| blob | 5f56a82f3c511099aa6134669305f897923a0b50 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
2 ; RUN: llc < %s -mtriple=thumbv8.1m.main-none-eabihf -mattr=+mve.fp | FileCheck %s --check-prefix=CHECKLE
3 ; RUN: llc < %s -mtriple=thumbebv8.1m.main-none-eabihf -mattr=+mve.fp | FileCheck %s --check-prefix=CHECKBE
6 define <8 x i8> @inserti8_first(ptr %p) {
7 ; CHECKLE-LABEL: inserti8_first:
8 ; CHECKLE: @ %bb.0:
9 ; CHECKLE-NEXT: vldrb.u16 q0, [r0]
10 ; CHECKLE-NEXT: bx lr
11 ;
12 ; CHECKBE-LABEL: inserti8_first:
13 ; CHECKBE: @ %bb.0:
14 ; CHECKBE-NEXT: vldrb.u16 q1, [r0]
15 ; CHECKBE-NEXT: vrev64.16 q0, q1
16 ; CHECKBE-NEXT: bx lr
17 %q = getelementptr inbounds i8, ptr %p, i32 1
18 %l1 = load <8 x i8>, ptr %q
19 %l2 = load i8, ptr %p
20 %s = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
21 %ins = insertelement <8 x i8> %s, i8 %l2, i32 0
22 ret <8 x i8> %ins
23 }
25 define <8 x i8> @inserti8_last(ptr %p) {
26 ; CHECKLE-LABEL: inserti8_last:
27 ; CHECKLE: @ %bb.0:
28 ; CHECKLE-NEXT: vldrb.u16 q0, [r0, #1]
29 ; CHECKLE-NEXT: bx lr
30 ;
31 ; CHECKBE-LABEL: inserti8_last:
32 ; CHECKBE: @ %bb.0:
33 ; CHECKBE-NEXT: vldrb.u16 q1, [r0, #1]
34 ; CHECKBE-NEXT: vrev64.16 q0, q1
35 ; CHECKBE-NEXT: bx lr
36 %q = getelementptr inbounds i8, ptr %p, i32 8
37 %l1 = load <8 x i8>, ptr %p
38 %l2 = load i8, ptr %q
39 %s = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
40 %ins = insertelement <8 x i8> %s, i8 %l2, i32 7
41 ret <8 x i8> %ins
42 }
44 define <8 x i16> @inserti8_first_sext(ptr %p) {
45 ; CHECKLE-LABEL: inserti8_first_sext:
46 ; CHECKLE: @ %bb.0:
47 ; CHECKLE-NEXT: vldrb.s16 q0, [r0]
48 ; CHECKLE-NEXT: bx lr
49 ;
50 ; CHECKBE-LABEL: inserti8_first_sext:
51 ; CHECKBE: @ %bb.0:
52 ; CHECKBE-NEXT: vldrb.s16 q1, [r0]
53 ; CHECKBE-NEXT: vrev64.16 q0, q1
54 ; CHECKBE-NEXT: bx lr
55 %q = getelementptr inbounds i8, ptr %p, i32 1
56 %l1 = load <8 x i8>, ptr %q
57 %s1 = sext <8 x i8> %l1 to <8 x i16>
58 %l2 = load i8, ptr %p
59 %s2 = sext i8 %l2 to i16
60 %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
61 %ins = insertelement <8 x i16> %s, i16 %s2, i32 0
62 ret <8 x i16> %ins
63 }
65 define <8 x i16> @inserti8_last_sext(ptr %p) {
66 ; CHECKLE-LABEL: inserti8_last_sext:
67 ; CHECKLE: @ %bb.0:
68 ; CHECKLE-NEXT: vldrb.s16 q0, [r0, #1]
69 ; CHECKLE-NEXT: bx lr
70 ;
71 ; CHECKBE-LABEL: inserti8_last_sext:
72 ; CHECKBE: @ %bb.0:
73 ; CHECKBE-NEXT: vldrb.s16 q1, [r0, #1]
74 ; CHECKBE-NEXT: vrev64.16 q0, q1
75 ; CHECKBE-NEXT: bx lr
76 %q = getelementptr inbounds i8, ptr %p, i32 8
77 %l1 = load <8 x i8>, ptr %p
