xref: /llvm-project/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-rev.ll (revision 8e0cd7382adacd8bc1741dc26bc0be6bdf8e238a) 
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
6
7target triple = "aarch64-unknown-linux-gnu"
8
9;
10; RBIT
11;
12
13define <4 x i8> @bitreverse_v4i8(<4 x i8> %op) {
14; CHECK-LABEL: bitreverse_v4i8:
15; CHECK:       // %bb.0:
16; CHECK-NEXT:    ptrue p0.h, vl4
17; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
18; CHECK-NEXT:    rbit z0.h, p0/m, z0.h
19; CHECK-NEXT:    lsr z0.h, z0.h, #8
20; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
21; CHECK-NEXT:    ret
22;
23; NONEON-NOSVE-LABEL: bitreverse_v4i8:
24; NONEON-NOSVE:       // %bb.0:
25; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
26; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
27; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #6]
28; NONEON-NOSVE-NEXT:    rbit w8, w8
29; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
30; NONEON-NOSVE-NEXT:    strh w8, [sp, #14]
31; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #4]
32; NONEON-NOSVE-NEXT:    rbit w8, w8
33; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
34; NONEON-NOSVE-NEXT:    strh w8, [sp, #12]
35; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #2]
36; NONEON-NOSVE-NEXT:    rbit w8, w8
37; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
38; NONEON-NOSVE-NEXT:    strh w8, [sp, #10]
39; NONEON-NOSVE-NEXT:    ldrh w8, [sp]
40; NONEON-NOSVE-NEXT:    rbit w8, w8
41; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
42; NONEON-NOSVE-NEXT:    strh w8, [sp, #8]
43; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
44; NONEON-NOSVE-NEXT:    add sp, sp, #16
45; NONEON-NOSVE-NEXT:    ret
46  %res = call <4 x i8> @llvm.bitreverse.v4i8(<4 x i8> %op)
47  ret <4 x i8> %res
48}
49
50define <8 x i8> @bitreverse_v8i8(<8 x i8> %op) {
51; CHECK-LABEL: bitreverse_v8i8:
52; CHECK:       // %bb.0:
53; CHECK-NEXT:    ptrue p0.b, vl8
54; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
55; CHECK-NEXT:    rbit z0.b, p0/m, z0.b
56; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
57; CHECK-NEXT:    ret
58;
59; NONEON-NOSVE-LABEL: bitreverse_v8i8:
60; NONEON-NOSVE:       // %bb.0:
61; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
62; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
63; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
64; NONEON-NOSVE-NEXT:    rbit w8, w8
65; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
66; NONEON-NOSVE-NEXT:    strb w8, [sp, #15]
67; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
68; NONEON-NOSVE-NEXT:    rbit w8, w8
69; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
70; NONEON-NOSVE-NEXT:    strb w8, [sp, #14]
71; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
72; NONEON-NOSVE-NEXT:    rbit w8, w8
73; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
74; NONEON-NOSVE-NEXT:    strb w8, [sp, #13]
75; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
76; NONEON-NOSVE-NEXT:    rbit w8, w8
77; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
78; NONEON-NOSVE-NEXT:    strb w8, [sp, #12]
79; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
80; NONEON-NOSVE-NEXT:    rbit w8, w8
81; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
82; NONEON-NOSVE-NEXT:    strb w8, [sp, #11]
83; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
84; NONEON-NOSVE-NEXT:    rbit w8, w8
85; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
86; NONEON-NOSVE-NEXT:    strb w8, [sp, #10]
87; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
88; NONEON-NOSVE-NEXT:    rbit w8, w8
89; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
90; NONEON-NOSVE-NEXT:    strb w8, [sp, #9]
91; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
92; NONEON-NOSVE-NEXT:    rbit w8, w8
93; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
94; NONEON-NOSVE-NEXT:    strb w8, [sp, #8]
95; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
96; NONEON-NOSVE-NEXT:    add sp, sp, #16
97; NONEON-NOSVE-NEXT:    ret
98  %res = call <8 x i8> @llvm.bitreverse.v8i8(<8 x i8> %op)
99  ret <8 x i8> %res
100}
101
102define <16 x i8> @bitreverse_v16i8(<16 x i8> %op) {
103; CHECK-LABEL: bitreverse_v16i8:
104; CHECK:       // %bb.0:
105; CHECK-NEXT:    ptrue p0.b, vl16
106; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
107; CHECK-NEXT:    rbit z0.b, p0/m, z0.b
108; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
109; CHECK-NEXT:    ret
110;
111; NONEON-NOSVE-LABEL: bitreverse_v16i8:
112; NONEON-NOSVE:       // %bb.0:
113; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
114; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
115; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
116; NONEON-NOSVE-NEXT:    rbit w8, w8
117; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
118; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
119; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
120; NONEON-NOSVE-NEXT:    rbit w8, w8
121; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
122; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
123; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
124; NONEON-NOSVE-NEXT:    rbit w8, w8
125; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
126; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
127; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
128; NONEON-NOSVE-NEXT:    rbit w8, w8
129; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
130; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
131; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
132; NONEON-NOSVE-NEXT:    rbit w8, w8
133; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
134; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
135; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
136; NONEON-NOSVE-NEXT:    rbit w8, w8
137; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
138; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
139; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
140; NONEON-NOSVE-NEXT:    rbit w8, w8
141; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
142; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
143; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
144; NONEON-NOSVE-NEXT:    rbit w8, w8
145; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
146; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
147; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
148; NONEON-NOSVE-NEXT:    rbit w8, w8
149; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
150; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
151; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
152; NONEON-NOSVE-NEXT:    rbit w8, w8
153; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
154; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
155; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
156; NONEON-NOSVE-NEXT:    rbit w8, w8
157; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
158; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
159; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
160; NONEON-NOSVE-NEXT:    rbit w8, w8
161; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
162; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
163; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
164; NONEON-NOSVE-NEXT:    rbit w8, w8
165; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
166; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
167; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
168; NONEON-NOSVE-NEXT:    rbit w8, w8
169; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
170; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
171; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
172; NONEON-NOSVE-NEXT:    rbit w8, w8
173; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
174; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
175; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
176; NONEON-NOSVE-NEXT:    rbit w8, w8
177; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
178; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
179; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
180; NONEON-NOSVE-NEXT:    add sp, sp, #32
181; NONEON-NOSVE-NEXT:    ret
182  %res = call <16 x i8> @llvm.bitreverse.v16i8(<16 x i8> %op)
183  ret <16 x i8> %res
184}
185
186define void @bitreverse_v32i8(ptr %a) {
187; CHECK-LABEL: bitreverse_v32i8:
188; CHECK:       // %bb.0:
189; CHECK-NEXT:    ldp q0, q1, [x0]
190; CHECK-NEXT:    ptrue p0.b, vl16
191; CHECK-NEXT:    rbit z0.b, p0/m, z0.b
192; CHECK-NEXT:    rbit z1.b, p0/m, z1.b
193; CHECK-NEXT:    stp q0, q1, [x0]
194; CHECK-NEXT:    ret
195;
196; NONEON-NOSVE-LABEL: bitreverse_v32i8:
197; NONEON-NOSVE:       // %bb.0:
198; NONEON-NOSVE-NEXT:    ldp q1, q0, [x0]
199; NONEON-NOSVE-NEXT:    stp q1, q0, [sp, #-64]!
200; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
201; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #31]
202; NONEON-NOSVE-NEXT:    rbit w8, w8
203; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
204; NONEON-NOSVE-NEXT:    strb w8, [sp, #63]
205; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #30]
206; NONEON-NOSVE-NEXT:    rbit w8, w8
207; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
208; NONEON-NOSVE-NEXT:    strb w8, [sp, #62]
209; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #29]
210; NONEON-NOSVE-NEXT:    rbit w8, w8
211; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
212; NONEON-NOSVE-NEXT:    strb w8, [sp, #61]
213; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #28]
214; NONEON-NOSVE-NEXT:    rbit w8, w8
215; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
216; NONEON-NOSVE-NEXT:    strb w8, [sp, #60]
217; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #27]
218; NONEON-NOSVE-NEXT:    rbit w8, w8
219; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
220; NONEON-NOSVE-NEXT:    strb w8, [sp, #59]
221; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #26]
222; NONEON-NOSVE-NEXT:    rbit w8, w8
223; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
224; NONEON-NOSVE-NEXT:    strb w8, [sp, #58]
225; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #25]
226; NONEON-NOSVE-NEXT:    rbit w8, w8
227; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
228; NONEON-NOSVE-NEXT:    strb w8, [sp, #57]
229; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #24]
230; NONEON-NOSVE-NEXT:    rbit w8, w8
231; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
232; NONEON-NOSVE-NEXT:    strb w8, [sp, #56]
233; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #23]
234; NONEON-NOSVE-NEXT:    rbit w8, w8
235; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
236; NONEON-NOSVE-NEXT:    strb w8, [sp, #55]
237; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #22]
238; NONEON-NOSVE-NEXT:    rbit w8, w8
239; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
240; NONEON-NOSVE-NEXT:    strb w8, [sp, #54]
241; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #21]
242; NONEON-NOSVE-NEXT:    rbit w8, w8
243; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
244; NONEON-NOSVE-NEXT:    strb w8, [sp, #53]
245; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #20]
246; NONEON-NOSVE-NEXT:    rbit w8, w8
247; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
248; NONEON-NOSVE-NEXT:    strb w8, [sp, #52]
249; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #19]
250; NONEON-NOSVE-NEXT:    rbit w8, w8
251; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
252; NONEON-NOSVE-NEXT:    strb w8, [sp, #51]
253; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #18]
254; NONEON-NOSVE-NEXT:    rbit w8, w8
255; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
256; NONEON-NOSVE-NEXT:    strb w8, [sp, #50]
257; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #17]
258; NONEON-NOSVE-NEXT:    rbit w8, w8
259; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
260; NONEON-NOSVE-NEXT:    strb w8, [sp, #49]
261; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #16]
262; NONEON-NOSVE-NEXT:    rbit w8, w8
263; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
264; NONEON-NOSVE-NEXT:    strb w8, [sp, #48]
265; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
266; NONEON-NOSVE-NEXT:    rbit w8, w8
267; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
268; NONEON-NOSVE-NEXT:    strb w8, [sp, #47]
269; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
270; NONEON-NOSVE-NEXT:    rbit w8, w8
271; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
272; NONEON-NOSVE-NEXT:    strb w8, [sp, #46]
273; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
274; NONEON-NOSVE-NEXT:    rbit w8, w8
275; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
276; NONEON-NOSVE-NEXT:    strb w8, [sp, #45]
277; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
278; NONEON-NOSVE-NEXT:    rbit w8, w8
279; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
280; NONEON-NOSVE-NEXT:    strb w8, [sp, #44]
281; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
282; NONEON-NOSVE-NEXT:    rbit w8, w8
283; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
284; NONEON-NOSVE-NEXT:    strb w8, [sp, #43]
285; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
286; NONEON-NOSVE-NEXT:    rbit w8, w8
287; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
288; NONEON-NOSVE-NEXT:    strb w8, [sp, #42]
289; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
290; NONEON-NOSVE-NEXT:    rbit w8, w8
291; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
292; NONEON-NOSVE-NEXT:    strb w8, [sp, #41]
293; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
294; NONEON-NOSVE-NEXT:    rbit w8, w8
295; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
296; NONEON-NOSVE-NEXT:    strb w8, [sp, #40]
297; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
298; NONEON-NOSVE-NEXT:    rbit w8, w8
299; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
300; NONEON-NOSVE-NEXT:    strb w8, [sp, #39]
301; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
302; NONEON-NOSVE-NEXT:    rbit w8, w8
303; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
304; NONEON-NOSVE-NEXT:    strb w8, [sp, #38]
305; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
306; NONEON-NOSVE-NEXT:    rbit w8, w8
307; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
308; NONEON-NOSVE-NEXT:    strb w8, [sp, #37]
309; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
310; NONEON-NOSVE-NEXT:    rbit w8, w8
311; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
312; NONEON-NOSVE-NEXT:    strb w8, [sp, #36]
313; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
314; NONEON-NOSVE-NEXT:    rbit w8, w8
315; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
316; NONEON-NOSVE-NEXT:    strb w8, [sp, #35]
317; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
318; NONEON-NOSVE-NEXT:    rbit w8, w8
319; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
320; NONEON-NOSVE-NEXT:    strb w8, [sp, #34]
321; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
322; NONEON-NOSVE-NEXT:    rbit w8, w8
323; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
324; NONEON-NOSVE-NEXT:    strb w8, [sp, #33]
325; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
326; NONEON-NOSVE-NEXT:    rbit w8, w8
327; NONEON-NOSVE-NEXT:    lsr w8, w8, #24
328; NONEON-NOSVE-NEXT:    strb w8, [sp, #32]
329; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
330; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
331; NONEON-NOSVE-NEXT:    add sp, sp, #64
332; NONEON-NOSVE-NEXT:    ret
333  %op = load <32 x i8>, ptr %a
334  %res = call <32 x i8> @llvm.bitreverse.v32i8(<32 x i8> %op)
335  store <32 x i8> %res, ptr %a
336  ret void
337}
338
339define <2 x i16> @bitreverse_v2i16(<2 x i16> %op) {
340; CHECK-LABEL: bitreverse_v2i16:
341; CHECK:       // %bb.0:
342; CHECK-NEXT:    ptrue p0.s, vl2
343; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
344; CHECK-NEXT:    rbit z0.s, p0/m, z0.s
345; CHECK-NEXT:    lsr z0.s, z0.s, #16
346; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
347; CHECK-NEXT:    ret
348;
349; NONEON-NOSVE-LABEL: bitreverse_v2i16:
350; NONEON-NOSVE:       // %bb.0:
351; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
352; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
353; NONEON-NOSVE-NEXT:    ldr w8, [sp, #4]
354; NONEON-NOSVE-NEXT:    rbit w8, w8
355; NONEON-NOSVE-NEXT:    lsr w9, w8, #16
356; NONEON-NOSVE-NEXT:    ldr w8, [sp]
357; NONEON-NOSVE-NEXT:    rbit w8, w8
358; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
359; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #8]
360; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
361; NONEON-NOSVE-NEXT:    add sp, sp, #16
362; NONEON-NOSVE-NEXT:    ret
363  %res = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %op)
364  ret <2 x i16> %res
365}
366
367define <4 x i16> @bitreverse_v4i16(<4 x i16> %op) {
368; CHECK-LABEL: bitreverse_v4i16:
369; CHECK:       // %bb.0:
370; CHECK-NEXT:    ptrue p0.h, vl4
371; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
372; CHECK-NEXT:    rbit z0.h, p0/m, z0.h
373; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
374; CHECK-NEXT:    ret
375;
376; NONEON-NOSVE-LABEL: bitreverse_v4i16:
377; NONEON-NOSVE:       // %bb.0:
378; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
379; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
380; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #6]
381; NONEON-NOSVE-NEXT:    rbit w8, w8
382; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
383; NONEON-NOSVE-NEXT:    strh w8, [sp, #14]
384; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #4]
385; NONEON-NOSVE-NEXT:    rbit w8, w8
386; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
387; NONEON-NOSVE-NEXT:    strh w8, [sp, #12]
388; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #2]
389; NONEON-NOSVE-NEXT:    rbit w8, w8
390; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
391; NONEON-NOSVE-NEXT:    strh w8, [sp, #10]
392; NONEON-NOSVE-NEXT:    ldrh w8, [sp]
393; NONEON-NOSVE-NEXT:    rbit w8, w8
394; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
395; NONEON-NOSVE-NEXT:    strh w8, [sp, #8]
396; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
397; NONEON-NOSVE-NEXT:    add sp, sp, #16
398; NONEON-NOSVE-NEXT:    ret
399  %res = call <4 x i16> @llvm.bitreverse.v4i16(<4 x i16> %op)
400  ret <4 x i16> %res
401}
402
403define <8 x i16> @bitreverse_v8i16(<8 x i16> %op) {
404; CHECK-LABEL: bitreverse_v8i16:
405; CHECK:       // %bb.0:
406; CHECK-NEXT:    ptrue p0.h, vl8
407; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
408; CHECK-NEXT:    rbit z0.h, p0/m, z0.h
409; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
410; CHECK-NEXT:    ret
411;
412; NONEON-NOSVE-LABEL: bitreverse_v8i16:
413; NONEON-NOSVE:       // %bb.0:
414; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
415; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
416; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #14]
417; NONEON-NOSVE-NEXT:    rbit w8, w8
418; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
419; NONEON-NOSVE-NEXT:    strh w8, [sp, #30]
420; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #12]
421; NONEON-NOSVE-NEXT:    rbit w8, w8
422; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
423; NONEON-NOSVE-NEXT:    strh w8, [sp, #28]
424; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #10]
425; NONEON-NOSVE-NEXT:    rbit w8, w8
426; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
427; NONEON-NOSVE-NEXT:    strh w8, [sp, #26]
428; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #8]
429; NONEON-NOSVE-NEXT:    rbit w8, w8
430; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
431; NONEON-NOSVE-NEXT:    strh w8, [sp, #24]
432; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #6]
433; NONEON-NOSVE-NEXT:    rbit w8, w8
434; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
435; NONEON-NOSVE-NEXT:    strh w8, [sp, #22]
436; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #4]
437; NONEON-NOSVE-NEXT:    rbit w8, w8
438; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
439; NONEON-NOSVE-NEXT:    strh w8, [sp, #20]
440; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #2]
441; NONEON-NOSVE-NEXT:    rbit w8, w8
442; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
443; NONEON-NOSVE-NEXT:    strh w8, [sp, #18]
444; NONEON-NOSVE-NEXT:    ldrh w8, [sp]
445; NONEON-NOSVE-NEXT:    rbit w8, w8
446; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
447; NONEON-NOSVE-NEXT:    strh w8, [sp, #16]
448; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
449; NONEON-NOSVE-NEXT:    add sp, sp, #32
450; NONEON-NOSVE-NEXT:    ret
451  %res = call <8 x i16> @llvm.bitreverse.v8i16(<8 x i16> %op)
452  ret <8 x i16> %res
453}
454
455define void @bitreverse_v16i16(ptr %a) {
456; CHECK-LABEL: bitreverse_v16i16:
457; CHECK:       // %bb.0:
458; CHECK-NEXT:    ldp q0, q1, [x0]
459; CHECK-NEXT:    ptrue p0.h, vl8
460; CHECK-NEXT:    rbit z0.h, p0/m, z0.h
461; CHECK-NEXT:    rbit z1.h, p0/m, z1.h
462; CHECK-NEXT:    stp q0, q1, [x0]
463; CHECK-NEXT:    ret
464;
465; NONEON-NOSVE-LABEL: bitreverse_v16i16:
466; NONEON-NOSVE:       // %bb.0:
467; NONEON-NOSVE-NEXT:    ldp q1, q0, [x0]
468; NONEON-NOSVE-NEXT:    stp q1, q0, [sp, #-64]!
469; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
470; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #30]
471; NONEON-NOSVE-NEXT:    rbit w8, w8
472; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
473; NONEON-NOSVE-NEXT:    strh w8, [sp, #62]
474; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #28]
475; NONEON-NOSVE-NEXT:    rbit w8, w8
476; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
477; NONEON-NOSVE-NEXT:    strh w8, [sp, #60]
478; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #26]
479; NONEON-NOSVE-NEXT:    rbit w8, w8
480; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
481; NONEON-NOSVE-NEXT:    strh w8, [sp, #58]
482; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #24]
483; NONEON-NOSVE-NEXT:    rbit w8, w8
484; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
485; NONEON-NOSVE-NEXT:    strh w8, [sp, #56]
486; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #22]
487; NONEON-NOSVE-NEXT:    rbit w8, w8
488; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
489; NONEON-NOSVE-NEXT:    strh w8, [sp, #54]
490; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #20]
491; NONEON-NOSVE-NEXT:    rbit w8, w8
492; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
493; NONEON-NOSVE-NEXT:    strh w8, [sp, #52]
494; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #18]
495; NONEON-NOSVE-NEXT:    rbit w8, w8
496; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
497; NONEON-NOSVE-NEXT:    strh w8, [sp, #50]
498; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #16]
499; NONEON-NOSVE-NEXT:    rbit w8, w8
500; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
501; NONEON-NOSVE-NEXT:    strh w8, [sp, #48]
502; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #14]
503; NONEON-NOSVE-NEXT:    rbit w8, w8
504; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
505; NONEON-NOSVE-NEXT:    strh w8, [sp, #46]
506; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #12]
507; NONEON-NOSVE-NEXT:    rbit w8, w8
508; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
509; NONEON-NOSVE-NEXT:    strh w8, [sp, #44]
510; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #10]
511; NONEON-NOSVE-NEXT:    rbit w8, w8
512; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
513; NONEON-NOSVE-NEXT:    strh w8, [sp, #42]
514; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #8]
515; NONEON-NOSVE-NEXT:    rbit w8, w8
516; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
517; NONEON-NOSVE-NEXT:    strh w8, [sp, #40]
518; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #6]
519; NONEON-NOSVE-NEXT:    rbit w8, w8
520; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
521; NONEON-NOSVE-NEXT:    strh w8, [sp, #38]
522; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #4]
523; NONEON-NOSVE-NEXT:    rbit w8, w8
524; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
525; NONEON-NOSVE-NEXT:    strh w8, [sp, #36]
526; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #2]
527; NONEON-NOSVE-NEXT:    rbit w8, w8
528; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
529; NONEON-NOSVE-NEXT:    strh w8, [sp, #34]
530; NONEON-NOSVE-NEXT:    ldrh w8, [sp]
531; NONEON-NOSVE-NEXT:    rbit w8, w8
532; NONEON-NOSVE-NEXT:    lsr w8, w8, #16
533; NONEON-NOSVE-NEXT:    strh w8, [sp, #32]
534; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
535; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
536; NONEON-NOSVE-NEXT:    add sp, sp, #64
537; NONEON-NOSVE-NEXT:    ret
538  %op = load <16 x i16>, ptr %a
539  %res = call <16 x i16> @llvm.bitreverse.v16i16(<16 x i16> %op)
540  store <16 x i16> %res, ptr %a
541  ret void
542}
543
544define <2 x i32> @bitreverse_v2i32(<2 x i32> %op) {
545; CHECK-LABEL: bitreverse_v2i32:
546; CHECK:       // %bb.0:
547; CHECK-NEXT:    ptrue p0.s, vl2
548; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
549; CHECK-NEXT:    rbit z0.s, p0/m, z0.s
550; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
551; CHECK-NEXT:    ret
552;
553; NONEON-NOSVE-LABEL: bitreverse_v2i32:
554; NONEON-NOSVE:       // %bb.0:
555; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
556; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
557; NONEON-NOSVE-NEXT:    ldr w8, [sp, #4]
558; NONEON-NOSVE-NEXT:    rbit w9, w8
559; NONEON-NOSVE-NEXT:    ldr w8, [sp]
560; NONEON-NOSVE-NEXT:    rbit w8, w8
561; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #8]
562; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
563; NONEON-NOSVE-NEXT:    add sp, sp, #16
564; NONEON-NOSVE-NEXT:    ret
565  %res = call <2 x i32> @llvm.bitreverse.v2i32(<2 x i32> %op)
566  ret <2 x i32> %res
567}
568
569define <4 x i32> @bitreverse_v4i32(<4 x i32> %op) {
570; CHECK-LABEL: bitreverse_v4i32:
571; CHECK:       // %bb.0:
572; CHECK-NEXT:    ptrue p0.s, vl4
573; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
574; CHECK-NEXT:    rbit z0.s, p0/m, z0.s
575; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
576; CHECK-NEXT:    ret
577;
578; NONEON-NOSVE-LABEL: bitreverse_v4i32:
579; NONEON-NOSVE:       // %bb.0:
580; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
581; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
582; NONEON-NOSVE-NEXT:    ldr w8, [sp, #12]
583; NONEON-NOSVE-NEXT:    rbit w9, w8
584; NONEON-NOSVE-NEXT:    ldr w8, [sp, #8]
585; NONEON-NOSVE-NEXT:    rbit w8, w8
586; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #24]
587; NONEON-NOSVE-NEXT:    ldr w8, [sp, #4]
588; NONEON-NOSVE-NEXT:    rbit w9, w8
589; NONEON-NOSVE-NEXT:    ldr w8, [sp]
590; NONEON-NOSVE-NEXT:    rbit w8, w8
591; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #16]
592; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
593; NONEON-NOSVE-NEXT:    add sp, sp, #32
594; NONEON-NOSVE-NEXT:    ret
595  %res = call <4 x i32> @llvm.bitreverse.v4i32(<4 x i32> %op)
596  ret <4 x i32> %res
597}
598
599define void @bitreverse_v8i32(ptr %a) {
600; CHECK-LABEL: bitreverse_v8i32:
601; CHECK:       // %bb.0:
602; CHECK-NEXT:    ldp q0, q1, [x0]
603; CHECK-NEXT:    ptrue p0.s, vl4
604; CHECK-NEXT:    rbit z0.s, p0/m, z0.s
605; CHECK-NEXT:    rbit z1.s, p0/m, z1.s
606; CHECK-NEXT:    stp q0, q1, [x0]
607; CHECK-NEXT:    ret
608;
609; NONEON-NOSVE-LABEL: bitreverse_v8i32:
610; NONEON-NOSVE:       // %bb.0:
611; NONEON-NOSVE-NEXT:    ldp q1, q0, [x0]
612; NONEON-NOSVE-NEXT:    stp q1, q0, [sp, #-64]!
613; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
614; NONEON-NOSVE-NEXT:    ldr w8, [sp, #28]
615; NONEON-NOSVE-NEXT:    rbit w9, w8
616; NONEON-NOSVE-NEXT:    ldr w8, [sp, #24]
617; NONEON-NOSVE-NEXT:    rbit w8, w8
618; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #56]
619; NONEON-NOSVE-NEXT:    ldr w8, [sp, #20]
620; NONEON-NOSVE-NEXT:    rbit w9, w8
621; NONEON-NOSVE-NEXT:    ldr w8, [sp, #16]
622; NONEON-NOSVE-NEXT:    rbit w8, w8
623; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #48]
624; NONEON-NOSVE-NEXT:    ldr w8, [sp, #12]
625; NONEON-NOSVE-NEXT:    rbit w9, w8
626; NONEON-NOSVE-NEXT:    ldr w8, [sp, #8]
627; NONEON-NOSVE-NEXT:    rbit w8, w8
628; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #40]
629; NONEON-NOSVE-NEXT:    ldr w8, [sp, #4]
630; NONEON-NOSVE-NEXT:    rbit w9, w8
631; NONEON-NOSVE-NEXT:    ldr w8, [sp]
632; NONEON-NOSVE-NEXT:    rbit w8, w8
633; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #32]
634; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
635; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
636; NONEON-NOSVE-NEXT:    add sp, sp, #64
637; NONEON-NOSVE-NEXT:    ret
638  %op = load <8 x i32>, ptr %a
639  %res = call <8 x i32> @llvm.bitreverse.v8i32(<8 x i32> %op)
640  store <8 x i32> %res, ptr %a
641  ret void
642}
643
644define <1 x i64> @bitreverse_v1i64(<1 x i64> %op) {
645; CHECK-LABEL: bitreverse_v1i64:
646; CHECK:       // %bb.0:
647; CHECK-NEXT:    ptrue p0.d, vl1
648; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
649; CHECK-NEXT:    rbit z0.d, p0/m, z0.d
650; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
651; CHECK-NEXT:    ret
652;
653; NONEON-NOSVE-LABEL: bitreverse_v1i64:
654; NONEON-NOSVE:       // %bb.0:
655; NONEON-NOSVE-NEXT:    sub sp, sp, #16
656; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
657; NONEON-NOSVE-NEXT:    fmov x8, d0
658; NONEON-NOSVE-NEXT:    rbit x8, x8
659; NONEON-NOSVE-NEXT:    str x8, [sp, #8]
660; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
661; NONEON-NOSVE-NEXT:    add sp, sp, #16
662; NONEON-NOSVE-NEXT:    ret
663  %res = call <1 x i64> @llvm.bitreverse.v1i64(<1 x i64> %op)
664  ret <1 x i64> %res
665}
666
667define <2 x i64> @bitreverse_v2i64(<2 x i64> %op) {
668; CHECK-LABEL: bitreverse_v2i64:
669; CHECK:       // %bb.0:
670; CHECK-NEXT:    ptrue p0.d, vl2
671; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
672; CHECK-NEXT:    rbit z0.d, p0/m, z0.d
673; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
674; CHECK-NEXT:    ret
675;
676; NONEON-NOSVE-LABEL: bitreverse_v2i64:
677; NONEON-NOSVE:       // %bb.0:
678; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
679; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
680; NONEON-NOSVE-NEXT:    ldr x8, [sp, #8]
681; NONEON-NOSVE-NEXT:    rbit x9, x8
682; NONEON-NOSVE-NEXT:    ldr x8, [sp]
683; NONEON-NOSVE-NEXT:    rbit x8, x8
684; NONEON-NOSVE-NEXT:    stp x8, x9, [sp, #16]
685; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
686; NONEON-NOSVE-NEXT:    add sp, sp, #32
687; NONEON-NOSVE-NEXT:    ret
688  %res = call <2 x i64> @llvm.bitreverse.v2i64(<2 x i64> %op)
689  ret <2 x i64> %res
690}
691
692define void @bitreverse_v4i64(ptr %a) {
693; CHECK-LABEL: bitreverse_v4i64:
694; CHECK:       // %bb.0:
695; CHECK-NEXT:    ldp q0, q1, [x0]
696; CHECK-NEXT:    ptrue p0.d, vl2
697; CHECK-NEXT:    rbit z0.d, p0/m, z0.d
698; CHECK-NEXT:    rbit z1.d, p0/m, z1.d
699; CHECK-NEXT:    stp q0, q1, [x0]
700; CHECK-NEXT:    ret
701;
702; NONEON-NOSVE-LABEL: bitreverse_v4i64:
703; NONEON-NOSVE:       // %bb.0:
704; NONEON-NOSVE-NEXT:    ldp q1, q0, [x0]
705; NONEON-NOSVE-NEXT:    stp q1, q0, [sp, #-64]!
706; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
707; NONEON-NOSVE-NEXT:    ldr x8, [sp, #24]
708; NONEON-NOSVE-NEXT:    rbit x9, x8
709; NONEON-NOSVE-NEXT:    ldr x8, [sp, #16]
710; NONEON-NOSVE-NEXT:    rbit x8, x8
711; NONEON-NOSVE-NEXT:    stp x8, x9, [sp, #48]
712; NONEON-NOSVE-NEXT:    ldr x8, [sp, #8]
713; NONEON-NOSVE-NEXT:    rbit x9, x8
714; NONEON-NOSVE-NEXT:    ldr x8, [sp]
715; NONEON-NOSVE-NEXT:    rbit x8, x8
716; NONEON-NOSVE-NEXT:    stp x8, x9, [sp, #32]
717; NONEON-NOSVE-NEXT:    ldp q0, q1, [sp, #32]
718; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
719; NONEON-NOSVE-NEXT:    add sp, sp, #64
720; NONEON-NOSVE-NEXT:    ret
721  %op = load <4 x i64>, ptr %a
722  %res = call <4 x i64> @llvm.bitreverse.v4i64(<4 x i64> %op)
723  store <4 x i64> %res, ptr %a
724  ret void
725}
726
727;
728; REVB
729;
730
731define <2 x i16> @bswap_v2i16(<2 x i16> %op) {
732; CHECK-LABEL: bswap_v2i16:
733; CHECK:       // %bb.0:
734; CHECK-NEXT:    ptrue p0.s, vl2
735; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
736; CHECK-NEXT:    revb z0.s, p0/m, z0.s
737; CHECK-NEXT:    lsr z0.s, z0.s, #16
738; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
739; CHECK-NEXT:    ret
740;
741; NONEON-NOSVE-LABEL: bswap_v2i16:
742; NONEON-NOSVE:       // %bb.0:
743; NONEON-NOSVE-NEXT:    str d0, [sp, #-32]!
744; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
745; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
746; NONEON-NOSVE-NEXT:    strb w8, [sp, #15]
747; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
748; NONEON-NOSVE-NEXT:    strb w8, [sp, #14]
749; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
750; NONEON-NOSVE-NEXT:    strb w8, [sp, #13]
751; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
752; NONEON-NOSVE-NEXT:    strb w8, [sp, #12]
753; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
754; NONEON-NOSVE-NEXT:    strb w8, [sp, #11]
755; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
756; NONEON-NOSVE-NEXT:    strb w8, [sp, #10]
757; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
758; NONEON-NOSVE-NEXT:    strb w8, [sp, #9]
759; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
760; NONEON-NOSVE-NEXT:    strb w8, [sp, #8]
761; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
762; NONEON-NOSVE-NEXT:    str d0, [sp, #16]
763; NONEON-NOSVE-NEXT:    ldrh w9, [sp, #22]
764; NONEON-NOSVE-NEXT:    ldrh w8, [sp, #18]
765; NONEON-NOSVE-NEXT:    stp w8, w9, [sp, #24]
766; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
767; NONEON-NOSVE-NEXT:    add sp, sp, #32
768; NONEON-NOSVE-NEXT:    ret
769  %res = call <2 x i16> @llvm.bswap.v2i16(<2 x i16> %op)
770  ret <2 x i16> %res
771}
772
773define <4 x i16> @bswap_v4i16(<4 x i16> %op) {
774; CHECK-LABEL: bswap_v4i16:
775; CHECK:       // %bb.0:
776; CHECK-NEXT:    ptrue p0.h, vl4
777; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
778; CHECK-NEXT:    revb z0.h, p0/m, z0.h
779; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
780; CHECK-NEXT:    ret
781;
782; NONEON-NOSVE-LABEL: bswap_v4i16:
783; NONEON-NOSVE:       // %bb.0:
784; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
785; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
786; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
787; NONEON-NOSVE-NEXT:    strb w8, [sp, #15]
788; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
789; NONEON-NOSVE-NEXT:    strb w8, [sp, #14]
790; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
791; NONEON-NOSVE-NEXT:    strb w8, [sp, #13]
792; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
793; NONEON-NOSVE-NEXT:    strb w8, [sp, #12]
794; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
795; NONEON-NOSVE-NEXT:    strb w8, [sp, #11]
796; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
797; NONEON-NOSVE-NEXT:    strb w8, [sp, #10]
798; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
799; NONEON-NOSVE-NEXT:    strb w8, [sp, #9]
800; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
801; NONEON-NOSVE-NEXT:    strb w8, [sp, #8]
802; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
803; NONEON-NOSVE-NEXT:    add sp, sp, #16
804; NONEON-NOSVE-NEXT:    ret
805  %res = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %op)
806  ret <4 x i16> %res
807}
808
809define <8 x i16> @bswap_v8i16(<8 x i16> %op) {
810; CHECK-LABEL: bswap_v8i16:
811; CHECK:       // %bb.0:
812; CHECK-NEXT:    ptrue p0.h, vl8
813; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
814; CHECK-NEXT:    revb z0.h, p0/m, z0.h
815; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
816; CHECK-NEXT:    ret
817;
818; NONEON-NOSVE-LABEL: bswap_v8i16:
819; NONEON-NOSVE:       // %bb.0:
820; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
821; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
822; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
823; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
824; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
825; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
826; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
827; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
828; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
829; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
830; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
831; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
832; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
833; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
834; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
835; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
836; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
837; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
838; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
839; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
840; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
841; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
842; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
843; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
844; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
845; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
846; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
847; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
848; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
849; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
850; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
851; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
852; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
853; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
854; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
855; NONEON-NOSVE-NEXT:    add sp, sp, #32
856; NONEON-NOSVE-NEXT:    ret
857  %res = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %op)
858  ret <8 x i16> %res
859}
860
861define void @bswap_v16i16(ptr %a) {
862; CHECK-LABEL: bswap_v16i16:
863; CHECK:       // %bb.0:
864; CHECK-NEXT:    ldp q0, q1, [x0]
865; CHECK-NEXT:    ptrue p0.h, vl8
866; CHECK-NEXT:    revb z0.h, p0/m, z0.h
867; CHECK-NEXT:    revb z1.h, p0/m, z1.h
868; CHECK-NEXT:    stp q0, q1, [x0]
869; CHECK-NEXT:    ret
870;
871; NONEON-NOSVE-LABEL: bswap_v16i16:
872; NONEON-NOSVE:       // %bb.0:
873; NONEON-NOSVE-NEXT:    ldr q0, [x0, #16]
874; NONEON-NOSVE-NEXT:    str q0, [sp, #-64]!
875; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
876; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
877; NONEON-NOSVE-NEXT:    ldr q0, [x0]
878; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
879; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
880; NONEON-NOSVE-NEXT:    str q0, [sp, #32]
881; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
882; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
883; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
884; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
885; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
886; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
887; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
888; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
889; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
890; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
891; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
892; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
893; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
894; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
895; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
896; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
897; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
898; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
899; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
900; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
901; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
902; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
903; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
904; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
905; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
906; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
907; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
908; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
909; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
910; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #46]
911; NONEON-NOSVE-NEXT:    ldr q1, [sp, #16]
912; NONEON-NOSVE-NEXT:    strb w8, [sp, #63]
913; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #47]
914; NONEON-NOSVE-NEXT:    strb w8, [sp, #62]
915; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #44]
916; NONEON-NOSVE-NEXT:    strb w8, [sp, #61]
917; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #45]
918; NONEON-NOSVE-NEXT:    strb w8, [sp, #60]
919; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #42]
920; NONEON-NOSVE-NEXT:    strb w8, [sp, #59]
921; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #43]
922; NONEON-NOSVE-NEXT:    strb w8, [sp, #58]
923; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #40]
924; NONEON-NOSVE-NEXT:    strb w8, [sp, #57]
925; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #41]
926; NONEON-NOSVE-NEXT:    strb w8, [sp, #56]
927; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #38]
928; NONEON-NOSVE-NEXT:    strb w8, [sp, #55]
929; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #39]
930; NONEON-NOSVE-NEXT:    strb w8, [sp, #54]
931; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #36]
932; NONEON-NOSVE-NEXT:    strb w8, [sp, #53]
933; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #37]
934; NONEON-NOSVE-NEXT:    strb w8, [sp, #52]
935; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #34]
936; NONEON-NOSVE-NEXT:    strb w8, [sp, #51]
937; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #35]
938; NONEON-NOSVE-NEXT:    strb w8, [sp, #50]
939; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #32]
940; NONEON-NOSVE-NEXT:    strb w8, [sp, #49]
941; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #33]
942; NONEON-NOSVE-NEXT:    strb w8, [sp, #48]
943; NONEON-NOSVE-NEXT:    ldr q0, [sp, #48]
944; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
945; NONEON-NOSVE-NEXT:    add sp, sp, #64
946; NONEON-NOSVE-NEXT:    ret
947  %op = load <16 x i16>, ptr %a
948  %res = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> %op)
949  store <16 x i16> %res, ptr %a
950  ret void
951}
952
953define <2 x i32> @bswap_v2i32(<2 x i32> %op) {
954; CHECK-LABEL: bswap_v2i32:
955; CHECK:       // %bb.0:
956; CHECK-NEXT:    ptrue p0.s, vl2
957; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
958; CHECK-NEXT:    revb z0.s, p0/m, z0.s
959; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
960; CHECK-NEXT:    ret
961;
962; NONEON-NOSVE-LABEL: bswap_v2i32:
963; NONEON-NOSVE:       // %bb.0:
964; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
965; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
966; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
967; NONEON-NOSVE-NEXT:    strb w8, [sp, #15]
968; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
969; NONEON-NOSVE-NEXT:    strb w8, [sp, #14]
970; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
971; NONEON-NOSVE-NEXT:    strb w8, [sp, #13]
972; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
973; NONEON-NOSVE-NEXT:    strb w8, [sp, #12]
974; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
975; NONEON-NOSVE-NEXT:    strb w8, [sp, #11]
976; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
977; NONEON-NOSVE-NEXT:    strb w8, [sp, #10]
978; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
979; NONEON-NOSVE-NEXT:    strb w8, [sp, #9]
980; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
981; NONEON-NOSVE-NEXT:    strb w8, [sp, #8]
982; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
983; NONEON-NOSVE-NEXT:    add sp, sp, #16
984; NONEON-NOSVE-NEXT:    ret
985  %res = call <2 x i32> @llvm.bswap.v2i32(<2 x i32> %op)
986  ret <2 x i32> %res
987}
988
989define <4 x i32> @bswap_v4i32(<4 x i32> %op) {
990; CHECK-LABEL: bswap_v4i32:
991; CHECK:       // %bb.0:
992; CHECK-NEXT:    ptrue p0.s, vl4
993; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
994; CHECK-NEXT:    revb z0.s, p0/m, z0.s
995; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
996; CHECK-NEXT:    ret
997;
998; NONEON-NOSVE-LABEL: bswap_v4i32:
999; NONEON-NOSVE:       // %bb.0:
1000; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
1001; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
1002; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
1003; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
1004; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
1005; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
1006; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
1007; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
1008; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
1009; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
1010; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
1011; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
1012; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
1013; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
1014; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
1015; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
1016; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
1017; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
1018; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
1019; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
1020; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
1021; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
1022; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
1023; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
1024; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
1025; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
1026; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
1027; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
1028; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
1029; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
1030; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
1031; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
1032; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
1033; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
1034; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
1035; NONEON-NOSVE-NEXT:    add sp, sp, #32
1036; NONEON-NOSVE-NEXT:    ret
1037  %res = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %op)
1038  ret <4 x i32> %res
1039}
1040
1041define void @bswap_v8i32(ptr %a) {
1042; CHECK-LABEL: bswap_v8i32:
1043; CHECK:       // %bb.0:
1044; CHECK-NEXT:    ldp q0, q1, [x0]
1045; CHECK-NEXT:    ptrue p0.s, vl4
1046; CHECK-NEXT:    revb z0.s, p0/m, z0.s
1047; CHECK-NEXT:    revb z1.s, p0/m, z1.s
1048; CHECK-NEXT:    stp q0, q1, [x0]
1049; CHECK-NEXT:    ret
1050;
1051; NONEON-NOSVE-LABEL: bswap_v8i32:
1052; NONEON-NOSVE:       // %bb.0:
1053; NONEON-NOSVE-NEXT:    ldr q0, [x0, #16]
1054; NONEON-NOSVE-NEXT:    str q0, [sp, #-64]!
