xref: /llvm-project/llvm/test/CodeGen/AArch64/sve-streaming-mode-fixed-length-fp-select.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
6target triple = "aarch64-unknown-linux-gnu"
7
8define <2 x half> @select_v2f16(<2 x half> %op1, <2 x half> %op2, i1 %mask) {
9; CHECK-LABEL: select_v2f16:
10; CHECK:       // %bb.0:
11; CHECK-NEXT:    mov z2.h, w0
12; CHECK-NEXT:    ptrue p0.h
13; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
14; CHECK-NEXT:    // kill: def $d1 killed $d1 def $z1
15; CHECK-NEXT:    and z2.h, z2.h, #0x1
16; CHECK-NEXT:    cmpne p0.h, p0/z, z2.h, #0
17; CHECK-NEXT:    sel z0.h, p0, z0.h, z1.h
18; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
19; CHECK-NEXT:    ret
20;
21; NONEON-NOSVE-LABEL: select_v2f16:
22; NONEON-NOSVE:       // %bb.0:
23; NONEON-NOSVE-NEXT:    sub sp, sp, #32
24; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
25; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #8]
26; NONEON-NOSVE-NEXT:    tst w0, #0x1
27; NONEON-NOSVE-NEXT:    ldr h0, [sp, #22]
28; NONEON-NOSVE-NEXT:    ldr h1, [sp, #14]
29; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
30; NONEON-NOSVE-NEXT:    ldr h1, [sp, #12]
31; NONEON-NOSVE-NEXT:    str h0, [sp, #30]
32; NONEON-NOSVE-NEXT:    ldr h0, [sp, #20]
33; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
34; NONEON-NOSVE-NEXT:    ldr h1, [sp, #10]
35; NONEON-NOSVE-NEXT:    str h0, [sp, #28]
36; NONEON-NOSVE-NEXT:    ldr h0, [sp, #18]
37; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
38; NONEON-NOSVE-NEXT:    ldr h1, [sp, #8]
39; NONEON-NOSVE-NEXT:    str h0, [sp, #26]
40; NONEON-NOSVE-NEXT:    ldr h0, [sp, #16]
41; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
42; NONEON-NOSVE-NEXT:    str h0, [sp, #24]
43; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
44; NONEON-NOSVE-NEXT:    add sp, sp, #32
45; NONEON-NOSVE-NEXT:    ret
46  %sel = select i1 %mask, <2 x half> %op1, <2 x half> %op2
47  ret <2 x half> %sel
48}
49
50define <4 x half> @select_v4f16(<4 x half> %op1, <4 x half> %op2, i1 %mask) {
51; CHECK-LABEL: select_v4f16:
52; CHECK:       // %bb.0:
53; CHECK-NEXT:    mov z2.h, w0
54; CHECK-NEXT:    ptrue p0.h
55; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
56; CHECK-NEXT:    // kill: def $d1 killed $d1 def $z1
57; CHECK-NEXT:    and z2.h, z2.h, #0x1
58; CHECK-NEXT:    cmpne p0.h, p0/z, z2.h, #0
59; CHECK-NEXT:    sel z0.h, p0, z0.h, z1.h
60; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
61; CHECK-NEXT:    ret
62;
63; NONEON-NOSVE-LABEL: select_v4f16:
64; NONEON-NOSVE:       // %bb.0:
65; NONEON-NOSVE-NEXT:    sub sp, sp, #32
66; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
67; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #8]
68; NONEON-NOSVE-NEXT:    tst w0, #0x1
69; NONEON-NOSVE-NEXT:    ldr h0, [sp, #22]
70; NONEON-NOSVE-NEXT:    ldr h1, [sp, #14]
71; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
72; NONEON-NOSVE-NEXT:    ldr h1, [sp, #12]
73; NONEON-NOSVE-NEXT:    str h0, [sp, #30]
74; NONEON-NOSVE-NEXT:    ldr h0, [sp, #20]
75; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
76; NONEON-NOSVE-NEXT:    ldr h1, [sp, #10]
77; NONEON-NOSVE-NEXT:    str h0, [sp, #28]
78; NONEON-NOSVE-NEXT:    ldr h0, [sp, #18]
79; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
80; NONEON-NOSVE-NEXT:    ldr h1, [sp, #8]
81; NONEON-NOSVE-NEXT:    str h0, [sp, #26]
82; NONEON-NOSVE-NEXT:    ldr h0, [sp, #16]
83; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
84; NONEON-NOSVE-NEXT:    str h0, [sp, #24]
85; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
86; NONEON-NOSVE-NEXT:    add sp, sp, #32
87; NONEON-NOSVE-NEXT:    ret
88  %sel = select i1 %mask, <4 x half> %op1, <4 x half> %op2
89  ret <4 x half> %sel
90}
91
92define <8 x half> @select_v8f16(<8 x half> %op1, <8 x half> %op2, i1 %mask) {
93; CHECK-LABEL: select_v8f16:
94; CHECK:       // %bb.0:
95; CHECK-NEXT:    mov z2.h, w0
96; CHECK-NEXT:    ptrue p0.h
97; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
98; CHECK-NEXT:    // kill: def $q1 killed $q1 def $z1
99; CHECK-NEXT:    and z2.h, z2.h, #0x1
100; CHECK-NEXT:    cmpne p0.h, p0/z, z2.h, #0
101; CHECK-NEXT:    sel z0.h, p0, z0.h, z1.h
102; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
103; CHECK-NEXT:    ret
104;
105; NONEON-NOSVE-LABEL: select_v8f16:
106; NONEON-NOSVE:       // %bb.0:
107; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #-48]!
108; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 48
109; NONEON-NOSVE-NEXT:    ldr h0, [sp, #30]
110; NONEON-NOSVE-NEXT:    ldr h1, [sp, #14]
111; NONEON-NOSVE-NEXT:    tst w0, #0x1
112; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
113; NONEON-NOSVE-NEXT:    ldr h1, [sp, #12]
114; NONEON-NOSVE-NEXT:    str h0, [sp, #46]
115; NONEON-NOSVE-NEXT:    ldr h0, [sp, #28]
116; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
117; NONEON-NOSVE-NEXT:    ldr h1, [sp, #10]
118; NONEON-NOSVE-NEXT:    str h0, [sp, #44]
119; NONEON-NOSVE-NEXT:    ldr h0, [sp, #26]
120; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
121; NONEON-NOSVE-NEXT:    ldr h1, [sp, #8]
122; NONEON-NOSVE-NEXT:    str h0, [sp, #42]
123; NONEON-NOSVE-NEXT:    ldr h0, [sp, #24]
124; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
125; NONEON-NOSVE-NEXT:    ldr h1, [sp, #6]
126; NONEON-NOSVE-NEXT:    str h0, [sp, #40]
127; NONEON-NOSVE-NEXT:    ldr h0, [sp, #22]
128; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
129; NONEON-NOSVE-NEXT:    ldr h1, [sp, #4]
130; NONEON-NOSVE-NEXT:    str h0, [sp, #38]
131; NONEON-NOSVE-NEXT:    ldr h0, [sp, #20]
132; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
133; NONEON-NOSVE-NEXT:    ldr h1, [sp, #2]
134; NONEON-NOSVE-NEXT:    str h0, [sp, #36]
135; NONEON-NOSVE-NEXT:    ldr h0, [sp, #18]
136; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
137; NONEON-NOSVE-NEXT:    ldr h1, [sp]
138; NONEON-NOSVE-NEXT:    str h0, [sp, #34]
139; NONEON-NOSVE-NEXT:    ldr h0, [sp, #16]
140; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
141; NONEON-NOSVE-NEXT:    str h0, [sp, #32]
142; NONEON-NOSVE-NEXT:    ldr q0, [sp, #32]
143; NONEON-NOSVE-NEXT:    add sp, sp, #48
144; NONEON-NOSVE-NEXT:    ret
145  %sel = select i1 %mask, <8 x half> %op1, <8 x half> %op2
146  ret <8 x half> %sel
147}
148
149define void @select_v16f16(ptr %a, ptr %b, i1 %mask) {
150; CHECK-LABEL: select_v16f16:
151; CHECK:       // %bb.0:
152; CHECK-NEXT:    mov z0.h, w2
153; CHECK-NEXT:    ptrue p0.h