78 %s1 = sext <8 x i8> %l1 to <8 x i16>
79 %l2 = load i8, ptr %q
80 %s2 = sext i8 %l2 to i16
81 %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
82 %ins = insertelement <8 x i16> %s, i16 %s2, i32 7
83 ret <8 x i16> %ins
84 }
86 define <8 x i16> @inserti8_first_zext(ptr %p) {
87 ; CHECKLE-LABEL: inserti8_first_zext:
88 ; CHECKLE: @ %bb.0:
89 ; CHECKLE-NEXT: vldrb.u16 q0, [r0]
90 ; CHECKLE-NEXT: bx lr
91 ;
92 ; CHECKBE-LABEL: inserti8_first_zext:
93 ; CHECKBE: @ %bb.0:
94 ; CHECKBE-NEXT: vldrb.u16 q1, [r0]
95 ; CHECKBE-NEXT: vrev64.16 q0, q1
96 ; CHECKBE-NEXT: bx lr
97 %q = getelementptr inbounds i8, ptr %p, i32 1
98 %l1 = load <8 x i8>, ptr %q
99 %s1 = zext <8 x i8> %l1 to <8 x i16>
100 %l2 = load i8, ptr %p
101 %s2 = zext i8 %l2 to i16
102 %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
103 %ins = insertelement <8 x i16> %s, i16 %s2, i32 0
104 ret <8 x i16> %ins
105 }
107 define <8 x i16> @inserti8_last_zext(ptr %p) {
108 ; CHECKLE-LABEL: inserti8_last_zext:
109 ; CHECKLE: @ %bb.0:
110 ; CHECKLE-NEXT: vldrb.u16 q0, [r0, #1]
111 ; CHECKLE-NEXT: bx lr
112 ;
113 ; CHECKBE-LABEL: inserti8_last_zext:
114 ; CHECKBE: @ %bb.0:
115 ; CHECKBE-NEXT: vldrb.u16 q1, [r0, #1]
116 ; CHECKBE-NEXT: vrev64.16 q0, q1
117 ; CHECKBE-NEXT: bx lr
118 %q = getelementptr inbounds i8, ptr %p, i32 8
119 %l1 = load <8 x i8>, ptr %p
120 %s1 = zext <8 x i8> %l1 to <8 x i16>
121 %l2 = load i8, ptr %q
122 %s2 = zext i8 %l2 to i16
123 %s = shufflevector <8 x i16> %s1, <8 x i16> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
124 %ins = insertelement <8 x i16> %s, i16 %s2, i32 7
125 ret <8 x i16> %ins
126 }
128 define <8 x i32> @inserti32_first(ptr %p) {
129 ; CHECKLE-LABEL: inserti32_first:
130 ; CHECKLE: @ %bb.0:
131 ; CHECKLE-NEXT: vldrw.u32 q2, [r0, #20]
132 ; CHECKLE-NEXT: vldr s4, [r0, #16]
133 ; CHECKLE-NEXT: vldrw.u32 q0, [r0]
134 ; CHECKLE-NEXT: vmov.f32 s5, s8
135 ; CHECKLE-NEXT: vmov.f32 s6, s9
136 ; CHECKLE-NEXT: vmov.f32 s7, s10
137 ; CHECKLE-NEXT: bx lr
138 ;
139 ; CHECKBE-LABEL: inserti32_first:
140 ; CHECKBE: @ %bb.0:
141 ; CHECKBE-NEXT: vldrw.u32 q3, [r0, #20]
142 ; CHECKBE-NEXT: vldrb.u8 q1, [r0]
143 ; CHECKBE-NEXT: vldr s8, [r0, #16]
144 ; CHECKBE-NEXT: vmov.f32 s9, s12
145 ; CHECKBE-NEXT: vrev64.8 q0, q1
146 ; CHECKBE-NEXT: vmov.f32 s10, s13
147 ; CHECKBE-NEXT: vmov.f32 s11, s14
148 ; CHECKBE-NEXT: vrev64.32 q1, q2
149 ; CHECKBE-NEXT: bx lr
150 %q = getelementptr inbounds i8, ptr %p, i32 4
151 %l1 = load <8 x i32>, ptr %q
152 %l2 = load i32, ptr %p
153 %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