1055; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
1056; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
1057; NONEON-NOSVE-NEXT:    ldr q0, [x0]
1058; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
1059; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
1060; NONEON-NOSVE-NEXT:    str q0, [sp, #32]
1061; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
1062; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
1063; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
1064; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
1065; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
1066; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
1067; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
1068; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
1069; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
1070; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
1071; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
1072; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
1073; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
1074; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
1075; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
1076; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
1077; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
1078; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
1079; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
1080; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
1081; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
1082; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
1083; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
1084; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
1085; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
1086; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
1087; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
1088; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
1089; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
1090; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #44]
1091; NONEON-NOSVE-NEXT:    ldr q1, [sp, #16]
1092; NONEON-NOSVE-NEXT:    strb w8, [sp, #63]
1093; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #45]
1094; NONEON-NOSVE-NEXT:    strb w8, [sp, #62]
1095; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #46]
1096; NONEON-NOSVE-NEXT:    strb w8, [sp, #61]
1097; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #47]
1098; NONEON-NOSVE-NEXT:    strb w8, [sp, #60]
1099; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #40]
1100; NONEON-NOSVE-NEXT:    strb w8, [sp, #59]
1101; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #41]
1102; NONEON-NOSVE-NEXT:    strb w8, [sp, #58]
1103; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #42]
1104; NONEON-NOSVE-NEXT:    strb w8, [sp, #57]
1105; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #43]
1106; NONEON-NOSVE-NEXT:    strb w8, [sp, #56]
1107; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #36]
1108; NONEON-NOSVE-NEXT:    strb w8, [sp, #55]
1109; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #37]
1110; NONEON-NOSVE-NEXT:    strb w8, [sp, #54]
1111; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #38]
1112; NONEON-NOSVE-NEXT:    strb w8, [sp, #53]
1113; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #39]
1114; NONEON-NOSVE-NEXT:    strb w8, [sp, #52]
1115; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #32]
1116; NONEON-NOSVE-NEXT:    strb w8, [sp, #51]
1117; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #33]
1118; NONEON-NOSVE-NEXT:    strb w8, [sp, #50]
1119; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #34]
1120; NONEON-NOSVE-NEXT:    strb w8, [sp, #49]
1121; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #35]
1122; NONEON-NOSVE-NEXT:    strb w8, [sp, #48]
1123; NONEON-NOSVE-NEXT:    ldr q0, [sp, #48]
1124; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
1125; NONEON-NOSVE-NEXT:    add sp, sp, #64
1126; NONEON-NOSVE-NEXT:    ret
1127  %op = load <8 x i32>, ptr %a
1128  %res = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> %op)
1129  store <8 x i32> %res, ptr %a
1130  ret void
1131}
1132
1133define <1 x i64> @bswap_v1i64(<1 x i64> %op) {
1134; CHECK-LABEL: bswap_v1i64:
1135; CHECK:       // %bb.0:
1136; CHECK-NEXT:    ptrue p0.d, vl1
1137; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
1138; CHECK-NEXT:    revb z0.d, p0/m, z0.d
1139; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
1140; CHECK-NEXT:    ret
1141;
1142; NONEON-NOSVE-LABEL: bswap_v1i64:
1143; NONEON-NOSVE:       // %bb.0:
1144; NONEON-NOSVE-NEXT:    str d0, [sp, #-16]!
1145; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
1146; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
1147; NONEON-NOSVE-NEXT:    strb w8, [sp, #15]
1148; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
1149; NONEON-NOSVE-NEXT:    strb w8, [sp, #14]
1150; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
1151; NONEON-NOSVE-NEXT:    strb w8, [sp, #13]
1152; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
1153; NONEON-NOSVE-NEXT:    strb w8, [sp, #12]
1154; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
1155; NONEON-NOSVE-NEXT:    strb w8, [sp, #11]
1156; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
1157; NONEON-NOSVE-NEXT:    strb w8, [sp, #10]
1158; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
1159; NONEON-NOSVE-NEXT:    strb w8, [sp, #9]
1160; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
1161; NONEON-NOSVE-NEXT:    strb w8, [sp, #8]
1162; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
1163; NONEON-NOSVE-NEXT:    add sp, sp, #16
1164; NONEON-NOSVE-NEXT:    ret
1165  %res = call <1 x i64> @llvm.bswap.v1i64(<1 x i64> %op)
1166  ret <1 x i64> %res
1167}
1168
1169define <2 x i64> @bswap_v2i64(<2 x i64> %op) {
1170; CHECK-LABEL: bswap_v2i64:
1171; CHECK:       // %bb.0:
1172; CHECK-NEXT:    ptrue p0.d, vl2
1173; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
1174; CHECK-NEXT:    revb z0.d, p0/m, z0.d
1175; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
1176; CHECK-NEXT:    ret
1177;
1178; NONEON-NOSVE-LABEL: bswap_v2i64:
1179; NONEON-NOSVE:       // %bb.0:
1180; NONEON-NOSVE-NEXT:    str q0, [sp, #-32]!
