, x9, #63
276; CHECK-NEXT:    stp x9, x10, [x8]
277; CHECK-NEXT:    fmov x9, d1
278; CHECK-NEXT:    asr x12, x11, #63
279; CHECK-NEXT:    stp x10, x10, [x8, #16]
280; CHECK-NEXT:    stp x11, x12, [x8, #64]
281; CHECK-NEXT:    fmov x11, d0
282; CHECK-NEXT:    asr x10, x9, #63
283; CHECK-NEXT:    stp x12, x12, [x8, #80]
284; CHECK-NEXT:    stp x10, x10, [x8, #48]
285; CHECK-NEXT:    asr x12, x11, #63
286; CHECK-NEXT:    stp x9, x10, [x8, #32]
287; CHECK-NEXT:    stp x12, x12, [x8, #112]
288; CHECK-NEXT:    stp x11, x12, [x8, #96]
289; CHECK-NEXT:    ret
290;
291; NONEON-NOSVE-LABEL: load_sext_v4i32i256:
292; NONEON-NOSVE:       // %bb.0:
293; NONEON-NOSVE-NEXT:    ldr q0, [x0]
294; NONEON-NOSVE-NEXT:    str q0, [sp, #-96]!
295; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 96
296; NONEON-NOSVE-NEXT:    ldp d1, d0, [sp]
297; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #16]
298; NONEON-NOSVE-NEXT:    ldpsw x9, x10, [sp, #24]
299; NONEON-NOSVE-NEXT:    stp x9, x10, [sp, #48]
300; NONEON-NOSVE-NEXT:    ldpsw x9, x10, [sp, #16]
301; NONEON-NOSVE-NEXT:    stp x9, x10, [sp, #32]
302; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
303; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #64]
304; NONEON-NOSVE-NEXT:    ldp x11, x9, [sp, #64]
305; NONEON-NOSVE-NEXT:    ldp x12, x13, [sp, #80]
306; NONEON-NOSVE-NEXT:    asr x10, x9, #63
307; NONEON-NOSVE-NEXT:    asr x14, x11, #63
308; NONEON-NOSVE-NEXT:    stp x10, x10, [x8, #112]
309; NONEON-NOSVE-NEXT:    stp x9, x10, [x8, #96]
310; NONEON-NOSVE-NEXT:    asr x9, x13, #63
311; NONEON-NOSVE-NEXT:    asr x10, x12, #63
312; NONEON-NOSVE-NEXT:    stp x14, x14, [x8, #80]
313; NONEON-NOSVE-NEXT:    stp x11, x14, [x8, #64]
314; NONEON-NOSVE-NEXT:    stp x9, x9, [x8, #48]
315; NONEON-NOSVE-NEXT:    stp x13, x9, [x8, #32]
316; NONEON-NOSVE-NEXT:    stp x10, x10, [x8, #16]
317; NONEON-NOSVE-NEXT:    stp x12, x10, [x8]
318; NONEON-NOSVE-NEXT:    add sp, sp, #96
319; NONEON-NOSVE-NEXT:    ret
320  %a = load <4 x i32>, ptr %ap
321  %val = sext <4 x i32> %a to <4 x i256>
322  ret <4 x i256> %val
323}
324
325define <2 x i256> @load_sext_v2i64i256(ptr %ap) {
326; CHECK-LABEL: load_sext_v2i64i256:
327; CHECK:       // %bb.0:
328; CHECK-NEXT:    ldr q0, [x0]
329; CHECK-NEXT:    mov z1.d, z0.d[1]
330; CHECK-NEXT:    fmov x8, d0
331; CHECK-NEXT:    fmov x9, d1
332; CHECK-NEXT:    asr x8, x8, #63
333; CHECK-NEXT:    fmov d3, x8
334; CHECK-NEXT:    mov z2.d, x8
335; CHECK-NEXT:    asr x9, x9, #63
336; CHECK-NEXT:    fmov d4, x9
337; CHECK-NEXT:    zip1 z0.d, z0.d, z3.d
338; CHECK-NEXT:    mov z3.d, x9
339; CHECK-NEXT:    fmov x2, d2
340; CHECK-NEXT:    zip1 z1.d, z1.d, z4.d
341; CHECK-NEXT:    mov z4.d, z2.d[1]
342; CHECK-NEXT:    mov z5.d, z0.d[1]
343; CHECK-NEXT:    mov z6.d, z3.d[1]
344; CHECK-NEXT:    fmov x0, d0
345; CHECK-NEXT:    fmov x6, d3
346; CHECK-NEXT:    mov z2.d, z1.d[1]
347; CHECK-NEXT:    fmov x3, d4
348; CHECK-NEXT:    fmov x1, d5
349; CHECK-NEXT:    fmov x4, d1
350; CHECK-NEXT:    fmov x7, d6
351; CHECK-NEXT:    fmov x5, d2
352; CHECK-NEXT:    ret
353;
354; NONEON-NOSVE-LABEL: load_sext_v2i64i256:
355; NONEON-NOSVE:       // %bb.0:
356; NONEON-NOSVE-NEXT:    sub sp, sp, #144
357; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 144
358; NONEON-NOSVE-NEXT:    ldr q0, [x0]
359; NONEON-NOSVE-NEXT:    str q0, [sp]
360; NONEON-NOSVE-NEXT:    ldp x8, x10, [sp]
361; NONEON-NOSVE-NEXT:    asr x9, x8, #63
362; NONEON-NOSVE-NEXT:    stp x8, x9, [sp, #32]