154 %ins = insertelement <8 x i32> %s, i32 %l2, i32 0
155 ret <8 x i32> %ins
156 }
158 define <8 x i32> @inserti32_last(ptr %p) {
159 ; CHECKLE-LABEL: inserti32_last:
160 ; CHECKLE: @ %bb.0:
161 ; CHECKLE-NEXT: vldrw.u32 q2, [r0]
162 ; CHECKLE-NEXT: vldr s3, [r0, #16]
163 ; CHECKLE-NEXT: vldrw.u32 q1, [r0, #20]
164 ; CHECKLE-NEXT: vmov.f32 s0, s9
165 ; CHECKLE-NEXT: vmov.f32 s1, s10
166 ; CHECKLE-NEXT: vmov.f32 s2, s11
167 ; CHECKLE-NEXT: bx lr
168 ;
169 ; CHECKBE-LABEL: inserti32_last:
170 ; CHECKBE: @ %bb.0:
171 ; CHECKBE-NEXT: vldrw.u32 q3, [r0]
172 ; CHECKBE-NEXT: vldrb.u8 q0, [r0, #20]
173 ; CHECKBE-NEXT: vldr s11, [r0, #16]
174 ; CHECKBE-NEXT: vmov.f32 s8, s13
175 ; CHECKBE-NEXT: vrev64.8 q1, q0
176 ; CHECKBE-NEXT: vmov.f32 s9, s14
177 ; CHECKBE-NEXT: vmov.f32 s10, s15
178 ; CHECKBE-NEXT: vrev64.32 q0, q2
179 ; CHECKBE-NEXT: bx lr
180 %q = getelementptr inbounds i8, ptr %p, i32 32
181 %l1 = load <8 x i32>, ptr %p
182 %l2 = load i32, ptr %q
183 %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
184 %ins = insertelement <8 x i32> %s, i32 %l2, i32 7
185 ret <8 x i32> %ins
186 }
188 define <8 x i32> @inserti32_first_multiuse(ptr %p) {
189 ; CHECKLE-LABEL: inserti32_first_multiuse:
190 ; CHECKLE: @ %bb.0:
191 ; CHECKLE-NEXT: vldrw.u32 q0, [r0, #20]
192 ; CHECKLE-NEXT: vldrw.u32 q2, [r0, #4]
193 ; CHECKLE-NEXT: vmov.f32 s4, s11
194 ; CHECKLE-NEXT: vmov.f32 s5, s0
195 ; CHECKLE-NEXT: vmov.f32 s6, s1
196 ; CHECKLE-NEXT: vmov.f32 s7, s2
197 ; CHECKLE-NEXT: vadd.i32 q1, q0, q1
198 ; CHECKLE-NEXT: vldrw.u32 q0, [r0]
199 ; CHECKLE-NEXT: vadd.i32 q0, q2, q0
200 ; CHECKLE-NEXT: bx lr
201 ;
202 ; CHECKBE-LABEL: inserti32_first_multiuse:
203 ; CHECKBE: @ %bb.0:
204 ; CHECKBE-NEXT: vldrw.u32 q0, [r0, #20]
205 ; CHECKBE-NEXT: vldrw.u32 q2, [r0, #4]
206 ; CHECKBE-NEXT: vmov.f32 s4, s11
207 ; CHECKBE-NEXT: vmov.f32 s5, s0
208 ; CHECKBE-NEXT: vmov.f32 s6, s1
209 ; CHECKBE-NEXT: vmov.f32 s7, s2
210 ; CHECKBE-NEXT: vadd.i32 q0, q0, q1
211 ; CHECKBE-NEXT: vrev64.32 q1, q0
212 ; CHECKBE-NEXT: vldrw.u32 q0, [r0]
213 ; CHECKBE-NEXT: vadd.i32 q2, q2, q0
214 ; CHECKBE-NEXT: vrev64.32 q0, q2
215 ; CHECKBE-NEXT: bx lr
216 %q = getelementptr inbounds i8, ptr %p, i32 4
217 %l1 = load <8 x i32>, ptr %q
218 %l2 = load i32, ptr %p
219 %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 undef, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
220 %ins = insertelement <8 x i32> %s, i32 %l2, i32 0
221 %a = add <8 x i32> %l1, %ins
222 ret <8 x i32> %a
223 }
225 define <8 x i32> @inserti32_last_multiuse(ptr %p) {
226 ; CHECKLE-LABEL: inserti32_last_multiuse:
227 ; CHECKLE: @ %bb.0:
228 ; CHECKLE-NEXT: vldrw.u32 q0, [r0]
229 ; CHECKLE-NEXT: vldrw.u32 q1, [r0, #16]
230 ; CHECKLE-NEXT: vmov.f32 s8, s1
231 ; CHECKLE-NEXT: vmov.f32 s9, s2
232 ; CHECKLE-NEXT: vmov.f32 s10, s3
233 ; CHECKLE-NEXT: vmov.f32 s11, s4
234 ; CHECKLE-NEXT: vadd.i32 q0, q0, q2
235 ; CHECKLE-NEXT: vldrw.u32 q2, [r0, #20]
236 ; CHECKLE-NEXT: vadd.i32 q1, q1, q2
237 ; CHECKLE-NEXT: bx lr
238 ;
239 ; CHECKBE-LABEL: inserti32_last_multiuse:
240 ; CHECKBE: @ %bb.0:
241 ; CHECKBE-NEXT: vldrw.u32 q0, [r0]
242 ; CHECKBE-NEXT: vldrw.u32 q1, [r0, #16]
243 ; CHECKBE-NEXT: vmov.f32 s8, s1
244 ; CHECKBE-NEXT: vmov.f32 s9, s2
245 ; CHECKBE-NEXT: vmov.f32 s10, s3
246 ; CHECKBE-NEXT: vmov.f32 s11, s4
247 ; CHECKBE-NEXT: vadd.i32 q2, q0, q2
248 ; CHECKBE-NEXT: vrev64.32 q0, q2
249 ; CHECKBE-NEXT: vldrw.u32 q2, [r0, #20]
250 ; CHECKBE-NEXT: vadd.i32 q2, q1, q2
251 ; CHECKBE-NEXT: vrev64.32 q1, q2
252 ; CHECKBE-NEXT: bx lr
253 %q = getelementptr inbounds i8, ptr %p, i32 32
254 %l1 = load <8 x i32>, ptr %p
255 %l2 = load i32, ptr %q
256 %s = shufflevector <8 x i32> %l1, <8 x i32> undef, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 undef>
257 %ins = insertelement <8 x i32> %s, i32 %l2, i32 7
258 %a = add <8 x i32> %l1, %ins
259 ret <8 x i32> %a
260 }
262 define <4 x float> @insertf32_first(ptr %p) {
263 ; CHECKLE-LABEL: insertf32_first:
264 ; CHECKLE: @ %bb.0:
265 ; CHECKLE-NEXT: vldrw.u32 q0, [r0]
266 ; CHECKLE-NEXT: bx lr
267 ;
268 ; CHECKBE-LABEL: insertf32_first:
269 ; CHECKBE: @ %bb.0:
270 ; CHECKBE-NEXT: vldrb.u8 q1, [r0]
271 ; CHECKBE-NEXT: vrev64.8 q0, q1
272 ; CHECKBE-NEXT: bx lr
273 %q = getelementptr inbounds i8, ptr %p, i32 4
274 %l1 = load <4 x float>, ptr %q
275 %l2 = load float, ptr %p
276 %s = shufflevector <4 x float> %l1, <4 x float> undef, <4 x i32> <i32 undef, i32 0, i32 1, i32 2>
277 %ins = insertelement <4 x float> %s, float %l2, i32 0
278 ret <4 x float> %ins
279 }
281 define <4 x float> @insertf32_last(ptr %p) {
282 ; CHECKLE-LABEL: insertf32_last:
283 ; CHECKLE: @ %bb.0:
284 ; CHECKLE-NEXT: vldrw.u32 q0, [r0, #4]
285 ; CHECKLE-NEXT: bx lr
286 ;
287 ; CHECKBE-LABEL: insertf32_last:
288 ; CHECKBE: @ %bb.0:
289 ; CHECKBE-NEXT: vldrb.u8 q1, [r0, #4]
290 ; CHECKBE-NEXT: vrev64.8 q0, q1
291 ; CHECKBE-NEXT: bx lr
292 %q = getelementptr inbounds i8, ptr %p, i32 16
293 %l1 = load <4 x float>, ptr %p
294 %l2 = load float, ptr %q
295 %s = shufflevector <4 x float> %l1, <4 x float> undef, <4 x i32> <i32 1, i32 2, i32 3, i32 undef>
296 %ins = insertelement <4 x float> %s, float %l2, i32 3
297 ret <4 x float> %ins
298 }