1181; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
1182; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
1183; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
1184; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
1185; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
1186; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
1187; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
1188; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
1189; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
1190; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
1191; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
1192; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
1193; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
1194; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
1195; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
1196; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
1197; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
1198; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
1199; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
1200; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
1201; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
1202; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
1203; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
1204; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
1205; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
1206; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
1207; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
1208; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
1209; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
1210; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
1211; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
1212; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
1213; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
1214; NONEON-NOSVE-NEXT:    ldr q0, [sp, #16]
1215; NONEON-NOSVE-NEXT:    add sp, sp, #32
1216; NONEON-NOSVE-NEXT:    ret
1217  %res = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %op)
1218  ret <2 x i64> %res
1219}
1220
1221define void @bswap_v4i64(ptr %a) {
1222; CHECK-LABEL: bswap_v4i64:
1223; CHECK:       // %bb.0:
1224; CHECK-NEXT:    ldp q0, q1, [x0]
1225; CHECK-NEXT:    ptrue p0.d, vl2
1226; CHECK-NEXT:    revb z0.d, p0/m, z0.d
1227; CHECK-NEXT:    revb z1.d, p0/m, z1.d
1228; CHECK-NEXT:    stp q0, q1, [x0]
1229; CHECK-NEXT:    ret
1230;
1231; NONEON-NOSVE-LABEL: bswap_v4i64:
1232; NONEON-NOSVE:       // %bb.0:
1233; NONEON-NOSVE-NEXT:    ldr q0, [x0, #16]
1234; NONEON-NOSVE-NEXT:    str q0, [sp, #-64]!
1235; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 64
1236; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #8]
1237; NONEON-NOSVE-NEXT:    ldr q0, [x0]
1238; NONEON-NOSVE-NEXT:    strb w8, [sp, #31]
1239; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #9]
1240; NONEON-NOSVE-NEXT:    str q0, [sp, #32]
1241; NONEON-NOSVE-NEXT:    strb w8, [sp, #30]
1242; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #10]
1243; NONEON-NOSVE-NEXT:    strb w8, [sp, #29]
1244; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #11]
1245; NONEON-NOSVE-NEXT:    strb w8, [sp, #28]
1246; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #12]
1247; NONEON-NOSVE-NEXT:    strb w8, [sp, #27]
1248; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #13]
1249; NONEON-NOSVE-NEXT:    strb w8, [sp, #26]
1250; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #14]
1251; NONEON-NOSVE-NEXT:    strb w8, [sp, #25]
1252; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #15]
1253; NONEON-NOSVE-NEXT:    strb w8, [sp, #24]
1254; NONEON-NOSVE-NEXT:    ldrb w8, [sp]
1255; NONEON-NOSVE-NEXT:    strb w8, [sp, #23]
1256; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #1]
1257; NONEON-NOSVE-NEXT:    strb w8, [sp, #22]
1258; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #2]
1259; NONEON-NOSVE-NEXT:    strb w8, [sp, #21]
1260; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #3]
1261; NONEON-NOSVE-NEXT:    strb w8, [sp, #20]
1262; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #4]
1263; NONEON-NOSVE-NEXT:    strb w8, [sp, #19]
1264; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #5]
1265; NONEON-NOSVE-NEXT:    strb w8, [sp, #18]
1266; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #6]
1267; NONEON-NOSVE-NEXT:    strb w8, [sp, #17]
1268; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #7]
1269; NONEON-NOSVE-NEXT:    strb w8, [sp, #16]
1270; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #40]
1271; NONEON-NOSVE-NEXT:    ldr q1, [sp, #16]
1272; NONEON-NOSVE-NEXT:    strb w8, [sp, #63]
1273; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #41]
1274; NONEON-NOSVE-NEXT:    strb w8, [sp, #62]
1275; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #42]
1276; NONEON-NOSVE-NEXT:    strb w8, [sp, #61]
1277; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #43]
1278; NONEON-NOSVE-NEXT:    strb w8, [sp, #60]
1279; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #44]
1280; NONEON-NOSVE-NEXT:    strb w8, [sp, #59]
1281; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #45]
1282; NONEON-NOSVE-NEXT:    strb w8, [sp, #58]
1283; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #46]
1284; NONEON-NOSVE-NEXT:    strb w8, [sp, #57]
1285; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #47]
1286; NONEON-NOSVE-NEXT:    strb w8, [sp, #56]
1287; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #32]
1288; NONEON-NOSVE-NEXT:    strb w8, [sp, #55]
1289; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #33]
1290; NONEON-NOSVE-NEXT:    strb w8, [sp, #54]
1291; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #34]
1292; NONEON-NOSVE-NEXT:    strb w8, [sp, #53]
1293; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #35]
1294; NONEON-NOSVE-NEXT:    strb w8, [sp, #52]
1295; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #36]
1296; NONEON-NOSVE-NEXT:    strb w8, [sp, #51]
1297; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #37]
1298; NONEON-NOSVE-NEXT:    strb w8, [sp, #50]
1299; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #38]
1300; NONEON-NOSVE-NEXT:    strb w8, [sp, #49]
1301; NONEON-NOSVE-NEXT:    ldrb w8, [sp, #39]
1302; NONEON-NOSVE-NEXT:    strb w8, [sp, #48]
1303; NONEON-NOSVE-NEXT:    ldr q0, [sp, #48]
1304; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
1305; NONEON-NOSVE-NEXT:    add sp, sp, #64
1306; NONEON-NOSVE-NEXT:    ret
1307  %op = load <4 x i64>, ptr %a
1308  %res = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %op)
1309  store <4 x i64> %res, ptr %a
1310  ret void
1311}
1312
1313declare <4 x i8> @llvm.bitreverse.v4i8(<4 x i8>)
1314declare <8 x i8> @llvm.bitreverse.v8i8(<8 x i8>)
1315declare <16 x i8> @llvm.bitreverse.v16i8(<16 x i8>)
1316declare <32 x i8> @llvm.bitreverse.v32i8(<32 x i8>)
1317declare <2 x i16> @llvm.bitreverse.v2i16(<2 x i16>)
1318declare <4 x i16> @llvm.bitreverse.v4i16(<4 x i16>)
1319declare <8 x i16> @llvm.bitreverse.v8i16(<8 x i16>)
1320declare <16 x i16> @llvm.bitreverse.v16i16(<16 x i16>)
1321declare <2 x i32> @llvm.bitreverse.v2i32(<2 x i32>)
1322declare <4 x i32> @llvm.bitreverse.v4i32(<4 x i32>)
1323declare <8 x i32> @llvm.bitreverse.v8i32(<8 x i32>)
1324declare <1 x i64> @llvm.bitreverse.v1i64(<1 x i64>)
1325declare <2 x i64> @llvm.bitreverse.v2i64(<2 x i64>)
1326declare <4 x i64> @llvm.bitreverse.v4i64(<4 x i64>)
1327
1328declare <2 x i16> @llvm.bswap.v2i16(<2 x i16>)
1329declare <4 x i16> @llvm.bswap.v4i16(<4 x i16>)
1330declare <8 x i16> @llvm.bswap.v8i16(<8 x i16>)
1331declare <16 x i16> @llvm.bswap.v16i16(<16 x i16>)
1332declare <2 x i32> @llvm.bswap.v2i32(<2 x i32>)
1333declare <4 x i32> @llvm.bswap.v4i32(<4 x i32>)
1334declare <8 x i32> @llvm.bswap.v8i32(<8 x i32>)
1335declare <1 x i64> @llvm.bswap.v1i64(<1 x i64>)
1336declare <2 x i64> @llvm.bswap.v2i64(<2 x i64>)
1337declare <4 x i64> @llvm.bswap.v4i64(<4 x i64>)
1338