154; CHECK-NEXT:    and z0.h, z0.h, #0x1
155; CHECK-NEXT:    cmpne p0.h, p0/z, z0.h, #0
156; CHECK-NEXT:    ldr q0, [x0]
157; CHECK-NEXT:    ldr q1, [x0, #16]
158; CHECK-NEXT:    ldr q2, [x1]
159; CHECK-NEXT:    ldr q3, [x1, #16]
160; CHECK-NEXT:    sel z0.h, p0, z0.h, z2.h
161; CHECK-NEXT:    sel z1.h, p0, z1.h, z3.h
162; CHECK-NEXT:    stp q0, q1, [x0]
163; CHECK-NEXT:    ret
164;
165; NONEON-NOSVE-LABEL: select_v16f16:
166; NONEON-NOSVE:       // %bb.0:
167; NONEON-NOSVE-NEXT:    sub sp, sp, #96
168; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 96
169; NONEON-NOSVE-NEXT:    ldr q0, [x0]
170; NONEON-NOSVE-NEXT:    ldr q1, [x0, #16]
171; NONEON-NOSVE-NEXT:    tst w2, #0x1
172; NONEON-NOSVE-NEXT:    ldr q2, [x1]
173; NONEON-NOSVE-NEXT:    ldr q3, [x1, #16]
174; NONEON-NOSVE-NEXT:    str q0, [sp]
175; NONEON-NOSVE-NEXT:    stp q1, q3, [sp, #16]
176; NONEON-NOSVE-NEXT:    ldr h0, [sp, #30]
177; NONEON-NOSVE-NEXT:    ldr h1, [sp, #46]
178; NONEON-NOSVE-NEXT:    str q2, [sp, #48]
179; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
180; NONEON-NOSVE-NEXT:    ldr h1, [sp, #44]
181; NONEON-NOSVE-NEXT:    str h0, [sp, #78]
182; NONEON-NOSVE-NEXT:    ldr h0, [sp, #28]
183; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
184; NONEON-NOSVE-NEXT:    ldr h1, [sp, #42]
185; NONEON-NOSVE-NEXT:    str h0, [sp, #76]
186; NONEON-NOSVE-NEXT:    ldr h0, [sp, #26]
187; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
188; NONEON-NOSVE-NEXT:    ldr h1, [sp, #40]
189; NONEON-NOSVE-NEXT:    str h0, [sp, #74]
190; NONEON-NOSVE-NEXT:    ldr h0, [sp, #24]
191; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
192; NONEON-NOSVE-NEXT:    ldr h1, [sp, #38]
193; NONEON-NOSVE-NEXT:    str h0, [sp, #72]
194; NONEON-NOSVE-NEXT:    ldr h0, [sp, #22]
195; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
196; NONEON-NOSVE-NEXT:    ldr h1, [sp, #36]
197; NONEON-NOSVE-NEXT:    str h0, [sp, #70]
198; NONEON-NOSVE-NEXT:    ldr h0, [sp, #20]
199; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
200; NONEON-NOSVE-NEXT:    ldr h1, [sp, #34]
201; NONEON-NOSVE-NEXT:    str h0, [sp, #68]
202; NONEON-NOSVE-NEXT:    ldr h0, [sp, #18]
203; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
204; NONEON-NOSVE-NEXT:    ldr h1, [sp, #32]
205; NONEON-NOSVE-NEXT:    str h0, [sp, #66]
206; NONEON-NOSVE-NEXT:    ldr h0, [sp, #16]
207; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
208; NONEON-NOSVE-NEXT:    ldr h1, [sp, #62]
209; NONEON-NOSVE-NEXT:    str h0, [sp, #64]
210; NONEON-NOSVE-NEXT:    ldr h0, [sp, #14]
211; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
212; NONEON-NOSVE-NEXT:    ldr h1, [sp, #60]
213; NONEON-NOSVE-NEXT:    str h0, [sp, #94]
214; NONEON-NOSVE-NEXT:    ldr h0, [sp, #12]
215; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
216; NONEON-NOSVE-NEXT:    ldr h1, [sp, #58]
217; NONEON-NOSVE-NEXT:    str h0, [sp, #92]
218; NONEON-NOSVE-NEXT:    ldr h0, [sp, #10]
219; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
220; NONEON-NOSVE-NEXT:    ldr h1, [sp, #56]
221; NONEON-NOSVE-NEXT:    str h0, [sp, #90]
222; NONEON-NOSVE-NEXT:    ldr h0, [sp, #8]
223; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
224; NONEON-NOSVE-NEXT:    ldr h1, [sp, #54]