363; NONEON-NOSVE-NEXT:    asr x8, x10, #63
364; NONEON-NOSVE-NEXT:    stp x9, x9, [sp, #16]
365; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #16]
366; NONEON-NOSVE-NEXT:    stp x10, x8, [sp, #64]
367; NONEON-NOSVE-NEXT:    stp x8, x8, [sp, #48]
368; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #80]
369; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #48]
370; NONEON-NOSVE-NEXT:    ldp x0, x1, [sp, #96]
371; NONEON-NOSVE-NEXT:    ldp x2, x3, [sp, #80]
372; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #112]
373; NONEON-NOSVE-NEXT:    ldp x4, x5, [sp, #128]
374; NONEON-NOSVE-NEXT:    ldp x6, x7, [sp, #112]
375; NONEON-NOSVE-NEXT:    add sp, sp, #144
376; NONEON-NOSVE-NEXT:    ret
377  %a = load <2 x i64>, ptr %ap
378  %val = sext <2 x i64> %a to <2 x i256>
379  ret <2 x i256> %val
380}
381
382define <16 x i64> @load_zext_v16i16i64(ptr %ap)  {
383; CHECK-LABEL: load_zext_v16i16i64:
384; CHECK:       // %bb.0:
385; CHECK-NEXT:    ptrue p0.d, vl2
386; CHECK-NEXT:    mov x8, #2 // =0x2
387; CHECK-NEXT:    mov x9, #4 // =0x4
388; CHECK-NEXT:    mov x10, #6 // =0x6
389; CHECK-NEXT:    ld1h { z1.d }, p0/z, [x0, x8, lsl #1]
390; CHECK-NEXT:    mov x8, #8 // =0x8
391; CHECK-NEXT:    ld1h { z2.d }, p0/z, [x0, x9, lsl #1]
392; CHECK-NEXT:    ld1h { z3.d }, p0/z, [x0, x10, lsl #1]
393; CHECK-NEXT:    mov x9, #10 // =0xa
394; CHECK-NEXT:    ld1h { z4.d }, p0/z, [x0, x8, lsl #1]
395; CHECK-NEXT:    mov x8, #12 // =0xc
396; CHECK-NEXT:    mov x10, #14 // =0xe
397; CHECK-NEXT:    ld1h { z0.d }, p0/z, [x0]
398; CHECK-NEXT:    ld1h { z5.d }, p0/z, [x0, x9, lsl #1]
399; CHECK-NEXT:    ld1h { z6.d }, p0/z, [x0, x8, lsl #1]
400; CHECK-NEXT:    ld1h { z7.d }, p0/z, [x0, x10, lsl #1]
401; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
402; CHECK-NEXT:    // kill: def $q1 killed $q1 killed $z1
403; CHECK-NEXT:    // kill: def $q2 killed $q2 killed $z2
404; CHECK-NEXT:    // kill: def $q3 killed $q3 killed $z3
405; CHECK-NEXT:    // kill: def $q4 killed $q4 killed $z4
406; CHECK-NEXT:    // kill: def $q5 killed $q5 killed $z5
407; CHECK-NEXT:    // kill: def $q6 killed $q6 killed $z6
408; CHECK-NEXT:    // kill: def $q7 killed $q7 killed $z7
409; CHECK-NEXT:    ret
410;
411; NONEON-NOSVE-LABEL: load_zext_v16i16i64:
412; NONEON-NOSVE:       // %bb.0:
413; NONEON-NOSVE-NEXT:    sub sp, sp, #336
414; NONEON-NOSVE-NEXT:    str x29, [sp, #320] // 8-byte Folded Spill
415; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 336
416; NONEON-NOSVE-NEXT:    .cfi_offset w29, -16
417; NONEON-NOSVE-NEXT:    ldp q0, q1, [x0]
418; NONEON-NOSVE-NEXT:    str wzr, [sp, #316]
419; NONEON-NOSVE-NEXT:    str wzr, [sp, #308]
420; NONEON-NOSVE-NEXT:    ldr x29, [sp, #320] // 8-byte Folded Reload
421; NONEON-NOSVE-NEXT:    str wzr, [sp, #300]
422; NONEON-NOSVE-NEXT:    stp q0, q1, [sp]
423; NONEON-NOSVE-NEXT:    ldp d1, d0, [sp]
424; NONEON-NOSVE-NEXT:    str wzr, [sp, #292]
425; NONEON-NOSVE-NEXT:    str wzr, [sp, #284]
426; NONEON-NOSVE-NEXT:    str wzr, [sp, #276]
427; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #40]
428; NONEON-NOSVE-NEXT:    ldp d1, d0, [sp, #16]
429; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #50]
430; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #48]
431; NONEON-NOSVE-NEXT:    str wzr, [sp, #268]
432; NONEON-NOSVE-NEXT:    str wzr, [sp, #260]
433; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #96]
434; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #54]
435; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #52]
436; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #56]
437; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #88]
438; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #42]
439; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #40]
440; NONEON-NOSVE-NEXT:    ldp d0, d1, [sp, #88]
441; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #80]
442; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #46]
443; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #44]
444; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #72]
445; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #66]
446; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #64]
447; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #152]
448; NONEON-NOSVE-NEXT:    ldp d0, d1, [sp, #72]
449; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #128]
450; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #70]
451; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #68]
452; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #120]
453; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #58]
454; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #56]
455; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #136]
456; NONEON-NOSVE-NEXT:    ldp d2, d1, [sp, #120]
457; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #112]
458; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #62]
459; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #60]
460; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #104]
461; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #160]
462; NONEON-NOSVE-NEXT:    str d1, [sp, #328]
463; NONEON-NOSVE-NEXT:    ldp d0, d1, [sp, #104]
464; NONEON-NOSVE-NEXT:    stp w9, wzr, [sp, #248]
465; NONEON-NOSVE-NEXT:    stp w8, wzr, [sp, #240]
466; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #152]
467; NONEON-NOSVE-NEXT:    stp d1, d2, [sp, #176]
468; NONEON-NOSVE-NEXT:    str d0, [sp, #168]
469; NONEON-NOSVE-NEXT:    stp w9, wzr, [sp, #232]
470; NONEON-NOSVE-NEXT:    stp w8, wzr, [sp, #224]
471; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #144]
472; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #224]
473; NONEON-NOSVE-NEXT:    stp w9, wzr, [sp, #216]
474; NONEON-NOSVE-NEXT:    stp w8, wzr, [sp, #208]
475; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #136]
476; NONEON-NOSVE-NEXT:    stp w8, wzr, [sp, #192]
477; NONEON-NOSVE-NEXT:    ldr w8, [sp, #332]
478; NONEON-NOSVE-NEXT:    stp w9, wzr, [sp, #200]
479; NONEON-NOSVE-NEXT:    str w8, [sp, #312]
480; NONEON-NOSVE-NEXT:    ldr w8, [sp, #328]
481; NONEON-NOSVE-NEXT:    ldp q3, q2, [sp, #192]
482; NONEON-NOSVE-NEXT:    str w8, [sp, #304]
483; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #184]
484; NONEON-NOSVE-NEXT:    str w9, [sp, #296]
485; NONEON-NOSVE-NEXT:    str w8, [sp, #288]
486; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #176]
487; NONEON-NOSVE-NEXT:    ldp q5, q4, [sp, #288]
488; NONEON-NOSVE-NEXT:    str w9, [sp, #280]
489; NONEON-NOSVE-NEXT:    str w8, [sp, #272]
490; NONEON-NOSVE-NEXT:    ldp w8, w9, [sp, #168]
491; NONEON-NOSVE-NEXT:    str w9, [sp, #264]
492; NONEON-NOSVE-NEXT:    str w8, [sp, #256]
493; NONEON-NOSVE-NEXT:    ldp q7, q6, [sp, #256]
494; NONEON-NOSVE-NEXT:    add sp, sp, #336
495; NONEON-NOSVE-NEXT:    ret
496  %a = load <16 x i16>, ptr %ap
497  %val = zext <16 x i16> %a to <16 x i64>
498  ret <16 x i64> %val
499}
500