225; NONEON-NOSVE-NEXT:    str h0, [sp, #88]
226; NONEON-NOSVE-NEXT:    ldr h0, [sp, #6]
227; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
228; NONEON-NOSVE-NEXT:    ldr h1, [sp, #52]
229; NONEON-NOSVE-NEXT:    str h0, [sp, #86]
230; NONEON-NOSVE-NEXT:    ldr h0, [sp, #4]
231; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
232; NONEON-NOSVE-NEXT:    ldr h1, [sp, #50]
233; NONEON-NOSVE-NEXT:    str h0, [sp, #84]
234; NONEON-NOSVE-NEXT:    ldr h0, [sp, #2]
235; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
236; NONEON-NOSVE-NEXT:    ldr h1, [sp, #48]
237; NONEON-NOSVE-NEXT:    str h0, [sp, #82]
238; NONEON-NOSVE-NEXT:    ldr h0, [sp]
239; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
240; NONEON-NOSVE-NEXT:    str h0, [sp, #80]
241; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #64]
242; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
243; NONEON-NOSVE-NEXT:    add sp, sp, #96
244; NONEON-NOSVE-NEXT:    ret
245  %op1 = load volatile <16 x half>, ptr %a
246  %op2 = load volatile <16 x half>, ptr %b
247  %sel = select i1 %mask, <16 x half> %op1, <16 x half> %op2
248  store <16 x half> %sel, ptr %a
249  ret void
250}
251
252define <2 x float> @select_v2f32(<2 x float> %op1, <2 x float> %op2, i1 %mask) {
253; CHECK-LABEL: select_v2f32:
254; CHECK:       // %bb.0:
255; CHECK-NEXT:    and w8, w0, #0x1
256; CHECK-NEXT:    ptrue p0.s
257; CHECK-NEXT:    // kill: def $d0 killed $d0 def $z0
258; CHECK-NEXT:    // kill: def $d1 killed $d1 def $z1
259; CHECK-NEXT:    mov z2.s, w8
260; CHECK-NEXT:    cmpne p0.s, p0/z, z2.s, #0
261; CHECK-NEXT:    sel z0.s, p0, z0.s, z1.s
262; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $z0
263; CHECK-NEXT:    ret
264;
265; NONEON-NOSVE-LABEL: select_v2f32:
266; NONEON-NOSVE:       // %bb.0:
267; NONEON-NOSVE-NEXT:    sub sp, sp, #32
268; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 32
269; NONEON-NOSVE-NEXT:    stp d0, d1, [sp, #8]
270; NONEON-NOSVE-NEXT:    tst w0, #0x1
271; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #8]
272; NONEON-NOSVE-NEXT:    ldr s0, [sp, #20]
273; NONEON-NOSVE-NEXT:    fcsel s3, s2, s0, ne
274; NONEON-NOSVE-NEXT:    ldr s0, [sp, #16]
275; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
276; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #24]
277; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
278; NONEON-NOSVE-NEXT:    add sp, sp, #32
279; NONEON-NOSVE-NEXT:    ret
280  %sel = select i1 %mask, <2 x float> %op1, <2 x float> %op2
281  ret <2 x float> %sel
282}
283
284define <4 x float> @select_v4f32(<4 x float> %op1, <4 x float> %op2, i1 %mask) {
285; CHECK-LABEL: select_v4f32:
286; CHECK:       // %bb.0:
287; CHECK-NEXT:    and w8, w0, #0x1
288; CHECK-NEXT:    ptrue p0.s
289; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
290; CHECK-NEXT:    // kill: def $q1 killed $q1 def $z1
291; CHECK-NEXT:    mov z2.s, w8
292; CHECK-NEXT:    cmpne p0.s, p0/z, z2.s, #0
293; CHECK-NEXT:    sel z0.s, p0, z0.s, z1.s
294; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
295; CHECK-NEXT:    ret
296;
297; NONEON-NOSVE-LABEL: select_v4f32:
298; NONEON-NOSVE:       // %bb.0:
299; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #-48]!
300; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 48
301; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #8]
302; NONEON-NOSVE-NEXT:    tst w0, #0x1
303; NONEON-NOSVE-NEXT:    ldr s0, [sp, #28]
304; NONEON-NOSVE-NEXT:    fcsel s3, s2, s0, ne
305; NONEON-NOSVE-NEXT:    ldr s0, [sp, #24]
306; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
307; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp]
308; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #40]
309; NONEON-NOSVE-NEXT:    ldr s0, [sp, #20]
310; NONEON-NOSVE-NEXT:    fcsel s3, s2, s0, ne
311; NONEON-NOSVE-NEXT:    ldr s0, [sp, #16]
312; NONEON-NOSVE-NEXT:    fcsel s0, s1, s0, ne
313; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #32]
314; NONEON-NOSVE-NEXT:    ldr q0, [sp, #32]
315; NONEON-NOSVE-NEXT:    add sp, sp, #48
316; NONEON-NOSVE-NEXT:    ret
317  %sel = select i1 %mask, <4 x float> %op1, <4 x float> %op2
318  ret <4 x float> %sel
319}
320
321define void @select_v8f32(ptr %a, ptr %b, i1 %mask) {
322; CHECK-LABEL: select_v8f32:
323; CHECK:       // %bb.0:
324; CHECK-NEXT:    and w8, w2, #0x1
325; CHECK-NEXT:    ptrue p0.s
326; CHECK-NEXT:    mov z0.s, w8
327; CHECK-NEXT:    cmpne p0.s, p0/z, z0.s, #0
328; CHECK-NEXT:    ldr q0, [x0]
329; CHECK-NEXT:    ldr q1, [x0, #16]
330; CHECK-NEXT:    ldr q2, [x1]
331; CHECK-NEXT:    ldr q3, [x1, #16]
332; CHECK-NEXT:    sel z0.s, p0, z0.s, z2.s
333; CHECK-NEXT:    sel z1.s, p0, z1.s, z3.s
334; CHECK-NEXT:    stp q0, q1, [x0]
335; CHECK-NEXT:    ret
336;
337; NONEON-NOSVE-LABEL: select_v8f32:
338; NONEON-NOSVE:       // %bb.0:
339; NONEON-NOSVE-NEXT:    sub sp, sp, #96
340; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 96
341; NONEON-NOSVE-NEXT:    ldr q0, [x0]
342; NONEON-NOSVE-NEXT:    ldr q1, [x0, #16]
343; NONEON-NOSVE-NEXT:    tst w2, #0x1
344; NONEON-NOSVE-NEXT:    ldr q2, [x1]
345; NONEON-NOSVE-NEXT:    ldr q3, [x1, #16]
346; NONEON-NOSVE-NEXT:    str q0, [sp]
347; NONEON-NOSVE-NEXT:    stp q1, q3, [sp, #16]
348; NONEON-NOSVE-NEXT:    str q2, [sp, #48]
349; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #40]
350; NONEON-NOSVE-NEXT:    ldr s0, [sp, #28]
351; NONEON-NOSVE-NEXT:    fcsel s3, s0, s2, ne
352; NONEON-NOSVE-NEXT:    ldr s0, [sp, #24]
353; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
354; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #32]
355; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #72]
356; NONEON-NOSVE-NEXT:    ldr s0, [sp, #20]
357; NONEON-NOSVE-NEXT:    fcsel s3, s0, s2, ne
358; NONEON-NOSVE-NEXT:    ldr s0, [sp, #16]
359; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
360; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #56]
361; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #64]
362; NONEON-NOSVE-NEXT:    ldr s0, [sp, #12]
363; NONEON-NOSVE-NEXT:    fcsel s3, s0, s2, ne
364; NONEON-NOSVE-NEXT:    ldr s0, [sp, #8]
365; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
366; NONEON-NOSVE-NEXT:    ldp s1, s2, [sp, #48]
367; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #88]
368; NONEON-NOSVE-NEXT:    ldr s0, [sp, #4]
369; NONEON-NOSVE-NEXT:    fcsel s3, s0, s2, ne
370; NONEON-NOSVE-NEXT:    ldr s0, [sp]
371; NONEON-NOSVE-NEXT:    fcsel s0, s0, s1, ne
372; NONEON-NOSVE-NEXT:    stp s0, s3, [sp, #80]
373; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #64]
374; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
375; NONEON-NOSVE-NEXT:    add sp, sp, #96
376; NONEON-NOSVE-NEXT:    ret
377  %op1 = load volatile <8 x float>, ptr %a
378  %op2 = load volatile <8 x float>, ptr %b
379  %sel = select i1 %mask, <8 x float> %op1, <8 x float> %op2
380  store <8 x float> %sel, ptr %a
381  ret void
382}
383
384define <1 x double> @select_v1f64(<1 x double> %op1, <1 x double> %op2, i1 %mask) {
385; CHECK-LABEL: select_v1f64:
386; CHECK:       // %bb.0:
387; CHECK-NEXT:    tst w0, #0x1
388; CHECK-NEXT:    fcsel d0, d0, d1, ne
389; CHECK-NEXT:    ret
390;
391; NONEON-NOSVE-LABEL: select_v1f64:
392; NONEON-NOSVE:       // %bb.0:
393; NONEON-NOSVE-NEXT:    sub sp, sp, #16
394; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 16
395; NONEON-NOSVE-NEXT:    tst w0, #0x1
396; NONEON-NOSVE-NEXT:    fcsel d0, d0, d1, ne
397; NONEON-NOSVE-NEXT:    str d0, [sp, #8]
398; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
399; NONEON-NOSVE-NEXT:    add sp, sp, #16
400; NONEON-NOSVE-NEXT:    ret
401  %sel = select i1 %mask, <1 x double> %op1, <1 x double> %op2
402  ret <1 x double> %sel
403}
404
405define <2 x double> @select_v2f64(<2 x double> %op1, <2 x double> %op2, i1 %mask) {
406; CHECK-LABEL: select_v2f64:
407; CHECK:       // %bb.0:
408; CHECK-NEXT:    // kill: def $w0 killed $w0 def $x0
409; CHECK-NEXT:    and x8, x0, #0x1
410; CHECK-NEXT:    ptrue p0.d
411; CHECK-NEXT:    // kill: def $q0 killed $q0 def $z0
412; CHECK-NEXT:    // kill: def $q1 killed $q1 def $z1
413; CHECK-NEXT:    mov z2.d, x8
414; CHECK-NEXT:    cmpne p0.d, p0/z, z2.d, #0
415; CHECK-NEXT:    sel z0.d, p0, z0.d, z1.d
416; CHECK-NEXT:    // kill: def $q0 killed $q0 killed $z0
417; CHECK-NEXT:    ret
418;
419; NONEON-NOSVE-LABEL: select_v2f64:
420; NONEON-NOSVE:       // %bb.0:
421; NONEON-NOSVE-NEXT:    stp q0, q1, [sp, #-48]!
422; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 48
423; NONEON-NOSVE-NEXT:    ldp d1, d2, [sp]
424; NONEON-NOSVE-NEXT:    tst w0, #0x1
425; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
426; NONEON-NOSVE-NEXT:    fcsel d3, d2, d0, ne
427; NONEON-NOSVE-NEXT:    ldr d0, [sp, #16]
428; NONEON-NOSVE-NEXT:    fcsel d0, d1, d0, ne
429; NONEON-NOSVE-NEXT:    stp d0, d3, [sp, #32]
430; NONEON-NOSVE-NEXT:    ldr q0, [sp, #32]
431; NONEON-NOSVE-NEXT:    add sp, sp, #48
432; NONEON-NOSVE-NEXT:    ret
433  %sel = select i1 %mask, <2 x double> %op1, <2 x double> %op2
434  ret <2 x double> %sel
435}
436
437define void @select_v4f64(ptr %a, ptr %b, i1 %mask) {
438; CHECK-LABEL: select_v4f64:
439; CHECK:       // %bb.0:
440; CHECK-NEXT:    // kill: def $w2 killed $w2 def $x2
441; CHECK-NEXT:    and x8, x2, #0x1
442; CHECK-NEXT:    ptrue p0.d
443; CHECK-NEXT:    mov z0.d, x8
444; CHECK-NEXT:    cmpne p0.d, p0/z, z0.d, #0
445; CHECK-NEXT:    ldr q0, [x0]
446; CHECK-NEXT:    ldr q1, [x0, #16]
447; CHECK-NEXT:    ldr q2, [x1]
448; CHECK-NEXT:    ldr q3, [x1, #16]
449; CHECK-NEXT:    sel z0.d, p0, z0.d, z2.d
450; CHECK-NEXT:    sel z1.d, p0, z1.d, z3.d
451; CHECK-NEXT:    stp q0, q1, [x0]
452; CHECK-NEXT:    ret
453;
454; NONEON-NOSVE-LABEL: select_v4f64:
455; NONEON-NOSVE:       // %bb.0:
456; NONEON-NOSVE-NEXT:    sub sp, sp, #96
457; NONEON-NOSVE-NEXT:    .cfi_def_cfa_offset 96
458; NONEON-NOSVE-NEXT:    ldr q0, [x0]
459; NONEON-NOSVE-NEXT:    ldr q1, [x0, #16]
460; NONEON-NOSVE-NEXT:    tst w2, #0x1
461; NONEON-NOSVE-NEXT:    ldr q2, [x1]
462; NONEON-NOSVE-NEXT:    ldr q3, [x1, #16]
463; NONEON-NOSVE-NEXT:    str q0, [sp]
464; NONEON-NOSVE-NEXT:    stp q1, q3, [sp, #16]
465; NONEON-NOSVE-NEXT:    str q2, [sp, #48]
466; NONEON-NOSVE-NEXT:    ldp d1, d2, [sp, #32]
467; NONEON-NOSVE-NEXT:    ldr d0, [sp, #24]
468; NONEON-NOSVE-NEXT:    fcsel d3, d0, d2, ne
469; NONEON-NOSVE-NEXT:    ldr d0, [sp, #16]
470; NONEON-NOSVE-NEXT:    fcsel d0, d0, d1, ne
471; NONEON-NOSVE-NEXT:    ldp d1, d2, [sp, #48]
472; NONEON-NOSVE-NEXT:    stp d0, d3, [sp, #64]
473; NONEON-NOSVE-NEXT:    ldr d0, [sp, #8]
474; NONEON-NOSVE-NEXT:    fcsel d3, d0, d2, ne
475; NONEON-NOSVE-NEXT:    ldr d0, [sp]
476; NONEON-NOSVE-NEXT:    fcsel d0, d0, d1, ne
477; NONEON-NOSVE-NEXT:    stp d0, d3, [sp, #80]
478; NONEON-NOSVE-NEXT:    ldp q1, q0, [sp, #64]
479; NONEON-NOSVE-NEXT:    stp q0, q1, [x0]
480; NONEON-NOSVE-NEXT:    add sp, sp, #96
481; NONEON-NOSVE-NEXT:    ret
482  %op1 = load volatile <4 x double>, ptr %a
483  %op2 = load volatile <4 x double>, ptr %b
484  %sel = select i1 %mask, <4 x double> %op1, <4 x double> %op2
485  store <4 x double> %sel, ptr %a
486  ret void
487}
488